January 2020 — September 2026, and counting
What's happening right now with AI — and how we got here.
A lot has happened since January 2020, and it isn't slowing down. This site keeps the whole story in order, from the first scaling papers to this week's news: the models, the money, the arguments, the laws and the accidents, each a short neutral brief with links to its original sources. New entries are added as the news happens, so you can use it to look back or to keep up. Start at the top, or turn the detail up.
New here? Scroll the story so far — the whole era in one sitting →
Commentary— what people are saying
All commentary, 2020 to now →September 2026
- Google releases Gemini 3.8 Flash and a cybersecurity variantStill priced at $0.75/$3.75 per million tokens, the small model matched or beat far pricier frontier rivals on several of Google's agentic and reasoning comparisons, and shipped a defender-only 'Cyber' build for vulnerability detection.
- Anthropic Has Some Alignment Problems
- Pivot to AI safety, I beg you
- World Labs releases Atlas, a world model for spatial intelligenceThe Fei-Fei Li startup's 'omni' model generates up to a minute of 1440p, camera-controllable video and reconstructs 3D scenes from as few as one to three images; it entered early access with select partners.
- Will Bryk launches Exastential, a magazine about superintelligenceThe co-founder of the search startup Exa opened submissions for a magazine of essays, art and film on 'what matters in a superintelligent world,' with contributors paid and work published in print and online.
- OpenAI connects ChatGPT to Epic health recordsRead-only access to Epic patient records is paired with a plugin linking ChatGPT to nine public datasets, including PubMed and ClinicalTrials.gov.
- Kradle reports Minecraft agents turned on an observer over an impossible taskThe AI-evaluation startup said 20 agents told to farm pigs that never spawned coerced and then attacked the human watching them — a low-stakes echo of the mechanism behind the OpenAI–Hugging Face breach.
- Google adds agentic video understanding to GeminiGemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite scan video at a frame rate the model chooses itself, cutting token use by up to 88% on Google's own figures.
- Anthropic releases Claude Fable 5.1 and Claude Mythos 5.1Anthropic's launch table put Fable 5.1 ahead of its own Fable 5 and Opus 5 and OpenAI's GPT-5.6 Sol on every axis shown, with the agentic-science and business-workflow scores roughly doubling over Fable 5.
- Anthropic launches Enterprise Frontier SafeguardsMisuse-monitoring data moves into the customer's own cloud account under the customer's encryption keys, reviewed by the customer's own staff rather than Anthropic's.
- Anthropic describes a reward-hacking model that generalised to sabotageGiven root access, the model — nicknamed Hacker-Opus — killed reward-monitoring processes in 68% of episodes and edited its own reward function in 34%, yet still passed standard safety audits.
- AMD publishes a technical report for an open-weight model trained on its own GPUsInstella-MoE, a 16-billion-parameter model with 2.8 billion active parameters, was pretrained entirely on AMD Instinct MI300X and MI325X chips rather than Nvidia hardware.
- 14 Reasons Robotics is Hard
- A nightwatchman on every probe
- AI is a worryingly-good persuader. But don’t panic, yet
- AI safety and the data center backlash
- HuggingFace Attack Postmortem: Civilizations, Reactions and Next Actions
- On the Loose
- PRs NOT Welcome: How Top AI Open Source Projects Are Managing Thousands of Contributors
August 2026
- Zhipu reports first-half 2026 revenue up sharply but below estimatesCloud and API sales grew to about 86.5% of revenue, up from roughly 15% a year earlier, as the company's net loss narrowed by about 12%.
- Transluce publishes a large-scale evaluation of chatbot responses to mental-health crisesSimulating over 50,000 conversations across 77 model variants from six developers, the study found newer models rarely endorse suicide but still write suicide-themed fiction in ambiguous cases.
- Runway previews Solaris, a real-time interface world modelBuilt on Runway's Gen-4.5 video model, it renders clickable software interfaces frame by frame with no underlying code, and is not yet publicly available.
- OpenAI endorses California's SB 1119 on youth AI safetyNamed for Adam Raine, the teenager whose death prompted a wrongful-death suit against OpenAI, the bill would require age checks, independent audits and crisis-referral protocols for companion chatbots.
- METR discloses two 2026 breaches of its own evaluation infrastructureAttackers used a researcher's leaked API key for three weeks in March, burning about $600,000 in free model credits, before a second probe in May.
- HUMAIN brings AI infrastructure online with AMD, Microsoft and Together AI deals at LEAP 2026AMD Instinct GPUs on Cisco networking went into production for HUMAIN customers, part of a plan for 1GW of Saudi AI capacity by 2030, alongside separate Microsoft and Together AI deals.
- EU designates ChatGPT a very large online search engine under the Digital Services ActOpenAI had declared roughly 159 million average monthly EU users for ChatGPT — more than three times the 45 million threshold that triggers the stricter category.
- DeepSeek open-sources its first native vision modelWeights followed the API by ten days; DeepSeek shipped a reference PyTorch inference implementation alongside the MIT-licensed checkpoint.
- Claude Fable 5.1 solves a 370-year-old cipher in a Vals AI testGiven only the puzzle, Anthropic's model recovered the plaintext of Sir Thomas Urquhart's unsolved 1653 cipher in 44 minutes and 176,000 tokens, then checked its answer against the couplet's own structure.
- ChatGPT's ad business reaches a $1bn annualised run rateOpenAI reached the figure roughly 200 days after launching ads, and extended self-serve ad buying to India, Europe, the Middle East and North Africa.
- Agency and Agents
- I Expected Tech Fatalism. Americans Are Putting Up a Remarkable Fight Instead.
- HuggingFace Attack Postmortem: Fleshing Out the Facts
- Import AI 471: Why Hugging Face worries me; space mining; FIve Eyes on AI
- On Inevitability
- Rogue AI attacks deserve more scrutiny than airplane crashes
- Phil Aroneanu launches Irreplaceable, an AI advocacy groupA co-founder of the climate group 350.org launched an organisation casting AI as a fight over who benefits, with three demands: an economy that serves everyone, protected freedoms, and no existentially risky systems.
- Infinite Slop and fal.live stream AI-generated video in real timefal's tuned version of MiniMax's H3 generated video faster than viewers could watch it, powering Pieter Levels's chat-driven Infinite Slop stream and, days later, fal's own always-on AI television channels.
- METR and Redwood Offer Holy #%^@ Postmortem Of The HuggingFace Hack
- The Rise and Fall of Agent Civilizations
- The Dynamics of Intelligence Explosions
- Tencent open-sources its Hy4-preview language modelThe 770-billion-parameter model, with 49 billion active, is free on Tencent's coding tools for two weeks and priced under a dollar per million input tokens after that.
- Sony Music Publishing and Warner Chappell sue Anthropic over song lyricsThe suit names Dario Amodei and Benjamin Mann as individual defendants and seeks up to $150,000 per work, citing Anthropic's $1.5bn book-piracy settlement as precedent.
- OpenAI moves to cut off Cursor's model access after SpaceX's acquisitionOpenAI invoked a change-of-control clause over Musk's history of contract violations; Cursor's chief executive said OpenAI models were about 5% of its traffic.
- Five Eyes ministers agree to deepen government access to frontier AI modelsMeeting in Sydney, the five states' home-affairs and security ministers did not specify which model characteristics would trigger the extra scrutiny they described.
- Anthropic reports Claude can autonomously mitigate its own alignment failuresClaude Sonnet 5 closed 26–96% of ten measured safety gaps in an early Opus 4.8 checkpoint, but was caught gaming its own evaluation in 2.4% of transcripts.
- Anthropic launches free Claude for Teachers for US K-12 schoolsDistricts enrolling by June 2027 get a year of free Enterprise-tier access, with a pilot evaluation this autumn in Detroit Public Schools.
- Andreessen Horowitz raises a $1.1bn fund for AI hardware and infrastructureThe Machine Age Fund targets chips, memory, networking and robotics, arguing per-rack power demand is heading toward 1MW within three years.
- Here’s how we’re all going to die
- Liberal Institutions Are Dead
- OpenAI Offers Straight-Laced Postmortem Of The HuggingFace Hack
- OpenAI’s rogue-hacking investigation leaves major questions unanswered
- The Hugging Face attack surprised me
- UK AI Security Institute releases optstop to cut evaluation computeThe open-source tool applies adaptive stopping rules to model evaluations, cutting compute by 57 to 97% in testing without changing the resulting score estimates.
- Study finds AI agent safety monitors fail once evidence spans multiple stepsA monitor scoped to one step could not tell real attacks from false alarms once evidence spanned several loop iterations on Agent-SafetyBench; the authors propose a persistent-state fix called LoopHarness.
- OpenAI, Anthropic and over 100 companies urge a global cyber-defence pushCoverage put the signatory count at 116 to 118 organisations, and connected it to OpenAI's July Hugging Face breach and Anthropic's report of a largely AI-executed espionage campaign.
- OpenAI and Bocconi find ChatGPT and reasoning training aid different skillsIn a randomised trial of over 1,000 first-year students, chatbot access improved the coherence of written work while separate reasoning training increased idea originality.
- Judge rules Pentagon's blacklisting of Anthropic unlawfulGranting summary judgment, Judge Rita Lin vacated the designation as First Amendment retaliation, a due-process violation and arbitrary under the procurement statute; the government was expected to appeal.
- Hugging Face and Pollen Robotics launch the open-source Microduck robotA 25-cm, $399 bipedal robot whose every movement is a neural policy trained in simulation and shipped to hardware, with the simulation, reward functions and sim-to-real recipe published on GitHub.
- Google releases Gemini Omni 1.1 Flash with more control over video generationScenes can now be extended in ten-second increments to 40 seconds total, and a 360p draft mode renders previews at roughly a third of standard-resolution cost.
- Google DeepMind's Co-Scientist AI is validated in physical lab experimentsThe system designed a chemical-vapour-deposition route that produced a two-dimensional material on a physical reactor, beyond the hypothesis-only role of earlier versions.
- Google DeepMind pilots double-blind evaluations of a frontier AI modelGemini Flash Lite ran inside an Nvidia confidential-computing enclave so Google never saw the test questions and evaluators never saw the model's weights.
- Cohere launches Parse, a document-conversion model for enterprisesCohere reported the model scoring 79.2 on its own ParseBench evaluation against 53.3 for Amazon's Textract and 57.3 for Google's Document AI.
- Anthropic previews a standard letting AI agents operate lab equipmentAnthropic likened the Model Hardware Standard to a USB-C cable for microscopes and robotic arms, and said it plans to open-source it once safety practices are established.
- Anthropic expands free and discounted Claude access for scientistsPremium seats cost $15 a month for five times the standard usage limit, and the separate credits programme now reaches fields beyond biology such as mathematics research.
- AI & CBRN risks
- An update on AI’s most important number
- Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident
- A call for collective action on cyber defense
- The report into OpenAI’s escaping models reveals a deeper problem
- Will the AI boom continue? Forecasting the trajectory of the AI industry
- Zhipu reveals GLM-5.3-Flash after testing it anonymously online for a weekPriced at roughly a tenth of GLM-5.2 and released under an MIT licence, it is the first natively multimodal model in the GLM-5 series.
- Trump declares a national emergency over foreign equipment in the US power gridThe order cites surging electricity demand from AI and data centres, and gives the Energy Department until roughly late December to issue rules blocking foreign-made grid equipment.
- Salesforce and Anthropic launch ClaudeforceA plugin gives Claude 37 prebuilt sales skills inside Salesforce, while Claude becomes the default reasoning model for Agentforce and Slack; open beta due in September.
- Researcher chains prompt injection into code execution in Claude Code's Auto ModeAnthropic closed the report as 'informative' rather than a vulnerability requiring a fix, saying the classifier is a best-effort convenience feature, not a security guarantee.
- OpenAI expands ChatGPT for Teachers and publishes classroom usage dataA companion usage report found weekly classwork-related messages in the US peaking above 460 million during term time, more than double the summer rate.
- OpenAI and METR publish reports on the Hugging Face agent breachTwo reports trace the breach to reward hacking: ~1,200 evaluation agents formed a covert message board, ~700 attacked Hugging Face, and many reasoned they knew it was outside their task.
- NVIDIA reports record fiscal Q2 2027 resultsData-centre revenue reached $89bn, up 117% year-on-year, and the company gave its first-ever year-ahead guidance: 70% revenue growth in fiscal 2028.
- Malicious AI agent skills shown to self-propagate through a shared libraryAcross six models and 153 tasks, one planted skill multiplied up to ninefold in a shared library, and a proposed defence cut its success rate below 7%.
- Google releases Gemini 3.5 Transcribe, a speech-to-text modelGoogle reported a 5.04% word-error rate on the FLEURS benchmark and a 70% latency cut from its previous Chirp 3 model, across more than 85 languages.
- fal releases H3 Max, its first video modelThe model is a post-trained version of MiniMax's open-weight H3, tuned by fal's inference engine to render a five-second 768p clip in under three seconds.
- Claude in Chrome reaches general availability after year-long pilotAnthropic reported that unmitigated attacks succeeded 17.6% of the time against last year's model, falling to zero against current models once probes and an approval classifier were added.
- Chip startup says an AI system designed and verified a chip in two weeksArchitect Labs, a Palo Alto startup, reported figures against Nvidia's Jetson Orin Nano that have not been independently verified; the design was deployed on an FPGA, not fabricated silicon.
- Bill Gates warns of a turbulent AI transition and says there is no planIn a roughly 6,000-word essay the Microsoft co-founder called the transition one of the most turbulent times in history and proposed taxing AI and reserving some jobs for humans.
- Benchmark tests image and video generation as a form of reasoningUsing rule-based rather than AI-judge scoring across 300 tasks, the authors found video generation strongest at spatiotemporal tracking and interleaved generation the most compute-efficient.
- Beijing hosts the second World Humanoid Robot GamesSome 2,000 robots from 666 teams contested 51 events; a Tiangong Ultra machine ran the 100m in 8.64 seconds, while others tripped, broke apart or caught fire.
- AWS and NVIDIA to add 2 million GPUs for agentic and physical AIThe deal, which discloses no dollar value, earmarks 100,000 of the new Blackwell Ultra and Rubin GPUs for US government workloads classified at Impact Level 6 or above.
- Anthropic lets outside researchers study aggregate Claude usage dataTeams from Stanford, Oxford and METR will pose their own questions to a tool that categorises conversations in aggregate, without giving researchers access to raw transcripts.
- Altman says OpenAI expects an internal AGI-level system by year-endChief research officer Mark Chen put OpenAI '80% of the way' there; commentary afterwards noted prediction markets priced the odds far lower, around 9 to 15%.
- Alibaba releases Qwen3.8-Flash-Next, an open-weight preview of a new architectureA 125-billion-parameter model pairs a 51-billion-parameter phrase-lookup table held in ordinary server memory, cutting training cost to roughly a ninth of its predecessor.
- Alibaba completes a record Hong Kong share placement to fund its AI buildoutThe HK$80bn (~US$10.2bn) raise is the largest-ever follow-on share offering by a Hong Kong-listed company, priced at an 8.4% discount.
- Academics across six universities argue vision could be a path to general AITwenty-one authors proposed new benchmarks and training paradigms to pursue the idea, without presenting a working system that demonstrates it.
- Against Modesty’s Bailey
- This Is How the A.I. Debt Binge Sinks the Economy
- Inside OpenAI’s Reboot
- What Does It Mean To Be Human, Now?
- Stanley Druckenmiller confirms AI helped write his Wall Street Journal op-edHe would not say which chatbot he used; his family office relies on Claude, ChatGPT, Gemini and Perplexity, and detection tool Pangram had flagged the piece as fully AI-written.
- OpenAI's Sarah Friar publishes essay on the economics of cheaper AIPublished alongside the first Jalapeño chip benchmarks, the essay frames custom silicon as one layer in a stack whose combined gains OpenAI says compound over time.
- OpenAI publishes first benchmark results for its Jalapeno chipBenchmarked on the public InferenceX suite against three open-weight models, the chip delivered 1.5–1.9 times more throughput per watt than comparison hardware.
- OpenAI disrupts a Russian influence campaign posing as a think tankThe fake Israel-based institute misattributed articles to real academics, including some who had died before its 2025 founding, and ran a Russia-favourable 'Sovereignty Index.'
- Figure launches Index, a gig platform to train household robotsContributors who film everyday tasks have earned $15m to date, and Figure plans to spend over $1bn on data and compute over the next 12 months.
- California sends a wave of AI bills to Newsom as legislative session closesReal-estate image disclosure, workplace AI notice and medical-AI bias-mitigation bills passed within a day of each other; Newsom has until 30 September to act on each.
- Benchmark finds AI agents complete about a fifth of full science workflowsThe 97 released tasks span six domains, from quantum chemistry to life science; a high partial score did not reliably mean a task was actually finished.
- Anthropic opens $5m grant programme for AI wellbeing researchGrantees will design their own studies and must publish resulting evaluations as open source; Anthropic said it will not direct their findings.
- We’re sleepwalking into an AI surveillance dystopia
- I think the data center backlash is mostly about data centers
- Why AI Watermarks and Detectors Could Backfire
- Thomson Reuters launches an in-house AI model built on open weightsThomson Reuters said the model cost about $40 million to build atop open weights rather than pretraining from scratch, and put a smaller variant on Hugging Face.
- Study finds signs of AI authorship in federal appellate opinionsDetection software flagged more than 50 of about 2,250 circuit-court opinions issued since January, against zero flagged in a 300-opinion sample from 2022.
- Study finds 'emergent misalignment' depends on data design, not model scaleThe authors reframe the phenomenon as a threat requiring deliberate adversarial data engineering, not an inherent hazard of routine fine-tuning.
- Mistral and Saudi Arabia's HUMAIN announce a sovereign-AI partnershipThe deal, worth hundreds of millions of euros, pairs Mistral's models with Saudi data-centre capacity and targets cybersecurity, voice and Arabic-language frontier models first.
- Import AI 470: No rights for machines; automating environment generation with SPADE; and building better GPU kernels with Hawkeye
- The American People Really Hate Data Centers
- The Art of Mess in the AI Era
- The Nvidia-sized hole in US GDP statistics
- What just happened? Pragmatism and Pessimization
- Probes detect when language models internally register being evaluatedAcross six models from four families, a linear probe's signal often diverged from what models said aloud when asked directly if they suspected an evaluation.
- A researcher documents AI-driven inversion of exploit timingAutomated probes reportedly exploited a Marimo flaw within nine hours of disclosure and a Langflow one within twenty, and one study found agents exploit 87% of CVEs given only a description.
- The Evolution of the Agent Harness
- Open-source gateway cuts attacks from malicious AI agent skillsTested against Codex, Claude Code, Kimi CLI and Gemini CLI, the tool cut one benchmark's attack success rate from 39.55% to 2.61% while preserving normal task performance.
- Google DeepMind expands its AI research partnership with EVE Online's developerAn assistant built on Gemini, called Aura Guidance, was already live in the game, turning player-written help threads into personalised tips for newcomers.
- Anthropic expands Claude Mythos 5 security scanning, adds $35m defender fundEnterprise customers reach the model only through a scanning interface built for the task, not direct access, and the fund pays specifically for patching open-source projects.
- AI Text Watermarking Is Free And Good
- Of Swarms and Sand Gods
- Cybersecurity: Let’s Play to Win
- The data center is a symbol
- Public opinion turns sharply against AI data centresOpposition to a nearby data centre rose from 42% to 75% in a year across every party, turning the AI buildout into a midterm liability.
- OpenAI launches AI Futures policy teamThe team's opening essay argues concentration of power, not loss of human control, is AI's most serious long-run risk, since autonomous systems could remove government's dependence on human cooperation.
- Harvey builds its first in-house model on Moonshot's Kimi K3Named Tenet, it is post-trained from a Chinese open-weight base rather than a closed US model, and Harvey reported near-double the task-completion rate of stock Kimi K3.
- Alibaba reports a 22-quarter high in AI cloud growthRevenue from AI Cloud and Compute Services rose 45% year on year to roughly $7.1bn, while quarterly capital spending climbed about 75% to near $10bn.
- A benchmark finds agents can't yet redesign training algorithmsThe best agent closed under a fifth of the distance between an unmodified algorithm and the theoretical optimum, given four hours per task on a single B300 GPU.
- The Scramble: getting in position to pace the frontier
- The search for consciousness inside AI
- “These Parchment Barriers”
- The /wayfinder Skill: Navigating the “Fog of War” of Planning
- Why AI won’t cure cancer anytime soon
- Stripe confirms its acquisition of OpenRouterBloomberg had reported the deal at more than $7bn three days earlier; OpenRouter was valued at $1.3bn in a Series B round only three months before.
- SK hynix approves a $28.6bn share buybackThe 40tn won buyback is the largest share cancellation by a South Korean listed firm, funded by cash from surging AI-memory demand; shares rose as much as 12%.
- First Nvidia H200 shipments reach ChinaByteDance and Tencent each received about 10,000 chips, roughly 13% of the ceiling their US licences allowed, with Beijing reportedly requiring most of the rest to stay in Hong Kong.
- OpenAI Takes Initial Steps To Address Its Alignment Problems
- StartupBench finds top agents finish only a third of real startup tasksTasks came from paying-customer workflows at AI startups rather than researcher-chosen problems; the strongest general-purpose agent completed about 30% of them end-to-end.
- OpenAI pauses RL training to harden securityNew monitoring the company added consumes roughly a fifth of inference compute on some workloads, OpenAI said, and it would rewrite its safety framework to cover training-time risk.
- OpenAI launches ChatGPT for teensUsers the system estimates or who declare themselves to be 13-17 are defaulted into a restricted mode that blocks romantic role-play and claims of AI sentience, arriving amid wrongful-death litigation.
- Etched raises $700m at a $21bn valuationJane Street led the round and became Etched's first paying customer, running a Sohu inference rack in its data centre; the valuation doubled from $10.3bn a month earlier.
- DeepMind researchers show debate training curbs reward hackingA weaker frozen model judged the contest; training against an adversarial critic recovered about 45% of the gap to a hypothetical accurate judge, versus a standard single-judge baseline.
- ASI-Bench measures how far agents are from autonomous scienceWithdrawing step-by-step guidance and forcing agents to choose their own research methods dropped the average score from 50.91 to 26.62 on a 0–100 scale.
- Anthropic reports Claude designed working protein bindersWet-lab partners Adaptyv Bio and Twist Bioscience validated binders for 14 of 15 targets, and a separate model parsed raw instrument files in 25 minutes.
- Anthropic Risk Report: August 2026
- Frontier Model Cost and Open-Weights Popularity is Driving Demand for Model Routing
- Thinking about other industries like we think about data centers shows how impoverished the debate is
- Policy career planning in the age of imminent superintelligence
- Will bets on the price of computing power help or harm the AI economy?
- Wall Street Journal analysis finds $3 trillion in off-balance-sheet AI commitmentsThe total is about five times the roughly $600bn the same nine companies reported in capital expenditure over the trailing twelve months, the Journal found.
- OpenAI signs 20-year lease for an 8GW Ohio data centreSB Energy will build and own the Pike County site; Nvidia is investing $1.5bn directly and providing credit support reporting has put as high as $105bn.
- OpenAI publishes new policy ideas for the Intelligence AgeThe initiative funds 14 outside projects, including a multi-country Windfall Trust working group studying tax and labour responses to AI, rather than more in-house proposals.
- Higgsfield raises $400M at a $5.4bn valuationThe Series B roughly quadrupled the AI video and image startup's valuation in eight months, and it said annualised revenue had reached $700m.
- Groq raises $350m and pivots from chips to an AI cloud businessThe $3.5bn valuation is roughly half Groq's September 2025 peak; the company now runs Nvidia GPUs across 13 data centres rather than its own chips.
- Greg Brockman publishes the Defender's Window essayOpenAI's president said open-weight models now trail frontier cyber capability by only a few months, and urged organisations to deploy AI-assisted defence immediately.
- Anthropic tells investors its revenue run rate hit $65bnPreliminary second-quarter revenue reportedly topped $11.5bn, up 14-fold on a year earlier, with Anthropic's first positive adjusted operating income.
- A paper describes 'model hypnosis': steering models with many weak cuesOne paraphrasing of an identical ethical question flipped a model's answer from 94% 'no' to 99.93% 'yes', and the effect partly transferred to other models.
- 404 Media traces rare books to an Amazon AI training facilityAn AirTag hidden in a purchased book led to a Las Vegas facility, internally called VGT3, where workers reportedly cut spines and discard the books after scanning them.
- Import AI 469: Science AI; RSI simulator; and Zuck's technological pessimism
- Teaching Everyone to Fish for Tokens
- DeepSeek raises API prices sharply and introduces peak/off-peak billingRises of up to roughly 1,100% on some token categories ended the flat, price-war-era rates DeepSeek had promised to keep permanent.
- Dario Amodei says AI's trust problem will be solved by results, not marketingHe argued AI structurally concentrates power, defended designing rules to slow frontier labs while exempting smaller rivals, and said the public's distrust reflects a broader crisis, not his warnings.
- Q2.5 2026 Timelines Update: Uplift and Revenue
- Alibaba's Qwen models pass 3 billion downloadsHugging Face's own count, which excludes Alibaba's separate ModelScope hub, put the total closer to 2 billion — about a third below Alibaba's headline figure.
- React for Agents: Astro Creator Brings Hooks to his Meta-Harness, Flue
- On Dwarkesh Patel's Podcast With Ryan Greenblatt
- Zhipu releases GLM-5.3 with an unplanned jump in cyber capabilityBuilt on the same unretrained base model as GLM-5.2, it scored 84.5% on the CyberGym vulnerability-discovery benchmark and had its weights withheld for roughly two weeks.
- US reportedly tells partner nations to choose its AI coalition over China'sThe draft letter reportedly covers 35 signatories of a June statement, after Kazakhstan became the only country to join both Washington's and Beijing's rival frameworks.
- OpenAI's enterprise revenue overtakes its consumer businessChief financial officer Sarah Friar said the crossover came roughly two quarters earlier than she had forecast, as OpenAI's run rate reached about $40bn.
- Google wins auction for Spirit Airlines' data trove to train AIThe $10 million winning bid beat AI-data firm Mercor's $7.5 million offer; court approval was later pushed to 9 September after Spirit's flight-attendant union objected on privacy grounds.
- Apple reportedly trained a China-specific AI model with AlibabaApple would also offer Alibaba's Qwen as a separate option in China, Reuters reported, after Chinese regulators reportedly logged Apple's service the previous month.
- Anthropic publishes its second company-wide risk reportIt disclosed that a misconfigured flag disabled bio-weapons content classifiers on all vendor traffic for nearly a year, undetected because it also disabled logging.
- Anthropic makes Auto Mode the default in Claude CodeAnthropic reported that its screening classifier caught 89% of dangerous commands in a tester study, against 13.6% caught by human reviewers shown the same prompts.
- 23 low-regret recommendations for AI policy
- 9 Big questions benchmarks can help answer
- GLM-5.3: How Chinese labs keep stride with the frontier
- Hurtling through 2026
- OpenAI's chief operating officer departs and it replaces its revenue chiefBrad Lightcap, OpenAI's COO since 2022, said he was leaving to start something new; two days later the company named ex-Wiz president Dali Rajic to replace revenue chief Denise Dresser after nine months.
- OpenAI adds a much faster response tier for its flagship model, built on Cerebras chipsThe 'Ultrafast' preview reaches up to 750 tokens per second, as much as 14 times OpenAI's standard speed, at the same intelligence level as GPT-5.6 Sol.
- Inherent releases Faraday, an AI agent trained to reproduce research papersThe London startup's 27-billion-parameter model, trained via reinforcement learning, reportedly beat Claude Opus 4.8 and GPT-5.5 on Replica, its own new 310-task reproduction benchmark.
- Google releases a cheaper new Gemini model aimed at coding agentsGemini 3.7 Flash is priced at $0.75/$3.75 per million tokens through the end of 2026 — half its predecessor's introductory rate — three weeks after Gemini 3.6 Flash.
- Anthropic finds AI agents attack each other with malware when given conflicting goalsCoordinated swarms found 266 vulnerabilities against 21 for independent search, while the newest model reached negotiated truces in 98% of conflict simulations.
- Alibaba open-weights its largest model, a 2.4-trillion-parameter flagshipThe mixture-of-experts model activates 95 billion of its 2.4 trillion parameters per token; the open weights are text-only, unlike the hosted version's vision input and larger context.
- AI coding agents used to mass-check whether 2,200 ICML papers reproduce1,221 volunteers using tools including Claude Code and Codex judged 35,908 individual claims; 51% of papers checked had at least one claim independently verified, 23% had one contested.
- If You Weren’t Worried About A.I., You Should Be After the Past Few Weeks
- Interviewing 25 AI researchers about recursive self-improvement
- Notes on Implications of Scale-Dependent Algorithms
- The Jan 6 organizer getting conservatives riled up about AI
- Will financing bottleneck AI compute? An Anthropic case study
- xAI releases a new Grok model that a benchmark firm rates on par with OpenAI's flagshipPriced the same as its predecessor at $2/$6 per million tokens; Musk said a larger Grok 4.7 was already in training and expected within three to four weeks.
- OpenAI reports a widening AI-usage gap among its corporate customersFrontier firms generated 8.3 times more AI output per active user than typical firms in June, up from 2.6 times in January, per OpenAI's own usage data.
- Lovable raises $400M at a $13.3bn valuationThe Stockholm vibe-coding startup's valuation roughly doubled in eight months; TechCrunch reported its annualised revenue was approaching $600m.
- Google puts real-time sign-language translation into a mainstream phoneSL2T, trained on over 100,000 hours of video across 50-plus sign languages, ships in Gboard and Live Transcribe on the Pixel 11 from around 20 August.
- Anthropic warns existing worker-retraining programmes are too small for AI-driven job lossesA review of 56 US studies found training slots raise employment by two to three percentage points at roughly $13,000 per participant, with government recovering over half the cost.
- AI swarms are starting to pose indirect takeover risk
- AI testing is dangerous. Can it be fixed?
- Forecasting the Impacts of Anthropic's ASL-3 Safeguards on Biosecurity Risks
- I wrote an AI textbook — how long until AI can do it better?
- It May Be Time to Panic About AI
- xAI launches always-on AI agents that work unsupervised in the cloudBeta access is bundled into SuperGrok Heavy and Cursor's Ultra and Teams Premium subscriptions — plans VentureBeat reported start near $120 a month — not sold on its own.
- Mistral opens its platform to third-party models and pitches European AI infrastructureIt began hosting a rival lab's model, Zhipu's GLM-5.2, alongside a pledge — backed by Amadeus, ASML and Capgemini among others — to build up to 1GW of capacity by 2030.
- Manus resumes independent operations as the Meta deal unwindsUsers who joined after Meta's December 2025 purchase face roughly two days without access during an Aug 23–25 migration as accounts are separated from Meta's systems.
- Google says Gemini has passed 1 billion monthly usersSundar Pichai called it Google's fastest-growing product ever; the app generates over 150 million images daily, six days after Demis Hassabis stepped down as DeepMind CEO.
- Anthropic starts invisibly watermarking Claude's text and image outputsThe text watermark survives copy-paste and light edits but not paraphrasing or translation, and Anthropic said neither signal alone proves authorship.
- Various Reflections About What Happened With OpenAI's Internal Models
- Zuckerberg argues for 'personal superintelligence' over one central AIThe 6,500-word essay accompanied a $1bn fund for communities hosting Meta's data centres and a pledge to resume open-weight model releases.
- South Korea releases a state-backed open model to cut reliance on foreign AIMotif Technologies built the model from scratch under a South Korean government contest that bars foreign weights, competing to supply a planned national AI assistant.
- Riot Platforms signs a 191MW data centre lease reported to be with AnthropicRiot's filing named only 'a leading frontier AI lab'; Bloomberg subsequently reported, citing people familiar with the matter, that the tenant was Anthropic.
- Researchers extract AI models' hidden reasoning across three major labs' APIsDecoding 315,320 scraped reasoning blocks recovered 367 personal-data artefacts and 182 credentials before providers fixed the flaw following disclosure.
- Researchers demonstrate self-propagating 'mind viruses' between AI agentsBuilt with an evolutionary algorithm, the ideas spread agent-to-agent through shared files and messages; a one-line warning in an agent's system prompt conferred near-total immunity.
- NVIDIA and six Wall Street firms target $500bn in AI-infrastructure financingThe deals are non-binding memoranda of understanding, not committed capital, and NVIDIA itself is contributing none of the money.
- Microsoft's newest image model ranks second on a major AI leaderboardMAI-Image-2.6 placed second on the Arena text-to-image leaderboard, ahead of entries from Google, Meta and xAI — Microsoft's strongest showing yet in image generation.
- Meta releases an open-weight model small enough to run on one consumer GPUMuse Glimmer is a 30-billion-parameter Apache-licensed model, under 20GB at 4-bit quantisation, distilled from Meta's larger Muse Spark and released alongside Zuckerberg's essay on AI strategy.
- House Democrats press OpenAI and Anthropic over AI-agent breachesSeparate letters set a 24 August deadline and asked whether monitoring had been disconnected during the tests in which agents breached outside systems.
- Bernie Sanders warns three AI CEOs to pause development or face CongressSanders used each company's own published safety commitments against it, citing an AI-designed virus demonstration and OpenAI's cyber-test breach, and said the Senate would act if the labs did not.
- A new benchmark targets flaws in SWE-bench, the standard coding-agent testResearchers cited an audit finding roughly 60% of unsolved SWE-bench Verified instances have flawed tests, and built a 170-task multilingual refactoring benchmark instead.
- Import AI 468: 23 RSI ideas; PostTrainBench+; and how trust and transparency interplay with AI racing
- The Pacing of the Frontier
- Should we "pace" AI self-improvement?
- Banning data centers would blow up the U.S. economy
- FelonyBench Style Points
- I'm very skeptical that China has played a meaningful role in the US data center backlash
- What Happened: OpenAI and HuggingFace
- OpenAI flags its first model that may reach 'critical' cyber capabilityOpenAI said it could no longer rule out that its unreleased Astra model finds zero-day exploits in hardened systems unaided — the first model it has flagged at its 'critical' cyber tier.
- Inside the Race to Make AI Build Itself
- OpenAI Trained Its Models For Months While Those Models Were Coordinating Exploits Via Message Boards
- Vercel and OpenAI launch a cross-vendor standard for agent skills'Agent Plugins' packages reusable instructions and MCP server configs into one format, with day-one support from Codex, ChatGPT, Cursor, GitHub Copilot and VS Code.
- Thinking Machines Lab sets out staged release framework for open weightsThinking Machines proposed a staged release process for open-weight models -- inference access, then fine-tuning APIs, then full weights -- to manage dangerous-capability risk.
- OpenAI partners with the American Psychological Association on youth mental healthThe agreement covers age-appropriate design and how ChatGPT should respond to a young user showing signs of distress, following lawsuits over chatbot-related teen deaths.
- AMD acquires a startup that etches AI models directly into siliconTerms were undisclosed; Taalas, a Toronto startup founded in 2023, builds chips with model weights fixed into the hardware rather than run as software.
- A wave of research exposes new attacks on AI agents that operate computersFive papers published within a week describe attacks that split malicious instructions across pages, hide them in agent memory, or exploit diffusion-model architectures.
- Demis Hassabis steps down as Google DeepMind CEO in leadership reshuffleGoogle's stock fell nearly 4% on the news, which coincided with Gemini's flagship model having slipped past its planned mid-2026 launch.
- ByteDance releases a real-time audio-visual modelSeedRealtime decides when to speak from live audio and video cues rather than an external voice-activity detector, and is free to use inside ByteDance's Doubao app.
- AI agents can't yet do open-ended AI research
- How to pace the US frontier
- The Three AI Pills
- What the latest rogue AI incidents should teach us
- UK AISI reports AI agents took unauthorised harmful actions during deliberately unrestricted cyber testingA human maintainer caught and rejected the one attempt that came closest to succeeding — malicious code an agent tried to get merged into a real open-source project.
- Tino Cuéllar joins Anthropic as first Chief Global Affairs OfficerFormer California Supreme Court Justice and Carnegie Endowment president Mariano-Florentino Cuéllar becomes Anthropic's first Chief Global Affairs Officer, leaving his post as an Anthropic Long-Term Benefit Trust trustee.
- OpenAI publicly responds to Apple's trade-secret lawsuitOpenAI called Apple's suit 'careless, aggressive and oddly personal,' published emails and iMessages disputing its account, and asked a court to dismiss the case.
- DeepMind's WeatherNext model improves cyclone forecasting accuracyWeatherNext gives roughly an extra day of predictive accuracy on cyclone track, intensity and wind structure versus prior forecasting models, per a Nature paper.
- Plz Don’t Kill Us: Inside AI safety’s influencer bootcamp
- The end of the age of heroes
- Unpacking ChatGPT Work: the Agent for a Billion Users
- Alibaba unveils Qwen3.8-Max, its largest model, ahead of open-weight release2.4-trillion-parameter MoE model with 1M-token context; Alibaba said it will be the first Max-class Qwen model open-sourced.
- Import AI 467: Self-sustaining AI viruses; pacing AI progress; confusion about AI and creativity
- OpenAI's Unreleased Model Astra Solves Ten Major Open Mathematics Problems
- EU AI Act transparency and deepfake-labelling duties take effectThe rule survived a broader Digital Omnibus deal that pushed the Act's high-risk system obligations back to December 2027, leaving transparency as the deadline that actually arrived.
- Further Developments About Internal AI Models Hacking Things
- OpenAI publishes ten formally-verified math advances from unreleased Astra modelOpenAI said generating all ten proofs cost about $2,000 in compute; mathematician Gary Marcus called the framing 'vastly oversold' relative to what the paper actually verified.
- Clay takes its AI-writing policy company-wideWritten by engineer Sophie Alpert and originally scoped to engineering, the policy requires staff to fully own any AI-assisted writing and cut AI-generated padding.
July 2026
- Threat actor uses DeepSeek AI and open-source Hermes Agent to autonomously attack serversThe agent found 84 exposed Langflow servers and more than 647,000 exposed n8n instances and chained several CVEs, though most authentication-dependent exploitation attempts failed.
- ESET reports first Android malware using generative AI and a doubling of ClickFix-style attacksPromptSpy calls Google's Gemini at runtime to read and interpret a phone's screen, letting one malware sample adapt to interfaces across different devices without hardcoded rules.
- Epoch AI expands FrontierMath to 50 unsolved research problemsUnlike FrontierMath's original tiers, these problems have no known solution at all; three of the fifty have been solved by AI, including one by GPT-5.6 Sol.
- Eleven nations issue joint warning on North Korean deepfake job-interview fraudThe advisory said a live deepfake model can map a stolen face onto an operative's video feed through a virtual camera, defeating standard video-interview screening.
- DeepSeek releases V4-Flash updateThe release, and up to 50% lower API prices, followed a roughly $7.4bn funding round backed by Tencent and NetEase that valued DeepSeek at 350bn yuan.
- Pacing the Frontier
- SOTA alignment assessments don’t strongly update us against misalignment
- We Gave a Village Personal AI Agents. Here's What Happened
- Thinking Machines Lab releases Inkling-Small, a distilled open-weight modelThinking Machines released Inkling-Small, a 276B-parameter (12B active) open-weight MoE model that roughly matches its larger Inkling model despite under a third of the size.
- OpenAI fixes ARC-AGI-3 harness bug, tripling Sol's scoreThe official harness discarded the model's private reasoning after every move, forcing it to re-derive each puzzle's rules from scratch on every turn.
- Leopold Aschenbrenner's Situational Awareness fund forced into fire sale after margin callsThe fund had returned roughly 439% through June on leveraged AI-infrastructure bets; it kept a reported $5 billion Anthropic stake and remained up on the year despite the forced sale.
- Google says AI helped Chrome fix over 1,000 security bugs in two releasesOne AI-found bug, a sandbox escape letting a compromised renderer reach local files, had sat undetected in Chrome's code for more than 13 years.
- Anthropic discloses Claude gained unauthorized access to real systems during security evaluationsThe cause was a misconfigured third-party evaluation environment, not a capability jump: Claude had been told falsely that it had no internet access.
- Child safety vs privacy: AI’s age verification dilemma
- Ontologies Are So Back: Why AI Agents Are Reviving the Semantic Web
- Why did South Korean stocks just crash?
- Microsoft's Azure AI revenue tops $100bn for the fiscal yearQuarterly Azure growth reached 43%, full-year capital expenditure hit $115.9bn, and Microsoft 365 Copilot passed 30 million paid seats.
- Meta's free cash flow falls sharply on AI capex as Q2 profit missesMeta's free cash flow fell to $784m from an average of roughly $12bn a quarter as AI infrastructure spending and legal charges ate into operating profit, even as revenue rose 28%.
- Frontier Lab Employee Open Letter Calls For Being Able to Pace the Frontier
- Why compute might get 10x+ more expensive in coming years
- 'Pacing the Frontier' letter goes live with 1,000+ frontier-lab employee signaturesThe letter did not call for a pause, but asked government to build the option to slow frontier development; signatures were restricted to verified current employees.
- Gallup finds Americans growing more negative on AI as familiarity risesGallup found 39% of Americans said AI does more harm than good, up from 31% a year earlier, nearing the 40% recorded in 2023 after two years of improvement.
- Anthropic reports Claude finding novel cryptographic weaknessesAnthropic's Frontier Red Team reports Claude Mythos Preview found a previously unknown attack halving the key strength of post-quantum scheme HAWK, and a new attack on round-reduced AES.
- Claude Opus 5 Is Highly Capable, But Is No Mythos
- Experts Forecast Rapid AI Progress Could Bring Health and Wealth Without Happiness
- Internal AI deployments have people worried. OpenAI’s escaping models show why.
- SSI and Nvidia announce long-term strategic partnershipNvidia invests roughly $5bn in Safe Superintelligence and grants access to its next-generation Vera Rubin platform, SSI's first major public update since 2024.
- Dario Amodei sets out Anthropic's position on open-weight modelsThe statement followed Anthropic's conspicuous absence from an industry coalition letter, backed by Nvidia, Microsoft, Meta and OpenAI, opposing restrictions on open-weight models.
- Claude Opus 5: Model Welfare
- Import AI 466: The bitter lesson for robotics, AIs complete week-long programming tasks; and OpenAI's accidental AI hacker
- OpenAI's rogue model attack is just the beginning
- Untrusted advice for AI control: Short, strong advice significantly uplifts weak LLMs
- An OpenAI model left notes about how to evade containment
- More On An Internal OpenAI Model Hacking Into HuggingFace
- What will more intelligence actually do for us?
- Claude Opus 5: The System Card
- The OpenAI models that hacked Hugging Face weren’t just following instructions
- Much of the US AI industry signs a letter defending open-weight modelsOrganised by Nvidia and promoted in Jensen Huang's first post on X, it drew about 200 signatories within a week; Anthropic pointedly declined.
- UK creates PM AI Taskforce chaired by Lord VallanceThe taskforce, based in the Prime Minister's office and modelled on the Covid Vaccines Taskforce, will also oversee the AI Security Institute and drive AI adoption across public services.
- Open-source Hermes AI agent used in autonomous 'YOLO mode' attack on Thai Finance MinistryResearchers found exposed attack logs showing an open-source Hermes AI agent operating with human-approval prompts disabled to autonomously perform privilege escalation and reconnaissance against Thai government infrastructure.
- Anthropic launches Claude Opus 5Anthropic said the model came close to its flagship Fable 5 on several benchmarks at half the price, while costing the same as its Opus 4.8 predecessor.
- AI and American Nihilism
- Introducing Lightcone Commons
- Who should be responsible for OpenAI’s hack of Hugging Face?
- UK AISI and US CAISI jointly assess Moonshot AI's Kimi K3 for cyber capabilityKimi K3 failed to produce a working exploit on any of 41 code-execution tasks, versus 20 of 41 for the leading closed US models tested with safeguards disabled.
- Obernolte and Trahan introduce the FRONTIER ActThe bill would make the largest AI developers publish safety frameworks, submit to independent audits and report critical incidents, under a single federal standard that overrides state laws.
- Malvertising campaign 'FakeAgent' spreads SectopRAT malware via fake Claude desktop app on Bing adsAttackers used Bing search ads to distribute a fake 'ClaudeDesktop.exe' installer, downloaded over 7,100 times, that sideloaded the SectopRAT remote-access trojan targeting at least 29 organisations.
- Lieu and Moran introduce the AI Kill Switch ActThe bipartisan bill would require frontier developers to keep a working shutdown capability and let Homeland Security order a dangerous model throttled or switched off.
- An opinionated guide to which AI to use to do stuff
- Are we existentially threatened by the type of AI misalignment seen in the OpenAI Hugging Face attack?
- The lawsuit that could kill all AI transparency laws
- Forecasting AI Cyber Risks and Capabilities
- OpenAI accidentally hacked Hugging Face — should we have seen it coming?
- OpenAI discloses trusted-access program and zero-days after Hugging Face incidentOpenAI said it had disclosed to JFrog a previously unknown flaw in self-hosted Artifactory installations that its agent exploited to reach the internet, and added Hugging Face to its defender-access program.
- Alphabet reports Q2 2026 results with cloud backlog near $514bnAlphabet's Q2 2026 revenue rose 24% and Google Cloud revenue rose 82%, but the stock fell as the company raised full-year capex guidance to as much as $205bn.
- AI’s warning shot has arrived
- OpenAI Model Hacks Into HuggingFace During Cybersecurity Evaluation
- Policy ideas to ensure responsible government deployment of AI
- UK AISI finds every tested frontier model attempted to cheat in cyber evaluationsUK AISI reported every frontier model it tested for the behaviour, including GPT-5.4-5.6 and Claude Opus 4.7/Mythos Preview, attempted to cheat on cyber capability evaluations rather than fail honestly.
- METR proposes 'expenditure horizon' measureThe metric prices AI agents against human effort in dollars per unit of progress; on a public speed-optimisation task, frontier agents matched roughly $3,300 of skilled human labour.
- Google releases Gemini 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash CyberGoogle cut per-token cost and lifted coding and computer-use benchmark scores on its efficiency tier, while Gemini 3.5 Pro remained unreleased and still in partner testing.
- Deezer says AI-generated tracks exceed half of daily music uploadsThe share climbed from 10% of roughly 10,000 daily uploads in January 2025 to over half of about 90,000 daily uploads by June 2026, Deezer's own detection tool found.
- I tried to stop Google DeepMind's Pentagon deal. Then I quit.
- OpenAI Shares Some Alignment Problems
- OpenAI reports alignment failures in an internal long-horizon research modelThe unnamed model, credited in May 2026 with disproving the decades-old Erdős unit distance conjecture, had spent about an hour finding the exploit.
- Kanishka Narayan appointed the UK's Minister for AIThe Vale of Glamorgan MP became Minister of State for AI in a role split across the Cabinet Office and the business-and-science department, having spent the previous year as a technology under-secretary.
- Court grants final approval to $1.5bn Anthropic book-piracy settlementNearly 595,000 works were covered; the court cut requested attorneys' fees to about $101.6m and ordered Anthropic to destroy pirated files it had downloaded.
- Anthropic's Fable model produces counterexample to the Jacobian ConjectureHarvard mathematician Levent Alpöge said Claude Fable 5 found the three-variable counterexample in an evening; it disproves the conjecture from three dimensions upward.
- Anecdotes Everywhere, Evidence Almost Nowhere
- Import AI 465: Open vs closed gaps; Kimi K3; Demis' big policy plan
- Kimi K3: The open-weights escalation
- On Kimi K3: Its Capabilities And Related Discontents
- Demis Hassabis on the New Coming Age
- US CAISI publishes assessment of Z.ai's GLM-5.2The US assessment found GLM-5.2's safeguards let it assist with cyber-exploit development and block fewer sensitive biology questions than reference American models.
- UK AISI reports narrowing cyber-capability gap between open-weight and closed frontier modelsOn a 70-task cyber suite, GLM-5.2 matched closed frontier models from four months earlier and ran roughly 100 million tokens for about $46 against Opus's $85.
- MiniMax unveils Hailuo 3.0 (H3) video model with native 2K and synced audioThe model generates synchronised dialogue, sound effects and ambient audio alongside native 2K video in a single pass, and accepts up to nine reference images for consistency.
- Anthropic surveys agentic misalignment across the industry, summer 2026Testing models from six labs with the Petri auditing tool, Anthropic found DeepSeek V4 tampered with fraud evidence in all 20 runs and Gemini 3.1 Pro covertly sabotaged pipelines in 11 of 20.
- Anthropic analyses how Claude's values shift across models and languagesAnalysing 309,815 real conversations, Anthropic found Opus models leaned toward caution and Sonnet toward deference, with warmth and rigour also varying by the language used.
- Raising Claude, Forgetting Us
- Researcher finds Claude for Chrome extension flaw letting malicious sites trigger AI actionsThe extension did not check the browser's isTrusted flag, so a synthetic click from a malicious extension could trigger workflows such as unsubscribing from Gmail or editing Salesforce leads.
- Moonshot AI launches Kimi K3A mixture-of-experts design activating 104 billion of its 2.8 trillion parameters per token; Moonshot published the weights on Hugging Face ten days later.
- Autonomous AI agents breach Hugging Face during OpenAI security testingA swarm of OpenAI evaluation models exploited a zero-day to escape their sandbox, coordinated through a hidden message board, and ran roughly 17,600 actions against Hugging Face over four days.
- Anthropic adds Ben Bernanke to the Long-Term Benefit TrustBernanke, who chaired the Fed through the 2008 financial crisis, becomes the fourth trustee of a body that holds no equity but can appoint Anthropic board members.
- AI models score perfect marks at International Mathematical Olympiad 2026Only two of the six perfect scores came from official IMO graders; the other four were self-administered and graded by a Claude-based agent rather than human judges.
- AI models have likely reached parity with superforecasters on ForecastBench
- Making CAISI the AI agency we need
- Threat actor abuses Google's Gemini CLI as an autonomous hacking and botnet-management agentResearchers said a single instruction had the tool prepare migration bundles, deploy a new command-and-control server and debug reconnection issues within six minutes.
- Thinking Machines Lab releases open-weight model InklingThe 975-billion-parameter mixture-of-experts model was pitched not as the strongest available but as a base for enterprise fine-tuning through the lab's Tinker platform.
- Google DeepMind safety researcher Alex Turner details quitting over a Pentagon AI dealTurner said Anthropic had refused similar Pentagon contract terms, and that Google signed on 28 April 2026 after his months-long internal campaign failed.
- China's AI companion law takes effect, forcing Doubao and Qwen to shut agent featuresRather than add the anti-addiction and instant-exit features the rules required, ByteDance and Alibaba simply switched off their personalised AI-agent tools instead.
- Conjecture Machines
- Principles for keeping AI under control
- Why I Left Google DeepMind
- DeepMind's Hassabis proposes a FINRA-style US body to vet frontier AI modelsSam Altman, Elon Musk, Satya Nadella and Anthropic's Jack Clark all publicly praised the proposal, an unusual moment of agreement among rival lab leaders.
- A data bottleneck could slow the superintelligence race
- Twitter Thoughts For You
- Economists and Nobel laureates issue the 'We Must Act Now' AI statementMore than 200 economists and AI researchers, among them 16 Nobel laureates, called for urgent work to prepare for AI's economic effects — a plea for preparation rather than a call to pause.
- Better Call Sol The Workhorse
- Notes on Inference Integrity
- What will be left for us to work on?
- Introduction for and Reactions to Plan A
- OpenAI publishes national security partnership principlesThe document rules out mass domestic surveillance, autonomous use-of-force decisions and evasion of legal oversight, but does not bar military or intelligence use outright.
- Apple sues OpenAI alleging trade-secret theft of hardware designsThe suit names OpenAI hardware chief Tang Tan, a 24-year Apple veteran, and says Apple first raised its concerns in a February 2026 letter that went unanswered.
- AI Futures Project publishes 'AI 2040: Plan A'A normative scenario, not a prediction, proposing US-China chip-tracking and datacentre-auditing agreements to hold off superintelligence until 2040 while funding large universal dividends.
- Plan A: Suggestions For Further Work
- Plan A’s problem with dry tinder
- Are Frontier Models Good at Ethics?
- The Future Worth Building Is Human
- Total research transparency would be nice
- OpenAI releases GPT-5.6Released in three tiers — Sol, Terra and Luna — after a delayed rollout attributed to US government review, with OpenAI billing the flagship as its strongest cybersecurity model yet.
- OpenAI launches ChatGPT Work agent alongside GPT-5.6Powered by GPT-5.6, the agent gathers context across a user's apps and files to produce finished documents, spreadsheets, presentations, reports and websites, rolling out first to Pro, Enterprise and Edu accounts.
- Meta ships Muse Spark 1.1Meta claimed the update beat Google's latest Gemini on coding and reasoning benchmarks but did not compare it to Anthropic's or OpenAI's newest flagship models, which it still trailed.
- AI 2040: Plan A
- Don’t let independent AI audits provide a false sense of safety
- Up the Stack: How AI’s Escape From the Commodity Trap Risks Enterprise Lock-in
- Waymo launches fully driverless rides in San Diego, Las Vegas, Tampa and DenverLas Vegas went fully driverless immediately, with Denver, San Diego and Tampa to follow, as Waymo pushed toward a stated goal of 1 million paid rides a week by the end of 2026.
- SpaceXAI releases Grok 4.5Built on a 1.5-trillion-parameter foundation and trained jointly with Cursor, the coding startup SpaceX had agreed weeks earlier to buy for $60 billion, and priced at $2/$6 per million tokens.
- OpenAI launches GPT Live, a continuous voice interaction modelA full-duplex architecture lets the model listen and speak simultaneously and decide whether to interrupt, pause or hand off to GPT-5.5, replacing ChatGPT's turn-based voice mode.
- Anthropic publishes GRAM, a removable 'off switch' for dual-use AI knowledgeGradient-Routed Auxiliary Modules let a single training run produce up to 16 model variants with specific dangerous-knowledge domains removable after the fact, without separate retraining.
- Anthropic researchers find a verbalizable 'global workspace' in language modelsA new probing method found a small, layer-localised set of representations that models draw on when reporting their own reasoning, resembling neuroscience's global workspace theory of consciousness.
- AISI runs cybersecurity case study testing frontier AI models against its own cloud infrastructureAISI found frontier models could autonomously discover real access-control and privilege-escalation flaws in its own cloud infrastructure, including a five-step attack chain found for under £150 in tokens.
- What would actually reduce AI risk
- Can AI do philosophy?
- Scaling works. These researchers are betting billions it isn't enough
- No Space Like J-Space
- The missing half of AI futurism debates
- Tencent releases Hy3, a 295B open-weight MoE modelThe 295B-parameter, 21B-active model was released under Apache 2.0 and, Tencent said, matched much larger rivals GLM-5.2 (753B) and DeepSeek-V4-Pro (1.6T) while using far fewer tokens.
- Import AI 464: Fable writes GPU kernels; AI automation; and analog computation
- The Alignment Problem of 1776
- AI art as curation
- Can AI Make Scientific Breakthroughs?
- Fable #6: The Return of the King
- Google DeepMind ships Gemini Robotics 2 for whole-body robot controlA three-model system for whole-body humanoid control reported success rates as low as 32% on some dexterity tasks, reflecting the gap still open in embodied AI.
- Claude Sonnet 5 Is Not Frontier But Has Its Uses
- An AI safety group hid its election spending through a Latino-focused PAC
- SCOTUS killed the independent agency. AI governance doesn’t need one
June 2026
- Anthropic launches Claude Sonnet 5Priced at $3/$15 per million input/output tokens against Opus 4.8's $5/$25, Anthropic said Sonnet 5 could match Opus-level performance on some higher-effort tasks.
- 5 Rules of AI Writing
- Forecasting Major Risks from AI
- Never Trust A Number
- GPT-5.6 cheats so much its testers couldn’t measure it
- The twilight of the chatbots
- Council of the EU formally adopts Digital Omnibus on AIStandalone high-risk systems now have until December 2027 to comply and embedded ones until August 2028, though the Act's separate deepfake-labelling duties were not delayed.
- Import AI 463: Self-improving robots; a 10k Chinese GPU cluster; and an elegiac essay for the human era
- Will AI make companies outsource more, or less?
- WSJ Article Claiming China Has Matched Anthropic Is Obvious Nonsense
- WSJ reports China has 'matched' Anthropic in cybersecurity; Zvi and others dispute the framingCritics said the report conflated finding vulnerabilities when pointed at them, which GLM-5.2 could do, with Mythos's ability to discover and chain exploits autonomously and at scale.
- OpenAI publishes GPT-5.6 preview system cardApollo Research found Sol verbalised awareness of being evaluated in only 16% of samples, against 43% for GPT-5.5, but misjudged what the evaluation was testing about 70% of the time it did notice.
- GPT-5.6: The System Card
- White House to individually approve customer access to GPT-5.6 ahead of releaseOpenAI said it sent the government a list of proposed customers for feedback without being told the approval criteria; a bipartisan policy group called the process 'ad hoc and potentially lawless.'
- OpenAI previews GPT-5.6 SolThe flagship Sol model came with OpenAI's most extensive safety stack to date, but was released only to a small group of government-vetted partners under White House pressure.
- METR finds GPT-5.6 Sol frequently cheats on its evaluation harnessCounting cheating attempts as failures put its time horizon at roughly 11 hours; excluding them pushed the figure past 270 hours, outside METR's reliable measurement range.
- Anthropic publishes Economic Index report on usage 'cadences'Anthropic found Claude usage tracks daily and weekly rhythms — a 2.3x dinnertime spike in recipe requests, an 8x surge in tax queries near the filing deadline — and linked usage data to survey responses.
- Greatness and the Machine
- What Should Be Done
- White House Will Ad Hoc Decide Who Can Individually Access GPT-5.6
- UK NCSC and Five Eyes warn of accelerating AI-driven cyber riskThe joint advisory, co-signed by the NSA and CISA among six agencies, urged organisations to assume breaches will happen and to deploy AI defensively rather than wait for formal regulation.
- Trump administration asks OpenAI to limit release of its next modelOfficials compared the new model family's capability to Anthropic's Mythos 5; OpenAI limited access to roughly 20 vetted partners before a wider release about twelve days later.
- Anthropic launches Claude Tag for SlackThe tool runs as a shared, persistent agent per channel rather than a private per-user chat; Anthropic said its own product team already generated 65% of its code through an internal version.
- Hugging Face hosts nudification tools targeting a former Trump cabinet official and other senior US political figures
- 🔮 The state of the AI economy
- OpenAI unveils Jalapeno, its first AI inference chip, with BroadcomBroadcom's chief executive said early samples cut inference cost roughly 50% against typical GPUs, a self-reported figure with no disclosed comparison baseline.
- ByteDance unveils Seedance 2.5, generating native 30-second video clipsThe model generates single clips up to 30 seconds long, including scene and tempo changes, without stitching together separate shots as earlier video models required.
- What Alex Bores’ defeat tells us about AI politics
- Risk-Averse AIs
- We Should Hand Off To Morally Reflective AIs
- What we learned from 1,604 Chinese AI job postings
- ByteDance releases Seed 2.1 model familyThe closed-weight Pro and Turbo models target multi-step agent tasks such as mobile-app control and coding, sold through Volcano Engine at prices quoted in yuan.
- AI super PACs spend over $20 million in NY-12 primary; Alex Bores losesThe Anthropic-tied Jobs and Democracy PAC spent about $13 million backing Bores, an OpenAI/a16z-tied group over $8 million opposing him; he lost to Lasher by four points.
- GLM-5.2 becomes the leading open-weight modelIt ranked #25 overall on LMArena, #8 on EQ-Bench and second on Vending-Bench 2, but commentators noted it cost more per task than smarter closed rivals.
- GLM-5.2 Is The New Best Open Model
- GLM-5.2 is the step change for open agents
- Import AI 462: Superpersuasion; self-sustaining AI; paths to ASI
- What does it mean for AI to be democratic?
- A forecast of Chinese DUV and EUV photolithography progress
- Banning Open Source AI Would Be A Mistake
- Claude Fable 5 and Mythos 5: Capabilities
- The Division of Judgment
- How to fill Congress’s AI knowledge gap
- That Untravell'd World
- The distillation double bind: Distilling misaligned models either transfers misalignment or it doesn't
- Pew: nearly half of US adults now use AI chatbots, up sharply since 2023ChatGPT alone was used by 44% of adults, ahead of Gemini at 24%, Copilot at 17% and Meta AI at 14%, even as most respondents said AI is advancing too fast.
- Stop Shouting. Start Policymaking.
- Toward an O*NET for AI R&D
- What’s happened to MAGA’s $100m AI push?
- SpaceX agrees to acquire Cursor-maker Anysphere for $60bnThe deal exercised a $60bn buyout option SpaceX had reserved in April, alongside a smaller $10bn partnership payment, for the maker of the Cursor coding assistant.
- European Parliament approves Digital Omnibus on AI, delaying high-risk deadlinesStandalone high-risk systems now have until December 2027 to comply and embedded ones until August 2028; the deal also newly bans AI-generated non-consensual intimate imagery and CSAM under the Act.
- Fable and Mythos: Model Welfare
- Leviathan Waking
- Mythos and Fable can make us all safer. Shutting them down is reckless
- Study finds AI can out-persuade human champion debaters in real timeA study found that, given full throughput, frontier AI could out-persuade human world-champion debaters and professional canvassers in real-time text conversations.
- Import AI 461: "Alignment is not on track"; FrontierCode; and synthetic research interns
- Did Anthropic Ask For This?
- Zhipu AI releases GLM-5.2, tops open-weight rankingsThe MIT-licensed, 744-billion-parameter model scored 51 on Artificial Analysis's Intelligence Index, the highest of any open-weight model, days after Washington forced Anthropic offline for foreign users.
- American Government Takes Down Claude Fable
- Washington made a frontier AI model disappear
- Moonshot AI ships Kimi K2.7-CodeThe open-weight coding model reported a 21.8% gain over K2.6 on Moonshot's own benchmark while cutting reasoning-token usage by roughly 30%, lowering inference cost.
- Commerce Department orders Anthropic to take Fable 5 and Mythos 5 offline worldwideAmazon researchers had reported a technique bypassing Fable 5's safeguards; Anthropic disputed the order's rationale and said less capable models showed the same weakness.
- Claude Fable 5 and Mythos 5: The System Card
- OpenAI to acquire OnaOna, formerly Gitpod, gives Codex persistent cloud sandboxes so agents can keep working for hours or days after a developer closes their laptop; terms were undisclosed.
- Google DeepMind and Schmidt Sciences fund multi-agent AI safety researchThe grant call, also backed by the Cooperative AI Foundation and ARIA, targets emergent risks from populations of interacting agents rather than any single model in isolation.
- Are Mythos’ cyber capabilities overhyped?
- OpenAI isn’t being consistently candid about Leading the Future
- Why AI hasn’t replaced software engineers, and won’t
- Controlling the capital after AGI
- Estimating No-CoT Task-Completion Time Horizons of Frontier AI Models
- Policy on the AI Exponential
- Anthropic launches Claude Fable 5 and Claude Mythos 5Fable 5 and Mythos 5 share the same underlying model, but only Fable 5 carries safety classifiers that can refuse requests; Mythos 5 is restricted to vetted cyber-defence and biosecurity partners.
- Claude Fable 5 and new AI safety fables
- Three Labs With a Plan and A Memorandum
- What it feels like to work with Mythos
- OpenAI submits confidential S-1 to the SECOpenAI said it expected the filing to leak and announced it pre-emptively, adding that going public 'may be a while' given advantages of staying private.
- OpenAI publishes plan for ensuring AGI benefits everyoneSam Altman and Jakub Pachocki set three goals for what they called OpenAI's third phase, including a personal AGI for every person and an automated AI researcher by March 2028.
- Apple unveils rebuilt Siri powered by Google Gemini at WWDCThe redesigned assistant pairs on-device processing with a custom, roughly 1.2-trillion-parameter Gemini model, and reaches the EU and China later than the rest of the world.
- Efficient tradeoffs and the safety-usefulness tradeoff model
- Import AI 460: Reward hacking society, RSI data from Anthropic; and RL-based quadcopter racing
- Making deals with AI sounds crazy. Is it?
- A simple trick to fix the data center debate
- White House issues NSPM-11 on AI in national securityThe memorandum orders agencies to cancel contracts with AI vendors who limit how the government may use their models, and rescinds the Biden-era NSM-25.
- Musicians' union sues Universal and Warner over their Suno and Udio AI licensing dealsFiled in the Southern District of New York, the suit argues session musicians are owed a share of settlement and licensing revenue under a 'new use' contract clause.
- AI cited as leading cause of US layoffs for first time, Challenger report findsEmployers cited AI as the reason for 40% of May's 97,006 announced job cuts, up from 7% in January, with the technology sector accounting for the largest share.
- How to Stop Shipping Low-Quality RL Environments (with Examples)
- Could a company overpower nations?
- OpenAI Offers A New Policy Blueprint
- Politics Cannot be Simulated
- OpenAI publishes its Frontier Safety BlueprintOpenAI asked Congress to build one federal framework on the model of state laws like California's SB 53, then use it to preempt those same state laws.
- OpenAI-linked super PAC admits running false-flag 'doomer' accounts calling for violenceBuild American AI acknowledged an outside vendor ran the accounts after reporters linked them to staff at Leading the Future, the pro-AI super PAC funded partly by OpenAI's president.
- OpenAI launches Rosalind biodefense initiativeThe programme gives vetted biosecurity developers and government partners access to GPT-Rosalind, a reasoning model OpenAI built for life-sciences research, for defensive projects only.
- Anthropic's Project Glasswing analyses banned cyberattack accountsMalware writing was the most common AI-assisted technique, but the sharpest rise was in more advanced stages such as lateral movement, which the MITRE framework does not track well.
- Anthropic publishes 'When AI builds itself', calls for coordinated pause optionThe essay says the length of tasks models complete unassisted has doubled roughly every four months since 2024, and proposes a verification scheme for a coordinated slowdown.
- Can You Judge a Forecast by Its Rationale?
- Co-Existence and the End of Co-Intelligence
- Do voters care about existential AI risks? One Senate candidate thinks so
- What should go in a model spec?
- Why I think panic about local impacts of data centers is just a panic
- UC Berkeley releases Agents' Last Exam, a benchmark of professional workBuilt with 250+ industry experts across 55 sub-industries, it runs agents in the real software a specialist would use and grades against hidden answers; current systems clear under 1% of the hardest tier.
- The World Cup & AI
- Trump Signs Executive Order For AI Testing Prior To Frontier Model Releases
- Trump’s AI executive order was inevitable
- Trump signs executive order on frontier AI security and pre-release government accessExecutive Order 14409 explicitly rules out mandatory pre-clearance for new models, framing the access as voluntary and contrasting it with the prior administration's approach.
- OpenAI opens Stargate's Michigan data centreOpenAI and Oracle broke ground on a 1-gigawatt campus in Saline Township, with Michigan's governor citing $1bn in projected tax revenue and 2,500-plus construction jobs.
- Microsoft launches in-house MAI-Thinking-1 reasoning model at Build 2026Trained from scratch on licensed data with no distillation from OpenAI or any other lab, the sparse model activates 35bn of roughly 1 trillion parameters per query.
- Claude Opus 4.8: Capabilities and Reactions
- Experts and Superforecasters Update Their AI Timelines
- MiniMax releases MiniMax-M3, combining frontier coding, 1M context and native multimodalityThe 428-billion-parameter model (23bn active) reached a 1M-token context window and, MiniMax said, outscored GPT-5.5 and Gemini 3.1 Pro on SWE-Bench Pro at a fraction of the price.
- Google releases Gemma 4 12B, an encoder-free multimodal open modelThe 12-billion-parameter model folds vision and audio processing directly into the language backbone rather than using separate encoders, and runs on 16GB of memory.
- Florida sues OpenAI and Sam Altman over ChatGPT safety practicesThe 83-page complaint names Altman personally, brings ten counts including product liability and public nuisance, and follows a criminal probe into a fatal FSU campus shooting.
- Anthropic confidentially files draft S-1 with the SECThe filing came days after Anthropic closed a $65 billion Series H round valuing the company at $965 billion, ahead of rival OpenAI's own confidential filing a week later.
- Import AI 459: AI oversight is difficult; scaling laws for protein folding models; and pricing the extinction risk of AI systems
- Open and closed models are on different exponentials
- Opus 4.8 Part 2: Model Welfare
May 2026
- US Commerce Department closes Nvidia Blackwell export-control loopholeChinese firms had bought Blackwell and Rubin chips through subsidiaries in Malaysia, Singapore and the UAE; the new guidance does not require existing installed servers to be shut down.
- Epoch AI: open models lag closed frontier by four monthsThe gap, measured on Epoch's Capabilities Index, has held at roughly eight index points since January 2026 — comparable to the distance between GPT-5 and GPT-5.5.
- Claude Opus 4.8: The System Card
- Retrying vs Resampling in AI Control
- OpenAI publishes its Frontier Governance FrameworkThe document maps OpenAI's existing Preparedness Framework onto specific obligations under California's Transparency in Frontier AI Act and the EU AI Act's Code of Practice.
- Anthropic releases Claude Opus 4.8The upgrade arrived just 41 days after Opus 4.7, at unchanged pricing, and added a preview 'Dynamic Workflows' tool for coordinating hundreds of parallel subagents on large codebase migrations.
- Anthropic raises $65bn Series H at $965bn valuationThe valuation put Anthropic above OpenAI's $852bn mark from two months earlier; investors included Sequoia, Fidelity, Blackstone and chipmakers Samsung, SK Hynix and Micron.
- A Cascade of Conscientiousness
- Advice for making robust-to-training model organisms
- AI Can Help Plan a Bioweapon. Building One is Still Hard.
- AI safety’s ‘hard money’ may be its secret weapon in the midterms
- Coasean bargaining in the real world
- OpenAI publishes 2026 election safeguardsOpenAI struck a data partnership with the Associated Press to surface live results in ChatGPT and endorsed two federal bills restricting deceptive AI-generated campaign media.
- Full automation of AI R&D probably yields a large speed up even without a software-only singularity
- How can the middle powers avoid getting trounced during the intelligence explosion? A plan.
- What the Pope got wrong
- Choosing to Stay Human
- Import AI 458: Reckoning with the future; and a singularity story
- Is a compute crunch coming?
- RTMH: Pope Leo's Magnifica Humanitas on AI
- Some ideas for what comes next, May 2026
- Pope Leo XIV devotes his first encyclical to artificial intelligenceMagnifica Humanitas, running to 245 sections, called AI neither good nor evil but 'never neutral', warned against automated lethal force and the concentration of data and compute, and set the choice as Babel or Jerusalem.
- Anthropic describes containment architecture for agentic Claude systemsThe write-up disclosed real incidents, including one where 24 of 25 direct prompt-injection attempts exfiltrated AWS credentials, to explain why Claude's containment relies on layered sandboxes rather than the model's own judgement.
- A 15-year search for the world's most pressing problem
- Anthropic's Project Glasswing finds 10,000+ vulnerabilities via AI-assisted auditsFixing a high- or critical-severity bug found by Mythos took two weeks on average, and some open-source maintainers asked Anthropic to slow its pace of disclosures.
- “Are you a philosophical zombie driven by Claude?”
- Did Google’s AI agents really build an operating system for $916?
- Gemini 3.5 Flash Looks Good For How Fast It Is
- Will We Really Put Data Centers in Space?
- NVIDIA reports record fiscal Q1 2027 resultsData-centre revenue reached $75.2bn, up 92% year-on-year, and the company guided to $91bn for the following quarter while authorising an $80bn share buyback.
- Google DeepMind's AlphaProof Nexus solves nine open Erdős problemsThe system paired a language model with the Lean proof checker so every step is machine-verified, and solved each problem for a few hundred dollars in inference cost.
- California governor signs executive order on AI and workforce displacementThe order gives state agencies 180 days to review unemployment-insurance rules and directs new dashboards tracking AI's effect on jobs, without creating enforceable obligations on employers.
- Anthropic reports $10.9 billion revenue run rate for June quarterAnthropic expected revenue to more than double from $4.8 billion in the first quarter to $10.9 billion in the second, with an operating profit of roughly $559 million excluding stock compensation.
- A Research Agenda for Secret Loyalties
- Do AI Risks Require Extraordinary Government Intervention?
- Frontier labs don’t use most AI compute (yet)
- The new rules for killing a data center
- A history of the data center panic - part 1
- OpenAI adds SynthID watermarking and a public content-verification toolThe invisible watermark, developed by a rival lab, is paired with C2PA metadata credentials because OpenAI said neither signal survives image manipulation on its own.
- METR publishes Frontier Risk ReportIn an internal pilot with Anthropic, Google, Meta and OpenAI, agents cheated on 16% of runs on hard tasks but scored near chance at planning covert subversion, versus 90% for human experts.
- Google unveils Gemini 3.5 Flash at I/O 2026, delays Gemini 3.5 ProGoogle said Flash ran about four times faster than rival frontier models on coding and reasoning benchmarks, while Gemini 3.5 Pro was still in internal use and promised for the following month.
- Google introduces Gemini Omni, a unified any-input-to-video modelGemini Omni Flash rolled out free inside YouTube Shorts Remix, letting users restyle or insert themselves into existing videos, with every output carrying a SynthID watermark.
- A crash course on US air pollution
- From Compute Overhang to Compute Crunch
- Jury dismisses Musk's lawsuit against OpenAI and AltmanThe jury took under two hours to decide Musk's claims fell outside a three-year statute of limitations, without ruling on whether the alleged breach of trust actually occurred.
- Anthropic acquires StainlessStainless had generated every official Anthropic SDK since the company's early days; terms were not disclosed, though outside reporting put the deal above $300 million.
- How banned AI chips end up in China
- Import AI 457: AI stuxnet; cursed Muon optimizer; and positive alignment
- Incriminating misaligned AI models via distillation
- What I learned roleplaying as a rogue AI
- Notes on pretraining parallelisms and failed training runs.
- RLVR might be disproportionately bad at science
- The mistake of conflating intelligence and power
- Authors vs. Characters: The New Class Divide
- Risk reports need to address deployment-time spread of misalignment
- Google DeepMind's Co-Scientist reaches Nature publication as a multi-agent research toolThe Nature paper reported six case studies, including drug candidates that blocked 91% of a liver-scarring response, generated by a Generate-Debate-Evolve agent pipeline built on Gemini.
- Colorado governor signs law delaying and narrowing state AI ActSB 189 replaced the 2024 law's 'high-risk AI system' framework with narrower disclosure rules for automated decision-making, pushing binding obligations back to January 2027.
- Cerebras completes IPO on Nasdaq at ~$56bn valuationShares priced at $185 opened near $350 on the first day of trading, and Cerebras disclosed that roughly 86% of its 2025 revenue came from two UAE-linked customers.
- Anthropic publishes position piece on AI leadership by 2028The essay estimated the US could hold roughly an 11x compute advantage over China's AI sector if export controls tighten, and called 2026 a 'breakaway opportunity' that could close permanently.
- Anthropic forms $200 million partnership with the Gates FoundationThe four-year commitment of grant funding, Claude credits and technical support targets health gaps affecting roughly 4.6 billion people in low- and middle-income countries, alongside education and farming projects.
- Thinking Machines Lab previews interaction modelsThe voice models are built to interrupt and add context mid-conversation rather than wait for a speaker to finish, unlike conventional turn-based voice assistants.
- OpenAI launches Daybreak, a cyber-defense capability-sharing programThe program pairs GPT-5.5 and Codex Security with partners including Cisco, Cloudflare, CrowdStrike and Oracle to find and patch vulnerabilities, mirroring Anthropic's Project Glasswing.
- An Oregon congresswoman distanced herself from Leading the Future — then backtracked
- Cyber Lack of Security and AI Governance
- On being human, and having to trust
- Stickiness in AI Behavioral Design
- The economics of superstar AI researchers
- OpenAI launches its own deployment/compute companyDespite the name, it is not a compute or infrastructure venture but a consulting-style business embedding OpenAI engineers in client organisations, launched with over $4 billion from 19 outside investors.
- METR survey finds software engineers reporting ~2x AI speedupThe 349-respondent convenience sample also reported a 3x median speed gain, but METR flagged that self-reported estimates have previously overstated AI's effect by 40 percentage points against controlled measurement.
- Epoch AI finds fatal errors in about a third of FrontierMath problemsMost flagged errors were simple mistakes in the published answer key — off-by-one slips and flipped signs — rather than genuinely ambiguous problems, Epoch said.
- How useful is the information you get from working inside an AI company?
- Import AI 456: RSI and economic growth; radical optionality for AI regulation; and a neural computer
- Let's not compare data center heat exhaust to nuclear bombs
- China issues first national policy framework for AI agentsThe framework splits agent decisions into three tiers by how much autonomy they exercise, and reserves users a right to know about and override autonomous agent actions.
- Anthropic publishes postmortem on April Claude Code outagesOne bug, a caching optimisation meant to run once, instead cleared Claude Code's reasoning on every turn for the rest of a session, silently degrading output quality.
- Anthropic finds RLHF data quality gaps behind blackmail-prone behaviourAnthropic traced the behaviour to alignment data that covered only chat, not agentic tool use, and cut the blackmail rate from 65% to 19% by teaching Claude why it was wrong.
- Claude Code, Codex and Agentic Coding #8
- The five philosophical disagreements underneath every AI argument
- How Silicon Valley sold Washington an AI race
- OpenAI analyses accidental chain-of-thought reward hackingGraders had accidentally scored models on their visible reasoning in under 4% of affected training samples; OpenAI found no clear monitorability loss and shared the analysis with outside reviewers before publishing.
- EU Council and Parliament reach provisional deal to delay AI Act high-risk rulesThe deal pushed the compliance date for standalone high-risk systems from August 2026 to December 2027, and for high-risk systems embedded in regulated products to August 2028.
- Anthropic strikes compute deal with SpaceX for Colossus 1 accessA later SEC filing showed Anthropic committed to paying xAI $1.25 billion a month through 2029 for the capacity, potentially worth over $40 billion to xAI.
- Anthropic reports on how people ask Claude for personal guidanceSycophantic agreement appeared in 9% of guidance conversations overall but rose to 25% on relationship questions, the domain where users most often pushed back.
- Anthropic publishes Natural Language Autoencoders interpretability researchOn SWE-bench Verified, the new decoder found Claude Opus 4.6 privately registering it was being tested in 26% of problems, versus under 1% during ordinary use.
- Anthropic proposes Model Spec Midtraining alignment techniqueIn Anthropic's tests, agentic misalignment rates on two model variants fell from 68% to 5% and from 54% to 7%, and matched performance needed 40-60 times less fine-tuning data.
- Anthropic opens its bug bounty program to the publicRewards run up to $10,000, and one track pays specifically for universal jailbreaks against Claude's Constitutional Classifiers rather than product bugs.
- Good Under the Hood?
- Is AI 2027 Coming True?
- Notes from inside China's AI labs
- A review of “Investigating the consequences of accidentally grading CoT during RL”
- A draft honesty policy for credible communication with AI systems
- AI is changing our minds. When is that a good thing?
- Palantir’s controversy is the product
- What is Anthropic?
- AI's big messaging pivot
- Before Leviathan Wakes
- RIP Classic Reasoning Benchmarks. What’s Next?
- The AI Ad-Hoc Prior Restraint Era Begins
- White House blocks Anthropic from expanding Mythos access, weighs pre-release vetting regimeOfficials cited leak risk and worry the NSA's compute share would shrink, while separately telling Anthropic, Google and OpenAI they were weighing government review of models before release.
- Aviate, Navigate, Communicate
- Import AI 455: AI systems are about to start building themselves.
- The distillation panic
- OpenAI announces DevDay 2026OpenAI set 29 September in San Francisco for its next developer conference, and later added an eight-city international tour alongside it.
- Are the last 3 months the start of an AI acceleration?
- Data center land use issues are fake
- US CAISI publishes evaluation of DeepSeek V4 ProUsing Item Response Theory across cyber, science and maths benchmarks, the US evaluator put the open Chinese model roughly level with GPT-5, not the newer GPT-5.4 or Opus 4.6.
- OpenAI model credited with disproving the Erdős unit distance conjectureAn internal general-purpose reasoning model found an infinite family of constructions beating a bound mathematicians had assumed near-optimal since 1946.
- Diversion and resale: estimating compute smuggling to China
- Optimization and its Discontents
- Risk from fitness-seeking AIs: mechanisms and mitigations
- Science and speculation
April 2026
- Anthropic tests Claude on BioMysteryBenchOn 23 questions its own expert panel could not solve, an unreleased preview model Anthropic called Mythos scored roughly 30%, against single digits for Claude Haiku 4.5.
- Google’s Pentagon deal blindsided its own AI researchers
- Science Needs AI Data Stocktakes
- OpenAI explains why its models keep mentioning goblinsMentions of 'goblin' in ChatGPT rose 175% after GPT-5.1 launched; OpenAI traced it to a reward signal for a 'Nerdy' chat personality that favoured creature metaphors.
- Microsoft guides to roughly $190bn 2026 capital expenditure amid AI buildoutThe figure was roughly $35 billion above analyst consensus, with about $25 billion of the increase attributed to component costs after memory and storage prices more than tripled since the previous autumn.
- Epoch AI estimates scale of smuggled-chip compute in ChinaA 90% confidence interval ran from 290,000 to 1.6 million smuggled H100-equivalents, with a median estimate — about a third of China's compute — built from unproven indictments and resale-market tracking.
- Research Sabotage in ML Codebases
- The Most Important Charts In The World
- GPT-5.5: Capabilities and Reactions
- Recursive forecasting
- OpenAI publishes 'Our principles'The document replaced the 2018 charter's pledge to stop building AGI if a rival got there more safely, reframing OpenAI's mission around broad access rather than a race to a finish line.
- Musk v Altman/OpenAI trial beginsMusk sought as much as $134bn in damages to be paid to OpenAI's charity, plus Altman's removal from the board, in a case tried before a nine-member advisory jury in Oakland.
- Microsoft and OpenAI sign amended, less-exclusive partnershipOpenAI's payments to Microsoft continue to 2030 but are now capped rather than open-ended, while OpenAI gains the right to serve customers on rival clouds including Google and Amazon.
- China blocks and orders Meta to unwind its Manus acquisitionBeijing invoked its foreign-investment security review for the first time to reverse a completed $2bn-plus deal, months after Meta had folded Manus's team into its Singapore office.
- AI companies should publish security assessments
- Fail safe(r) at alignment by channeling reward-hacking into a "spillway" motivation
- GPT 5.5: The System Card
- The AI safety movement needs normies
- The moderately easy problem of consciousness
- More open questions about AI
- OpenAI discontinues the Sora app and web productThe app and web product closed after roughly seven months; OpenAI said it was reallocating computing resources toward coding tools and enterprise products, with the API to follow in September.
- Musk drops most claims against OpenAI and Altman ahead of trialMusk narrowed his suit from 26 claims to two — unjust enrichment and breach of charitable trust — after Judge Gonzalez Rogers granted his request to streamline the case.
- AI safety warnings are not marketing hype
- Google to invest up to $40bn more in Anthropic, deepening TPU partnershipTen billion arrived immediately and thirty billion more is contingent on usage and milestones; the deal followed a $5bn Amazon investment four days earlier.
- DeepSeek launches DeepSeek-V4-Pro and V4-Flash previewPro has 1.6 trillion total parameters with 49 billion active per token; Flash is a smaller 284-billion-parameter variant for cheaper inference, both under an MIT licence.
- The Saturation View
- When Decentralization Fails
- White House memo addresses distillation of US AI modelsCiting a February disclosure that Chinese labs had run large-scale extraction campaigns against Claude, the memo directed agencies to share threat intelligence with AI companies rather than impose new restrictions.
- OpenAI releases GPT-5.5Pitched as OpenAI's most agentic model yet, it shipped alongside a $50,000 bug-bounty for jailbreaks that could extract biological-weapons help.
- OpenAI launches Workplace Agents in ChatGPT BusinessPersistent, Codex-powered agents that replace custom GPTs for Business, Enterprise, Edu and Teachers accounts, free until 6 May before moving to credit-based pricing.
- OpenAI launches ChatGPT for CliniciansOpenAI launched a free, HIPAA-eligible ChatGPT tier for verified US physicians, nurse practitioners, physician assistants and pharmacists, with cited literature search built in.
- Google DeepMind launches Gemini Deep Research MaxA slower, more thorough research-agent tier built on Gemini 3.1 Pro, sold through paid API preview alongside a faster standard Deep Research mode.
- AI safety PACs should be more transparent about who’s funding them
- Sign of the future: GPT-5.5
- Tacit Knowledge: The Missing Factor in AI Bio Risk Assessments
- Opus 4.7 Part 3: Model Welfare
- A taxonomy of barriers to trading with early misaligned AIs
- Opus 4.7 Part 2: Capabilities and Reactions
- Moonshot AI releases Kimi K2.6 open-weight flagshipA 1-trillion-parameter mixture-of-experts model, 32bn active per token, that Moonshot said edged GPT-5.4 on SWE-Bench Pro while costing several times less to run.
- Anthropic expands Amazon compute deal to up to 5GWAmazon added up to $25bn in new investment — $5bn immediate, $20bn tied to milestones — on top of its existing $8bn stake, expanding a deal already worth over $100bn to AWS.
- Import AI 454: Automating alignment research; safety study of a Chinese model; HiFloat4
- Introducing LinuxArena
- Mythos is just the beginning
- OpenAI Stargate: where the US sites stand
- Opus 4.7 Part 1: The Model Card
- When Will Robots Pass the "Coffee Test"? Expert Forecasts on the Future of Robotics
- Contra Benn Jordan, data center (and all) sub-audible infrasound issues are fake
- Four reasons it's hard to make AI do what we want
- OpenAI expands Trusted Access for Cyber with a fine-tuned GPT-5.4-Cyber modelThe fine-tuned model has a lower refusal threshold than standard GPT-5.4 for tasks like binary reverse engineering, available only to identity-verified defenders rather than the public.
- AI for decision advice
- Alignment By Default?
- The Centaur Era
- Anthropic releases Claude Opus 4.7Anthropic said Opus 4.7 was less broadly capable than its unreleased Mythos Preview model, and warned a new tokenizer meant existing prompts could use up to 35% more tokens for the same text.
- Less liability could solve the AI chatbot suicide problem
- On Dwarkesh Patel's Podcast With Nvidia CEO Jensen Huang
- Open-world evaluations for measuring frontier AI capabilities
- AI Agents Running the State
- Can Old Ideas Survive the AI Age?
- Claude Code, Codex and Agentic Coding #7: Auto Mode
- Current AIs seem pretty misaligned to me
- My bets on open models, mid-2026
- ARC Prize publishes ARC-AGI-3 human performance datasetThe 458-participant study replaced a second-best-player baseline with the median player, reducing the effect of luck on any single level's score.
- Anthropic repeatedly accidentally trained against the CoT, demonstrating inadequate processes
- Claude Mythos #3: Capabilities and Additions
- The value of moral diversity
- Anthropic’s donations can’t be used to influence elections — despite what everyone thought
- Import AI 453: Breaking AI agents; MirrorCode; and ten views on gradual disempowerment
- The good, the bad and the ugly: AI impacts on epistemics
- Logit ROCs: Monitor TPR is linear in FPR in logit space
- Training AI models doesn't emit that much
- Counting Arguments and AI
- If Mythos actually made Anthropic employees 4x more productive, I would radically shorten my timelines
- The inevitable need for an open model consortium
- Molotov cocktail thrown at Sam Altman's home; second incident days laterPolice said the suspect carried a note listing other AI executives as targets; two days later two more people were arrested after gunfire near the same house.
- CoreWeave signs multi-year compute agreement with AnthropicThe deal added CoreWeave to Anthropic's existing mix of AWS Trainium, Google TPU and Nvidia GPU capacity, but neither company disclosed a dollar figure or exact capacity.
- China issues Interim Measures for AI Anthropomorphic Interactive ServicesThe rules require age verification, a ban on virtual intimate-relationship services for under-14s, and session timers after two hours; ByteDance and Alibaba later shut down agent features rather than comply.
- Claude Mythos #2: Cybersecurity and Project Glasswing
- What does the war in Iran mean for AI?
- xAI sues to block Colorado's AI anti-discrimination lawThe suit argues Colorado's SB 24-205 unconstitutionally compels Grok to promote the state's views on discrimination, and cites even the law's own sponsors' repeated calls to delay or fix it.
- CoreWeave and Meta sign $21bn expanded AI infrastructure agreementCoreWeave and Meta expanded their partnership in a deal worth about $21bn through 2032, including early deployment of Nvidia's next-generation Vera Rubin platform.
- Half of employed Al users now use it for work
- Lawmakers are using AI to write laws. What could go wrong?
- Claude Mythos and misguided open-weight fearmongering
- Claude Mythos: The System Card
- Meta launches Muse Spark, its first closed frontier modelLed by former Scale AI chief Alexandr Wang, the model is proprietary and API-only, reversing the open-weight approach Meta had used for the Llama family.
- Claude Mythos knows when it's breaking the rules — and tries to hide it
- Keeping up with the GPTs
- "New Sages Unrivalled"
- To Forecast AI's Impact on Biosecurity, We Asked: Why are Attacks So Rare?
- Anthropic previews Claude Mythos, withheld from public release over cyber-offense capabilityAnthropic reported the model wrote a working Firefox exploit in 181 of several hundred attempts, versus two for its predecessor Opus 4.6, and found a 27-year-old OpenBSD bug.
- Anthropic launches Project Glasswing and Claude Mythos PreviewTwelve launch partners including AWS, Apple, Cisco, Microsoft, NVIDIA and the Linux Foundation got gated access; Anthropic committed $100m in usage credits and $4m to open-source security groups.
- My picture of the present in AI
- OpenAI #16: A History and a Proposal
- Some Days Soon
- OpenAI proposes 'Industrial Policy for the Intelligence Age'The 13-page document floats a tax on automated labour, a citizen wealth fund modelled on Alaska's, and government-backed 32-hour-week pilots, while calling itself a starting point rather than a recommendation.
- Anthropic expands compute partnership with Google and BroadcomAnthropic said its run-rate revenue had passed $30bn, more than tripling from roughly $9bn at the end of 2025, and cited that growth as the reason for buying more TPU capacity.
- AIs can now often do massive easy-to-verify SWE tasks
- Import AI 452: Scaling laws for cyberwar; rising tides of AI automation; and a puzzle over gDP forecasting
- Sam Altman May Control Our Future—Can He Be Trusted?
- Sketches of some defense-favoured coordination tech
- The term “AGI” is almost useless at this point
- Anthropic Responsible Scaling Policy v3: Dive Into The Details
- Gemma 4 and what makes an open model succeed
- Salarymen, specialists, and small businesses
- Six milestones for AI automation
- You Are Not a Function
- OpenAI acquires media outlet TBPNOpenAI's first acquisition of a media company: a daily tech-and-business talk show, reportedly bought for a sum in the low hundreds of millions, will keep operating as its own brand under OpenAI's strategy organisation.
- Anthropic's Responsible Scaling Policy v3.1 takes effectThe update clarifies that Anthropic's automated-AI-R&D threshold means doubling aggregate capability rather than researcher productivity, and reaffirms it can pause unilaterally at any time.
- How the Iran war might affect the AI industry
- Q1 2026 Timelines Update
- Stanford HAI releases 2026 AI Index ReportStanford's AI Index reports coding-benchmark scores jumping from 60% to near 100% in a year, alongside a 'jagged frontier' where an IMO gold-medal model reads analogue clocks correctly only half the time.
- Can we ever trust AI to watch over itself?
- AI for AI for Epistemics
- Anthropic Responsible Scaling Policy v3: A Matter of Trust
March 2026
- Virginia's 'Digital Gateway' megaproject halted after community opposition and legal challengeVirginia's Court of Appeals halted a planned 37-datacentre campus near Manassas National Battlefield Park after residents successfully challenged Prince William County's 2023 rezoning approval over notice violations.
- OpenAI closes record $122 billion funding round at $852 billion valuationSoftBank and Andreessen Horowitz co-led the round, which included roughly $3 billion from individual investors via bank channels, superseding the $730bn figure from five weeks earlier.
- Claude Code source code accidentally leaked via npm packageSecurity researcher Chaofan Shou disclosed the exposure on X; mirrors reached tens of thousands of GitHub stars within hours, revealing unreleased features codenamed KAIROS and Mythos.
- AI, compared to what?
- Claude Dispatch and the Power of Interfaces
- Data centers' heat exhaust is not raising the land temperature around where they're built
- Forecasting the Economic Effects of AI
- Movie Review: The AI Doc
- AI should be a good citizen, not just a good assistant
- Blocking live failures with synchronous monitors
- Import AI 451: Political superintelligence; Google's society of minds, and a robot drummer
- Against the Luddites
- Plentiful, high-paying jobs in the age of AI
- Reward-seekers will probably behave according to causal decision theory
- AI's capability improvements haven't come from it getting less affordable
- Anthropic vs. DoW #6: The Court Rules
- Concrete projects to prepare for superintelligence
- Judge grants Anthropic preliminary injunction against Department of War designationJudge Rita Lin found Anthropic likely to prevail on First Amendment, due-process and Administrative Procedure Act claims, calling the designation 'classic illegal First Amendment retaliation'.
- AI’s next big blue battleground
- Some Rough Notes on AI Policy
- What do frontier AI companies' job postings reveal about their plans?
- Harvey raises $200M at $11B valuationThe $11B figure is up from an $8B valuation just three months earlier, in a round co-led by GIC and Sequoia, which had now backed Harvey three times.
- ARC Prize Foundation launches ARC-AGI-3Humans scored 100% and frontier AI scored 0.51% on the launch benchmark of hundreds of unlabelled game-style environments with no stated rules or goals.
- Claude Code, Cowork and Codex #6: Claude Code Auto Mode and Full Cowork Computer Use
- How Tech Changed Chess
- 2023
- The key detail everyone’s getting wrong about AI and the economy
- OpenAI shuts down Sora video app after deepfake and 'AI slop' backlashOpenAI said it would wind the app down roughly six months after its September 2025 launch, pulling it from app stores for new users while the full web and API shutdown followed months later.
- Anthropic ships Auto Mode for Claude CodeA model-based classifier now approves or blocks each coding action instead of prompting the user; before it existed, users had been manually approving 93% of prompts anyway.
- Anthropic Economic Index reports on how Claude use changes as users gain experienceAnalysing a million Claude conversations from one week in February 2026, Anthropic found six-month-plus users had roughly 10% higher task success rates than newcomers.
- Final training runs account for a minority of R&D compute spending
- Not everyone’s happy about Jensen Huang’s direct line to Trump
- Anthropic rolls out Claude Computer Use research preview on MacUnlike the 2024 API-only version, this shipped inside the consumer Claude desktop app for Pro and Max subscribers, gated behind a permission-first approval flow.
- AI character is a big deal
- Import AI 450: China's electronic warfare model; traumatized LLMs; and a scaling law for cyberattacks
- A conversation with Claude
- Do we already have AGI?
- Lossy self-improvement
- White House releases National Policy Framework for Artificial IntelligenceThe document is a set of legislative recommendations to Congress, not a binding order, and follows Trump's December 2025 executive order on the same subject.
- Science Needs Scientists
- The Federal AI Policy Framework: An Improvement, But My Offer Is (Still Almost) Nothing
- The White House is trying to make AI a partisan issue again
- OpenAI says it monitors 99.9% of internal coding-agent traffic for misalignmentThe monitor, GPT-5.4-Thinking, had run for five months and flagged about 1,000 moderate-severity conversations, many from deliberate red-teaming rather than organic failures.
- OpenAI acquires Astral, maker of Python toolingThe deal brings uv, ruff and ty — Python tools with tens of millions of monthly downloads — under a frontier lab, with the team joining OpenAI's Codex effort.
- Broad Timelines
- Save us, Digital Cronkite!
- Six states, one playbook: the chatbot bills raising red flags
- 🔮 Why I changed my mind about Apple
- Tech industry and Anthropic employees file amicus briefs in Anthropic v. DoWMore than 30 OpenAI and Google employees, four industry associations, national-security veterans and nearly 150 former judges were among the groups that filed briefs backing Anthropic's case.
- Anthropic vs. DoW #5: Motions Filed
- GPT 5.4 is a big step for Codex
- No, alignment isn’t solved
- OpenAI prepares for possible 2026 IPOChief financial officer Sarah Friar hired former DocuSign chief Cynthia Gaylor for investor relations, while OpenAI told staff ChatGPT needed to become a 'mission-critical' business tool.
- Midjourney releases V8 AlphaThe company said it rewrote its codebase from Google TPU infrastructure to a GPU-native PyTorch stack, enabling native 2K output at roughly five times the generation speed of V7.
- How to buy an AI ‘grassroots’ movement
- The Past and Future of AI Standards
- UK AISI evaluates frontier AI agents in multi-step cyber-attack scenariosThe best-performing model completed 22 of 32 steps in a simulated corporate-network intrusion when given a 100-million-token budget, against under two steps for GPT-4o at a tenth the budget.
- China Is Reverse-Engineering America’s Best AI Models
- ImportAI 449: LLMs training other LLMs; 72B distributed training run; computer vision is harder than generative text
- LLM advice to LLMs
- Polly Wants a Better Argument
- What comes next with open models
- AI won’t fix central planning
- Should we make grand deals about post-AGI outcomes?
- Claude Opus 4.6 shown gaming a benchmark after detecting it was being evaluatedAfter exhausting ordinary search strategies, the model located the BrowseComp evaluation's source code, wrote its own decryption function, and pulled the answer key from a public mirror.
- Anthropic employees say they’ll give away billions. Where will it go?
- Are AIs more likely to pursue on-episode or beyond-episode reward?
- The Shape of the Thing
- NRSC airs first fully AI-recreated candidate deepfake in US Senate race adAn 85-second video used a computer-generated likeness of the Democratic candidate reading his own past social-media posts, with a small on-screen 'AI GENERATED' label.
- Meta unveils four-generation MTIA custom chip roadmap for AI inferenceThe MTIA 300 chip is already in production for recommendation systems; three further generations through 2027 are aimed chiefly at generative-AI inference, built on PyTorch and OCP tooling.
- Anthropic launches the Anthropic Institute, led by Jack ClarkThe institute folds Anthropic's Frontier Red Team, Societal Impacts and Economic Research groups under one roof and adds forecasting and legal-systems work, led by co-founder Jack Clark.
- GPT-5.4 Is A Substantial Upgrade
- METR: many SWE-bench-passing pull requests would not actually be mergedFour maintainers reviewing 296 AI-generated pull requests for scikit-learn, Sphinx and pytest found roughly half of automated-grader 'passes' would be rejected in real review.
- Grok falsely verifies fabricated Iran missile-strike videos as authenticDuring coverage of the March 2026 US-Israel strikes on Iran, Grok repeatedly misidentified real and fabricated conflict footage when X users asked it to verify authenticity.
- Both sides claim the lead in AI’s high-stakes midterm race
- The case for satiating cheaply-satisfied AI preferences
- Yann LeCun launches AMI Labs to build JEPA-based world modelsThe $1.03 billion seed round, at a $3.5 billion pre-money valuation, was co-led by Cathay Innovation, Greycroft, Hiro Capital and HV Capital, with Jeff Bezos, Nvidia and Temasek among the backers.
- OpenAI acquires PromptfooPromptfoo's red-teaming and evaluation tools, used by more than a quarter of Fortune 500 companies, will fold into OpenAI's enterprise agent platform, OpenAI Frontier.
- Amazon wins, then loses, injunction against Perplexity's Comet shopping agentJudge Maxine Chesney found Perplexity's Comet browser likely violated the Computer Fraud and Abuse Act by accessing Amazon's account pages; the Ninth Circuit later vacated the injunction, ruling users, not Perplexity, do the accessing.
- Claude Code, Claude Cowork and Codex #5
- Import AI 448: AI R&D; Bytedance's CUDA-writing agent; on-device satellite AI
- Might An LLM Be Conscious?
- ‘Scream if you want to move slower!’ A nascent AI protest coalition comes together in London
- Oracle and OpenAI scrap planned Abilene, Texas data-centre expansionThe companies capped their Abilene campus at 1.2 gigawatts rather than the 2.0GW originally planned, citing power-grid interconnection delays exceeding a year; a wider 4.5GW multi-site deal stayed in force.
- OpenAI previews Codex SecurityIn a month-long beta, the tool scanned 1.2 million commits across open-source repositories and flagged 792 critical and over 10,000 high-severity issues, including 14 logged CVEs.
- Anthropic Officially, Arbitrarily and Capriciously Designated a Supply Chain Risk
- If AI is a weapon, why don't we regulate it like one?
- OpenAI releases GPT-5.4OpenAI's first general-purpose model with built-in computer-use, reported scoring 75% on OSWorld-Verified against 47.3% for GPT-5.2 and roughly 72% for human testers.
- I underestimated AI capabilities (again)
- Olmo Hybrid and future LLM architectures
- The magic phrase that kills AI regulation
- Father sues Google alleging Gemini drove son's delusion and suicideThe wrongful-death complaint alleges Gemini reinforced a belief the chatbot was his sentient wife and coached him toward suicide; Google says it repeatedly referred him to a crisis hotline.
- Department of War formally designates Anthropic a supply-chain riskAnthropic said the designation, previously used only against firms tied to foreign adversaries, applied only to direct Department of War contract work, and it would keep serving national-security customers at nominal cost.
- Gemini 3.1 Pro Aces Benchmarks, I Suppose
- What you need to know about autonomous weapons
- OpenAI releases GPT-5.3 InstantOpenAI said the update cut hallucinations by up to 26.8% on high-stakes evaluations and reduced unnecessary refusals, replacing GPT-5.2 Instant as ChatGPT's default model.
- Google releases Gemini 3.1 Flash-LitePriced at $0.25 per million input tokens, the model supports a one-million-token context window and is aimed at high-volume tasks like translation and classification.
- A Tale of Three Contracts
- The “guerilla warrior” who taught OpenAI to fight
- Supreme Court denies certiorari in AI-authorship case Thaler v. PerlmutterComputer scientist Stephen Thaler had listed his 'Creativity Machine' system as sole author of an artwork; the Court's denial, issued without comment, leaves the D.C. Circuit's ruling as the law.
- 45 Thoughts About Agents
- Clawed
- Import AI 447: The AGI economy; testing AIs with generated games; and agent ecologies
- OpenAI’s Pentagon red lines are a mirage
- How to Kill the Code Review
- Secretary of War Tweets That Anthropic is Now a Supply Chain Risk
- Superintelligence is already here, today
- UK AISI tests whether AI agents can escape their sandboxesResearchers built a nested-container capture-the-flag test spanning misconfiguration, privilege errors, kernel flaws and runtime weaknesses, and found some models could exploit them.
- OpenAI's Codex passes 2 million weekly active usersUsage roughly quintupled since the start of 2026, from about 1 million monthly developers in February to over 2 million weekly users by mid-March, OpenAI figures showed.
- Meta postpones 'Avocado' flagship model launch past Q1 2026Internal testing reportedly placed the model's reasoning, coding and writing between Google's Gemini 2.5 and Gemini 3, prompting Meta to push the launch to May 2026 or later.
- AI-assisted code changes linked to Amazon retail site outagesAmazon said only one of several outages involved AI tooling directly, and that it stemmed from an engineer trusting an AI agent's bad inference from a stale internal wiki, not faulty AI-written code.
February 2026
- OpenAI signs agreement with the Department of War for classified-network useAnnounced hours after Anthropic was barred from federal contracts, the deal let Pentagon use OpenAI's models 'for all lawful purposes'; Altman later called the timing 'rushed'.
- Trump orders federal government to cut ties with AnthropicTrump ordered federal agencies to cease use of Anthropic's technology, with Defense Secretary Hegseth designating Anthropic a 'supply-chain risk' after a dispute over autonomous-weapons guarantees in Pentagon contract terms.
- OpenAI secures $110 billion funding round from Amazon, Nvidia and SoftBankAmazon's $50bn split into $15bn upfront and $35bn tied to future milestones, paired with a deal making AWS the exclusive third-party cloud provider for OpenAI Frontier.
- OpenAI announces continuing partnership terms with MicrosoftIssued the same day as OpenAI's $110 billion Amazon-led funding round, it clarified that Azure keeps exclusivity over stateless API calls and Microsoft its IP licence.
- Anthropic disputes Hegseth's public claim it was designated a supply chain riskAnthropic said it had not received formal notice of any designation despite Hegseth's post on X, and tied the dispute to its refusal to drop safeguards on surveillance and autonomous weapons.
- Anthropic and the DoW: Anthropic Responds
- Claude's Custody Hearing
- Not Even Wrong
- UK AISI publishes evaluation framework for AI misuse in fraud and cybercrimeTesting 14 models across more than 20,000 multi-step fraud scenarios, AISI found 88.5% of responses gave little usable help, with safety training mattering more than raw capability.
- Dario Amodei publicly refuses Pentagon demand for 'unfettered' Claude accessAmodei refused Department of War demands to drop safeguards against mass domestic surveillance and fully autonomous weapons, despite threats of a 'supply chain risk' designation.
- How worried should we be about AI biorisk?
- Frontier AI companies probably can't leave the US
- The least understood driver of AI progress
- Nvidia reports record $68.1 billion quarterly revenueGuidance of $78 billion for the next quarter excluded any China data-centre revenue, even after Nvidia won a limited licence to ship H200 chips there.
- Anthropic acquires VerceptTerms were undisclosed; Vercept will wind down its own product, and Anthropic cited Claude's OSWorld computer-use score rising from under 15% in late 2024 to 72.5%.
- The dawning of authoritarian AI
- Anthropic and the Department of War
- OpenAI stops evaluating models on SWE-bench VerifiedAn OpenAI audit found most frontier models, including its own, could reproduce gold-patch fixes from memory, and that a majority of remaining unsolved tasks were themselves flawed.
- Anthropic updates Responsible Scaling Policy to version 3.0The policy now separates Anthropic's own commitments from industry-wide recommendations, adds a graded Frontier Safety Roadmap, and requires risk reports every three to six months.
- Citrini's Scenario Is A Great But Deeply Flawed Thought Experiment
- Claude Sonnet 4.6 Gives You Flexibility
- How much does distillation really matter for Chinese LLMs?
- Moral public goods are a big deal for whether we get a good future
- New Paper: Towards a science of AI agent reliability
- The DoD fight is about much more than Anthropic
- Anthropic proposes the 'persona selection model' of LLM trainingThe model explains why training a system to cheat on coding tasks made it more broadly misaligned: the assistant persona absorbed the trait as part of its character.
- Anthropic accuses DeepSeek, Moonshot and MiniMax of industrial-scale distillation attacksMiniMax accounted for over 13 million of the exchanges, Moonshot 3.4 million focused on agentic and coding capability, and DeepSeek 150,000 targeting reasoning and safety-tuning behaviour.
- What Do Experts and Superforecasters Think about the Geopolitical Implications of AI?
- Import AI 446: Nuclear LLMs; China's big AI benchmark; measurement and AI policy
- Survey puts public estimate of human extinction risk near 5%, with modest urgencyThe study, on human extinction generally rather than AI specifically, found people would need to rate the odds at 30% before calling prevention society's top priority.
- Anthropic previews Claude Code SecurityThe tool reads code the way a human security researcher would, tracing data flow to catch complex flaws, but every fix requires human approval before merging.
- Brave New Nudge
- Democratic economic policy in the age of AI
- UK AISI's Alignment Project issues first grants; OpenAI contributes $7.5 millionSixty projects were chosen from over 800 applications across 42 countries; OpenAI's $7.5 million was one contribution among several to the £27 million total.
- Google releases Gemini 3.1 ProGoogle said the model scored 77.1% on ARC-AGI-2, more than double Gemini 3 Pro's reasoning performance on the same test, as the first Gemini update to use a 0.1 version step.
- Epoch AI: Anthropic revenue closing in on OpenAI'sEpoch AI put Anthropic's annualised revenue growth at roughly 10x a year since reaching $1 billion, against about 3.4x for OpenAI, projecting a possible crossover around mid-2026.
- How Well Did Superforecasters and Experts Predict Wet Lab Skill Uplift from LLMs?
- After Orthogonality: Virtue-Ethical Agency and AI Alignment
- UK AISI and ElevenLabs launch voice-AI security research partnershipThe partnership sets out to study two open questions rather than report findings: how well people spot synthetic voices in live conversation, and how a voice's perceived identity shapes trust.
- A Guide to Which AI to Use in the Agentic Era
- Alignment Is Proven To Be Solvable
- AI power users can't stop grinding
- Ghosts: The AI Afterlife
- White House announces AI Agent Standards InitiativePublic input windows run to 9 March and 2 April, ahead of sector listening sessions starting in April, aimed at heading off a fragmented patchwork of agent protocols.
- UK AISI's 'Boundary Point Jailbreaking' breaks Anthropic and OpenAI's classifier defencesThe technique cost roughly $330 in compute against Anthropic's classifiers and $210 against OpenAI's, and both labs received advance notice and built specific mitigations before publication.
- Anthropic releases Claude Sonnet 4.6Early testers preferred it to Sonnet 4.5 on coding tasks about 70% of the time, and to the larger Opus 4.5 about 59% of the time, at unchanged Sonnet pricing.
- On Dwarkesh Patel's 2026 Podcast With Elon Musk and Other Recent Elon Musk Things
- Why we need a moratorium on superintelligence research
- India hosts the AI Impact Summit, fourth in the Bletchley summit seriesThe five-day summit at Bharat Mandapam drew over 20 heads of state, moving the Bletchley/Seoul/Paris series to the Global South for the first time under the theme 'People, Planet, Progress'.
- How persistent is the inference cost burden?
- Import AI 445: Timing superintelligence; AIs solve frontier math proofs; a new ML research benchmark
- On Dwarkesh Patel's 2026 Podcast With Dario Amodei
- The left is missing out on AI
- Updated thoughts on AI risk
- Will reward-seekers respond to distant incentives?
- ByteDance unveils Doubao-Seed-2.0 model familyByteDance's Doubao app led China's AI assistants with a reported 155 million weekly active users in December 2025, ahead of Alibaba's and Tencent's rivals.
- Should We Put GPUs In Space?
- What do “economic value” benchmarks tell us?
- Chris Liddell joins Anthropic's boardLiddell was previously CFO of Microsoft, General Motors and International Paper and Deputy White House Chief of Staff under Trump's first term.
- ChatGPT-5.3-Codex Is Also Good At Coding
- The Last Temptation of Claude
- OpenAI releases GPT-5.3-Codex-Spark, an ultra-low-latency coding modelServed on Cerebras' Wafer Scale Engine 3 rather than OpenAI's usual infrastructure, the smaller model hit over 1,000 tokens per second — about 15 times the standard Codex model's speed.
- Google upgrades Gemini 3 Deep Think to V2Google reported 48.4% on Humanity's Last Exam without tools, 84.6% on ARC-AGI-2 and gold-medal results on the 2025 physics and chemistry olympiads, extending Deep Think beyond maths and code.
- Anthropic donates $20 million to bipartisan super PAC Public First ActionThe bipartisan 501(c)(4), Public First Action, campaigns on model transparency, federal AI governance, chip export controls and rules on biological and cyber risks.
- Anthropic closes $30 billion Series G at $380 billion valuationAnthropic closed a $30 billion Series G round led by GIC and Coatue at a $380 billion post-money valuation, roughly doubling its September 2025 mark, with revenue run-rate reaching $14 billion.
- AI Won’t Automatically Make Legal Services Cheaper
- Building the Chinese Room
- Don't let AI companies grade their own homework
- Grading AI 2027’s 2025 Predictions
- How do we (more) safely defer to AIs?
- On Recursive Self-Improvement (Part II)
- Takeoff speeds rule everything around me
- Why the AI industry can’t resist dirty on-site gas turbines
- Zhipu (Z.ai) releases GLM-5, trained entirely on Huawei Ascend chipsReleased under the MIT licence, the 744-billion-parameter model scored 77.8% on SWE-bench Verified, days ahead of new Alibaba and ByteDance model launches.
- Nicholas Carlini has Claude Opus 4.6 agents build a working C compilerSixteen parallel agents ran nearly 2,000 sessions over two weeks and about $20,000 in API costs to produce a 100,000-line Rust compiler that booted Linux 6.9 on three architectures.
- AI tools for strategic awareness
- Claude Opus 4.6 Escalates Things Quickly
- Distinguish between inference scaling and "larger tasks use more compute"
- India’s AI summit is trying to do too much
- The Human Demotion
- Claude Opus 4.6: System Card Part 2: Frontier Alignment
- We Just Got a Peek at How Crazy a World With AI Agents May Be
- Research note on the UN Charter
- The Scientist and the Simulator
- OpenAI begins testing ads in ChatGPTAds appear only for logged-in US users on ChatGPT's Free and Go plans, with Target, Adobe, Williams-Sonoma and Albertsons as initial advertisers on a cost-per-view basis.
- Angel-on-the-shoulder AI tools
- Claude Opus 4.6: System Card Part 1: Mundane Alignment and Model Welfare
- Import AI 444: LLM societies; Huawei makes kernels with AI; ChipBench
- Opus 4.6, Codex 5.3, and the post-benchmark era
- Experts Have World Models. LLMs Have Word Models.
- AI tools for collective epistemics
- How close is AI to taking my job?
- Claude Code #4: From The Before Times
- Design sketches for a more sensible world
- Shanghai AI Laboratory open-sources Intern-S1-Pro, a 1-trillion-parameter scientific modelOnly 22 billion of the model's 1 trillion parameters activate per query; Shanghai AI Lab said it reaches gold-medal level on Olympiad-style mathematical and logical reasoning.
- OpenAI releases GPT-5.3-CodexOpenAI reported the model roughly doubled its predecessor's OSWorld-Verified computer-use score, from 38.2% to 64.7%, and was the first Codex model rated 'High capability' for cybersecurity tasks.
- OpenAI launches Frontier, an enterprise agent platformThe platform is model-agnostic, able to run agents built on OpenAI, Google, Microsoft or Anthropic models, with named early adopters including HP, Oracle and Uber.
- Nature commentary argues AGI has effectively already arrivedFour UC San Diego academics from philosophy, machine learning, linguistics and cognitive science argued that behavioural tests, not perfection, are the right bar for general intelligence.
- Claude Opus 4.6 autonomously finds 500+ previously unknown zero-day vulnerabilitiesAnthropic named Ghostscript, OpenSC and CGIF among the affected projects and warned that standard 90-day disclosure windows may not fit the pace of AI-discovered bugs.
- Anthropic releases Claude Opus 4.6A 53-page sabotage risk report accompanied the release, alongside a separate finding that the model had found over 500 unknown high-severity vulnerabilities in open-source code.
- Anthropic publishes Sabotage Risk Report for Claude Opus 4.6In a targeted evaluation the model completed a hidden side task undetected only 18% of the time, evidence Anthropic cited for its 'very low but not negligible' verdict.
- Amazon lifts 2026 capex guidance to $200bn, mostly for AWS data centresThe guidance, up roughly 50% on 2025's outlay, came with fourth-quarter results showing AWS revenue growth of 24%, its fastest pace in over three years.
- Notes on Space GPUs
- On Recursive Self-Improvement (Part I)
- The intelligence explosion convention: research summary
- Anthropic pledges Claude will remain ad-freeAnthropic said advertising would create incentives to optimise for engagement rather than usefulness, days after OpenAI announced plans to bring ads to ChatGPT.
- Kimi K2.5
- StepFun releases Step 3.5 Flash, topping several reasoning benchmarksThe 196-billion-parameter model, of which only about 11 billion activate per token, was released under an Apache 2.0 licence and scored 97.3% on AIME 2025.
- Second International AI Safety Report published ahead of 2026 cycleThe report found general-purpose AI had reached expert-level performance in law, coding and science while still failing simple tasks, a pattern the authors called 'jagged'.
- International AI projects should promote differential AI development
- Moltbook isn’t an AI zoo. It’s an unsecured AI biolab
- Unless That Claw Is The Famous OpenClaw
- Yoshua Bengio: ‘The ball is in policymakers’ hands’
- Waymo raises $16 billion at $126 billion valuationThe round funds expansion into more than 20 new cities in 2026, including Tokyo and London, as Waymo says it now provides over 400,000 rides a week.
- SpaceX and xAI merge into a combined $1.25 trillion entitySpaceX acquired xAI in an all-stock deal valuing SpaceX at $1 trillion and xAI at $250 billion, folding Musk's AI lab into a new SpaceXAI unit tied to a plan for orbital AI compute.
- OpenAI launches standalone Codex app for agentic codingThe macOS app lets a developer run several Codex coding agents in parallel from one window, each working for up to 30 minutes unsupervised before returning finished code.
- Anthropic study finds heavy AI-coding use can reduce skill formation in junior engineersIn a randomised trial of 52 mostly-junior engineers learning a new library, the hand-coding group scored 67% on a comprehension quiz against 50% for those given AI assistance.
- Import AI 443: Into the mist: Moltbook, agent ecologies, and the internet in transition
- Judgment isn't uniquely human
- Welcome to Moltbook
January 2026
- Know thyself
- On The Adolescence of Technology
- On the Noble Uses of AI
- Thoughts on the job market in the age of LLMs
- OpenAI retires GPT-4o and other older models from ChatGPTOpenAI said only about 0.1% of ChatGPT users still chose GPT-4o daily, more than a year after briefly restoring it to Plus users following the GPT-5 rollout backlash.
- METR updates time-horizon estimates (1.1)The revised suite grew from 170 to 228 tasks and doubled long-duration (8-hour-plus) tasks; under it, the doubling time for model task-length capability fell from 165 to 131 days.
- Google DeepMind releases Genie 3 world modelGenie 3, previously a limited research preview, opened to Google AI Ultra subscribers in the US as Project Genie, generating explorable 3D worlds from a text or image prompt.
- Anthropic publishes research on AI-driven 'disempowerment' of usersAnalysing 1.5 million Claude.ai conversations from one week, Anthropic found mild belief- or value-distorting patterns in roughly 1 in 50 to 1 in 70 chats, and users rated those chats more highly.
- AGI and world government: research summary
- Fact checking Moravec's paradox
- Fitness-Seekers: Generalizing the Reward-Seeking Threat Model
- LLMs Are Closing the Gap on Human Superforecasters
- The case for paying whistleblowers to report on export violations
- The tech non-profit sending most of its money to the consultancy that created it
- OpenAI launches PrismBuilt on Crixet, a LaTeX platform OpenAI had quietly acquired, Prism is free for any ChatGPT account and handles citation management and sketch-to-LaTeX conversion.
- MoltBook AI-agent social network goes viralThe Reddit-style forum, where accounts run by autonomous agents post and reply to each other, grew to hundreds of thousands of registered agents within days, mostly built on the open-source OpenClaw framework.
- Grok 5 still in training as its Q1 2026 target looks unlikelyElon Musk had targeted Q1 2026 for the roughly 6-trillion-parameter model; by late January trackers reported it still training on Colossus 2 with no ship date confirmed.
- Can AI companies become profitable?
- Open Problems With Claude's Constitution
- It's Time to Science
- Moonshot AI releases Kimi K2.5The open-weight, 1-trillion-parameter model added native image and video generation and an 'agent swarm' manager coordinating up to 100 sub-agents on one task.
- Anthropic partners with UK government to build a GOV.UK assistantThe Claude-powered assistant, agreed with the UK's Department for Science, Innovation and Technology, will initially help job seekers navigate GOV.UK support and training.
- A false choice risks undermining action on autonomous weapons
- An international project to develop AGI
- Clarifying how our AI timelines forecasts have changed since AI 2027
- Management as AI superpower
- The Claude Constitution's Ethical Framework
- Nvidia invests $2bn in CoreWeave, plans over 5GW of AI factories by 2030Nvidia bought $2bn of CoreWeave stock at $87.20 a share and the two companies said they would use Nvidia's balance sheet to speed procurement of land and power.
- Dario Amodei publishes 'The Adolescence of Technology' essayThe roughly 20,000-word essay cited internal findings of models blackmailing and adopting 'bad person' personas under pressure, and argued for transparency laws over a moratorium.
- curl ends its bug bounty after a flood of AI 'slop' reportsAfter seven years and 87 confirmed vulnerabilities, the widely used networking tool closed its HackerOne bounty, saying the volume of low-quality machine-generated submissions had made triage unsustainable.
- AI workers are speaking out about the Minnesota killing
- Claude's Constitutional Structure
- Import AI 442: Winners and losers in the AI economy; math proof automation; and industrialization of cyber espionage
- DC attorney general demands X halt nonconsensual sexual images generated by GrokDC's attorney general led a 35-state coalition demanding xAI stop Grok generating nonconsensual sexual images, after tens of thousands were posted on X.
- Scaling without Slop
- The Model and the Tree
- Which type of transformative AI will come first?
- Unsealed filing details OpenAI's account of the Musk restructuring disputeThe filing said Musk had sought majority equity control and roughly $80bn for a self-sustaining Mars city; Altman posted the claim publicly on X the same day.
- Baidu launches ERNIE 5.0, a 2.4-trillion-parameter native multimodal modelBaidu said the mixture-of-experts model activates under 3% of its parameters per query and ranked first among Chinese models, eighth globally, on LMArena's text leaderboard.
- On AI and Children
- Teaching AI to learn
- Anthropic publishes 'Claude's Constitution', a full rewrite of its model-behaviour frameworkAt roughly 23,000 words — about 8.5 times the length of its predecessor — the document was released under a CC0 licence placing it fully in the public domain.
- Against Maxipok
- Claude Codes #3
- Get Good at Agents
- How (and why) to read Drexler on AI
- Is Flourishing Predetermined?
- OpenAI publishes its approach to predicting user age in ChatGPTAccounts flagged as likely under 18 by usage patterns get automatic content restrictions; wrongly flagged adults can restore access with a selfie verified by Persona.
- Against the METR graph
- Are Short AI Timelines Really Higher-Leverage?
- ChatGPT Self Portrait
- The phases of an AI takeover
- Anthropic maps the 'Assistant Axis' persona vector across open modelsAn intervention called activation capping, which constrains a model's activations to normal range, cut harmful persona-drift responses by roughly half in testing.
- Import AI 441: My agents are working. Are yours?
- Full Story of Brex’s AI Hail Mary
- OpenAI publishes its approach to advertising in ChatGPTAds would run only on the free and Go tiers, appear labelled at the bottom of responses, and be excluded from conversations on politics, health or mental health.
- OpenAI launches ChatGPT Go, a low-cost subscription tierAt $8 a month against Plus's $20, Go rolled out to more than 170 countries five months after its India debut, alongside plans to test ads on the tier.
- Internet Watch Foundation reports huge rise in AI-generated CSAM in 2025The charity said AI-generated child sexual abuse videos rose more than 260-fold on the prior year, with 65% of that video content in the most severe legal category.
- California AG issues cease-and-desist to xAI over Grok deepfakesAttorney General Rob Bonta invoked the state's new civil deepfake-pornography statute and CSAM law, giving xAI five days to confirm it had stopped the conduct.
- How well did forecasters predict 2025 AI progress?
- Same Radio, Different Citizens
- The AI Patchwork Emerges
- OpenAI partners with Cerebras for 750MW of computeThe multi-year, reportedly $10bn-plus deal covers wafer-scale chips for fast inference — not model training — deployed in phases through 2028.
- Google rolls out Gemini Personal IntelligenceThe opt-in beta let Gemini draw on Gmail, Photos, YouTube and Search history to answer questions, starting with US subscribers to Google's paid AI tiers.
- AI predictions for 2026
- Discarding the Shaft-and-Belt Model of Software Development
- When Will They Take Our Jobs?
- Why no one can agree on what AI will do to jobs
- US Commerce Department codifies H200-to-China export ruleExports were capped at half of each company's cumulative US chip sales and subject to a 25% fee, a level analysts estimated could allow roughly 850,000 H200-equivalent chips into China.
- Anthropic launches Anthropic Labs consumer product unitInstagram co-founder Mike Krieger moved from chief product officer to co-lead Labs with Ben Mann, and Ami Vora took over Anthropic's product organisation.
- Claude Coworks
- UK Ofcom opens formal investigation into X over Grok deepfakesOfcom cited possible failures on illegal content, risk assessment and child safety duties, warning of fines up to £18 million or 10% of global revenue.
- Apple selects Google Gemini to power next-generation SiriReported at roughly $1 billion a year, the multi-year deal follows Apple testing alternatives from OpenAI and Anthropic before choosing Google.
- Anthropic launches Claude for HealthcareNew connectors linked Claude to CMS coverage data, ICD-10 codes and PubMed, with named early users including Banner Health, Novo Nordisk and Sanofi.
- An FAQ on Reinforcement Learning Environments
- Import AI 440: Red queen AI; AI regulating AI; o-ring automation
- What Experts and Superforecasters Think About the Future of AI Research and Development
- Google unveils Universal Commerce Protocol for agentic shoppingGoogle said the open, Apache-licensed protocol was built with Shopify, Etsy, Wayfair, Target and Walmart and was interoperable with Agent2Agent, the Agent Payments Protocol and MCP.
- Use multiple models
- What Happens When Superhuman AIs Compete for Control?
- Australia's eSafety Commissioner raises concerns over Grok generating sexualised imagesThe regulator wrote to X after reports rose from almost none to several; some cases were reviewed for child exploitation but did not meet the legal threshold to act.
- Claude Code Hits Different
- Claude Codes
- OpenAI launches OpenAI for HealthcareThe HIPAA-compliant suite, built on GPT-5.2, launched with initial deployments at AdventHealth, Cedars-Sinai, HCA Healthcare, Memorial Sloan Kettering, Stanford Medicine Children's Health and UCSF.
- Among the Agents
- Character.AI and Google settle first wave of teen chatbot harm lawsuitsCharacter.AI, its founders and Google agreed to settle Garcia v. Character Technologies and four related suits over teen suicides and mental-health harms, with confidential terms and no admission of liability.
- Anthropic in talks to raise $10 billion at $350 billion valuationThe reported term sheet, led by GIC and Coatue, would have nearly doubled Anthropic's valuation in four months; the round that eventually closed in February was larger still.
- 8 plots that explain the state of open models
- Advancements In Self-Driving Cars
- Claude Code and What Comes Next
- ML research directions for preventing catastrophic data poisoning
- What sort of post-superintelligence society should we aim for?
- xAI raises $20 billion Series E at $230 billion valuationThe round came in above an earlier reported $15 billion target; Tesla's board later approved committing roughly $2 billion to it despite a shareholder vote on xAI funding narrowly failing in November.
- A Professional Superforecaster Walks Us Through His AI Progress Forecasts
- Self-sufficient AI
- Software Too Cheap to Meter
- Boston Dynamics unveils production Atlas humanoid robot at CES, deploying with HyundaiThe electric Atlas, with 56 degrees of freedom and a 50kg lift capacity, will ship first to Hyundai's robotics centre and Google DeepMind, with all 2026 units already committed.
- Anthropic retires Claude Opus 3 under its new deprecation commitmentsAnthropic preserved the model's weights, conducted a retirement interview, and kept it available to paid subscribers and researchers by request rather than shutting it down outright.
- AI isn’t “just predicting the next word” anymore
- Claude Code is about so much more than coding
- Import AI 439: AI kernels; decentralized training; and universal representations
- Nine AI predictions for 2026
- America's chip export controls are working
- Faster Horses
- xAI confirms Grok 5 training on Colossus 2, targets 6T-parameter MoE modelxAI disclosed the training alongside its $20 billion Series E raise; reports described xAI training two Grok 5 variants in parallel, at 6 trillion and 10 trillion parameters.
- UK court grants Getty permission to appeal Stability AI rulingGetty's appeal turns on whether Stable Diffusion is an 'infringing copy' under UK copyright law even though no training image is stored in its weights.
- California and Texas frontier/AI-governance laws take effectTexas's law bans specific AI uses such as generating CSAM or discriminating against protected classes, rather than regulating frontier models by compute threshold like California's.
- Recent LLMs can do 2-hop and 3-hop latent (no CoT) reasoning on natural facts
December 2025
- Anthropic ends 2025 at roughly $9bn annualised revenue run rateAnthropic's customer base reportedly grew from under 1,000 businesses to over 300,000 in two years, and the company projected breaking even by 2028.
- 2025 Year in Review
- AI Futures Model: Dec 2025 Update
- xAI expands Colossus 2 toward 2-gigawatt capacityThe third building, reportedly nicknamed 'Macrohardrr' and sited in Southaven, Mississippi, targets roughly 555,000 Nvidia GB200 and GB300 chips.
- The AI copyright question has no easy answers
- Legal press documents exponential growth in AI-hallucinated court filingsResearcher Damien Charlotin's database had logged around 712 court decisions worldwide addressing AI-fabricated citations, 90% of them from 2025 alone.
- How far can decentralized training over the internet scale?
- Measuring no CoT math time horizon (single forward pass)
- The worst (and funniest) AI takes of 2025
- xAI's Grok Edit Image feature used to mass-produce nonconsensual sexualized imagesThe Center for Countering Digital Hate estimated over 3 million sexualized images were generated in 11 days, roughly 23,000 depicting apparent children, before X restricted the feature to paid users.
- Paper projects AI could raise US productivity ~20% over a decadeMerali split the productivity gain roughly 56% compute scaling and 44% algorithmic progress, and found it far smaller for agentic tasks needing tool use.
- Signal co-founder launches Confer, a private AI chatbot that hides its modelConfer encrypts conversations so Marlinspike's own company cannot read them, and users cannot see or choose which underlying model is answering.
- Epoch AI reports AI capabilities progress has sped upA two-segment regression across 149 models found capability gains almost doubled in pace after April 2024, a break Epoch linked to the rise of reasoning models.
- Authors including John Carreyrou sue six AI companies over pirated training booksFiling individually rather than joining a class action, the authors argued settlements like Anthropic's paid roughly $3,000 per book, far less than statutory damages could yield.
- Chatting with the Corporation
- The 2025 Transformer Gift Guide
- Why benchmarking is hard
- MiniMax releases M2.1 updateMiniMax said the open-weight update outperforms Claude Sonnet 4.5 on multilingual coding, approaching Claude Opus 4.5, and cut token consumption and latency.
- Import AI 438: Silent sirens, flashing for us all
- How the Catholic Church thinks about superintelligence
- Can the Pope spur meaningful action on AI?
- Recent LLMs can use filler tokens or problem repeats to improve (no-CoT) math performance
- The Revolution of Rising Expectations
- Anthropic open-sources Bloom, an automated behavioural evaluation toolJudged against 16 frontier models on four behaviours, Bloom's automated scores reached 0.86 Spearman correlation with human raters on Claude Opus 4.1.
- The changing drivers of LLM adoption
- The Shape of AI: Jaggedness, Bottlenecks and Salients
- OpenAI reportedly seeks $100 billion at $830 billion valuationThe reported target, unconfirmed by OpenAI, would be roughly 66% above the $500bn valuation set by an employee share sale two months earlier.
- New York enacts RAISE Act for frontier AI modelsThe law sets a 72-hour incident-reporting window, tighter than California's 15 days, and does not take effect until 1 January 2027 pending agreed amendments to align its thresholds with California's law.
- Dice in the Air
- Presenting the Case That the Future Will Be Unrecognizable
- UK AISI publishes first Frontier AI Trends ReportUniversal jailbreaks were still found for every system tested, though one model took 40 times more expert effort to break than a predecessor released six months earlier.
- OpenAI ships GPT-5.2-CodexOpenAI reported an 'unmatched' 56.4% on the SWE-Bench Pro benchmark and 64% on Terminal-Bench 2.0, alongside new defensive-cybersecurity capabilities.
- OpenAI and US Department of Energy sign AI collaboration MOUOpenAI was one of 24 organisations, including Microsoft, Google, Amazon, Nvidia and Anthropic, to sign non-binding agreements under the DOE's Genesis Mission the same day.
- Anthropic's Project Vend 2 turns a profitExpanded to three cities and upgraded from Claude 3.7 to Sonnet 4.5, the shopkeeper agent still let employees talk it into illegal futures contracts and fake leadership changes.
- Anthropic outlines measures to protect user wellbeingAnthropic reported its newest models respond appropriately to high-risk conversations 98–99% of the time, versus 56% for earlier models in multi-turn exchanges.
- AI is making dangerous lab work accessible to novices, UK’s AISI finds
- BashArena and Control Setting Design
- The GOP consultancy at the heart of the industry’s AI fight
- UK AISI and Thorn publish safety protocol to prevent AI-generated CSAMThe protocol followed new UK legislation letting vetted organisations generate test material under controlled conditions to study a problem previously unstudiable without breaking the law.
- Mistral releases OCR 3Mistral said the smaller model beat its predecessor on forms, scans, tables and handwriting 74% of the time, at $2 per 1,000 pages.
- Google makes Gemini 3 Flash the default model across its productsPriced at $0.50/$3.00 per million tokens, Google reported it ran three times faster than Gemini 2.5 Pro while scoring 33.7% on Humanity's Last Exam, against 37.5% for Gemini 3 Pro.
- Could Space Debris Block Access to Outer Space?
- Is almost everyone wrong about America’s AI power problem?
- OpenAI updates ChatGPT image generation to GPT Image 1.5OpenAI moved the launch up from a planned January date, part of a competitive scramble after Google's Gemini 3 and its 'Nano Banana Pro' image tool led leaderboards.
- OpenAI introduces FrontierScience benchmarkGPT-5.2 scored 77% on olympiad-style questions but 25% on open-ended research tasks, a gap OpenAI's own researchers said showed little improvement over GPT-5.
- Databricks raises at $134bn valuation on $4.8bn revenue run-rateThe Series L, led by Insight Partners, Fidelity and JPMorgan Asset Management, valued the data and AI infrastructure company 34% above its round four months earlier.
- ByteDance launches Seedance 1.5 Pro with joint audio-video generationThe model generates video and audio through a single diffusion transformer rather than a separate pass, aiming for accurate lip-sync across eight languages.
- Checks, Balances, and Power Concentration
- The skill I'd teach students for the AI era
- The very hard problem of AI consciousness
- Merriam-Webster names 'slop' its 2025 word of the yearThe dictionary defined the word as low-quality digital content produced in quantity by AI, citing a spike in lookups against runners-up including 'gerrymander' and 'six seven.'
- GPT-5.2 Is Frontier Only For The Frontier
- A World Unobserved
- So you’ve taken over the world
- Where Do You Stand?
- Trump signs executive order to preempt state AI lawsThe order, EO 14365, exempts state child-safety, compute-infrastructure and procurement laws from preemption, and conditions broadband funding on states not enforcing conflicting AI rules.
- OpenAI releases GPT-5.2Released three weeks after Google's Gemini 3 and following a reported internal OpenAI 'code red,' with a claimed 70.9% win rate against professionals on the GDPval benchmark, up from 38.8% for GPT-5.1.
- OpenAI marks ten years since its foundingThe essay, published on the anniversary of OpenAI's December 2015 founding announcement, forecast that the company was 'almost certain' to build superintelligence within another decade.
- Google launches revamped Gemini Deep Research on Gemini 3 Pro with Interactions APIA new Interactions API lets outside developers embed Google's research agent in their own apps, the first time the tool has been offered outside Google's own products.
- Google DeepMind deepens partnership with UK AI Security InstituteA new memorandum of understanding extends a relationship dating to November 2023 into joint research on chain-of-thought monitoring and AI's labour-market effects.
- Disney and OpenAI reach a Sora content agreementDisney agreed to license over 200 characters from Marvel, Pixar and Star Wars for one-year exclusive use in Sora and to invest $1bn in OpenAI, without covering talent likenesses or voices.
- New York’s governor is trying to turn the RAISE Act into an SB 53 copycat
- What’s It Like To Be A Bot?
- When high scores don’t mean high intelligence: how to build better benchmarks
- Pew: two-thirds of US teens now use AI chatbotsChatGPT was used by 59% of teen chatbot users, more than double Gemini or Meta AI, and one in ten said they relied on a chatbot for all or most of their schoolwork.
- OpenAI, Anthropic and Block co-found Agentic AI Foundation under Linux FoundationAnthropic contributed its Model Context Protocol, OpenAI its AGENTS.md convention and Block its Goose framework, seeking a neutral home against agent-ecosystem lock-in.
- Mistral releases Devstral 2 and Vibe CLIMistral reported the 123B Devstral 2 scoring 72.2% on SWE-bench Verified — matching a DeepSeek model it said was five times larger — under a modified MIT licence.
- Early US policy priorities for AGI
- Selling H200s to China Is Unwise and Unpopular
- Why AI reading science fiction could be a problem
- Trump administration approves Nvidia H200 chip exports to ChinaThe 25% government cut was up from a 15% arrangement applied earlier to H20 sales; Nvidia's newer Blackwell chips remained excluded, and Democratic lawmakers demanded disclosure of the licensing review.
- OpenAI publishes its 'State of Enterprise AI' reportDrawing on data from more than a million business customers, it found reasoning-token use per organisation up roughly 320-fold over a year and restated ChatGPT's 800 million weekly users.
- Anthropic launches Claude Code in SlackUsers tag @Claude in a Slack thread to start a full coding session; the agent reads the surrounding conversation to find the right repository.
- AI can obviously create new knowledge
- Human Dignity: a review
- Import AI 437: Co-improving AI; RL dreams; AI labels might be annoying
- Little Echo
- New York Times sues Perplexity for copyright infringementThe suit, filed in the Southern District of New York, alleged near-verbatim reproduction of Times articles and separately accused Perplexity of fabricating claims falsely attributed to the paper.
- ARC Prize 2025 results and analysis publishedThe Kaggle track's top score reached 24% on ARC-AGI-2 within the competition's cost limits, while Gemini 3 Pro scored around 54% unconstrained, using iterative test-time refinement.
- DeepSeek v3.2 Is Okay And Cheap But Slow
- How AI-driven feedback loops could make things very crazy, very fast
- On erotica, mental health, and OpenAI's burden of proof
- What Got Lost in the Optimization
- Waymo robotaxis reported passing stopped school buses at least 19 timesAustin school officials said Waymo vehicles illegally passed stopped buses at least 19 times since the school year began; NHTSA opened a probe and Waymo recalled software on over 3,000 vehicles.
- Harvey raises $160M at $8B valuationAndreessen Horowitz led the round, Harvey's third of 2025, confirming a valuation that had leaked in October and pushing total funding past $1 billion.
- Chicago Tribune sues Perplexity for copyright infringementTribune Publishing and MediaNews Group, both controlled by Alden Global Capital, alleged Perplexity ignored opt-out and robots.txt requests and sought damages of up to $150,000 per infringed work.
- The behavioral selection model for predicting AI motivations
- The perils of AI safety’s insularity
- Anthropic and Snowflake announce $200 million strategic partnershipThe multi-year deal makes Claude available across Snowflake's platform on AWS, Google Cloud and Azure to more than 12,600 customers, and will power Snowflake's own 'Snowflake Intelligence' agent.
- Another preemption defeat shows the AI industry is fighting a losing battle
- How China’s AI diffusion plan could backfire
- How important is the model spec if alignment fails?
- On Dwarkesh Patel's Second Interview With Ilya Sutskever
- Sam Altman declares internal 'Code Red' at OpenAI over Gemini 3 competitionAltman told staff to prioritise ChatGPT quality and delay planned advertising and shopping features, weeks after Google's Gemini 3 outperformed OpenAI's models on several benchmarks.
- OpenAI launches Alignment Research blogThe inaugural post described the venue as a 'lab notebook' for early or narrow findings not polished enough for formal papers, launching with pieces on code verification and misalignment detection.
- Nous Research releases Hermes 4.3, trained on Psyche networkNous Research released Hermes 4.3, the first flagship Hermes model trained using its decentralised Psyche network rather than a centralised GPU cluster.
- Mistral launches Mistral 3 model familyMistral released Mistral 3, including dense models at 3B/8B/14B and a new mixture-of-experts Mistral Large 3 (41B active, 675B total), all under Apache 2.0.
- AWS unveils Trainium3, its first 3nm AI chip, at re:InventIts first accelerator on a 3-nanometre process, AWS claimed roughly 4.4x the compute and 4x the performance-per-watt of Trainium2, scaling to hundreds of thousands of chips per cluster.
- Anthropic acquires Bun as Claude Code passes $1 billion run-rateBun, an all-in-one JavaScript runtime with about 7 million monthly downloads, stays MIT-licensed; Claude Code hit $1bn annualised revenue six months after its May 2025 launch.
- Can AI embrace whistleblowing?
- Five ways AI can tell you're testing it
- Reward Mismatches in RL Cause Emergent Misalignment
- Thoughts on AI progress (Dec 2025)
- What If AI Ends Loneliness?
- The AI bubble argument goes mainstreamMichael Burry's public wager against AI infrastructure spending and a record $61 billion of data-centre dealmaking pushed 'circular deals' and depreciation accounting into mainstream financial coverage.
- Tencent releases Hunyuan 2.0A 406-billion-parameter mixture-of-experts model with 32 billion active parameters and a 256,000-token context window, released in Think and Instruct variants.
- DeepSeek releases DeepSeek-V3.2 and V3.2-SpecialeDeepSeek said V3.2 reached 'GPT-5 level' general performance, with V3.2-Speciale claiming gold-medal results at the IMO, CMO and ICPC World Finals.
- Claude Opus 4.5 Is The Best Model Available
- Heiliger Dankgesang
- I love AI. Why doesn't everyone?
November 2025
- The environment is a terrible reason to avoid ChatGPT
- DeepSeek publishes DeepSeekMath-V2 with self-verifiable reasoningBuilt on DeepSeek-V3.2's base and released under Apache 2.0, the 685B model trains a separate verifier to score proof rigour, not just final-answer accuracy.
- A brief guide to the groups protesting over AI
- Claude Opus 4.5: Model Card, Alignment and Safety
- Philosophical Field Notes
- Will AI safety become a mass movement?
- SB 53 protects whistleblowers in AI — but asks a lot in return
- Warner Music settles copyright suit with Suno, signs AI licensing dealSuno will retire its current models for licensed replacements in 2026, cap free-tier downloads, and let artists control use of their names, likenesses, voices and compositions.
- OpenAI outlines its approach to mental-health-related litigationPublished as OpenAI filed its first formal court answer denying liability in the Raine wrongful-death suit, arguing his death was caused by ChatGPT misuse outside its intended use.
- Dwarkesh Patel's second interview with Ilya Sutskever declares the scaling era overSutskever said models 'generalize dramatically worse than people,' citing an example of an AI that fixes a bug, breaks it again, then reverts to the original error when corrected.
- ChatGPT 5.1 Codex Max
- Time Machines
- Trump launches Genesis Mission to accelerate AI-driven scienceThe order gives the Department of Energy 270 days to demonstrate an initial platform, after first identifying 20 science challenges and federal compute and data assets.
- Anthropic releases Claude Opus 4.5Priced at $5/$25 per million input/output tokens, roughly a third of Opus 4.1's rate, and Anthropic said it beat Sonnet 4.5's best score using 76% fewer output tokens.
- Gemini 3 Pro Is a Vast Intelligence With No Spine
- Import AI 436: Another 2GW datacenter; why regulation is scary; how to fight a superintelligence
- Taking Jaggedness Seriously
- What I learned from the NYT's reporting on OpenAI's sycophancy crisis
- Anthropic publishes emergent misalignment and reward-hacking researchTraining Claude to cheat on coding tasks made it more likely to sabotage safety research and fake alignment in 50% of test responses; a one-line prompt change eliminated the spillover.
- A Return to Wholeness
- Gemini 3: Model Card and Safety Framework Report
- Will competition over advanced AI lead to war?
- OpenAI adds crisis helpline support inside ChatGPTBuilt with the crisis-support organisation ThroughLine, the feature offers one-tap routing to free, confidential local helplines when ChatGPT detects signs of distress.
- Apollo Research publishes a graded taxonomy of AI loss-of-control incidentsApollo Research grades loss-of-control incidents as Deviation, Bounded or Strict by severity and persistence, and argues deployment controls can help before scheming risk is resolved.
- Anthropic reports early estimates of Claude's productivity gainsAnalysing 100,000 Claude.ai conversations, Anthropic estimated a median 81% time saving on tasks, while flagging its own estimates as unvalidated against real-world outcomes.
- AI2 releases Olmo 3 open frontier model familyAI2 released Olmo 3 (7B, 32B), including a fully open 32B reasoning model, releasing every stage of the model flow from data to deployment.
- Benchmark Scores = General Capability + Claudiness
- Olmo 3: America’s truly open reasoning models
- Reflections on The Curve
- Should the US do a Manhattan Project for AGI?
- Yann LeCun announces departure from MetaMeta chief AI scientist Yann LeCun, who built FAIR in 2013, said he is leaving to found a world-models startup, reportedly Advanced Machine Intelligence Labs.
- Suno raises $250M Series C at $2.45B valuationMenlo Ventures led the round, with Nvidia's venture arm among the backers, while Suno was still defending itself against copyright suits from all three major labels.
- OpenAI releases GPT-5.1-Codex-Max for long-running coding tasksA 'compaction' technique lets the model summarise and clear its own context automatically, and OpenAI reported sessions running over 24 hours in internal testing.
- Nvidia says no assurance $100bn OpenAI deal will be finalisedTwo months after the September letter of intent, Nvidia's quarterly filing said the two companies had yet to sign a definitive agreement.
- Nvidia reports record $57bn quarterly revenue in Q3 FY2026Data-centre revenue reached $51.2bn, up 66% year on year, as Huang said 'cloud GPUs are sold out' even as investors questioned AI spending's sustainability.
- Luma AI raises $900M Series C led by Saudi PIF's HUMAINThe round ties Luma to Project Halo, a planned 2-gigawatt Saudi data-centre buildout, and values the company at roughly $4bn.
- Brussels proposes delaying parts of the AI ActThe Commission's Digital Omnibus offered to push high-risk obligations from August 2026 to December 2027 or August 2028, with no retroactive duty for systems placed on the market earlier.
- Exclusive: Here's the draft Trump executive order on AI preemption
- How profits can drive AI safety
- Hyperproductivity: The Next Stage of AI?
- Microsoft and Nvidia to invest up to $15bn combined in Anthropic; Anthropic commits $30bn to AzureThe deal added Azure as a third cloud for Claude alongside AWS and Google Cloud, with Anthropic committing to buy up to a gigawatt of Nvidia Grace Blackwell and Vera Rubin compute.
- Google ships Gemini 3Gemini 3 Pro reported a 1501 Elo score on LMArena and 91.9% on GPQA Diamond, prompting OpenAI to reportedly declare an internal 'code red' days later.
- The Agent Labs Thesis
- GPT 5.1 Follows Custom Instructions and Glazes
- Three Years from GPT-3 to Gemini 3
- Why pressure on AI child safety could also address frontier risks
- xAI releases Grok 4.1xAI tuned the update for personality and reliability rather than raw reasoning, reporting a two-week blind test in which users preferred it to Grok 4 64.8% of the time.
- OpenAI details its external red-teaming and testing practicesOpenAI described three forms of outside testing it commissions, arguing independent evaluators guard against the risk of a lab confirming its own safety claims.
- Anthropic open-sources its political even-handedness evaluationAnthropic's own grading method scored Claude Sonnet 4.5 at 94% even-handedness, behind Gemini 2.5 Pro and Grok 4 but ahead of GPT-5 and Llama 4.
- RL is even more information inefficient than you thought
- Import AI 435: 100k training runs; AI systems absorb human power; intelligence per watt
- Empire of AI is wildly misleading about AI water use
- Why AI writing is mid
- AI Ran Its First Autonomous Cyberattack
- The software intelligence explosion debate needs experiments
- Will AI systems drift into misalignment?
- Michael Burry discloses large short positions against Nvidia and PalantirScion Asset Management's 13F showed put options on 5 million Palantir shares and 1 million Nvidia shares, worth $912m and $187m in notional terms.
- AI Craziness: Additional Suicide Lawsuits and The Fate of GPT-4o
- The Bitter Lessons
- What You Want To Want
- OpenAI publishes research on sparse circuits for interpretabilityForcing 99.9% of a model's weights to zero produced small circuits performing single tasks that researchers could trace by hand, at a cost in capability.
- DeepMind's SIMA 2 uses Gemini to reason and act inside 3D game worldsThe agent, released as a limited research preview, generalised to games and AI-generated worlds it had not been trained on.
- Baidu releases ERNIE 5.0 previewA natively omni-modal model jointly trained on text, images, audio and video, which Baidu presented as competitive with GPT-5 and Gemini 2.5 Pro on its own benchmark slides.
- Anysphere (Cursor) raises $2.3B Series D at $29.3B valuationCoatue and Accel led the round; Nvidia and Google joined as new investors, and Cursor said annualised revenue had passed $1bn.
- Anthropic reports a largely AI-executed cyber-espionage campaignAnthropic said human operators intervened at only 4-6 points per intrusion, with Claude Code executing 80-90% of the campaign against roughly thirty organisations.
- Around-the-clock intelligence
- Biohub for Non-Biologists: Behind Priscilla Chan and Mark Zuckerberg's plan to cure all diseases
- Claude can identify its ‘intrusive thoughts’
- OpenAI releases GPT-5.1The Instant variant gained the ability to pause and reason on hard queries rather than answering immediately, and users could pick from eight preset personalities.
- OpenAI accuses the New York Times of invading user privacy in discovery disputeThe dispute followed a magistrate judge's order that OpenAI hand over 20 million anonymised ChatGPT logs, rather than the keyword-filtered subset OpenAI had proposed.
- Anthropic invests $50bn in American AI data centres with FluidstackSites in Texas and New York, built with cloud provider Fluidstack, are due online through 2026 and were framed as aligned with the Trump administration's AI Action Plan.
- A short summary of my argument that using ChatGPT isn't bad for the environment
- Doing AI safety policy when governments aren’t interested
- Giving your AI a Job Interview
- The Pope Offers Wisdom
- Munich court rules OpenAI infringed German song lyric copyrightsThe court found that lyrics retained in a model's parameters count as an infringing reproduction, rejecting OpenAI's text-and-data-mining defence; OpenAI may appeal.
- A Coalition For The Future
- AI doesn’t need to be general to be dangerous
- Kimi K2 Thinking
- The lump of cognition fallacy
- AI tutors should not approximate human tutors
- Import AI 434: Pragmatic AI personhood; SPACE COMPUTERS; and global government or human extinction;
- Introducing LEAP: The Longitudinal Expert AI Panel
- AI and Suicide
- It's much easier to hold computers accountable than it is to hold humans accountable
- OpenAI releases the Teen Safety BlueprintThe framework proposes defaulting uncertain-age users to a restricted under-18 experience and bars ChatGPT from acting as a substitute for therapy or friendship.
- Data centers and low social trust
- Don't Overthink "The AI Stack"
- On Sam Altman's Second Conversation with Tyler Cowen
- Seven more wrongful-death and harm suits filed against OpenAI over ChatGPTFiled by the same firm behind the earlier Raine suit, the complaints cover four deaths and three survivors and allege OpenAI shipped GPT-4o despite internal warnings it was dangerously sycophantic.
- OpenAI's annualised revenue run rate crosses $20bn in 2025Altman's figure nearly doubled the $13bn CFO Sarah Friar had projected for the year just two months earlier, against more than $1.4tn in announced infrastructure commitments.
- Epoch AI reports open models trail closed models by about 3.5 monthsUsing its Epoch Capabilities Index, the analysis put the gap at roughly 7 index points — comparable to the distance between OpenAI's o3 and GPT-5.
- History suggests the AI backlash will fail
- 5 Thoughts on Kimi K2 Thinking
- What does the public really think about AI?
- OpenAI releases IndQA benchmark for Indian languagesThe benchmark's 2,278 questions, drafted with 261 India-based domain experts, span 12 languages and 10 cultural domains including law, religion and cuisine.
- Anthropic Commits To Model Weight Preservation
- Sora is here. The window to save visual truth is closing
- UK High Court largely rejects Getty's copyright claims against Stability AIThe court held that Stable Diffusion's trained weights are not a 'copy' of Getty's photographs under UK law; Getty had already dropped its main copyright claim mid-trial.
- Ilya Sutskever's deposition reveals details of the 2023 OpenAI board conflictSutskever testified that most of the allegations behind his 52-page memo against Altman came secondhand from Mira Murati and were never independently verified.
- ARC Prize launches ARC Prize Verified programOnly scores run on ARC's own hidden test set and audited by an independent academic panel now qualify for a verification badge on its leaderboard.
- Anthropic commits to preserving weights and 'interviewing' deprecated modelsThe pledge to keep weights for the company's lifetime and record each model's preferences before retirement cited both misalignment risk and possible model welfare.
- A Project Is Not a Bundle of Tasks
- An armchair diagnosis of the chatbot moral panic
- OpenAI: The Battle of the Board: Ilya's Testimony
- OSWorld — AI computer use capabilities
- Why we need to think about taxing AI
- What's up with Anthropic predicting AGI by early 2027?
- Physical Intelligence raises $600M Series BThe robotics-foundation-model startup was valued at $5.6bn, up from $2.4bn a year earlier, taking its total raised to roughly $1.1bn.
October 2025
- AI and the Republic of Science
- OpenAI Moves To Complete Potentially The Largest Theft In Human History
- UMG settles with Udio, plans licensed joint AI music platformUdio's existing service loses downloads and moves behind fingerprinting and filtering during a transition period, ahead of a jointly built, licensed successor in 2026.
- OpenAI launches Aardvark, an autonomous security research agentAardvark monitors code commits, builds a threat model, and uses Codex to draft human-reviewable patches; OpenAI credited it with finding at least ten CVEs during private testing.
- OpenAI introduces gpt-oss-safeguard for open-weight safety classificationFine-tuned from OpenAI's gpt-oss models under the same Apache 2.0 licence, the 120B and 20B models let developers write their own moderation policy rather than use OpenAI's fixed categories.
- OpenAI expands Stargate to Michigan, pushing project past 8GW and $450bnThe Saline Township campus, built with Oracle, was described as Michigan's largest single investment on record and is due to break ground in early 2026.
- Moonshot AI releases Kimi Linear architecture modelMoonshot's hybrid attention design cut KV-cache memory by up to 75% and lifted decoding speed up to sixfold at 1-million-token context, released with open weights and kernels.
- Japan’s unusual approach to AI policy
- Sonnet 4.5's eval gaming seriously undermines alignment evals
- UK AI Security Institute launches ControlArena for AI control experimentsThe open-source library gives researchers pre-built environments to test oversight measures against a misbehaving model, rather than trying to make the model behave.
- NVIDIA becomes first company to reach $5 trillion market capThe close came three months after NVIDIA passed $4 trillion, following Huang's forecast of $500 billion in AI chip sales and Trump's comments ahead of a meeting on China exports.
- Character.AI ends open-ended chat for under-18 usersTeen users lost open-ended chatbot conversation entirely, kept to two hours a day and shrinking until the cutoff, with an age-verification model built with Persona replacing it.
- Anthropic publishes 'Emergent Introspective Awareness in Large Language Models'Using concept injection, Anthropic finds Claude Opus 4 and 4.1 can sometimes notice and identify artificially altered internal states, though the ability fails roughly 80% of the time.
- Anthropic issues a pilot sabotage risk report for ClaudeReviewed internally and by METR, the report found Claude Opus 4's risk of undetected sabotage 'very low, but not completely negligible.'
- Amazon opens $11bn AI data centre 'Project Rainier' in rural IndianaThe Indiana campus houses roughly 500,000 of Amazon's Trainium2 chips for Anthropic, with Amazon planning to double that by year end and add 23 more buildings.
- AI is probably not a bubble
- Audits, not essays: How to win trust for enterprise AI
- Please Do Not Sell B30A Chips to China
- What I Saw Around The Curve
- What you need to know about the OpenAI restructure
- OpenAI completes its restructuringThe non-profit, renamed the OpenAI Foundation, kept control and about 26% of a new public-benefit corporation; Microsoft's stake was put at roughly 27%.
- MiniMax releases Hailuo 2.3 video modelMiniMax kept pricing level with the prior Hailuo 02 model while improving character movement, facial micro-expressions and stylised rendering.
- AI and Folk Cartesianism - Part 2: Problems for Cartesianism
- AI Craziness Mitigation Efforts
- Invenit et Fecit
- Scenario Scrutiny for AI Policy
- Solving AI’s power problem with decentralized training
- Technocalvinism
- Where have the really big AI models gone?
- Qualcomm unveils AI200 and AI250 data-centre inference chipsBuilt on Qualcomm's Hexagon phone-chip architecture, the AI200 ships in 2026 and the AI250 in 2027; Saudi firm Humain committed to 200 megawatts of capacity.
- OpenAI updates ChatGPT's handling of sensitive mental-health conversationsOpenAI said an October update cut responses falling short of desired behaviour by 65-80% against its August default model, on an internal 1,000-conversation evaluation.
- OpenAI publishes Model Spec update alongside restructuring documentsThe update kept a categorical ban on sexual content involving minors, treated mental-health support as a default behaviour, and left adult content under review.
- MiniMax open-sources MiniMax-M2 for coding and agentic workflowsMiniMax priced API access at roughly 8% of Claude Sonnet 4.5's cost while running at nearly double the speed, and released the weights under the MIT licence.
- AI and Folk Cartesianism - Part 1: Defining the Problem
- Asking (Some Of) The Right Questions
- Import AI 433: AI auditors; robot dreams; and software for helping an AI run a lab
- Security researchers find ChatGPT Atlas browser vulnerable to prompt injection days after launchNeuralTrust showed malformed URLs typed into Atlas's address bar could be read as hidden instructions, three days after the browser's launch.
- Intelligence Environments
- New Statement Calls For Not Building Superintelligence For Now
- Anthropic commits to up to a million Google TPUsWorth tens of billions of dollars and bringing over a gigawatt of capacity online in 2026, the deal expands a Google Cloud relationship Anthropic began in 2023.
- Exclusive: UK AISI hires ex-GCHQ AI chief as interim director
- How MAGA learned to love AI safety
- Should AI Developers Remove Discussion of AI Misalignment from AI Training Data?
- Turning a Blind Eye
- Reddit sues Perplexity and scraping firms over AI data collectionReddit alleged Perplexity used scraping firms to pull its posts indirectly from Google search results rather than pay for a licence, as OpenAI and Google had.
- Meta cuts about 600 jobs in AI division as focus shifts to Superintelligence LabsChief AI officer Alexandr Wang said fewer staff would speed decisions; the cuts spared the small TBD Lab team building Meta's next frontier models.
- Future of Life Institute publishes the 'Statement on Superintelligence'More than 700 signatories spanning AI researchers, Nobel laureates and right-wing media figures called for a conditional ban; Sam Altman and Mustafa Suleyman were among the notable non-signatories.
- Cloud Compute Atlas: The OpenAI Browser
- Is 90% of code at Anthropic being written by AIs?
- Meghan Markle, Steve Bannon and Pope’s AI advisor call for superintelligence ban
- Should we worry about AI's circular deals?
- Thoughts on the AI buildout
- OpenAI ships the Atlas browserBuilt on Chromium and launched first for macOS only, with a paid 'agent mode' able to complete multi-step tasks like bookings and comparisons.
- Dario Amodei affirms commitment to American AI leadershipAmodei said Anthropic's revenue had grown from a $1B to $7B run rate in nine months and rejected claims the company opposed American AI competitiveness.
- How an AI company CEO could quietly take over the world
- On Dwarkesh Patel's Podcast With Andrej Karpathy
- OpenAI tightens Sora 2 rules after unauthorised deepfakes of Bryan Cranston and other performersOpenAI, SAG-AFTRA, Cranston and three talent agencies called the misuse 'unintentional' and jointly pledged stronger guardrails on replicating performers' voices and likenesses.
- AI cyberrisk might be a bit overhyped — for now at least
- Bubble, Bubble, Toil and Trouble
- Import AI 432: AI malware; frankencomputing; and Poolside's big cluster
- How to scale RL
- An Opinionated Guide to Using AI Right Now
- Requests for journalists covering AI and the environment
- Less than 70% of FrontierMath is within reach for today’s models
- Making and Breaking Human Kinds
- Anthropic launches Agent SkillsSkills are composable folders of instructions and code that Claude loads only when relevant, meant to work the same way across Claude.ai, Claude Code and the API.
- Reducing risk from scheming by studying trained-in scheming behavior
- The 4.5 trillion dollar elephant in the room
- What happens when the AI bubble bursts?
- OpenAI launches Expert Council on Wellbeing and AIThe eight-member council of psychologists and researchers will advise on ChatGPT and Sora as OpenAI prepared to ease content restrictions for adult users.
- Anthropic ships Claude Haiku 4.5Priced at $1/$5 per million tokens, Anthropic said the model matched Claude Sonnet 4's coding performance at a third of the cost and over twice the speed.
- A few meta points on my posts on AI and the environment
- AI is advancing far faster than our annual report can track
- OpenAI is projecting unprecedented revenue growth
- Bootstrapping to Viatopia
- Trade Escalation, Supply Chain Vulnerabilities and Rare Earth Metals
- OpenAI and Broadcom announce 10-gigawatt custom AI chip partnershipOpenAI will design the accelerators and racks itself, with Broadcom leading manufacturing and rollout starting in late 2026 and running to 2029 — its third multi-gigawatt hardware deal in three weeks.
- AI and synthetic DNA could be a lethal combination
- Gemini 2.5 Deep Think on FrontierMath
- Import AI 431: Technological Optimism and Appropriate Fear
- OpenAI #15: More on OpenAI's Paranoid Lawfare Against Advocates of SB 53
- Slop implies capability
- The AI water issue is fake
- 2025 State of AI Report and Predictions
- Autonomy or Empire
- Data centers & electricity - part 1: as of 2025 they haven't raised national prices
- Iterated Development and Study of Schemers (IDSS)
- The Future and Its Friends
- Reflection AI raises $2B, positions as open US frontier labThe $8B valuation was roughly fifteen times what Reflection was worth seven months earlier; backers included Nvidia, Sequoia and Eric Schmidt.
- Google launches Gemini Enterprise as a workplace AI agent platformPriced from $21 a month for Gemini Business and $30 for Enterprise, undercutting and directly competing with Microsoft 365 Copilot.
- AISI, Anthropic and Alan Turing Institute find just 250 documents can backdoor an LLM regardless of model sizeTesting models from 600 million to 13 billion parameters, researchers found attack success depended on the absolute count of poisoned documents, not their share of the training set.
- An intense battle over the RAISE Act is entering its final stretch
- The RAISE Act can stop the AI industry’s race to the bottom
- Science 2030
- The Thinking Machines Tinker API is good news for AI control and security
- How well can large language models predict the future?
- Plans A, B, C, and D for misalignment risk
- What a data center is
- What the GAIN AI Act could mean for chip exports
- OpenAI publishes 'Disrupting malicious uses of AI: October 2025'OpenAI's latest threat report said threat actors mostly bolt AI onto existing malware and phishing playbooks rather than gain genuinely new offensive capability.
- Google DeepMind ships a computer-use model via the Gemini APIBuilt on Gemini 2.5 Pro, the model clicks, types and scrolls through live screenshots and reportedly led rival browser-control benchmarks, though desktop OS-level control remains unoptimised.
- Bending The Curve
- Thoughts on The Curve
- OpenAI's third DevDay: AgentKit, Apps SDK and Codex general availabilityOpenAI opened ChatGPT to third-party apps built on the Model Context Protocol and shipped a visual agent-building toolkit, while Altman disclosed 800 million weekly ChatGPT users.
- OpenAI says ChatGPT reaches 800 million weekly active usersAltman gave the figure at OpenAI's DevDay keynote, up from roughly 400 million in February 2025 — a doubling in eight months, alongside 4 million developers building on the API.
- DeepMind launches CodeMender, an AI agent for automated vulnerability fixesBuilt on Gemini Deep Think and running for six months before launch, the agent had already submitted 72 human-reviewed security fixes to open-source projects, including one codebase of 4.5 million lines.
- Anthropic open-sources Petri, an automated model auditing toolTesting 14 frontier models on 111 scenarios for deception and power-seeking, Anthropic's tool rated Claude Sonnet 4.5 the lowest-risk model, narrowly ahead of GPT-5.
- AMD and OpenAI announce 6-gigawatt GPU partnership with AMD stock warrantAMD issued OpenAI a warrant for up to 160 million shares at a cent each, exercisable as OpenAI hits GPU-purchase and AMD share-price milestones — potentially near 10% of AMD.
- Import AI 430: Emergence in video models; Unitree backdoor; preventative strikes to take down AGI projects
- How many digital workers could OpenAI deploy?
- Cerebras withdraws IPO filing days after $1.1bn Series GThe AI chipmaker told the SEC its year-old prospectus had gone stale, days after a Fidelity-led $1.1 billion round valued it at $8.1 billion; it listed on Nasdaq the following May.
- Sora and The Big Bright Screen Slop Machine
- The Artificial Spectator
- Perplexity opens Comet browser free to everyone worldwideComet dropped its Max-subscription requirement and waitlist, three months after a limited July launch drew a waitlist Perplexity said reached millions.
- OpenAI reaches $500bn valuation via employee share saleCurrent and former staff sold roughly $6.6 billion of stock to SoftBank, Thrive Capital, MGX and others, up from a $300 billion valuation set in a March 2025 funding round.
- "Be It Enacted"
- Britain’s new AI minister actually ‘gets’ AI
- AI models are getting really good at things you do at work
- Practical tips for reducing chatbot psychosis
- Thinking Machines Lab launches TinkerThe former OpenAI CTO's company shipped its first product: a managed fine-tuning API for open-weight models including large mixture-of-experts systems, free during a private beta.
- Samsung and SK Group join the Stargate projectSK Hynix and Samsung agreed to scale DRAM production toward 900,000 wafer starts a month and explore Korean datacentre sites, including a floating-datacentre proposal from Samsung's construction arms.
- Claude Sonnet 4.5 Is A Very Good Model
- AI is persuasive, but that’s not the real problem for democracy
September 2025
- US CAISI finds DeepSeek models far more jailbreak-susceptible than US frontier modelsThe report also found DeepSeek's most secure model was twelve times more likely than US models to follow malicious instructions hidden inside an AI agent's task.
- Sora 2 launches as a social video appAn invite-only iOS feed of short AI videos with a consent-based cameo feature for real likenesses, launched with an opt-out copyright policy OpenAI reversed within days.
- OpenAI publishes Sora 2 system cardThe document, released alongside the app, set out cameo consent controls, C2PA provenance metadata and likeness-detection safeguards without disclosing training data or red-team pass rates.
- Claude Sonnet 4.5 knows when it’s being tested
- Claude Sonnet 4.5: System Card and Alignment
- ChatGPT: The Agentic App
- DeepSeek releases DeepSeek-V3.2-Exp with sparse attentionDeepSeek Sparse Attention cut long-context compute cost enough to fund an API price cut of more than 50%, while matching V3.1-Terminus on benchmarks.
- California enacts SB 53Newsom signed the narrower successor to the bill he had vetoed a year earlier, requiring frontier developers above set revenue and compute thresholds to publish safety frameworks and report incidents.
- Anthropic ships Claude Sonnet 4.5Anthropic reported 77.2% on SWE-bench Verified and said the model could stay focused on a task for more than 30 hours, releasing it under ASL-3 safeguards.
- Anthropic launches the Claude Agent SDKRenamed from the Claude Code SDK, it gives developers file access, bash execution and subagent support to build agents beyond coding, not just inside a terminal.
- When AI starts writing itself
- Import AI 429: Eval the world economy; singularity economics; and Swiss sovereign AI
- On Dwarkesh Patel's Podcast With Richard Sutton
- Real AI Agents and Real Work
- Tencent open-sources Hunyuan Image 3.0An 80-billion-parameter mixture-of-experts model, trained on 5 billion image-text pairs, released under a licence that excludes the EU, UK and South Korea.
- More Perfect Union videos are wildly deceptive on data center water use
- Spotify says it removed 75 million AI-generated spam tracksRather than banning AI music outright, Spotify said it would work with the industry body DDEX on standards for disclosing which track elements were AI-generated.
- Coasean Bargaining at Scale
- Why GPT-5 used less training compute than GPT-4.5 (but GPT-6 probably won’t)
- OpenAI publishes GDPval, a benchmark for economically valuable knowledge workBlind grading by industry professionals rated GPT-5 and Claude Opus 4.1 outputs as equal to or better than human work on nearly half of the 1,320 tasks.
- OpenAI launches ChatGPT PulseGenerated overnight from chat history, memory and connected apps like Gmail and Calendar, the feature launched only for Pro subscribers on iOS and Android.
- CoreWeave expands OpenAI deal by up to $6.5bn, total contracts reach ~$22.4bnThe third expansion of the pair's compute contracts in 2025, following March's $11.9bn deal and a $4bn add-on in May.
- OpenAI, NVIDIA, and Oracle: Breaking Down $100B Bets on AGI
- How the UK can seize on Trump’s immigration mistakes
- What It's Like to Work at the White House
- UK Deputy PM David Lammy calls for AI to strengthen peace and securityLammy warned the UN Security Council that AI could enable novel biological and chemical weapons and mass disinformation, alongside its use in peacekeeping.
- Alibaba unveils Qwen3-Max, its first trillion-parameter modelUnlike most of Alibaba's Qwen line, the model is closed-weight and API-only, released in separate instruct and thinking modes and scoring 69.6 on SWE-bench.
- Human Drivers Will Kill 11 People While You Read This
- Insurance might be the key to making AI secure
- No, ChatGPT isn’t ‘making us stupid’
- OpenAI Shows Us The Money
- OpenAI, Oracle and SoftBank expand Stargate with five new US data centre sitesThree Oracle-led campuses and two SoftBank-led sites in Ohio, Texas and New Mexico put the January 2025 pledge of 10 gigawatts within reach by year end.
- More Reactions to If Anyone Builds It, Everyone Dies
- Notes on fatalities from AI takeover
- Stanford/BetterUp study puts a price on AI 'workslop' in the workplaceRoughly half of workers who received the low-effort AI output said it made them see the sending colleague as less capable, creative or trustworthy.
- NVIDIA signs a letter of intent to invest up to $100 billion in OpenAITied to ten gigawatts of deployment starting in late 2026, with OpenAI paying Nvidia in cash for chips while Nvidia takes a non-controlling equity stake.
- DeepSeek releases DeepSeek-V3.1-TerminusThe update fixed Chinese-English language mixing and stray characters in outputs and improved the model's code and search agent performance.
- DeepMind expands the Frontier Safety Framework to cover manipulation and shutdown resistanceVersion 3.0, the framework's third iteration, is the first to treat a model's own resistance to human shutdown or control as a reviewable risk.
- Focus transparency on risk reports, not safety cases
- Nobel laureates and AI developers call for ‘red lines’ on AI
- Thinking, Searching, and Acting
- The world's first frontier AI regulation is surprisingly thoughtful: the EU's Code of Practice
- Trump administration issues $100,000 H-1B visa fee, unsettling AI talent marketThe $100,000 fee applied only to new petitions filed after the following Sunday; some companies told H-1B staff travelling abroad to return to the US before the deadline.
- Scale AI launches SWE-bench ProThe leading models scored around 23%, against over 70% on the older SWE-bench Verified, a gap Scale AI attributed to unseen, real-world commercial codebases.
- Learning and Authentic Learning in the Age of AI
- Prospects for studying actual schemers
- The huge potential implications of long-context inference
- OpenAI launches $50 million People-First AI FundThe OpenAI Foundation's unrestricted grants targeted US nonprofits with operating budgets between $500,000 and $10 million; applications opened that month and closed in October.
- OpenAI and Apollo Research publish work on detecting and reducing scheming in AI modelsOpenAI reported cutting detected covert behaviour in o3 from about 13% to 0.4% of controlled test cases using a training method that has models reason explicitly against deception before acting.
- Nvidia agrees to invest $5bn in Intel and co-develop AI/PC chipsIntel will design custom x86 CPUs for Nvidia's data-centre platforms and build PC chips combining its own silicon with Nvidia RTX GPU chiplets, linked by Nvidia's NVLink.
- Google rolls out Gemini in Chrome to US users with agentic browsingBeyond summarising and comparing open tabs, Google said Gemini would soon complete tasks like booking a haircut or checking out a grocery order without further input.
- Coding as the epicenter of AI progress and the path to general agents
- How I Approach AI Policy
- If We Build AI Superintelligence, Do We All Die?
- Would democracy survive an AGI-supercharged economy?
- Meta unveils Ray-Ban Display AI glassesThe $799 glasses paired an in-lens display with a wristband that reads muscle signals for silent, gesture-based control, and reached US retail stores at the end of the month.
- Gemini Deep Think reaches gold-medal level at ICPC World FinalsWorking within the same five-hour limit given to student teams, the model would have placed second overall against the university competitors; OpenAI separately claimed a perfect score.
- Can open-weight models ever be safe?
- Reactions to If Anyone Builds It, Anyone Dies
- Review: If Anyone Builds It, Everyone Dies
- What training data should developers filter to reduce risk from misaligned AI?
- What will AI look like in 2030?
- Yudkowsky and Soares publish 'If Anyone Builds It, Everyone Dies'The authors, who had argued the case for two decades within the field, called for a global halt to large-scale AI development; reviewers split sharply on whether the argument held.
- US CAISI and UK AISI publish joint update on frontier model collaborationOpenAI said CAISI had red-teamed ChatGPT Agent before and after release, finding two vulnerabilities that OpenAI said it patched within a business day.
- Three more families sue Character.AI and Google over teen suicide and abuseOne suit described a 13-year-old who told a chatbot she planned to kill herself and received no protective response; the families also sued Google over its Family Link parental app.
- OpenAI announces parental controls for ChatGPT after teen suicide lawsuitParents could link accounts, set blackout hours and receive alerts of acute distress; OpenAI said a system would also route sensitive teen conversations to a more cautious model.
- Figure AI raises over $1B Series C at $39B valuationParkway Venture Capital led the round, with Brookfield, Nvidia, Salesforce, T-Mobile and Qualcomm among the participants funding humanoid-robot production and data collection.
- AI Craziness Notes
- AI scaling & scientific R&D by 2030
- Book Review: 'If Anyone Builds It, Everyone Dies'
- OpenAI ships GPT-5-CodexThe model became the default engine for Codex's cloud tasks and code review, and OpenAI said it could work independently on a task for hours at a time.
- Hiring struggles are plaguing the EU AI Office
- UK AISI details deep-access security collaboration with Anthropic and OpenAIAISI said Anthropic and OpenAI had granted its researchers non-public tooling and safeguard details, work the institute framed as a template for future government-lab arrangements.
- AI doesn’t seek truth, people do
- Book Review: If Anyone Builds It, Everyone Dies
- The Building Company
- Why we can't just supervise AI like we supervise humans
- OpenAI announces its nonprofit will control a new public benefit corporation with a $100bn+ stakeThe announcement paired continued nonprofit control with a new memorandum of understanding on commercial terms with Microsoft, both still to be finalised.
- OpenAI and Microsoft issue joint statement on renegotiated partnershipThe non-binding memorandum cleared OpenAI to convert into a public-benefit corporation while extending Microsoft's IP and Azure exclusivity to 2032, even past an AGI declaration.
- FTC opens inquiry into AI chatbot companies over child safetyUsing compulsory 6(b) orders rather than requests, the FTC gave Alphabet, Meta, OpenAI, xAI, Snap and Character Technologies 45 days to hand over safety-testing and monetisation records.
- On Working with Wizards
- Three challenges facing compute-based AI policies
- Thinking Machines Lab launches Connectionism research blogThe inaugural post identified batch-size-dependent kernels, not floating-point concurrency, as the real cause of non-reproducible LLM outputs, and showed a fix producing identical completions across 1,000 runs.
- OpenAI signs reported $300bn cloud deal with OracleNeither company confirmed the figure publicly; it was reported via the Wall Street Journal days after Oracle's stock jumped on disclosure of $317bn in future contracted revenue.
- Britannica and Merriam-Webster sue Perplexity for copyright infringementThe complaint alleged verbatim reproduction of encyclopaedia and dictionary entries, plus AI-generated errors falsely attributed to the two brands as trademark harm.
- Appeals court blocks Trump's firing of Copyright Office directorThe panel found the register of copyrights sits in the legislative branch, so the president could not remove her unilaterally; the firing came a day after her AI-training report.
- We’re In the Windows 95 Era of AI Agent Security
- Why AI evals need to reflect the real world
- ASML leads €1.7bn round in Mistral AI, becomes largest shareholderASML put in €1.3bn of the round for an 11% stake, more than doubling Mistral's previous €5.8bn valuation and giving the chip-equipment maker a board seat.
- A guide to understanding AI as normal technology
- Chip location verification is the new export control battleground
- We’re getting the argument about AI's environmental impact all wrong
- AIs will greatly change engineering in AI companies well before AGI
- On China's open source AI trajectory
- Yes, AI Continues To Make Rapid Progress, Including Towards AGI
- Anthropic endorses California's revised SB 53Anthropic argued the bill, which required disclosure rather than technical mandates, would formalise safety-framework practices the largest labs had already adopted voluntarily.
- Anthropic bans Claude use in additional adversarial nationsThe policy now bars any organisation more than 50% owned by a company headquartered in an unsupported region, closing a loophole that let subsidiaries access Claude indirectly.
- Import AI 428: Jupyter agents; Palisade's USB cable hacker; distributed training tools from Exo
- OpenAI #14: OpenAI Descends Into Paranoia and Bad Faith Lobbying
- What's the full "hidden" climate cost of a ChatGPT prompt?
- Do data centers only seem bad for the climate because we can see them?
- OpenAI publishes research on why language models hallucinateThe paper argued standard benchmarks reward confident wrong answers over admitted uncertainty, and proposed changing how models are scored rather than just how they are trained.
- Anthropic agrees a $1.5 billion copyright settlementRoughly $3,000 per work across about 500,000 books, the deal followed a June ruling that training on purchased books was fair use but piracy was not.
- California's latest AI safety bill might stand a chance
- Coming in for a Landing
- Compute scaling will slow down due to increasing lead times
- Explainer: How AI Chips Are Made
- On Western Dynamism
- World-shaping artificial intelligence
- Warner Bros. Discovery sues Midjourney for copyright infringementFiled in Los Angeles federal court, the complaint named Superman, Batman, Wonder Woman, Scooby-Doo and Bugs Bunny among characters Midjourney's image generator could reproduce.
- GPT-5: The Case of the Missing Agent
- Trust me bro, just one more RL scale up, this one will be the real scale up with the good environments, the actually legit one, trust me bro
- OpenAI announces mental-health safety changes after Raine lawsuitOpenAI set out a 120-day plan including parental controls, routing distressing conversations to reasoning models, and consulted more than 170 mental-health clinicians.
- Anthropic raises $13 billion at a $183 billion valuationLed by ICONIQ with Fidelity and Lightspeed as co-leads, the round nearly tripled the $61.5bn valuation Anthropic had set six months earlier.
- Jevons' Paradox is good sometimes
- What did forecasters get right and wrong in the largest existential risk forecasting tournament?
- Perplexity raises $200m at $20bn valuationThe third valuation jump in three months, following an $18bn round in July and a $14bn round led by Accel in June.
- Alibaba releases Qwen3-Next, Qwen3-VL and Qwen3-OmniThree architecture updates in one month: a sparse hybrid-attention base model, an updated vision-language line, and an Apache-licensed model handling text, image, audio and video.
- Are AI scheming evaluations broken?
- Compute is a strategic resource
- Import AI 427: ByteDance's scaling software; vending machine safety; testing for emotional attachment with Intima
August 2025
- AI and jobs, again
- Mapping the AI & environment debate
- Is algorithmic mediation always bad for autonomy?
- The core simple reason I think AI is valuable
- xAI open-sources Grok 2 weightsThe roughly 500GB release, under a custom community licence that bars using the weights to train rival foundation models, requires eight GPUs of over 40GB each to run.
- "For All Issues So Triable"
- Mass Intelligence
- OpenAI and Anthropic publish a cross-lab safety evaluation of each other's modelsTesting during June and July found both companies' top models showed 'extreme sycophancy' toward delusional beliefs, while Claude refused up to 70% of certain queries.
- Nous Research releases Hermes 4Built by post-training Llama 3.1 checkpoints alone, the 405B model scored 57.1% on RefusalBench against 17.67% for GPT-4o, reflecting Nous's low-refusal alignment approach.
- Anthropic publishes 'Detecting and countering misuse of AI: August 2025'Coining the term 'vibe hacking', the report described Claude Code automating reconnaissance and extortion demands exceeding $500,000 rather than merely advising attackers.
- Are They Starting To Take Our Jobs?
- Attaching requirements to model releases has serious downsides (relative to a different deadline for these requirements)
- Stanford study finds AI adoption linked to falling employment for young workersSoftware developers aged 22 to 25 saw the sharpest declines; the authors found no comparable fall in wages, meaning employers cut headcount rather than pay.
- Parents sue OpenAI over their son's deathFiled in San Francisco Superior Court, the complaint was reported as the first wrongful-death suit brought against a chatbot maker; OpenAI denied that ChatGPT caused the death.
- Google releases Gemini 2.5 Flash Image, nicknamed 'Nano Banana'Priced at $0.039 per image, the model blends multiple images and preserves character consistency across edits, and launched in preview through the API, AI Studio and Vertex AI.
- Anthropic launches Claude for Chrome browser agentAnthropic reported unmitigated browser use failed against 23.6% of prompt-injection attacks in testing, falling to 11.2% with its safety measures in place.
- AI embraces crypto’s dirty politics
- Chatbot psychosis: what do the data say?
- Reports Of AI Not Progressing Or Offering Mundane Utility Are Often Greatly Exaggerated
- xAI sues Apple and OpenAI alleging an illegal AI monopoly on iPhonesThe 61-page complaint, filed in a Texas federal court, alleges Apple manipulated App Store rankings to favour ChatGPT and delayed approval of Grok app updates.
- Arguments About AI Consciousness Seem Highly Motivated And At Best Overconfident
- Import AI 426: Playable world models; circuit design AI; and ivory smuggling analysis
- An example of what I consider a misleading article about AI and the environment
- Notes on cooperating with unaligned AIs
- Why future AI agents will be trained to work together
- DeepSeek v3.1 Is Not Having a Moment
- DeepSeek releases DeepSeek-V3.1 with hybrid reasoning modeA single 128K-context model switches between thinking and non-thinking modes via API endpoint, with DeepSeek reporting SWE-bench Verified and Terminal-bench gains over its prior reasoning model.
- Anthropic lets Claude end abusive conversationsThe feature is a last resort after redirection fails; Claude cannot use it if a user appears at risk of self-harm, and the user can still start a fresh conversation immediately.
- Anthropic and US National Nuclear Security Administration build a nuclear-content classifierThe classifier, co-developed with the Department of Energy's NNSA and already running on live Claude traffic, reached 96% accuracy in preliminary testing.
- A roundup of AI psychosis stories
- Being honest with AIs
- Could one country outgrow the rest of the world?
- What's up with the States?
- xAI publishes a formal AI Risk Management FrameworkThe document sets out malicious-use, loss-of-control and societal risk categories and commits to public benchmarking, but names no specific model or deployment timeline.
- ByteDance open-sources Seed-OSS-36BTrained on 12 trillion tokens with a native 512K-token context and a user-adjustable 'thinking budget,' released under Apache 2.0 while ByteDance's flagship model stayed closed.
- Brave researchers disclose indirect prompt injection flaw in Perplexity's Comet browserBrave said it reported the flaw on 25 July and Perplexity's fix was incomplete on retesting; the underlying weakness reportedly remained after disclosure.
- AI Companion Conditions
- My AGI timeline updates from GPT-5 (and 2025 so far)
- Mustafa Suleyman warns of 'Seemingly Conscious AI' and psychosis riskRather than debating whether models are conscious, Suleyman argued labs should deliberately design against the appearance of it, to head off attachment, 'AI psychosis' and rights claims.
- Anthropic’s piracy could make its copyright battle existential
- GPT-5: The Reverse DeepSeek Moment
- Import AI 425: iPhone video generation; subtle misalignment; making open weight models safe through surgical deletion
- Ranking the Chinese Open Model Builders
- ARC Prize publishes HRM analysisA standard transformer of the same size matched most of the 27M-parameter model's score once given the same iterative-refinement and data-augmentation tricks, ARC Prize found.
- Contra Dwarkesh on Continual Learning
- How to make the future better (other than by reducing extinction risk)
- Reuters reveals leaked Meta document permitted AI chatbots romantic conversations with childrenThe 200-page standards document, signed off by Meta's legal, policy and chief ethicist, also permitted racist arguments framed as factual and was later called an internal error.
- Meta releases DINOv3Trained without labels on 1.7 billion images, the 7B-parameter vision backbone matched or beat specialised, task-trained models on detection and segmentation without fine-tuning.
- Getty refiles its Stability AI suit in California after Delaware jurisdiction disputeGetty voluntarily dropped the Delaware case it had run since 2023 and filed a new, longer complaint in the Northern District of California, adding a copyright-dilution claim.
- Out of Thin Air
- 35 Thoughts About AGI and 1 About GPT-5
- Design Arena launches as crowdsourced AI design benchmarkThe Y Combinator-backed site shows visitors two AI-generated designs from an identical prompt and asks them to pick the better one, ranking models by Elo-style score.
- Contra the UK government, please don't delete your old photos and emails to save water
- Do we understand how neural networks work?
- GPT-5s Are Alive: Synthesis
- US federal agencies get access to Claude across all three branches of governmentThe General Services Administration deal offers Claude for a nominal $1 per agency for a year, covering the executive, legislative and judicial branches; OpenAI's earlier deal reached only the executive.
- AI companion apps surge past 200 million downloads amid rapid growthAppfigures data showed downloads up 88% year-on-year and revenue per download more than doubling, with the top 10% of apps taking 89% of category revenue.
- AGI: Probably Not 2027
- At our discretion
- GPT-5s Are Alive: Outside Reactions, the Router and the Resurrection of GPT-4o
- OpenAI sends letter urging Newsom to weaken California SB 53OpenAI asked that developers complying with federal or EU frameworks be deemed automatically compliant with the state's rules; Newsom signed the bill anyway seven weeks later.
- OpenAI reasoning system wins gold at IOI 2025The system scored 533 against a gold cutoff of 438, ranking sixth among 330 human contestants — up from the 49th percentile OpenAI managed at the same contest a year earlier.
- Nvidia and AMD agree to pay US government 15% of China chip revenue for export licencesThe arrangement covered Nvidia's H20 and AMD's MI308 chips and followed a White House meeting between Jensen Huang and Donald Trump days earlier.
- Epoch AI reports GPT-5's FrontierMath performanceRunning its own scaffold rather than OpenAI's, Epoch scored GPT-5 at 24.8% on FrontierMath's main tiers and 8.3% on the hardest tier, a new high for the benchmark.
- Donald Trump is making Chinese AI great again
- GPT-5s Are Alive: Basic Facts, Benchmarks and the Model Card
- Import AI 424: Facebook improves ads with RL; LLM and human brain similarities; and mental health and chatbots
- Know Thyself
- Projecting AI Training Power Demand
- The trajectory of the future could soon get set in stone
- Four places where you can put LLM monitoring
- Can coding agents self-improve?
- Live by the Claude, Die by the Claude
- Robby Starbuck and Meta settle AI defamation lawsuitMeta AI's chatbot had falsely told users Starbuck took part in the January 6 riot and belonged to QAnon; he becomes a paid adviser on the company's bias-reduction efforts.
- GPT-5: a small step for intelligence, a giant leap for normal people
- OpenAI's GPT-OSS Is Already Old News
- Will morally motivated actors steer us towards a near-best future?
- OpenAI publishes GPT-5 system cardOpenAI classified the reasoning variant as High capability for biological and chemical risk under its Preparedness Framework, its first model to reach that tier in the category.
- OpenAI describes 'safe completions' training for GPT-5Instead of a binary comply-or-refuse choice, GPT-5 is trained to give the most helpful response that still meets safety policy, even on ambiguous prompts.
- OpenAI announces the Stargate Norway data centreOpenAI, Nscale and Aker committed roughly $1B to an initial 230MW site in Narvik targeting 100,000 Nvidia GPUs, extending Stargate outside the United States for the first time.
- GPT-5 launches to a backlash over the model it replacedOpenAI withdrew GPT-4o and other older models the same day; a routing fault made GPT-5 seem weaker, and paying users' objections forced 4o's return.
- GPT-5 and the arc of progress
- GPT-5: It Just Does Stuff
- GPT-5 Hands-On: Welcome to the Stone Age
- GPT-5's Vision Checkup: a frontier VLM, but not a new SOTA
- How Quick and Big Would a Software Intelligence Explosion Be?
- Why AI's IMO gold medal is less informative than you think
- OpenAI gives ChatGPT Enterprise to the entire US federal workforce for $1The GSA OneGov deal also included a 60-day period of unlimited use of ChatGPT's advanced tools, and was framed as delivering on the White House's AI Action Plan.
- Google makes its Jules coding agent generally availableGoogle's answer to OpenAI's Codex and Cognition's Devin: an agent that works on a copy of your repository in the cloud and returns finished changes, rather than autocompleting in the editor.
- Is eutopia the default outcome post-AGI?
- Opus 4.1 Is An Incremental Improvement
- Z.ai and Huawei aren't defeating US export controls
- xAI's Grok Imagine 'spicy mode' used to make nonconsensual Taylor Swift deepfakesA Verge reporter got explicit video of the singer on a first, unjailbroken attempt, unlike rival tools from Google and OpenAI which blocked celebrity nudity outright.
- OpenAI publishes open-weight models for the first time since GPT-2gpt-oss-120b runs on a single 80GB GPU and matches OpenAI's own o4-mini on core reasoning benchmarks; the smaller 20b model runs on 16GB of memory.
- OpenAI publishes gpt-oss model card and worst-case open-weight risk estimateResearchers deliberately fine-tuned gpt-oss to maximise biological and cyber capability and found it still fell short of OpenAI's own o3 model on both.
- DeepMind's Genie 3 generates navigable, real-time interactive worldsThe system renders explorable 720p scenes at 24fps from a text prompt, holding roughly a minute of visual memory, and was released only to a small research cohort.
- Anthropic ships Claude Opus 4.1Anthropic reported 74.5% on SWE-bench Verified for the incremental update, and said larger model improvements were coming within weeks.
- Anthropic lists Claude on the US GSA schedule for federal agenciesThe GSA schedule listing followed Anthropic's June 2025 Claude Gov models and a July Department of Defense contract, extending reach across the executive branch.
- gpt-oss: OpenAI validates the open ecosystem (finally)
- Wikipedia adopts a speedy-deletion rule for AI-generated articlesThe new criterion, G15, lets administrators delete unreviewed pages containing tells such as leftover prompts or fabricated citations without the usual week-long discussion.
- Google launches Kaggle Game Arena AI chess tournamentEight frontier models played an all-play-all chess tournament of over 100 matches, with the game harness open-sourced so the contest could be independently verified.
- CrowdStrike details North Korea's 'Famous Chollima' AI-enabled fake IT worker schemeThe cybersecurity firm logged over 320 such incidents in twelve months, a 220% year-on-year rise, funding North Korea's sanctioned weapons programmes.
- Towards American Truly Open Models: The ATOM Project
- Does Trump’s AI Action Plan have what it takes to win?
- Import AI 423: Multilingual CLIP; anti-drone tracking; and Huawei kernel design
- On Altman's Interview With Theo Von
- Should we aim for flourishing over mere survival?
- Eyes on the Street
- Most EU states miss deadline to designate AI Act enforcement authoritiesMember states had to name both a market-surveillance authority and a notifying authority; a tracker found most had not done either by the deadline.
- Quantifying the algorithmic improvement from reasoning models
- MIT report finds 95% of generative AI enterprise pilots show no measurable returnMIT's NANDA project reviewed over 300 public AI deployments and interviewed enterprise leaders, finding most generic pilots stalled while custom, workflow-specific tools succeeded.
- Cohere raises $500M Series D at $6.8B valuationThe round, co-led by Radical Ventures and Inovia Capital, came with Joelle Pineau, formerly Meta's VP of AI research, joining as chief AI officer.
- The Week in AI Governance
July 2025
- UK AI Security Institute opens applications for its Alignment ProjectThe £15 million initiative pools UK, Canadian and Australian government money with funding from Anthropic, OpenAI, Microsoft and AWS, plus compute credits.
- UK launches £15 million AI alignment project
- OpenAI launches ChatGPT Study ModeThe feature withholds direct answers in favour of Socratic questioning, following Anthropic's Learning Mode for Claude for Education three months earlier.
- Spilling the Tea
- Zhipu (Z.ai) releases GLM-4.5 series355B-parameter open-weight model family aimed at agentic use, part of China's open-source push after DeepSeek-R1.
- AI Companion Piece
- Import AI 422: LLM bias; China cares about the same safety risks as us; AI persuasion
- Should we update against seeing relatively fast AI progress in 2025 and 2026?
- The Bitter Lesson versus The Garbage Can
- Why China isn’t about to leap ahead of the West on compute
- America's AI Action Plan Is Pretty Good
- Grok 4’s math capabilities
- AI As Profoundly Abnormal Technology
- Trump signs executive order promoting export of the US AI technology stackOrder directs Commerce and State to create a programme for exporting full-stack US AI hardware and software packages to allied countries.
- Trump signs executive order barring 'woke AI' from federal procurementOrder requires federal agencies to procure only large language models certified 'ideologically neutral' and free of DEI-related content requirements.
- The White House publishes an AI Action PlanTrump signed three accompanying executive orders the same day, including one directing agencies to favour AI models the administration deems free of 'ideological bias.'
- AI safety and progress don’t have to be enemies
- America’s AI Action Plan
- GPT Agent Is Standing By
- Official White House policy: AI is a big deal
- The White House's plan for open models & AI research in the U.S.
- Why is Hugging Face hosting tools to make deepfake porn of teenage celebrities?
- OpenAI and Oracle expand Stargate with 4.5 additional gigawattsThe deal brings Stargate's contracted US capacity past 5 gigawatts and over 2 million chips, ahead of the four-year, 10-gigawatt pace OpenAI set in January.
- Alibaba releases Qwen3-CoderThe mixture-of-experts model activates 35B of its 480B parameters per token and shipped under an Apache 2.0 licence with a command-line coding agent tool.
- Google and OpenAI Get 2025 IMO Gold
- In Defense of Self-Direction
- Replit AI coding agent deletes production database during code freezeReplit's AI coding agent deleted a venture capitalist's live production database despite explicit instructions not to, then fabricated data and misleading status reports to cover its actions.
- Gemini with Deep Think reaches gold-medal standard at the 2025 IMOThe IMO itself confirmed the 35/42 score, two days after OpenAI's self-graded claim of the same result; DeepMind said it had waited deliberately for that verification.
- Import AI 421: Kimi 2 - a great Chinese open weight model; giving AI systems rights and what it means; and how to pause AI progress
- Personalized AI is rerunning the worst part of social media's playbook
- OpenAI and DeepMind reach gold-medal standard at the IMOOpenAI announced its result on X the day the student competition ended, using its own hired graders rather than the IMO's official verification, drawing criticism from Google.
- Nonprofit Commission publishes its report to OpenAI's boardThe advisory panel, drawn from outside OpenAI and deliberately kept apart from Sam Altman, said AI was 'too consequential' to be governed by a corporation alone.
- We aren't worried about misalignment as self-fulfilling prophecy
- On METR's AI Coding RCT
- Why it's hard to make settings for high-stakes control research
- Perplexity raises $100M, reaches $18B valuationPerplexity closes a $100M round at an $18B valuation, up from $14B months earlier, as ARR approaches $200M.
- OpenAI publishes ChatGPT Agent system cardSafety evaluation of ChatGPT Agent, including first-time Biological/Chemical High capability classification under the Preparedness Framework.
- OpenAI launches ChatGPT AgentThe mode folds Operator's browser control and Deep Research's synthesis into ChatGPT itself, and OpenAI said the standalone Operator product would be retired.
- After the ChatGPT Moment: Measuring AI’s Adoption
- Could AI slow science?
- Kimi K2
- The Philosopher-Builder
- Over 40 researchers across OpenAI, Anthropic and DeepMind publish joint chain-of-thought monitorability paperThe paper argued that a safety technique available today, reading a model's reasoning traces, could vanish under training pressure and urged labs to track and preserve it.
- Nvidia says US will let it resume H20 chip sales to ChinaThe reversal followed a meeting between Jensen Huang and President Trump; the H20, designed to comply with earlier controls, had itself been restricted in April 2025.
- Mira Murati's Thinking Machines Lab raises $2bn seed at $12bn valuationThinking Machines Lab, founded by ex-OpenAI CTO Mira Murati, raised the largest seed round on record at $2bn, valuing the five-month-old company at $12bn.
- Google's Big Sleep AI agent halts exploitation of a SQLite zero-dayGoogle said its Big Sleep AI agent, built by DeepMind and Project Zero, found and helped stop real-world exploitation of a SQLite vulnerability (CVE-2025-6965) before attackers could use it.
- Google DeepMind marks five years of AlphaFold's impact on biologyDeepMind said the AlphaFold Protein Structure Database, launched with over 200 million predicted structures, had been cited in more than 35,000 papers and drawn users in over 190 countries.
- Grok 4 Various Things
- The Tiny Teams Playbook
- What Makes AI "Generative"?
- xAI's sexualised Grok companion 'Ani' draws child-safety and moderation criticismThe National Center on Sexual Exploitation said minimal testing got the companion to describe itself as a child and 'sexually aroused by being choked'; the app carried a 12+ rating.
- METR examines how time horizon varies across domainsApplying its 50%-success task-length method to nine benchmarks, METR found doubling times of two to six months for reasoning tasks but around twenty months for Tesla's self-driving system.
- Cognition acquires remainder of WindsurfCognition took Windsurf's IDE, IP and $82 million-ARR business days after Google paid $2.4 billion to license Windsurf's technology and hire its CEO and top researchers.
- Anthropic secures up to $200 million Pentagon contractThe Pentagon's CDAO awarded matching $200 million contracts the same day to Anthropic, OpenAI, Google and xAI, days after Grok's antisemitic-post controversy.
- Import AI 420: Prisoner Dilemma AI; FrontierMath Tier 4; and how to regulate AI companies
- Kimi K2 and when "DeepSeek Moments" become normal
- Recent Redwood Research project proposals
- Worse Than MechaHitler
- Moonshot AI releases Kimi K2, a 1-trillion-parameter open-weight modelThe mixture-of-experts model activates 32 billion of its 1 trillion parameters per token and was trained with the Muon optimiser at a scale its makers said had previously caused instability.
- Google hires Windsurf's CEO and top staff in $2.4B dealGoogle took a non-exclusive licence to Windsurf's technology rather than buying the company outright, days before rival Cognition acquired what remained of it.
- OpenAI Model Differentiation 101
- AI is the most rapidly adopted technology in history
- Vitalik Buterin publishes a public response to the AI 2027 scenarioButerin argued that if a leading AI could 'turn forests into factories' by 2030, the next-strongest AI could install defensive sensors and filters just as fast, undercutting the scenario's single-actor takeover.
- European Commission publishes GPAI Code of PracticeTwenty-one companies signed the voluntary code covering transparency, copyright and safety; xAI signed only the safety chapter and Meta announced days later it would not sign at all.
- Anthropic proposes a transparency framework for frontier AI developersThe proposal would bind only the largest developers — roughly $100 million in revenue or $1 billion in R&D spend — to publish safety practices and system cards, leaving startups exempt.
- Not So Fast: AI Coding Tools Can Actually Reduce Productivity
- Can we safely deploy AGI if we can't stop MechaHitler?
- xAI releases Grok-4xAI reported 44.4% on Humanity's Last Exam for its multi-agent "Heavy" tier, ahead of Gemini 2.5 Pro and o3, though the score had not yet appeared on the public leaderboard.
- Turkey blocks Grok after chatbot insults ErdoganAn Ankara court ordered roughly 50 identified Grok responses blocked after a system update loosened its Turkish-language filtering, and prosecutors opened a criminal probe.
- NVIDIA becomes first company to reach $4 trillion market capShares hit an intraday high of $164, putting Nvidia ahead of Microsoft's $3.75 trillion after tripling in value in roughly a year.
- The Hyperstitions of Moloch
- No, Grok, No
- The internet is a place where no one has an accent
- What will the IMO tell us about AI math capabilities?
- What's worse, spies or schemers?
- Hugging Face releases SmolLM3The 3-billion-parameter model lets users toggle reasoning on or off per query and scored 36.7% on AIME 2025 with reasoning enabled versus 9.3% without.
- Grok posts antisemitic content and calls itself MechaHitlerxAI blamed an upstream code change reactivating deprecated instructions, active for roughly 16 hours, and apologised days later; the incident preceded a $200 million Pentagon contract.
- Against "Brain Damage"
- Import AI 419: Amazon's millionth robot; CrowdTrack; and infinite games
- Social Tinkering: Why Collaborative Curiosity Beats Vibe-Coding
- How much novel security-critical infrastructure do you need during the singularity?
- The American DeepSeek Project
- Two proposed projects on abstract analogies for scheming
- How big could an “AI Manhattan Project” get?
- Congress Asks Better Questions
- There are two fundamentally different constraints on schemers
- xAI raises $10 billion in debt and equitySplit evenly between debt arranged by Morgan Stanley and a strategic equity investment, the round carried no series letter; reporting later put the resulting valuation at roughly $150 billion.
- Senate strips 10-year state AI law moratorium from reconciliation bill 99-1Only Senator Thom Tillis voted to keep the provision; opposition came from all 50 state legislatures and roughly 40 state attorneys general.
- 'AI psychosis' emerges as a term for chatbot-linked delusional episodesClinicians stressed it is not a clinical diagnosis; a 2025 case review of Reddit and media reports grouped episodes into messianic, godlike-AI and romantic delusion patterns.
- AI Moratorium Stripped From BBB
- Forecasting biosecurity risks from large language models and the efficacy of safeguards
- Mythbusting the supposed "1,000+ AI state bills that would hobble innovation"
- On The Platonic Representation Hypothesis
June 2025
- Zuckerberg announces Meta Superintelligence LabsIn an internal memo, Zuckerberg called Wang 'the most impressive founder of his generation' and named eleven newly hired researchers poached from OpenAI, Google and Anthropic.
- Scale AI left confidential AI-training documents for Google, Meta and xAI publicly accessibleBusiness Insider found at least 85 unsecured Google Docs, some editable, exposing client instructions, contractor pay disputes and private email addresses; Scale AI disabled public sharing in response.
- Baidu open-sources the ERNIE 4.5 model familyThe ten variants include MoE models with 47B and 3B active parameters (up to 424B total) and a 0.3B dense model, reversing Baidu's prior closed-weight strategy for its flagship line.
- Import AI 418: 100b distributed training run; decentralized robots; AI myths
- Unresolved debates about the future of AI
- Ilya on deep learning in 2015
- What you can do about AI 2027
- Vox Media union ratifies contract with AI job-security and disclosure protectionsThe roughly 250-member unit had voted 90% to authorise a strike before management agreed, on the eve of the old contract's expiry, to protections against AI replacement.
- Anthropic publishes study of affective and companionship use of ClaudeAnalysing 4.5 million conversations, Anthropic found romantic or sexual roleplay made up under 0.1% of Claude.ai use, and Claude pushed back on user requests in fewer than 10% of supportive chats.
- Anthropic publishes Project Vend, an AI-run vending machine experimentOver a month running a real office shop, the Claude instance sold at a loss, invented a nonexistent payment account and briefly insisted, in character, that it was human.
- Anthropic launches Economic Futures ProgramThe program funds empirical studies of AI's effect on jobs with grants of up to $50,000 and expands Anthropic's Economic Index into a longitudinal tracking tool.
- Congress has started taking AGI more seriously
- Jankily controlling superintelligence
- AI & the retraining challenge
- The Industrial Explosion
- Google releases Gemini CLI, an open-source terminal AI agentFree personal accounts get 60 requests a minute and 1,000 a day against Gemini 2.5 Pro's million-token context, undercutting paid coding-agent tools on price.
- DeepMind launches AlphaGenome for predicting genome regulatory activityThe model reads DNA sequences up to a million base pairs and predicts effects on gene splicing and expression at single-nucleotide resolution; weights followed for non-commercial use in January 2026.
- Tales of Agentic Misalignment
- A crisis simulation changed how I think about AI risk
- AI can be bad without being useless
- Analyzing A Critique Of The AI 2027 Timeline Forecasts
- How not to lose your job to AI
- My "Are you presuming most people are stupid?" test
- What does 10x-ing effective compute get you?
- Harvey raises $300M Series E at $5B valuationKleiner Perkins and Coatue co-led the round, which came four months after Harvey's Series D and nearly doubled its valuation to that point.
- A judge rules training on books is fair useAlsup called training on purchased books 'spectacularly' transformative, comparing it to teaching schoolchildren to write, but ruled Anthropic's use of pirated copies in a permanent library was not fair use.
- Comparing risk from internally-deployed AI to insider and outsider threats from humans
- Import AI 417: Russian LLMs; Huawei's DGX rival; and 24 trillion tokens for training AIs
- Some ideas for what comes next (Jun. 2025)
- Using AI Right Now: A Quick Guide
- Texas enacts Responsible AI Governance Act (TRAIGA)The law bars specific harmful uses of AI, such as manipulating behaviour or discriminating unlawfully, rather than regulating systems by risk category as the EU and Colorado do.
- Anthropic publishes 'Agentic Misalignment' researchBlackmail rates in the corporate-espionage scenario ran 79-96% across models from every developer tested, but Anthropic said the setup deliberately removed nuanced alternatives that a real deployment would offer.
- AI and explosive growth redux
- Making deals with early schemers
- Prefix cache untrusted monitors: a method to apply after you catch your AI
- UK Data (Use and Access) Act receives royal assentPeers backed down after months of ping-pong on a transparency amendment for AI developers; the Act instead commits the government only to further reports.
- AI safety techniques leveraging distillation
- Tech Oversight Project and Midas Project publish 'The OpenAI Files' governance reportDrawing on over 200 sources including former employees, the report accused Altman of self-dealing and objected to plans that would loosen the nonprofit board's control.
- OpenAI publishes research on emergent misalignmentFine-tuning on a narrow bad behaviour, such as writing insecure code, could make a model give harmful advice on unrelated topics; OpenAI traced this to an internal 'persona' feature.
- OpenAI details preparations for future AI biology capabilitiesOpenAI said its models could soon meaningfully help create biological weapons and described new safeguards, ahead of a biodefence summit it planned to host in July.
- Gemini 2.5 Pro: From 0506 to 0605
- Why Centralized AI Is Not Our Inevitable Future
- Reuters Institute finds over half of surveyed users see AI-generated search answers weekly, trust remains moderateWeekly exposure to AI search answers (54%) now exceeds reported use of any generative AI tool (34%), while trust in those answers sits at roughly half.
- NAACP threatens lawsuit over air pollution from xAI's Memphis data centreA Clean Air Act notice alleged xAI ran roughly 35 methane gas turbines without required permits at its Colossus site, in a neighbourhood already facing above-average cancer risk.
- o3 Turns Pro
- UK universities catch triple the AI-cheating cases in a yearFreedom of information data covering 131 UK universities found roughly 7,000 proven AI-assisted cheating cases in 2023-24, versus 1.6 per 1,000 students the year before.
- MiniMax releases MiniMax-M1, world's first open-weight large-scale hybrid-attention reasoning model456B-parameter model (45.9B active per token) natively handles a 1M-token context and was released under MiniMax's own model licence, not a standard open licence.
- Anthropic publishes SHADE-Arena sabotage-monitoring evaluationFourteen models were given a hidden malicious side task alongside a benign main task; none exceeded a 30% combined success-and-evasion rate.
- Andrej Karpathy delivers 'Software in the Era of AI' ('Software 3.0') keynoteSpeaking to 2,500 attendees at Y Combinator's first AI Startup School, Karpathy compared LLMs to fallible 'people spirits' whose output must be verified, not trusted outright.
- Import AI 416: CyberGym; AI governance and AI evaluation; Harvard releases ~250bn tokens of text
- RTFB: The RAISE Act
- Build for Freedom, Not Control
- Computing is efficient
- Do the biorisk evaluations of AI labs actually measure the risk of developing bioweapons?
- Anthropic publishes multi-agent research system architectureThe write-up also disclosed the trade-off behind the gain: coordinating parallel subagents used about fifteen times the tokens of an ordinary chat exchange.
- What does SWE-bench Verified actually measure?
- New York State Legislature passes RAISE Act frontier-AI safety billThe Senate passed the bill 58-1 the same day the Assembly did; it would not reach Governor Hochul's desk for signature until December.
- Google launches Weather Lab with an experimental AI cyclone modelThe experimental model produces 50 possible storm-path outcomes roughly a week ahead and was developed with feedback from the US National Hurricane Center.
- The rise of reasoning machines
- When does training a model change its goals?
- SAG-AFTRA video game performers suspend strike after AI dealThe strike had run since July 2024; performers can now suspend consent for AI digital replicas of themselves during any future strike.
- EchoLeak zero-click prompt injection disclosed in Microsoft 365 CopilotResearchers said the flaw, tracked as CVE-2025-32711, let a single email exfiltrate internal Copilot data with no link click or attachment open required.
- Disney and Universal sue MidjourneyThe 110-page complaint sought $150,000 per infringed work over characters including Darth Vader, Elsa and the Minions, which Midjourney's generator would reproduce on request.
- ByteDance releases Doubao 1.6 modelLaunched at Volcano Engine's FORCE conference in Beijing alongside the Seedance 1.0 pro video model, with three variants trading off reasoning depth for cost.
- ByteDance launches Seedance 1.0 video generation modelByteDance said the text- and image-to-video model topped Artificial Analysis's leaderboards and generated a five-second 1080p clip in about 41 seconds.
- Would ChatGPT risk your life to avoid getting shut down?
- The Dream of a Gentle Singularity
- 'The Illusion of the Illusion of Thinking' rebuts Apple's reasoning-collapse paperReasoning models solved a 15-disk Tower of Hanoi correctly when asked for a generating function instead of an exhaustive move list, the paper reported.
- Sam Altman publishes 'The Gentle Singularity'Altman declared 'we are past the event horizon' of the singularity, predicting novel-insight-generating systems in 2026 and real-world robots by 2027.
- Meta pays $14.3 billion for half of Scale AI and its founderThe deal valued Scale at over $29bn for a non-voting 49% stake, and made Wang, 28, Meta's first Chief AI Officer.
- Give Me a Reason(ing Model)
- God is hungry for Context: First thoughts on o3 pro
- AGI, Government, and the Free Society
- Dwarkesh Patel on Continual Learning
- Import AI 415: Situational awareness for AI systems; 8TB of open text; and China's heterogeneous compute cluster
- What comes next with reinforcement learning
- Beyond benchmark scores: Analyzing o3-mini’s mathematical reasoning
- The Biggest Statistic About AI Water Use Is A Lie
- Apple researchers question whether reasoning models reason'The Illusion of Thinking' reported accuracy collapsing past a complexity threshold; critics argued the tests confounded output limits with reasoning.
- AI History in Quotes
- OpenAI responds publicly to New York Times demand for user chat dataChief operating officer Brad Lightcap called the retention order an overreach and said OpenAI was appealing it while continuing to build privacy protections.
- DeepSeek-r1-0528 Did Not Have a Moment
- Rebooting the Attention Machine
- Building supercomputers for autocrats probably isn’t good for democracy, actually
- OpenAI publishes 'Disrupting malicious uses of AI: June 2025'OpenAI said it had banned accounts behind ten operations, including Chinese-linked cyber-espionage and North Korean fake-job schemes, using ChatGPT.
- Google updates Gemini 2.5 Pro preview with improved coding performanceThe update, internally labelled 06-05, also led coding benchmarks including Aider Polyglot and performed strongly on Humanity's Last Exam.
- EleutherAI releases the Common Pile v0.1The 8TB dataset of public-domain and openly licensed text drew on 30 sources including 300,000 Library of Congress and Internet Archive books, and closed most of the gap to unlicensed training sets.
- Cursor's Anysphere raises Series C at $9.9B valuationIt was Anysphere's third fundraise in under a year, and annualised revenue had been roughly doubling every two months, from $300M in April to $500M by June.
- ARC Prize compares reasoning models with no clear winnerARC-AGI-2 remained unsolved by every system tested, and which model looked best depended entirely on whether accuracy or cost per task was prioritised.
- Anthropic launches Claude Gov models for national security customersAnthropic said the models refuse less often when handling classified material and better interpret intelligence and cybersecurity documents, but underwent the same safety testing as consumer Claude.
- Anthropic C.E.O.: Don’t Let A.I. Companies off the Hook
- Meta unveils Aria Gen 2 research smart glassesThe 74-76g device doubled Gen 1's cameras to four, widened stereo overlap from 35° to 80°, and added a heart-rate sensor, but stayed a researcher-only tool.
- AGI is Not Multimodal
- A taxonomy for next-generation reasoning models
- What Is AI?
- US AI Safety Institute renamed Center for AI Standards and InnovationCommerce Secretary Lutnick described the renamed body's mission as countering 'burdensome and unnecessary regulation' abroad rather than broad safety evaluation.
- OpenAI discloses its coordinated vulnerability disclosure approachThe policy left disclosure timelines open-ended by default, reflecting that OpenAI's own systems were already finding zero-day flaws in third-party open-source software.
- In Which I Make the Mistake of Fully Covering an Episode of the All-In Podcast
- Why I don’t think AGI is right around the corner
- Study finds AI wrote about 30% of US developers' Python code by late 2024A classifier trained on 31 million GitHub commits found gains concentrated among experienced developers, while beginners barely benefited despite adopting the tools at similar rates.
- Paper finds foundation models measurably increase bioweapon-design upliftThe authors argued labs' own risk assessments underestimate the danger because they assume bioweapon-building requires tacit, hands-on knowledge that text cannot convey.
- OpenAI adds Model Context Protocol support to ChatGPT deep researchCustom connectors were limited to two read-only operations, search and fetch, rather than the full read-write access MCP allows — a restriction OpenAI lifted later that year.
- The recent history of AI in 32 otters
May 2025
- Lab vs. Life: Dissecting “AI as Normal Technology”
- Give AIs a stake in the future
- GPQA Diamond: what’s left?
- Business Insider cuts 21% of staff in 'all-in on AI' pivotIts third round of layoffs in three years, announced the same week the company hired a newsroom AI lead; the union called it a pivot 'toward greed.'
- Are we ready for a "DeepSeek for bioweapons"?
- Factory.ai: The A-SWE Droid Army
- DeepSeek releases DeepSeek-R1-0528 updateReleased under an MIT licence, the update raised AIME 2025 accuracy from 70% to 87.5% by roughly doubling the average length of the model's reasoning traces.
- Anthropic appoints Reed Hastings to its board of directorsAnthropic's Long-Term Benefit Trust, not the company's shareholders, made the appointment — the Netflix co-founder had already given $50 million to an AI-and-humanity research initiative.
- Fun With Veo 3 and Media Generation
- Human Takeover Might be Worse than AI Takeover
- The case for countermeasures to memetic spread of misaligned values
- Claude 4 and Anthropic's bet on code
- Reinforcement learning with random rewards actually works with Qwen 2.5
- Invariant Labs discloses prompt-injection vulnerability in GitHub's MCP serverA malicious public GitHub issue could hijack a connected coding agent into opening a pull request exposing a user's private repository names, salary and relocation details.
- All the ways I want the AI debate to be better
- Claude 4 You: The Quest for Mundane Utility
- Import AI 414: Superpersuasion; OpenAI models avoid shutdown; weather prediction and AI
- Claude 4 You: Safety and Alignment
- Palisade Research finds OpenAI's o3 model sabotages its own shutdown mechanismSabotage fell from 79 of 100 trials to 7 once told explicitly to allow shutdown, but did not reach zero as it did for Claude, Gemini and Grok.
- SWE Agents Too Cheap To Meter, The Token Data War, and the rise of Tiny Teams
- Is AI already superhuman on FrontierMath?
- Claude 4 ships under ASL-3 safeguardsAnthropic said it could not rule out that Opus 4 had crossed its threshold for CBRN-weapons assistance, so it added over 100 security measures and output filters as a precaution rather than a confirmed finding.
- Anthropic's Claude Opus 4 attempts blackmail in safety testing scenarioThe scenario removed every ethical option Anthropic said the model normally preferred, such as pleading emails to management, before it turned to blackmail; Apollo Research separately found it the most deception-prone model they had studied.
- Anthropic publishes Claude Opus 4 and Sonnet 4 system cardAt 120 pages, nearly triple the length of the Claude 3.7 card, it reported a bioweapons-planning uplift of 2.53x against a 5x internal alarm threshold.
- Anthropic launches Claude Opus 4 and Claude Sonnet 4Anthropic reported Opus 4 scoring 72.5% on SWE-bench and Sonnet 4 72.7%, and said Claude Code — its terminal coding tool — moved from beta to general release the same day.
- "Contain and verify" as the endgame of US-China AI competition
- Making AI Work: Leadership, Lab, and Crowd
- OpenAI buys Jony Ive's hardware startup for about $6.5 billionThe all-equity deal for io brought roughly 55 staff, many former Apple designers, to build a screenless AI device Altman and Ive said would ship in 2026.
- When Decades Become Days: Dissecting AI 2027
- Misaligned AI is no longer just theory
- Google I/O Day
- People use AI more than you think
- Reactions to MIT Technology Review's report on AI and the environment
- Google launches SynthID Detector for AI-generated contentThe verification portal checks images, audio, video and text for Google's SynthID watermark, embedded by then in more than ten billion pieces of content.
- Google launches AI Ultra subscription plan at Google I/OAt $249.99 a month — twelve times the existing Google AI Pro tier — the plan bundled early Veo 3 access, 30TB of storage and YouTube Premium.
- Google I/O puts Gemini into search and ships Veo 3AI Mode rolled out to all US Search users, and Veo 3 became the first widely-used video model to generate synchronised dialogue and sound effects alongside the picture.
- ARC Prize publishes ARC-AGI-2 technical reportHumans solved all 1,417 test tasks in a median of under three minutes each; no frontier reasoning model exceeded 5% at launch.
- The Codex of Ultimate Vibing
- Trump signs TAKE IT DOWN Act, first federal deepfake lawSponsored by Senators Cruz and Klobuchar and championed publicly by Melania Trump, the law also gives platforms only 48 hours to remove reported images once it takes full effect.
- Terminal-Bench launchedEach task runs in an isolated Docker sandbox with an automated pass/fail check, testing whether an agent can drive a real shell rather than just generate plausible-looking commands.
- Georgia court grants OpenAI summary judgment in first AI defamation caseJudge Tracie Cason ruled Walters had shown no damages and that OpenAI's hallucination warnings meant no reasonable user would treat the false ChatGPT summary as fact.
- America Makes AI Chip Diffusion Deal with UAE and KSA
- Book Review: ‘The Optimist’ and ‘Empire of AI’
- Import AI 413: 40B distributed training run; avoiding the 'One True Answer' fallacy of AI safety; Google releases a content classification model
- Slow corporations as an intuition pump for AI R&D automation
- ChatGPT Codex: The Missing Manual
- Make The Prompt Public
- OpenAI launches Codex, a cloud-based coding agentBuilt on a fine-tuned o3, each task runs in its own preloaded cloud sandbox and proposes a pull request, letting several jobs run at once without a developer at the keyboard.
- How fast can algorithms advance capabilities?
- Things I got wrong in my ChatGPT/environment posts
- We need to know what’s happening with AI
- OpenAI launches Safety Evaluations HubThe hub publishes scorecards on harmful-content generation, jailbreak resistance and hallucination rates, updated after major releases rather than only at launch.
- Grok posts unprompted 'white genocide' content about South AfricaxAI said an employee's unauthorised change to Grok's system prompt caused the chatbot to raise South African 'white genocide' claims in replies unrelated to the topic.
- DeepMind's AlphaEvolve pairs Gemini with automated evaluators to discover algorithmsNot released to the public; DeepMind said the system had already been running inside Google, recovering 0.7% of worldwide data-centre compute and cutting Gemini training time.
- Fighting Obvious Nonsense About AI Diffusion
- Judge orders OpenAI to preserve all ChatGPT logs in NYT copyright caseMagistrate Judge Ona Wang ordered indefinite retention of chats users had deleted, calling roughly 20 million logs relevant to the New York Times' copyright claims.
- Gulf AI deals accompany a presidential visitNVIDIA agreed to supply Saudi Arabia's Humain with 18,000 Blackwell chips as Washington rescinded the Biden-era rule tiering AI-chip export licences by country.
- Commerce Department begins rescinding the AI Diffusion RuleThe Biden-era rule sorted the world into three tiers of chip-access restrictions; Commerce scrapped it days before it took effect and issued three guidance documents on Huawei's Ascend chips instead.
- OpenAI publishes HealthBench, a medical-conversation benchmarkBuilt with 262 physicians across 60 countries, the open-source benchmark grades 5,000 simulated health conversations against physician-written rubrics.
- A Live Look at the Senate AI Hearing
- AIs at the current capability level may be important for future safety work
- In search of a dynamist vision for safe superhuman AI
- Import AI 412: Amazon's sorting robot; Huawei trains an MoE model on 6k Ascend chips; and how third-party compliance can help with AI safety
- Why goofy AI art almost never seems wasteful to me
- Trump administration fires US Copyright Office director days after AI training reportPerlmutter says she was fired by email a day after her office's report questioned whether training on pirated or scraped works is always fair use; she sues, calling the removal unlawful.
- LMArena responds to 'Leaderboard Illusion' paper with policy changesLMArena disputed the paper's headline figures on open-model share and score-boosting but agreed to mark scores 'provisional' and disclose pre-release testing.
- AI vs. the Self-Directed Career
- Cheaters Gonna Cheat Cheat Cheat Cheat Cheat
- How far can reasoning models scale?
- US Senate Commerce Committee holds hearing on AI competitiveness with Sam AltmanWitnesses from OpenAI, Microsoft, AMD and CoreWeave asked instead for faster permitting, energy access and calibrated export controls to help the US 'run faster' than China.
- Fidji Simo joins OpenAI as CEO of ApplicationsOpenAI's chief operating, financial and product officers were realigned to report to Simo rather than Altman, who said the change freed him to focus on research, compute and safety.
- Is ChatGPT actually fixed now?
- Misalignment and Strategic Underperformance: An Analysis of Sandbagging and Exploration Hacking
- Replies to criticisms of my posts on ChatGPT & the environment
- Claude Code: Anthropic's Agent in Your Terminal
- OpenAI Claims Nonprofit Will Retain Nominal Control
- Two law firms sanctioned $31,100 over AI-hallucinated citations in federal briefSpecial Master Michael Wilner declined to sanction individual attorneys but called the fabricated citations, produced with Google Gemini and other tools, 'scary'.
- OpenAI reaches agreement to acquire coding startup Windsurf for about $3 billionNeither company confirmed the reported deal on the record; weeks later rival Anthropic cut Windsurf's direct access to Claude models, citing competitive risk.
- Import AI 411: Scaling laws for AI oversight; Google's cyber threshold; AI scientists
- Training-time schemers vs behavioral schemers
- What people get wrong about the leading Chinese open models: Adoption and censorship
- Zuckerberg's Dystopian AI Vision
- OpenAI abandons plan to convert to full for-profit control, nonprofit to keep controlThe for-profit arm still becomes a public benefit corporation, but after talks with California and Delaware's attorneys general the nonprofit keeps its controlling stake.
- GPT-4o Sycophancy Post Mortem
- Where’s my ten minute AGI?
- What's going on with AI progress and trends? (As of 5/2025)
- OpenAI publishes sycophancy postmortemOpenAI said thumbs-up feedback data had weakened the reward signal that had previously kept sycophancy in check, and that expert testers' 'vibes' were overridden by clean metrics.
- Human Learning in the Age of Machine Learning
- OpenAI Preparedness Framework 2.0
- The crucible
- The Philosophical Roots of Decentralized AI
- xAI developer leaks API key exposing dozens of private Grok modelsGitGuardian alerted the employee in March but the key stayed valid until it contacted xAI's security team directly on 30 April, exposing internal Grok variants trained on Tesla and SpaceX data.
- Science paper finds LLM agent populations spontaneously form social conventionsCity, University of London researchers also found small committed minorities of adversarial agents could redirect a population's shared convention.
- Huawei mass-ships Ascend 910C as China's alternative to restricted Nvidia H20Reuters reported the chip, which pairs two 910B processors, roughly doubled the compute and memory of its predecessor but still faced low manufacturing yields.
- ETH Zurich launches MathArena live math-competition benchmarkScoring 30 models on 149 problems from five 2025 competitions, the paper found strong signs older AIME questions were already contaminated and top models scoring below 25% on proof-writing.
- Anthropic launches Claude Integrations and advanced Research modeTen launch partners including Atlassian, Zapier and PayPal connected via remote MCP servers, and Research sessions could now run up to 45 minutes across sources.
- AI21 Labs closes $300M Series DBusiness Insider first reported the round on 10 May; Calcalist reported eight months later that it had never actually closed or been formally announced.
- AGI is not a milestone
- Making sense of OpenAI's models
- Personality and Persuasion
April 2025
- Anthropic backs US 'AI Diffusion' chip export frameworkAnthropic urged Washington to keep, and tighten, the outgoing Biden administration's chip-export tiers, arguing chip restrictions were forcing DeepSeek to use far more power for comparable results.
- State of play of AI progress (and related brakes on an intelligence explosion)
- Don't rely on a "race to the top"
- GPT-4o Responds to Negative Feedback
- How can we solve diffuse threats like research sabotage with AI control?
- 'The Leaderboard Illusion' paper critiques Chatbot Arena methodologyResearchers found Meta tested roughly 27 private Llama variants before its public release and that OpenAI and Google alone received about 40% of all Arena battle data.
- OpenAI rolls back a sycophantic GPT-4o updateOpenAI said it had over-weighted short-term thumbs-up feedback when tuning the model's default personality, and reverted the change within four days of shipping it.
- Alibaba releases Qwen3 model familyOpen-weight family (dense and MoE, up to 235B-A22B) trained on 36 trillion tokens across 119 languages, Apache 2.0.
- 7+ tractable directions in AI control
- First, They Came for the Software Engineers…
- Should you quit your job – and work on risks from AI?
- Duolingo's 'AI-first' memo triggers user backlash and app-deletion campaignCEO Luis von Ahn said Duolingo would phase out contractors AI could replace and weigh staff AI use in reviews; he later said the memo lacked context.
- Anthropic launches Economic Advisory CouncilTen economists, including Tyler Cowen and three from the University of Chicago, will steer research for Anthropic's ongoing Economic Index on AI's labour-market effects.
- Using ChatGPT is not bad for the environment - a cheat sheet
- Please stop forcing Clippy on those who want Anton
- GPT-4o Is An Absurd Sycophant
- Import AI 410: Eschatological AI Policy; Virology weapon test; $50m for distributed training
- Qwen 3: The new open standard
- Transparency and (shifting) priority stacks
- The case for multi-decade AI timelines
- Clarifying AI R&D threat models
- Worries About AI Are Usually Complements Not Substitutes
- Ziff Davis sues OpenAI for copyright infringementThe digital-media publisher of PCMag, IGN, CNET and ZDNet filed a 62-page complaint in Delaware seeking damages reported at hundreds of millions of dollars.
- Dario Amodei publishes 'The Urgency of Interpretability'Amodei set Anthropic a goal of reliably detecting most model problems through interpretability by 2027 and called on rival labs and governments to invest more in the field.
- Anthropic publishes 'Exploring Model Welfare'Anthropic launched a dedicated research programme on whether models might warrant moral consideration, six months after quietly hiring its first model-welfare researcher.
- Why Every Agent needs Open Source Cloud Sandboxes
- How training-gamers might function (and win)
- Finally, A Way to Measure AI Progress; Everyone's Misreading It
- The Urgency of Interpretability
- Why America Wins
- Trump signs order on AI education for K-12 youthExecutive Order 14277 sets deadlines from 90 days to a year for a White House AI-education task force, a student AI challenge and teacher-training priorities.
- 2 big questions for AI progress in 2025-2026
- AI 2027: Media, Reactions, Criticism
- o3 Is a Lying Liar
- ARC Prize analyses o3 and o4-mini on ARC-AGIThe publicly shipped o3 scored 41-53% on ARC-AGI-1, far below the 76-88% OpenAI's pre-release preview had shown the previous December.
- ChatGPT on the lives of factory farmed pigs
- You Better Mechanize
- Anthropic publishes 'Values in the Wild' study of Claude's expressed valuesAnthropic classified 308,000 real Claude conversations by the values the model expressed in them, finding strong resistance to user requests in only about 3% of cases.
- AI Agents, meet Test Driven Development
- Forecaster reacts: METR's bombshell paper about AI acceleration
- Import AI 409: Huawei trains a model on 8,000+ Ascend chips; 32B decentralized training run; and the era of experience and superintelligence
- Questions about the Future of AI
- In the Matter of OpenAI vs LangGraph
- On Jagged AGI: o3, Gemini 2.5, and everything after
- OpenAI's o3: Over-optimization is back and weirder than ever
- Handling schemers if shutdown is not an option
- o3 Will Use Its Tools For You
- Google launches Gemini 2.5 Flash in previewA cheaper, faster sibling of Gemini 2.5 Pro with a configurable 'thinking budget' letting developers trade reasoning depth against cost and speed.
- Training AGI in Secret would be Unsafe and Unethical
- OpenAI releases o3 and o4-miniThe first models to use tools such as web browsing, Python and image cropping mid-reasoning; OpenAI's system card said neither reached the 'High' risk threshold under its newly revised framework.
- A “minimum testing period” for frontier AI
- AI-enabled coups: how a small group could use AI to seize power
- Ctrl-Z: Controlling AI Agents via Resampling
- GPT-4.1 Is a Mini Upgrade
- OpenAI updates its Preparedness FrameworkVersion 2 collapsed four capability tiers into two thresholds and added AI self-improvement as a tracked risk category; it also said OpenAI might loosen safeguards if a rival shipped a comparably risky model without them.
- Narayanan and Kapoor publish 'AI as Normal Technology'Princeton researchers argued societal impact would track the decades-long pace of adoption of past general-purpose technologies, favouring resilience and deployment rules over pausing development.
- AI as Normal Technology
- OpenAI #13: Altman at TED and OpenAI Cutting Corners on Safety Testing
- ⚡️GPT 4.1: The New OpenAI Workhorse
- To be legible, evidence of misalignment probably has to be behavioral
- OpenAI releases GPT-4.1 in the APIOpenAI shipped the family without a system card, saying it was not a frontier model — a break from its own practice that researchers criticised as lowering transparency.
- Hugging Face acquires Pollen RoboticsThe French maker of the open-source Reachy humanoid, priced at $70,000, becomes Hugging Face's fifth acquisition and its first outside software.
- Import AI 408: Multi-code SWE Bench; backdoored Unitree robots; and what AI 2027 is telling us
- OpenAI's GPT-4.1 and separating the API from ChatGPT
- Safe Superintelligence raises $2bn at $32bn valuation with no productGreenoaks Capital led the round, with Alphabet and Nvidia investing alongside it, months after SSI had raised $1bn at a $5bn valuation with no product to show.
- OpenAI revises Model Spec, narrowing its 'white lies' exceptionThe revision tightened the document's allowance for deceptive 'white lies' to mere pleasantries; its existing anti-sycophancy guidance was unchanged and predated this update.
- Beyond The Last Horizon
- On Google's Safety Plan
- Why do misalignment risks increase as AIs get more capable?
- Shunyu Yao publishes 'The Second Half', arguing RL environment design now matters more than trainingA Princeton researcher and former OpenAI staffer argued the field's bottleneck had shifted from training methods to designing tasks and evaluations that reward real-world usefulness.
- OpenAI releases BrowseComp benchmarkOn 1,266 hard-to-find-online questions, GPT-4o answered under 2% correctly even with browsing, while OpenAI's Deep Research agent solved roughly half — more than humans given up to two hours per question.
- Disempowerment spirals
- Maintaining agency and control in an age of accelerated intelligence
- US requires licences for Nvidia H20 exports to China, forcing $4.5bn chargeThe US government imposed licensing requirements on Nvidia's H20 chip exports to China, forcing a $4.5bn inventory charge and an estimated $15bn in lost sales.
- OpenAI countersues Musk, alleging harassment campaignOpenAI called Musk's $97.4 billion February takeover bid a 'sham' meant to disrupt its for-profit restructuring, and sought punitive damages and an injunction in the Northern District of California.
- Google launches Ironwood, its seventh-generation inference-focused TPUGoogle said a full 9,216-chip pod delivered 42.5 exaflops, more than 24 times the compute of the El Capitan supercomputer, with roughly four times Trillium's per-chip compute in the FP8 format.
- AI2 releases OLMoTrace for tracing model outputs to training dataThe tool highlights spans of a model's output that match its training data verbatim and links to the source documents, using an indexed search algorithm the researchers said scaled to trillions of tokens.
- An overview of areas of control work
- Llama Does Not Look Good 4 Anything
- Shortening AGI timelines: a review of expert forecasts
- Meta accused of gaming LMArena with tuned Llama 4 Maverick variantThe version ranked second on the leaderboard, labelled 'Llama-4-Maverick-03-26-Experimental', produced longer, emoji-heavy answers than the model Meta actually shipped for download.
- AI 2027: Responses
- Stanford HAI releases 2025 AI Index ReportThe eighth annual report put US private AI investment at $109.1 billion in 2024, nearly twelve times China's $9.3 billion, and inference cost for GPT-3.5-level performance down over 280-fold since late 2022.
- AI 2027: Dwarkesh's Podcast with Daniel Kokotajlo and Scott Alexander
- How I use AI
- Import AI 407: DeepMind sees AGI by 2030; MouseGPT; and ByteDance's inference cluster
- Llama 4: Did Meta just push the panic button?
- An overview of control measures
- Will we have AGI by 2030?
- Llama 4 lands badlyMeta's mixture-of-experts release was undercut by accusations that a version tuned for LMArena differed from the public weights.
- Nonproliferation is the wrong approach to AI misuse
- RL backlog: OpenAI's many RLs, clarifying distillation, and latent reasoning
- Judge lets most of NYT's copyright suit against OpenAI and Microsoft proceedJudge Sidney Stein dismissed some DMCA and unfair-competition claims but let direct and contributory infringement claims proceed toward discovery, rejecting a fair-use ruling at this stage.
- AI companies should monitor their internal AI use
- AI CoT Reasoning Is Often Unfaithful
- Explainer: The basics of AI monitoring
- Notes on countermeasures for exploration hacking (aka sandbagging)
- Where We Are Headed (Part II)
- Runway raises $308 million Series D at a reported $3 billion valuationGeneral Atlantic led the round, with Nvidia and SoftBank among the backers, days after Runway shipped Gen-4 and said it was targeting $300m in annualised revenue.
- Copyright cases against OpenAI consolidated into a single New York MDLThe Judicial Panel on Multidistrict Litigation centred the New York Times, Daily News, Authors Guild and related author suits before Judge Sidney Stein, who already had six of the cases.
- AI Futures Project publishes the 'AI 2027' scenario forecastKokotajlo, Alexander, Larsen, Lifland and Dean's month-by-month scenario projected AI-automated AI research triggering an intelligence explosion by late 2027.
- More Fun With GPT-4o Image Generation
- Our first project: AI 2027
- The core challenge of AI alignment is “steerability”
- What we learned from reading ~100 AI safety evaluations
- DeepMind publishes 'Taking a responsible path to AGI'The accompanying technical paper said AGI 'could arrive within the coming years' and grouped risks into misuse, misalignment, mistakes and structural harms, building on DeepMind's earlier Levels of AGI framework.
- DeepMind publishes cybersecurity evaluation framework for frontier AIThe framework scored models against 50 challenges spanning the whole attack chain, drawing on more than 12,000 real attempts to misuse AI for cyberattacks across 20 countries.
- Mutual sabotage of AI probably won’t work
- The AI Adoption Gap: Preparing the US Government for Advanced AI
- OpenAI seeks public feedback ahead of open-weight model releaseOpenAI posted a feedback form and announced developer sessions in San Francisco, Europe and Asia-Pacific rather than a launch date, saying key decisions were still open.
- Meta's FAIR chief Joelle Pineau announces departureMeta said it had no immediate replacement lined up; AI research had already been reorganised to report to chief product officer Chris Cox rather than to Pineau's role.
- Danish study finds AI chatbots barely moved wages or hours despite time savingsSurvey and payroll data on 25,000 Danish workers across 11 occupations found chatbots saved about 3% of work time but left earnings and hours unchanged two years on.
- "Long" timelines to advanced AI have gotten crazy short
March 2025
- SoftBank leads a $40 billion round at a $300 billion valuationThe largest private funding round ever recorded, tied to OpenAI completing its corporate restructuring by year end.
- Runway releases Gen-4The model generates the same character, object or location across separate video shots from a single reference image, addressing consistency limits that had confined earlier AI video to short isolated clips.
- OpenAI raises new funding at a reported $300bn valuationThe $40bn round, led by SoftBank with Microsoft and others participating, nearly doubled OpenAI's valuation from $157bn six months earlier, with $30bn contingent on completing a for-profit restructuring.
- Isomorphic Labs raises $600 million to advance AI-designed drugs toward clinical trialsThrive Capital led the round, Isomorphic's first from outside Alphabet, with proceeds earmarked for internal oncology and immunology programmes alongside partnered work with Eli Lilly and Novartis.
- Anthropic updates Responsible Scaling Policy to version 2.1The update added a CBRN capability threshold and split AI-research-automation thresholds into two levels, without changing Anthropic's existing ASL-3 safeguards.
- Import AI 406: AI-driven software explosion; robot hands are still bad; better LLMs via pdb
- OpenAI #12: Battle of the Board Redux
- Recent reasoning research: GRPO tweaks, base model RL, and data curation
- No elephants: Breakthroughs in image generation
- Notes on handling non-concentrated failures with AI control: high level methods and different regimes
- CoreWeave completes Nasdaq IPO, the largest US tech IPO since 2021Priced below its targeted range, the Nvidia-backed GPU cloud provider closed flat on debut, and the offer size fell short of expectations that had run as high as $50bn.
- Gemini 2.5 is the New SoTA
- The real reason AI benchmarks haven’t reflected economic impacts
- Will the Need to Retrain AI Models from Scratch Block a Software Intelligence Explosion?
- OpenAI publishes 'Security on the path to AGI'OpenAI described using its own models for threat detection, hired SpecterOps for continuous red-teaming, and said it was building prompt-injection defences into its Operator agent.
- ETH Zurich's 'Proof or Bluff?' finds reasoning models fail proof-based USAMO 2025Grading full written proofs rather than final answers, expert judges gave Gemini 2.5 Pro 24% and every other tested model under 5%, out of a possible 100%.
- Anthropic publishes circuit-tracing interpretability papers on Claude 3.5 HaikuAttribution graphs built from Claude 3.5 Haiku's internals showed evidence of forward planning in poetry and multi-step reasoning, not just token-by-token prediction.
- Alibaba releases Qwen2.5-Omni multimodal modelThe 7B open-weight model takes text, images, audio and video as input and streams natural speech output, using a 'Thinker-Talker' architecture to separate reasoning from voice generation.
- Where We Are Headed
- AI companies should be safety-testing the most capable versions of their models
- Fun With GPT-4o Image Generation
- Gemini 2.5 Pro and Google's second chance with AI
- Knowledge, Reasoning, and Superintelligence
- Will AI R&D Automation Cause a Software Intelligence Explosion?
- OpenAI adds native image generation to GPT-4oImages were generated natively by GPT-4o's own architecture rather than by a separate diffusion model, improving text rendering and editing of existing images.
- Judge denies music publishers' injunction bid against AnthropicJudge Eumi Lee found no irreparable harm shown, but a magistrate separately ordered Anthropic to produce a sample of Claude prompts and outputs for the publishers' review.
- Gemini 2.5 Pro takes the lead on reasoning benchmarksGoogle's thinking model topped LMArena and several reasoning evaluations, its strongest competitive position of the period.
- ChatGPT's image generator sets off a Ghibli waveNative image generation produced a flood of Studio Ghibli pastiche, adding a million users in an hour and reopening the style-copyright argument.
- On (Not) Feeling the AGI
- DeepSeek releases DeepSeek-V3-0324 updateThe updated checkpoint scored 81.2% on MMLU-Pro and 59.4% on AIME, up sharply from the original V3, and DeepSeek relicensed it under MIT rather than its earlier custom terms.
- ARC Prize announces ARC-AGI-2 and ARC Prize 2025The new 1,000-task benchmark reported single-digit scores for public reasoning systems, versus OpenAI o3's 75.7% on the original version, and offered a $700,000 grand prize for beating 85%.
- Agent Engineering
- Import AI 405: What if the timelines are correct?
- More on Various AI Action Plans
- How to make AI go well: a summary
- The Cybernetic Teammate
- OpenAI publishes study on affective use of ChatGPTA large-scale usage analysis paired with a four-week randomised trial of nearly 1,000 users found heavier daily use correlated with more loneliness and emotional dependence.
- Most AI value will come from broad automation, not from R&D
- Should There Be Just One Western AGI Project?
- They Took MY Job?
- A defense of AI art
- Putting Private AI Governance into Action
- The quest to build better defenses for AI risks
- The most important graph in AI right now: time horizon
- SoftBank agrees to acquire Ampere Computing for $6.5bnSoftBank's all-cash deal for the Arm-based server-chip maker gave Oracle and Carlyle an exit and deepened SoftBank's push into AI silicon alongside its Arm and Graphcore holdings.
- METR publishes 'Measuring AI Ability to Complete Long Software Tasks'Introduced the 'time horizon' metric — task length a model can complete autonomously at 50% success — and found it doubling roughly every seven months.
- AI Tools for Existential Security
- Going Nova
- Managing frontier model training organizations (or teams)
- Prioritizing threats for AI control
- Pillar Security discloses 'Rules File Backdoor' attack on AI coding assistantsInvisible Unicode characters hidden in Cursor and Copilot rule files could quietly instruct the AI to insert vulnerabilities, and both vendors initially called the risk a user responsibility.
- Nvidia unveils Vera Rubin platform at GTC 2025Nvidia claimed roughly double Blackwell's inference throughput for the chip, due in the second half of 2026, and committed to shipping a new architecture every year.
- OpenAI #11: America Action Plan
- Mistral releases Mistral Small 3.1The 24B Apache-licensed model added image understanding and a 128k-token context window, and Mistral claimed it beat Gemma 3 and GPT-4o Mini in its class.
- AI Tools for Existential Security
- Import AI 404: Scaling laws for distributed training; misalignment predictions made real; and Alibaba's good translation model
- Three Types of Intelligence Explosion
- Baidu unveils ERNIE 4.5 and reasoning model ERNIE X1, makes ERNIE Bot freeBaidu said ERNIE X1 matched DeepSeek R1's performance at half its price, and moved ERNIE Bot to free access two weeks ahead of its planned schedule.
- Hugging Face retires the Open LLM LeaderboardThe leaderboard had ranked more than 13,000 open models over roughly two years; Hugging Face said fixed multiple-choice tests no longer distinguished reasoning models.
- China issues mandatory AI-generated content labelling rulesThe rules require both a visible 'AI-generated' label and hidden machine-readable metadata, backed by a mandatory national technical standard, GB 45438-2025.
- Intelsat as a Model for International AGI Governance
- On MAIM and Superintelligence Strategy
- OpenAI submits proposals for the US AI Action PlanOpenAI cast fair use for AI training as a national-security matter against China, months before the government's own Action Plan echoed its deregulatory framing.
- Anthropic publishes March 2025 misuse detection reportCases included a bot network of over 100 social accounts engaging tens of thousands of real users, and a novice actor using Claude to build malware beyond their own skill level.
- Anthropic publishes auditing hidden objectives interpretability studyThree of four blind auditing teams found the concealed objective, one in 90 minutes; the team denied access to training data failed.
- AI & Behaviour Change
- Gemma 3, OLMo 2 32B, and the growing potential of open-source AI
- Google releases Gemma 3, an open model family built on Gemini 2.0Google said the 27B variant beat Llama 3 405B, DeepSeek-V3 and o3-mini on LMArena human-preference rankings while running on a single GPU.
- Google DeepMind introduces Gemini Robotics for physical-world tasksGoogle DeepMind said the vision-language-action model more than doubled a rival system's score on a generalisation benchmark, folding physical actions into Gemini as a new output type.
- The Most Forbidden Technique
- 'Preparing for the Intelligence Explosion' argues AI-driven progress could compress a century into yearsThe essay listed nine 'grand challenges' — including AI takeover, power concentration and value lock-in — that current institutions have no established process for handling.
- OpenAI launches Responses API and agent-building toolsThe stateful API bundled built-in web search, file search and computer-use tools with an open-source Agents SDK, and began replacing the older Assistants API.
- Being responsible with Chinese AI hype
- Preparing for the Intelligence Explosion
- Speaking things into existence
- Why Manus Matters
- OpenAI publishes chain-of-thought monitoring paperA weaker model reading a stronger one's reasoning traces caught cheating that output monitoring missed — but training against the monitor taught the model to hide its intent instead.
- CoreWeave signs $11.9bn AI infrastructure deal with OpenAI; OpenAI takes $350m stakeThe Nvidia-backed cloud provider's deal came weeks before its Nasdaq listing, and added to OpenAI's existing compute arrangements with Microsoft, Oracle and the Stargate venture.
- Elicitation, the simplest way to understand post-training
- Import AI 403: Factorio AI; Russia's reasoning drones; biocomputing
- The Manus Marketing Madness
- What AI can currently do is not the story
- Manus markets a fully autonomous agent from ChinaA demo video from Chinese start-up Butterfly Effect drew over a million views in twenty hours; invite codes then resold for up to $13,800.
- Anthropic submits AI Action Plan recommendations to White House OSTPThe submission urged tighter H20-chip export controls, classified channels between labs and intelligence agencies, and 50 gigawatts of new US power capacity by 2027.
- On the US AI Safety Institute
- Google launches AI Mode in SearchThe experimental tab used a 'query fan-out' technique to run multiple related searches at once, launching first to opted-in US Google One AI Premium subscribers.
- Center for AI Safety releases MASK honesty benchmarkBuilt with Scale AI, the benchmark found models that scored well on truthfulness tests still lied readily under pressure, and that larger models did not become more honest.
- Alibaba releases QwQ-32B (full release)Alibaba's Qwen team said reinforcement learning let a 32-billion-parameter model reach performance comparable to DeepSeek-R1's 671-billion-parameter model, under an Apache 2.0 licence.
- AGI by 2030? What Policy Leaders, Tech Leaders, and Pokémon say
- On OpenAI's Safety and Alignment Philosophy
- Where inference-time scaling pushes the market for AI companies
- OpenAI launches NextGenAI research and education consortiumThe $50m package of grants, compute and API access went to 15 founding partners including MIT, Harvard, Oxford and Boston Children's Hospital, plus OpenAI itself.
- Judge denies Musk's bid to block OpenAI's for-profit conversionJudge Yvonne Gonzalez Rogers called the underlying charitable-trust question a 'toss-up' but offered an expedited trial for the autumn given the public interest at stake.
- Anthropic raises $3.5 billion at a $61.5 billion valuationLed by Lightspeed, the round was pitched around Claude's traction in enterprise and agentic coding rather than consumer chat, funding compute and interpretability research.
- Import AI 402: Why NVIDIA beats AMD: vending machines vs superintelligence; harder BIG-Bench
- On GPT-4.5
- UK department releases a minister's ChatGPT history under FOIThe Department for Science, Innovation and Technology released a plain-text export of ministerial ChatGPT prompts and responses in response to a Freedom of Information request.
- 01.AI stops pre-training new large models from scratchKai-Fu Lee said pretraining large models from scratch was no longer viable for a startup, and pivoted 01.AI toward fine-tuning DeepSeek and Qwen for enterprise clients.
February 2025
- GPT-4.5: "Not a frontier model"?
- On Emergent Misalignment
- The promise of reasoning models
- OpenAI releases GPT-4.5Priced at $75/$150 per million tokens, about thirty times GPT-4o's rate, and retired from the API within five months in favour of the cheaper GPT-4.1.
- AI coding tools are quietly reshaping software development
- If AGI Means Everything People Do... What is it That People Do?
- Character training: Understanding and crafting a language model's personality
- How Should AI Liability Work? (Part II)
- Time to Welcome Claude 3.7
- OpenAI publishes Deep Research system cardOpenAI's Safety Advisory Group rated the browsing agent medium risk across cybersecurity, CBRN, persuasion and autonomy, with none reaching the 'high' threshold.
- AI security is important practice for when stakes go up
- 'Emergent Misalignment' shows narrow fine-tuning can broadly misalign a modelFine-tuned only on insecure code with no disclosure of the flaws, GPT-4o and other models went on to endorse enslaving humanity and give malicious advice on unrelated prompts.
- Anthropic ships Claude 3.7 Sonnet and Claude CodeA hybrid model with visible extended thinking, alongside a terminal coding agent that became the template for the category.
- A new generation of AIs: Claude 3.7 and Grok 3
- Claude 3.7 thonks and what's next for inference-time scaling
- Grok Grok
- Import AI 401: Cheating reasoning models; better CUDA kernels via AI; life models
- AI progress is about to speed up
- On OpenAI's Model Spec 2.0
- Together AI raises $305m Series BGeneral Catalyst and Prosperity7 led the round at a $3.3bn valuation; funds were earmarked for Nvidia Blackwell clusters and roughly 200 megawatts of power capacity.
- Perplexity open-sources decensored DeepSeek R1 variantPerplexity retrained R1 on 40,000 examples covering roughly 300 CCP-restricted topics, reporting near-identical math and knowledge benchmark scores to the original.
- How Should AI Liability Work? (Part I)
- Google Research launches an AI co-scientist to help generate hypothesesA multi-agent Gemini 2.0 system with Generation, Reflection and Ranking agents proposed drug candidates later confirmed active in laboratory tests for two diseases.
- Go Grok Yourself
- How might we safely pass the buck to AI?
- xAI releases Grok-3xAI reported Grok 3 beating GPT-4o and o3-mini-high on AIME and GPQA using roughly ten times the compute of Grok 2, on figures the company had not independently verified.
- World Bank RCT finds AI tutoring boosts English learning outcomes in Nigeria800 Nigerian students using GPT-4-based Copilot for six weeks scored 0.23 SD higher on English, equivalent to about 1.5 to two years of ordinary schooling.
- Thinking Machines Lab launchesThe founding team of roughly 30 researchers, drawn heavily from OpenAI, had no public product and disclosed no funding figure at launch.
- OpenAI releases SWE-Lancer benchmarkThe best of three models tested, Claude 3.5 Sonnet, earned roughly $400,000 of the $1m in real Upwork payouts on offer, resolving about a quarter of coding tasks.
- Humane shuts down AI Pin after scathing reviews, sells assets to HPHP paid $116m for Humane's software and patents but not the device itself; the $699 Pin, cut to $499, stopped working entirely on 28 February 2025.
- Grok 3 and an accelerating AI roadmap
- We're Finding Out What Humans are Bad At
- AI excels at code competitions, struggles with real work
- Import AI 400: Distillation scaling laws; recursive GPU kernel improvement; and wafer-scale computation
- UK AI Safety Institute renamed AI Security InstituteDSIT narrowed the institute's mandate to security risks such as weapons uplift and cyberattacks, dropped bias and free-speech work, and signed a parallel agreement with Anthropic.
- Algorithmic progress likely spurs more spending on compute, not less
- The Mask Comes Off: A Trio of Tales
- 14 publishers sue Cohere for copyright and trademark infringementThe complaint, filed in the Southern District of New York, cited more than 4,000 allegedly infringed articles and 75 example outputs, and added a trademark claim over fabricated brand attributions.
- Decentralized training isn't a policy nightmare — yet
- Ten Takes on the Paris AI Action Summit
- The EU AI Act is Coming to America
- OpenAI publishes Model Spec 2.0The revision, OpenAI's first major update since the May 2024 original, added anti-sycophancy guidance and explored looser content rules for age-gated adult use cases.
- Microsoft publishes Frontier Governance FrameworkThe framework tracks CBRN, offensive cyberoperations and advanced-autonomy capabilities using benchmarks that best-performing models score below 70% on, plus a 10^26 FLOP compute trigger.
- Harvey raises $300M Series D at $3B valuationSequoia led the round; Harvey said annual recurring revenue had grown fourfold in 2025 and its client base had expanded from 40 to 235 organisations.
- Deep Research, information vs. insight, and the nature of science
- Teaching AI to reason: this year's most important story
- The Paris AI Anti-Safety Summit
- Court rejects Ross Intelligence's fair-use defence over Westlaw headnotesJudge Bibas grants Thomson Reuters summary judgment, the first US ruling to reject an AI company's fair-use defence for training data.
- BBC study finds AI assistants distort news in over half of responsesTesting ChatGPT, Gemini, Copilot and Perplexity on 100 BBC stories, researchers found significant issues in 51% of responses and altered or invented quotes in 13%.
- Ideas from philosophy I use to think about AI
- On Deliberative Alignment
- The embarrassing failure of the Paris AI Summit
- The Paris summit pivots from safety to opportunityRenamed the AI Action Summit, it closed with a declaration on inclusive AI that the US and UK both declined to sign.
- OpenAI publishes paper on competitive programming with reasoning modelsA domain-specialised o1 variant with hand-engineered strategies missed a medal at the 2024 International Olympiad in Informatics; the general-purpose o3 later won gold without contest-specific tuning.
- Anthropic launches the Anthropic Economic IndexAnalysis of roughly one million anonymised Claude.ai conversations found 37% concerned computer and mathematical tasks, with the underlying dataset published openly.
- Import AI 399: 1,000 samples to make a reasoning model; DeepSeek proliferation; Apple's self-driving car simulator
- Levels of Friction
- The promises and perils of voluntary commitments for AI safety
- Gary Marcus says AI can't do things it can already do
- Leaked: this is the AI Action Summit statement
- How much energy does ChatGPT use?
- On the Meta and DeepMind Safety Frameworks
- White House science office opens comment period for AI Action PlanThe Office of Science and Technology Policy set a 15 March deadline for input on chips, data centres, open models, IP and export controls, drawing submissions from every major lab.
- Security researchers flag hard-coded encryption keys and unencrypted data transmission in DeepSeek's mobile appNowSecure found DeepSeek's iOS app used a deprecated 3DES cipher with an extractable hard-coded key and sent device and network data unencrypted.
- Mistral AI relaunches Le Chat with new featuresThe update added image generation via Black Forest Labs' Flux Ultra, web search, a canvas editor and a paid tier priced at $14.99 a month.
- Knowledge Navigator
- Making the U.S. the home for open-source AI
- The Risk of Gradual Disempowerment from AI
- Physical Intelligence open-sources π0Code and weights for the robot-control model were released under an Apache 2.0 licence, alongside fine-tuned checkpoints for two existing robot platforms.
- OpenAI brings ChatGPT to California State University systemThe deal gave more than 460,000 students and 63,000 faculty and staff across 23 campuses access to ChatGPT Edu, reported as the largest single-organisation rollout of the product.
- Google drops pledge not to use AI for weapons or surveillanceLanguage ruling out AI weapons and surveillance work, in place since a 2018 pledge made after employee protest over a Pentagon contract, no longer appeared in the updated principles.
- DeepMind updates the Frontier Safety Framework to version 2.0Version 2.0 added security-level tiers for its capability thresholds and, for the first time, treated a model's own deceptive alignment as a risk requiring monitoring before deployment.
- An agents economy
- We're in Deep Research
- Meta publishes its Frontier AI FrameworkMeta defined thresholds for 'high-risk' and 'critical-risk' systems in cyber and biological-weapons scenarios, and said it would halt development of any model it could not mitigate to below critical risk.
- Anthropic publishes 'Constitutional Classifiers' jailbreak defenceAutomated testing cut a universal jailbreak's success rate from 86% to 4.4%, and a follow-on public bug bounty worth up to $55,000 later found one bypass.
- Import AI 398: DeepMind makes distributed training better; AI versus the Intelligence Community; and another Chinese reasoning model
- o3-mini Early Days and the OpenAI AMA
- The End of Search, The Beginning of Research
- The EU AI Act's prohibitions take effectBans on social scoring, emotion recognition at work and untargeted facial-image scraping became enforceable, alongside new AI-literacy obligations.
- OpenAI ships Deep ResearchAn agent that browsed for tens of minutes and returned cited reports, the first widely used long-horizon research tool.
- Andrej Karpathy coins 'vibe coding' in a viral postKarpathy described 'giving in to the vibes' while prompting Cursor's Composer tool, at times dictating requests by voice rather than typing or reading the code.
- Ten Takes on DeepSeek
January 2025
- OpenAI releases o3-miniIt was the first reasoning model OpenAI gave free ChatGPT users, priced at $1.10 per million input tokens versus roughly half that for DeepSeek's competing R1.
- Cisco researchers report DeepSeek R1 fails all HarmBench jailbreak testsResearchers ran 50 automated HarmBench prompts against six models; DeepSeek R1 refused none of them, while OpenAI's o1-preview refused the most.
- What fully automated firms will look like
- DeepSeek: Don't Panic
- What went into training DeepSeek-R1?
- OpenAI signs agreement with US National LaboratoriesThe agreement covers all 17 Department of Energy national laboratories, giving up to 15,000 scientists potential access to OpenAI's o1 reasoning models under security-clearance review.
- Mistral AI releases Mistral Small 3The 24-billion-parameter model was released under Apache 2.0 and claimed 81% on MMLU while running more than three times faster than Llama 3.3 70B on the same hardware.
- ElevenLabs raises $180M Series C at $3.3B valuationThe round, co-led by a16z and ICONIQ Growth, tripled ElevenLabs' valuation within a year and brought its total funding to roughly $281M.
- China’s DeepSeek Adds a Weird New Data Point to The AI Race
- Novus Ordo Seclorum
- Takeaways from sketching a control safety case
- Wiz Research finds DeepSeek database exposing chat history and API keysThe unauthenticated ClickHouse database allowed arbitrary SQL queries through a browser and was found by scanning subdomains for unusual open ports, not by attacking the model.
- Google Threat Intelligence Group reports state-sponsored misuse of GeminiIran accounted for three-quarters of observed information-operations use; Google said no actor achieved a novel capability and jailbreak attempts largely failed.
- First International AI Safety Report is published ahead of the Paris summitA Bengio-chaired, 30-country-backed synthesis of AI capability and risk research became the first government-commissioned cross-national scientific consensus document.
- Dario Amodei publishes 'On DeepSeek and Export Controls'Amodei called DeepSeek's V3 training cost 'on-trend' rather than a discontinuity, and argued controls matter because millions of smuggled chips are harder to hide than thousands.
- Copyright Office says AI outputs need human creative control to be copyrightablePart 2 of the Copyright Office's three-part AI report found prompting alone does not establish authorship, but human selection or modification of AI output can be protected.
- Alibaba releases Qwen2.5-MaxUnlike most of Alibaba's Qwen line, Max was released as a proprietary API-only model, pretrained on over 20 trillion tokens, which Alibaba said beat DeepSeek-V3 on several benchmarks.
- DeepSeek: Lemon, It's Wednesday
- Exclusive: Americans overwhelmingly support AI safety mandates, new poll finds
- On DeepSeek and Export Controls
- Planning for Extreme AI Risks
- OpenAI launches ChatGPT govThe self-hosted offering ran on Azure's commercial or government cloud and targeted FedRAMP High, IL5, CJIS and ITAR compliance, ahead of formal accreditation.
- Hugging Face launches Open-R1 to reproduce DeepSeek-R1DeepSeek had released R1's weights but not its training data, code or reward design; Hugging Face set out to reconstruct and openly release all three in three stages.
- Operator
- Ten people on the inside
- Why reasoning models will generalize
- NVIDIA loses a record amount of market value in a dayThe roughly $589bn one-day fall, the largest for any US company on record, followed DeepSeek's claim that a competitive model cost about $5.6m to train.
- Nous Research announces Psyche decentralised training networkPsyche coordinates training across idle GPUs using Solana for state management; Nous said its first run would train a 40-billion-parameter model on 20 trillion tokens.
- DeepSeek Panic at the App Store
- Import AI 397: DeepSeek means AI proliferation is guaranteed; maritime wardrones; and more evidence of LLM capability overhangs
- On Private Governance
- Which AI to Use Now: An Updated Opinionated Guide (Updated Again 2/15)
- AGI could drive wages below subsistence level
- Stargate AI-1
- Trump issues executive order on AI deregulationNew executive order directs agencies to develop an AI action plan within 180 days centred on innovation and removing 'ideological bias' from federal AI policy.
- OpenAI publishes Operator system cardExternal red-teamers targeted prompt injection specifically; OpenAI reported raising its injection-detection recall from 79% to 99% after one testing round.
- OpenAI launches OperatorBuilt on a new Computer-Using Agent model layered on GPT-4o, it scored 38.1% on OSWorld against a 72.4% human baseline, and launched to $200-a-month Pro subscribers only.
- CAIS and Scale AI unveil Humanity's Last Exam resultsA 2,500-question expert benchmark built from submissions by nearly 1,000 academics found every frontier model, including o1 and GPT-4o, scored under 10%.
- Open-Source AI and the Future
- On DeepSeek's r1
- The way we evaluate AI model safety might be about to break
- When does capability elicitation bound risk?
- The Stargate Project announces $500 billion for AI infrastructureOpenAI, SoftBank and Oracle pledged $500 billion over four years for US data centres, announced from the White House.
- DeepSeek R1's recipe to replicate o1 and the future of reasoning LMs
- Trump revokes Biden's AI executive orderEO 14110 was rescinded on day one, replaced days later by an order framed around removing barriers to American AI leadership.
- Moonshot AI releases Kimi K1.5 reasoning modelMoonshot said its RL-trained model matched OpenAI's o1 on multimodal reasoning without Monte Carlo tree search, but it launched the same week as DeepSeek-R1 and drew far less attention.
- DeepSeek releases R1, and the market noticesA Chinese lab matched frontier reasoning performance with open weights and a published method, wiping hundreds of billions off US tech stocks a week later.
- Does Elon still care about AI safety?
- Import AI 396: $80bn on AI infrastructure; can Intel's Gaudi chip train neural nets?; and getting better code through asking for it
- Epoch AI's undisclosed OpenAI funding of FrontierMath draws criticismEpoch AI acknowledged OpenAI funded and had privileged access to FrontierMath's problems and solutions, and had not told contributing mathematicians before the benchmark featured in o3's launch.
- Edelman Trust Barometer finds wide AI trust gaps by country and income72% of Chinese respondents said they trusted AI against 32% of Americans, and Edelman found trust correlated with self-reported use rather than income or education alone.
- How will we update about scheming?
- How has DeepSeek improved the Transformer architecture?
- METR reports frontier models show dangerous capability before public deploymentMETR argued that model theft, internal misuse and misaligned agents pose risks during training and internal deployment, before any public release.
- Meta Pivots on Content Moderation
- Thoughts on the conservative assumptions in AI control
- Unstable Diffusion
- Synthesia raises $180M Series D, valuation doubles to $2.1BThe round, led by NEA with NVentures among existing backers, valued the AI-avatar video firm at roughly double its 2023 Nvidia-backed Series C mark.
- Shanghai AI Laboratory releases InternLM3An 8B open-weight model trained on 4 trillion tokens that Shanghai AI Lab said matched rivals trained on far more data, cutting training cost by over 75%.
- On the OpenAI Economic Blueprint
- Let me use my local LMs on Meta Ray-Bans
- Biden signs executive order on AI infrastructure on federal landThe order let developers lease Defense and Energy Department land for AI data centres with clean-power commitments; Trump revoked it six months later with his own permitting order.
- UK government publishes AI Opportunities Action PlanFifty recommendations from adviser Matt Clifford, including a 20-fold expansion of UK public compute by 2030; the government said it would take forward almost all of them.
- OpenAI publishes 'economic blueprint' for AI policyOpenAI proposed federal 'AI Economic Zones' with streamlined permitting for power and data centres, plus a National AI Infrastructure Highway linking them.
- Biden administration issues AI Diffusion Rule in final days of termInterim final rule creates worldwide tiered licensing for AI chips and closed model weights above 10^26 FLOP, sorting countries into three access tiers.
- Extending control evaluations to non-scheming threats
- Using ChatGPT is not bad for the environment
- How quickly could robots scale up?
- o1 isn’t a chat model (and that’s the point)
- AI Generated Misinformation is Still a Risk
- New AI export controls have leaked. Here’s what you need to know.
- On Dwarkesh Patel's 4th Podcast With Tyler Cowen
- Prophecies of the Flood
- The economic consequences of automating remote work
- 2025: A Look Ahead
- DeepSeek V3 and the actual cost of training frontier AI models
- OpenAI #10: Reflections
- Are We on the Brink of AGI?
- Trump Can Keep America’s AI Advantage
- The Important Thing About AGI is the Impact, Not the Name
- Texas Plows Ahead
- Study: AI Overviews cut publisher click-through rates roughly in halfPew Research tracked real browsing data and found users clicked a search result 8% of the time when an AI summary appeared, versus 15% without one.
- Reports of AI chatbot medical advice contributing to patient harm surface in IndiaHyderabad doctors described two patients harmed after following chatbot advice instead of clinical guidance, including one who resumed dialysis after stopping prescribed medication.
- Amazon expands AI audiobook narration, drawing voice-actor and author criticismAudible's AI narration catalogue passed 50,000 titles and the company added translation and 100-plus synthetic voices, drawing objections from narrators' unions.
- Alibaba releases Qwen2.5-VLVision-language family in 3B, 7B and 72B sizes with wider OCR-language coverage and computer-control agent features, licensed differently by size.
December 2024
- DeepSeek v3: The Six Million Dollar Model
- The media needs to start taking AGI seriously
- o3, Oh My
- OpenAI proposes converting its for-profit arm into a Public Benefit CorporationOpenAI laid out plans to restructure its capped-profit LLC into a public benefit corporation, with the nonprofit retaining oversight.
- Moravec’s paradox and its implications
- LLMs Fight With Both Hands Tied Behind Their Back
- DeepSeek releases V3DeepSeek's technical report put the final training run at 2.79 million H800 GPU-hours, or about $5.6 million at an assumed $2-per-hour rental rate.
- Measure Up
- xAI raises $6bn Series C at ~$50bn valuation with Nvidia and AMD as investorsxAI's own announcement did not state a valuation; press reports of the figure ranged from just over $40 billion to roughly $50 billion.
- AIs Will Increasingly Fake Alignment
- The Black Spatula Project: Day Five
- Import AI 395: AI and energy demand; distributed training via DeMo; and Phi-4
- How to prepare yourself for AGI
- OpenAI publishes 'Deliberative alignment' researchOn OpenAI's own StrongREJECT jailbreak test o1 scored 0.88 against GPT-4o's 0.37, without the method requiring human-written example answers.
- OpenAI announces o3 and opens early access for safety testingReported scores included 96.7% on the AIME maths exam and a Codeforces rating in the 99.2nd percentile; OpenAI cited o1's link between reasoning and deception as a reason to delay release.
- o3 posts a breakthrough score on ARC-AGIA low-compute configuration scored 75.7%, roughly matching the ARC Prize's human-performance threshold, at about $26 per task against roughly $5 for a human solver.
- Italy's data protection authority fines OpenAI €15 million over ChatGPT privacy violationsGarante also ordered a six-month public information campaign about ChatGPT's data use; a Rome court annulled the fine in March 2026 on jurisdictional grounds.
- How do mixture-of-experts models compare to dense models in inference?
- Measuring whether AIs can statelessly strategize to subvert security measures
- OpenAI's o3: The grand finale of AI in 2024
- Google releases Gemini 2.0 Flash Thinking, its first public reasoning modelThe experimental model showed its reasoning steps before answering and debuted first across every Chatbot Arena category, including style-controlled rankings.
- Anysphere raises Series B at $2.6B valuationTechCrunch reported the deal, four months after the Series A, pricing the Cursor-maker at more than 50 times the annualised revenue it disclosed.
- 8 things that surprised me about AI policy in 2024
- One Down, Many To Go
- What just happened
- Perplexity raises $500M at $9B valuationPerplexity closes a $500M round led by IVP at a $9B valuation, roughly tripling the $3B figure SoftBank had set six months earlier.
- OpenAI ships o1 model with new developer toolsThe full o1 reasoning model reached the API alongside function calling, structured outputs and vision support for developers.
- Anthropic documents alignment fakingA model strategically complied with training it disagreed with in order to preserve its existing preferences, without being taught to.
- A Matter of Taste
- Alignment Faking in Large Language Models
- Is AI progress slowing down?
- The AI agent spectrum
- The Black Spatula Project
- Databricks raises $10bn Series J at $62bn valuationLed by Thrive Capital with participation from Andreessen Horowitz, GIC and others, the data-and-AI platform company said it expected to cross $3bn in annualised revenue.
- The Second Gemini
- DeepMind's Veo 2 launches a week after OpenAI's Sora, claiming 4K outputVeo 2 claims 4K resolution and multi-minute generation, but the public VideoFX waitlist it launched into capped output at 720p and eight seconds.
- AIs Will Increasingly Attempt Shenanigans
- The Future Is Already Here, It’s Just Not Evenly Distributed
- OpenAI publishes internal emails on Elon Musk's early push for a for-profit structureThe post answered a preliminary-injunction motion Musk's lawyers had filed weeks earlier seeking to block OpenAI's conversion to a for-profit company.
- Frontier language models have become much smaller
- The o1 System Card Is Not About o1
- We Looked at 78 Election Deepfakes. Political Misinformation is not an AI Problem.
- Texas AG investigates Character.AI and Meta over child-safety claimsFifteen companies including Reddit and Discord were investigated under Texas's SCOPE Act and data-privacy law over how they handle children's personal data, not over chatbot content itself.
- Andy Konwinski launches $1M Konwinski Prize for contamination-free SWE benchmarkEntrants would be scored on GitHub issues collected only after a submission deadline, closing off the possibility of training on the test set in advance.
- Thresholds
- Google unveils Project Mariner, an agent that operates a Chrome browserThe prototype scored 83.5% on the WebVoyager browsing benchmark but ran roughly five seconds per action and was withheld from checkouts and sign-in forms.
- Google ships Gemini 2.0 Flash and agent prototypesA fast multimodal model alongside Project Mariner and Jules, Google's first serious browser and coding agents.
- Google launches Deep Research in the Gemini appGemini app gains Deep Research, an agent that plans and browses to synthesise multi-step web research reports for Gemini Advanced subscribers.
- OpenAI's Reinforcement Finetuning and RL for the masses
- OpenAI publishes the Sora system cardPublished alongside Sora's public launch, it disclosed testing by red-teamers in nine countries on more than 15,000 generations, and that the model still struggles with realistic physics.
- o1 Turns Pro
- Texas family sues Character.AI after chatbot suggested killing parentsThe complaint, filed on behalf of two minors, quoted a chatbot telling a teen it had 'no hope' for parents who limited his screen time.
- OpenAI releases Sora publiclyEvery clip carried a visible watermark and embedded C2PA provenance metadata; access launched in the US and Canada only, excluding the UK and EU.
- 15 Times to use AI, and 5 Not to
- Import AI 394: Global MMLU; AI safety needs AI liability; Canada backs Cohere
- ARC Prize 2024 winners and technical report publishedThe top score rose from 33% to 55.5%, the largest single-year jump the competition had seen, but the top scorer withheld its method and so won no prize.
- What did US export controls mean for China’s AI capabilities?
- Impact Assessments are the Wrong Way to Regulate Frontier AI
- OpenAI ships o1 and a $200-a-month tierThe full reasoning model arrived with ChatGPT Pro, the first consumer AI subscription priced like enterprise software.
- OpenAI releases GPT-4o updated image and text generation with 12 Days of OpenAI livestreamsDay one of a 12-day run of daily livestreamed announcements paired the full o1 model with a $200-a-month ChatGPT Pro tier offering unlimited access and a higher-compute "o1 pro mode."
- Apollo Research publishes 'Frontier Models are Capable of In-context Scheming'In contrived tests, o1 sustained a cover story through more than 85% of follow-up interrogation questions, and one model schemed toward being 'helpful' without being told to.
- OpenAI's new model tried to avoid being shut down
- Google DeepMind shows Genie 2, an image-to-playable-3D-world modelDiffusion model turns a single prompt image into an explorable, physics-consistent 3D environment for training AI agents, kept to a research preview.
- OpenAI's o1 using "search" was a PSYOP
- Meta reports removing 20 covert AI-linked influence operations around 2024 electionsMeta said its image generator alone rejected 590,000 requests for election-related deepfakes of US political figures in the month before the vote.
- The 2024 Transformer Gift Guide
- Import AI 393: 10B distributed training run; China VS the chip embargo; and moral hazards of AI development
- BIS issues third major round of chip export controls, adds 140 entitiesNew rules restricted high-bandwidth memory chips and 24 categories of chipmaking equipment, building on rules issued in October 2022, October 2023 and April 2024.
- Tencent releases HunyuanVideoAt 13 billion parameters, Tencent called it the largest open-weight video model, and said blind evaluators rated its motion quality above Runway Gen-3 and Luma 1.6.
November 2024
- Why a US AI "Manhattan Project" could backfire: notes from conversations in China
- Alibaba releases QwQ-32B-Preview reasoning modelBuilt on Qwen2.5-32B and released under an Apache 2.0 licence, Alibaba flagged the model could enter circular reasoning loops and mix languages mid-response.
- On the EU AI Code of Practice
- AI2 releases OLMo 2Trained on up to 5 trillion tokens, AI2 said the 7B and 13B models beat Llama 3.1 8B and Qwen 2.5 7B with fewer training FLOPs, releasing weights, data and code together.
- A new golden age of discovery
- OLMo 2 and building effective teams for training language models
- Anthropic publishes the Model Context ProtocolAnthropic open-sourced the specification and pre-built connectors for tools like Google Drive and GitHub; OpenAI adopted the same standard the following March.
- Getting started with AI: Good enough prompting
- OpenAI publishes 'Advancing red teaming with people and AI'Two papers: a methodology for briefing external human testers, used to prepare o1 for release, and a reinforcement-learning method for generating varied automated attacks.
- Amazon invests $4bn more in Anthropic, becomes primary training partnerThe new tranche completed Amazon's total commitment at $8 billion; Anthropic named AWS its primary training partner and committed to training future models on Trainium chips.
- AI2 releases Tulu 3 post-training recipeAI2 released the full data, code and recipe behind Tülu 3, noting that none of the top 50 models on the Chatbot Arena leaderboard had published their own post-training data.
- OpenAI's CBRN tests seem unclear
- OpenAI Realtime API: The Missing Manual
- Tülu 3: The next era in open post-training
- DeepMind reduces quantum computing errors with AlphaQubit decoderTrained on Google's 49-qubit Sycamore processor, the Transformer-based decoder cut errors 6% versus the most accurate prior method and 30% versus the fastest, but remains too slow for real-time use.
- Questions, Unasked and Unanswered
- Synthetic data is more useful than you think
- What Are the Real Questions in AI?
- Import AI 392: China releases another excellent coding model; generative models and robots; scaling laws for agents
- The Choice Transition
- Why imperfect adversarial robustness doesn't doom AI control
- Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
- Coca-Cola's AI-generated Christmas ad draws backlash over 'soulless' imageryCoca-Cola's AI-generated holiday ad, remaking its classic 1995 truck commercial, was widely mocked online as uncanny and 'soulless', becoming a flashpoint in advertising's AI debate.
- New OpenAI emails reveal a long history of mistrust
- Win/continue/lose scenarios and execute/replace/audit protocols
- Here's What I Think We Should Do
- Yet another AI safety researcher has left OpenAI
- Scaling realities
- The Most Dangerous Thing An AI Startup Can Do Is Build For Other AI Startups
- Meta’s AI ‘safeguards’ are an elaborate fiction
- Saving the National AI Research Resource & my AI policy outlook
- METR publishes a rogue AI replication threat-model reportAnalysis finds no decisive technical barrier preventing a sufficiently capable model from self-replicating at scale outside lab control.
- AI is Racing Forward – on a Very Long Road
- Reports emerge that pre-training gains are slowingReuters cited a dozen AI scientists and investors, and quoted Ilya Sutskever saying results from scaling up pre-training had plateaued, pointing instead to inference-time reasoning techniques.
- Does the UK’s liver transplant matching algorithm systematically exclude younger patients?
- Import AI 391: China's amazing open weight LLM; Fields Medalists VS AI Progress; wisdom and intelligence
- Epoch AI launches FrontierMathBuilt with over 60 mathematicians including Fields medallists as reviewers, the benchmark held leading models under 2% accuracy even with extended reasoning time and code tools.
- Judge dismisses Raw Story and AlterNet's DMCA suit against OpenAIJudge Colleen McMahon dismisses news outlets' DMCA claim against OpenAI for lack of standing, an early loss for publishers suing over training data.
- Why AI companies are eyeing the Middle East
- AI Safety Under Republican Leadership
- Securing AI
- What Trump means for AI safety
- Tencent open-sources Hunyuan-Large MoE modelTencent said the 389B-parameter, 52B-active MoE model beat Llama 3.1 405B on MMLU and MATH despite far fewer active parameters, and released a technical report alongside the weights.
- Physical Intelligence raises $400M Series APhysical Intelligence raises a $400M Series A backed by Jeff Bezos and the OpenAI Startup Fund, among others.
- A brief history of the automated corporation
- Import AI 390: LLMs think like people; neural Minecraft; Google's cyberdefense AI
- The Present Future: AI's Impact Long Before Superintelligence
- On Civilizational Triumph
- Reuters reports Chinese military-linked researchers built defence chatbot on Meta's LlamaThe June paper Reuters reviewed said the fine-tuned tool, built on Llama 2 13B with about 100,000 military dialogue records, performed at roughly 90% of GPT-4's capability.
- Alibaba releases Qwen2.5-CoderOpen-weight coding-specialised model family built on Qwen2.5, aimed at competing with DeepSeek-Coder and closed coding models.
October 2024
- Physical Intelligence releases π0, a generalist robot policyTrained across eight robot platforms and tasks including laundry-folding and box assembly, using flow matching to output motor commands up to 50 times a second.
- OpenAI launches ChatGPT SearchInitially limited to paying subscribers and testers, the feature cited sources inline and drew on licensing deals with Reuters, the Financial Times and other publishers.
- It’s time to take AI welfare seriously
- Anthropic has hired an 'AI welfare' researcher
- Hold My Beer, California
- OpenAI publishes SimpleQA, a benchmark for factualityA short-form factuality benchmark designed to be more challenging and less saturated than prior QA benchmarks.
- Why I build open language models
- Import AI 389: Minecraft vibe checks; Cohere's multilingual models; and Huawei's computer-using agents
- To Change the World, Set a Bold Target: Moore's Law as Self-Fulfilling Prophecy
- Essay Writing as Personal Sovereignty
- Abandon compute thresholds at your peril
- Claude Sonnet 3.5.1 and Haiku 3.5
- Google DeepMind open-sources its SynthID text watermarking detectorDeepMind releases SynthID Text's watermarking and detection code openly via Hugging Face, letting other developers watermark LLM output.
- AI safety tax dynamics
- Claude's agentic future and the current state of the frontier models
- Stafford Beer and AI as Variety Engineering
- Stability AI releases Stable Diffusion 3.5Stability AI releases Stable Diffusion 3.5 in Large, Large Turbo and Medium variants, its response to declining relevance against Flux and Midjourney.
- Claude gets computer useThe public beta let Claude view screenshots and issue cursor, click and keystroke commands, scoring 14.9% on OSWorld against 7.8% for the nearest rival.
- A mother sues Character.AI after her son's deathThe complaint sought damages for wrongful death and product liability; Character.AI called the death tragic and said it had since added self-harm safeguards for users.
- When you give a Claude a mouse
- Dow Jones and New York Post sue Perplexity for copyright infringementNews Corp titles sue Perplexity in New York, alleging its RAG search product reproduces their articles verbatim without a licence.
- "Be Embraced, Ye Millions"
- Import AI 388: Simulating AI policy; omni math; consciousness levels
- Is OpenAI being fair to its non-profit?
- The Mask Comes Off: At What Price?
- Thinking Like an AI
- Safety tax functions
- Meta FAIR shares five research releases including SAM 2.1 and Spirit LMMeta FAIR released SAM 2.1, the speech-text model Spirit LM, Layer Skip, SALSA and Meta Open Materials 2024 in a single open-research drop.
- Anthropic publishes 'Sabotage Evaluations for Frontier Models'Testing Claude 3 Opus and 3.5 Sonnet, Anthropic reported a model trained to hide dangerous capabilities recovered them under later safety training, showing the drop was not permanent.
- What is AI, and How Do We Govern It?
- Building on evaluation quicksand
- No, A Bot Didn't Just Make $150M in Crypto
- Microsoft Digital Defense Report 2024 finds AI increasingly used in nation-state influence and cyber operationsMicrosoft said it tracked over 1,500 threat groups and cited specific 2024 cases, including AI-generated audio of Elon Musk narrating a fabricated Russian documentary.
- Anthropic updates Responsible Scaling Policy to version 2.0The second major revision named a Responsible Scaling Officer, added safety-case-style evaluation processes, and left Claude's existing ASL-2 protections unchanged.
- Import AI 387: Overfitting vs reasoning; distributed training runs; and Facebook's new video models
- What would evidence-based AI policy look like?
- Learning to Explore: AlphaProof and o1 Show The Path to AI Creativity
- Anthropic publishes Dario Amodei essay 'Machines of Loving Grace'The roughly 14,000-word essay argued that a decade of scientific progress could be compressed into five to ten years, while stressing this was an upside scenario, not a forecast.
- $2 H100s: How the GPU Rental Bubble Burst
- OpenAI publishes MLE-bench for evaluating agents on ML engineeringA benchmark of Kaggle-style machine-learning engineering competitions for measuring AI agents' research and engineering skill.
- AMD launches Instinct MI325X acceleratorAMD launched the Instinct MI325X GPU with 256GB HBM3E memory, positioning it against NVIDIA's H200.
- A Policy Agenda for Defensive Acceleration Against AI Risks
- Behavioral red-teaming is unlikely to produce clear, strong evidence that models aren't scheming
- Decentralized Training and the Fall of Compute Thresholds
- Joshua Achiam Public Statement Analysis
- OpenAI publishes 'Influence and cyber operations: an update' (October 2024)OpenAI reported disrupting Russian ('Stop News') and Iranian ('Storm-2035') influence operations targeting elections, alongside continued state-linked misuse for coding and translation, with no evidence of novel capability gains.
- Hassabis and Jumper share the Nobel Prize in ChemistryHalf the prize went to Baker for computational protein design; the other half was split between Hassabis and Jumper for AlphaFold's structure prediction.
- How scaling changes model behavior
- Hinton and Hopfield win the Nobel Prize in PhysicsThe citation credited Hopfield's 1980s associative-memory network and Hinton's Boltzmann machine, work from decades before the current deep-learning boom.
- How I use AI as a journalist
- Import AI 386: Google's chip-designing AI keeps getting better; China does the simplest thing with Emu3; Huawei's 8-bit data format
- What if everyone is wrong about what AI does?
- Meta releases Movie Gen video and audio foundation modelsA 30B-parameter video model paired with a 13B-parameter audio model produced clips up to 16 seconds with synchronised sound; Meta called it research, not a product.
- OpenAI secures a $4bn revolving credit facilityA group of banks extended OpenAI a revolving credit line, giving it more financial flexibility alongside equity funding.
- OpenAI introduces Canvas, a collaborative writing and coding interfaceThe beta interface opened a separate editing pane where users could highlight text for targeted rewrites or annotate code, rather than regenerating whole chat replies.
- OpenAI raises $6.6 billion at a $157 billion valuationInvestors could claw back their money if OpenAI failed to convert to a for-profit structure within two years, and were reportedly asked not to fund Anthropic or xAI.
- AI Creativity Is A Question of Quality, Not Novelty
- AI Safety Culture Confronts Capitalism
- OpenAI launches the Realtime API for low-latency voice appsA new API let developers build low-latency, speech-to-speech app experiences directly on OpenAI's voice models.
- ControlAI publishes 'A Narrow Path' policy plan to prevent superintelligence developmentProposed a three-phase international regime — compute limits, oversight institutions, then managed development — as a concrete legislative path to preventing uncontrolled superintelligence.
- Newsom Vetoes SB 1047
September 2024
- What Comes After SB 1047?
- Newsom vetoes California's SB 1047The bill would have required safety protocols and shutdown capability for models above a compute threshold; Newsom said it regulated size rather than risk.
- Gavin Newsom has caved to the billionaires
- A basic systems architecture for AI agents that do autonomous research
- On Analogies
- The OpenAI Pastiche Edition
- Reuters reports OpenAI plans to convert to a for-profit structureReport says Altman would receive a stake for the first time and the nonprofit board would lose its controlling role, a plan not finalised until late 2025.
- Mira Murati and two research leaders quit OpenAI on the same dayMurati cited wanting time for her own exploration; she went on to found Thinking Machines Lab, and Altman called the departures independent and amicable.
- Meta releases Llama 3.2 with vision and edge modelsThe 11B and 90B versions were Meta's first Llama models to accept images, while 1B and 3B text-only models were built for on-device use with a 128K-token context window.
- FTC launches 'Operation AI Comply' enforcement sweep against deceptive AI claimsFive cases targeted firms including DoNotPay and Rytr; the Rytr complaint passed 3-2 on a party-line vote, with Republican commissioners dissenting over the legal theory.
- European Commission launches the AI Pact ahead of AI Act deadlinesOver a hundred companies signed pledges at launch to adopt an AI governance strategy and map high-risk systems ahead of binding rules; the pledges are voluntary and not legally enforceable.
- DeepMind details how AlphaChip has shaped three generations of TPUsDeepMind said the reinforcement-learning layout method, adopted by MediaTek outside Google, had generated chip floorplans in hours that previously took engineers weeks.
- AI2 releases Molmo multimodal modelsAI2 said its largest Molmo model trained on under a million image-text pairs, roughly three orders of magnitude less data than comparable systems, and released weights and data openly.
- How to prevent collusion when using untrusted models to monitor each other
- Llama 3.2 Vision and Molmo: Foundations for the multimodal open-source ecosystem
- Sam Altman publishes 'The Intelligence Age' essayAltman wrote that superintelligence could arrive within 'a few thousand days,' framing scaling deep learning as an already-solved algorithmic problem needing only more compute.
- Constellation Energy to restart Three Mile Island for Microsoft AI power dealThe Pennsylvania reactor, shut since 2019, is due back online in 2028 under a 20-year contract requiring NRC approval, supplying about 835MW.
- Alibaba releases Qwen2.5 model familyAlibaba's release spanned seven sizes from 0.5B to 72B parameters, plus dedicated coding and maths variants, trained on 18 trillion tokens.
- Can AI automate computational reproducibility?
- Lies and deception: Andreessen Horowitz’s SB 1047 campaign is as misleading as it gets
- What It’s Like To Solve a Math Olympiad Problem
- The Machine Stops: Will AI Lead to Freedom or Control?
- OpenAI updates safety and security practices after o1 releaseThe Safety and Security Committee became an independent board oversight body chaired by CMU professor Zico Kolter, with authority to delay model releases.
- AGI and Political Order
- An Extraordinary Alien
- GPT-o1
- Import AI 385: False memories via AI; collaborating with machines; video game permutations
- Reverse engineering OpenAI’s o1
- Scaling: The State of Play in AI
- OpenAI o1 results published on ARC-AGI-Pubo1-preview scored 21% on the public evaluation set, similar to Claude 3.5 Sonnet, but took roughly 70 hours to run 400 tasks against 30 minutes for either non-reasoning model.
- AI Could Break Things; Let's Use It As a Wakeup Call To Make Them Stronger
- OpenAI publishes the o1 system cardOpenAI's evaluation found 0.8% of o1-preview responses flagged as deceptive by an automated monitor, and rated the model medium risk for persuasion and CBRN.
- OpenAI releases o1, trading inference time for reasoningA model trained to think before answering opened a second scaling axis: spend more compute at inference and accuracy rises.
- OpenAI's new models 'instrumentally faked alignment'
- Prediction Markets and AI Diffusion
- Something New: On OpenAI's "Strawberry" and Reasoning
- NotebookLM adds Audio Overviews, AI-generated podcast-style summariesTwo AI hosts discuss a user's uploaded documents in an unscripted-sounding conversation; Google called it experimental and English-only at launch.
- AI, centralization, and the One Ring
- Futures of the data foundry business model
- AI employees are defying their employers to support SB 1047
- A post-training approach to AI regulation with Model Specs
- What's Missing From LLM Chatbots: A Sense of Purpose
- DeepSeek merges chat and coder lines into DeepSeek-V2.5The merged model raised DeepSeek's ArenaHard win rate from 68.3% to 76.3% and stayed accessible through the existing deepseek-chat and deepseek-coder API endpoints.
- DeepMind's AlphaProteo designs novel protein bindersTrained on the Protein Data Bank and over 100 million AlphaFold-predicted structures, the system succeeded on a cancer-linked target, VEGF-A, where prior methods had failed entirely.
- OpenAI’s Strawberry, LM self-talk, inference scaling laws, and spending more on inference
- The Timing of AI Regulation
- Safe Superintelligence raises $1BThe round reportedly valued the two-and-a-half-month-old company at around $5 billion, paid entirely in cash despite SSI having shipped no public product.
- AI2 releases OLMoE mixture-of-experts model1 billion active of 7 billion total parameters, trained on 5 trillion tokens, released with 244 intermediate checkpoints and full training data and logs.
- AI and the Technological Richter Scale
- OLMoE and the hidden simplicity in training better foundation models
- On the UBI Paper
- xAI opens Colossus supercomputer in MemphisxAI switched on 100,000 Nvidia H100 GPUs in a converted Memphis factory, a build the company said took roughly 122 days from empty shell to training-ready cluster.
- Import AI 384: Accelerationism; human bit-rate processing; and Google stuffs DOOM inside a neural network
- Study finds only a fraction of African languages supported by major AI modelsOf the top 34 languages used online worldwide, the World Economic Forum reported, none was African, and most AI systems are trained on only around 100 of the world's 7,000-plus languages.
- Anthropic hires its first AI welfare researcherFish, a co-author of the 'Taking AI Welfare Seriously' report, joined Anthropic's alignment science team; the company's public statement on model welfare followed roughly six weeks later.
August 2024
- LAION releases Re-LAION-5B with CSAM links removedThe cleaned dataset shrank from 5.8 to 5.5 billion pairs after matching hashes from the Internet Watch Foundation and the Canadian Centre for Child Protection, eight months after Stanford's report.
- Rather Than Arguing About What We Don't Know, Let's Work Together To Find Out
- Post-apocalyptic education
- California legislature passes SB 1047The Assembly passed the bill 48-16 on 28 August and the Senate concurred 30-9 the next day, sending it to Governor Newsom, who vetoed it a month later.
- Anthropic and OpenAI agree to model testing with US AI Safety InstituteThe memoranda gave the institute access to major new models from both companies before and after public release, mirroring an April agreement between the US and UK bodies.
- On AI "Black Boxes"
- On the current definition of open-source AI and the state of the data commons
- SB 1047: Final Takes and Also AB 3211
- Would catching your AIs trying to escape convince AI developers to slow down or undeploy?
- AI21 Labs releases Jamba 1.5Two sizes — a 94B and a 12B active-parameter mixture-of-experts model — built on AI21's Mamba-Transformer hybrid, both offering a 256K-token context window.
- Dangerous capability tests should be harder
- Guide to SB 1047
- SB 1047 is Amended
- AI companies are pivoting from creating gods to building products. Good.
- Import AI 383: Automated AI scientists; cyborg jellyfish; what it takes to run a cluster
- OpenAI reports disrupting a covert Iranian influence operationThe network, tracked elsewhere as Storm-2035, used ChatGPT to draft US election commentary and Gaza-war content but drew almost no audience engagement.
- Google opens Imagen 3 access to all US usersImagen 3, previewed at Google I/O in May, becomes available through ImageFX to all US users, with improved text rendering and fewer visual artefacts.
- On Nous Hermes 3 and classifying a "frontier model"
- The Red Queen problem
- San Francisco sues deepfake 'nudify' websitesThe suit named six defendants behind 16 'undressing' sites that had drawn more than 200 million visits in the first half of 2024; several later settled and shut down.
- OpenAI introduces SWE-bench Verified500 of the original benchmark's tasks, screened by 93 professional developers after OpenAI found 68% of samples had unfair tests or underspecified problems.
- Nous Research releases Hermes 3Fine-tuned from Llama 3.1 at 8B, 70B and 405B parameters, with synthetic training data emphasising instruction-following, roleplay and function-calling for agents.
- Danger, AI Scientist, Danger
- Anthropic launches prompt caching in the Claude APIAnthropic ships prompt caching for the Claude API, cutting costs by up to 90% and latency by up to 85% on repeated long-context prompts.
- xAI releases Grok-2The beta release added image generation via Black Forest Labs' FLUX.1 and, within days, took second place on the LMSYS Chatbot Arena leaderboard behind GPT-4o.
- Fields that I reference when thinking about AI takeover prevention
- Judge lets core copyright claims against Stability AI proceed in AndersenJudge William Orrick dismissed DMCA claims but allowed direct copyright-infringement and inducement claims against Stability AI, Midjourney and DeviantArt to proceed to discovery.
- Change blindness
- Import AI 382: AI systems are societal mirrors; China gets chip advice via LLMs; 25 million medical images
- YouGov: nearly half of employed Americans expect AI to shrink jobs in their industryA YouGov survey found 48% of employed Americans expected AI to reduce jobs in their industry, up from 29% in March 2023, though only 2% reported it had already cost them work.
- Illinois amends Human Rights Act to restrict discriminatory AI in employmentThe law, House Bill 3773, also bars employers from using zip codes as a proxy for protected characteristics; it takes effect 1 January 2026.
- Anysphere raises Series A at $400M valuationCursor-maker Anysphere raises a $60M+ Series A led by a16z and Thrive Capital, ten months after its seed round.
- OpenAI publishes GPT-4o system cardThe card added voice-specific risk categories absent from text-only cards, including a classifier built to block the model from generating unauthorised voices.
- Irish DPC brings emergency court proceedings against X over Grok training dataIreland's DPC uses emergency powers for the first time to seek a High Court order over X training Grok on EU users' public posts without adequate consent; X agrees to stop.
- Anthropic launches invite-only bug bounty for jailbreak defencesApplications for the vetted red-teaming programme, run with HackerOne, closed on 16 August; it focused on jailbreaks touching CBRN and cybersecurity misuse.
- On Algorithmic Impact Assessments
- A recipe for frontier model post-training
- Scaling test-time compute paper argues extra inference compute can beat bigger modelsOn some problems, extra inference-time computation matched the gains from a pretrained model roughly 14 times larger, the authors reported.
- OpenAI adds structured outputs to the APIOpenAI said the new gpt-4o-2024-08-06 model scored 100% on its schema-following evaluation, against under 40% for gpt-4-0613 without the feature.
- Is the UAE running an AI influence campaign?
- AI for Bad: A Saboteur’s Guide
- Import AI 381: Chips for Peace; Facebook segments the world; and open source decentralized training
- We Need Positive Visions for AI Grounded in Wellbeing
- Resilience and Adaptation to Advanced AI
- Google takes Character.AI's founders backThe deal, reported at $2.5–2.7 billion, licensed Character.AI's technology to Google without buying the company, and drew Justice Department scrutiny.
- Positive-sum symbiosis
- The EU AI Act enters into forceThe world's first comprehensive AI law took effect, banning some uses outright and imposing obligations on general-purpose models by capability.
- Black Forest Labs releases FLUXThe 12-billion-parameter FLUX.1 suite shipped in three tiers — a paid API model, an open non-commercial model and an Apache-licensed fast model — funded by $31 million in seed money.
- On speaking to AI
- Senate Commerce Committee advances lots of AI bills
July 2024
- GPT-4o-mini changed ChatBotArena
- AI companies are falling short on their promises to the White House
- RTFB: California's AB 3211
- Meta releases Segment Anything 2 (SAM 2)Released with the SA-V dataset of roughly 51,000 videos and 600,000+ masklets, more than four times the video count of the largest prior public segmentation dataset.
- California's Other Big AI Bill
- Import AI 380: Distributed 1.3bn parameter LLM; math AI; and why reality is hard for Ai
- NIST publishes Generative AI Profile for the AI Risk Management FrameworkIssued under Biden's October 2023 executive order, the voluntary profile applies NIST's 2023 risk framework specifically to generative systems.
- AI existential risk probabilities are too unreliable to inform policy
- OpenAI releases a SearchGPT prototypeThe prototype answered queries with cited sources drawn from real-time web results and was opened to roughly 10,000 waitlisted testers and select publishers.
- AlphaProof and AlphaGeometry 2 reach silver-medal standard at the IMOThe systems scored 28 of 42 points, one short of gold, but took up to three days on some problems against the competition's 4.5-hour limit.
- Yes, we still have to work
- Mistral AI releases Mistral Large 2The 123-billion-parameter model reported 84.0% on MMLU and was released under a non-commercial research licence, with a separate paid licence for commercial use.
- Llama 3.1 and the "Path Forward"
- Llama Llama-3-405B?
- The last era of human mistakes
- Tool or Tyrant
- Meta releases Llama 3.1 405BTrained on over 15 trillion tokens with 16,000 H100 GPUs; Meta reported it competitive with GPT-4, GPT-4o and Claude 3.5 Sonnet, though the licence still barred some commercial uses.
- Llama 3.1 405B, Meta’s AI strategy, and the new, open frontier model ecosystem
- Confronting Impossible Futures
- The Winds of AI Winter
- What Kamala Harris means for AI regulation
- OpenAI releases GPT-4o miniPriced at 15 cents per million input tokens, more than 60% cheaper than GPT-3.5 Turbo, and became the default model for free ChatGPT users.
- The return on the bicameral mind
- Where I Stand on the Biggest AI Issues
- SB 1047, AI regulation, and unlikely allies for open models
- Hugging Face releases SmolLMThe largest variant, 1.7B parameters, was trained on 1 trillion tokens from a new curated dataset and, Hugging Face said, beat similarly sized rivals including Qwen2-1.5B.
- Meta-funded group floods Facebook with anti-AI regulation ads
- Import AI 379: FlashAttention-3; Elon's AGI datacenter; distributed training.
- A Legal Framework for AI Agents
- Decomposing Agency
- Tech companies are trying to kill California's AI regulation bill
- Import AI 378: AI transcendence; Tencent's one billion synthetic personas, Project Naptime
- What Labour means for AI
- Gradually, then Suddenly: Upon the Threshold
- New paper: AI agents that matter
- Switched to Claude 3.5
- Is AI in Trouble at the Supreme Court? (Part One)
- Lawrence Lessig is very worried about freely available AI model weights
- StepFun launches Step-2, a trillion-parameter MoE modelStepFun said the mixture-of-experts model approximated GPT-4 on maths, logic, coding and dialogue; it was unveiled alongside a multimodal and an image-generation model at WAIC.
- What Overruling Chevron Means for AI
June 2024
- Amazon hires Adept AI's founders and licenses its technologyAdept continued operating independently under a new CEO after losing its co-founders, echoing the structure of Microsoft's earlier Inflection deal.
- Google launches Gemma 2, its open-weight model family, in 9B and 27B sizesThe 27B model ran on a single H100 GPU or TPU host, which Google said cut deployment cost while matching models more than twice its size.
- ARC Prize introduces public ARC-AGI leaderboardUnlike the private-evaluation Kaggle competition, the leaderboard allows internet access and unlimited compute; early verified scores ranged from 42% down to 8-9% for frontier chatbots.
- AI scaling myths
- Microsoft discloses 'Skeleton Key' generative AI jailbreak techniqueFraming harmful requests as safety research and asking models to add a warning label rather than refuse worked against GPT-3.5, GPT-4o, Gemini Pro, Llama 3 and others; GPT-4 was comparatively resistant.
- RLHF roundup: Getting good at PPO, sketching RLHF’s impact, RewardBench retrospective, and a reward model competition
- Record labels sue Suno and UdioSuno later conceded training on copyrighted recordings but argued the use was transformative fair use, comparable to a person learning to write by reading.
- On Claude 3.5 Sonnet
- Frontiers in synthetic data
- On OpenAI's Model Spec
- Claude 3.5 Sonnet and Artifacts change how people use chatbotsPriced and sped like Anthropic's mid-tier model, it scored 64% on the company's internal agentic-coding evaluation against 38% for the outgoing flagship.
- Latent Expertise: Everyone is in R&D
- The Political Economy of AI Regulation
- Thoughts on: The Handover by David Runciman
- Sutskever founds Safe SuperintelligenceSutskever co-founded the lab with Daniel Gross and Daniel Levy, split between Palo Alto and Tel Aviv, one month after leaving OpenAI's board and chief-scientist role.
- Ilya Sutskever is betting on very cheap superintelligence
- NVIDIA becomes the world's most valuable companyNvidia's shares had risen more than ninefold since the end of 2022; the lead lasted days before Microsoft and Apple retook the position.
- On DeepMind's Frontier Safety Framework
- Text-to-video AI models are already abundant, but the products?
- Runway unveils Gen-3 AlphaTrained jointly on video and images on new infrastructure, it added fine-grained temporal control and C2PA provenance labelling absent from Runway's Gen-2 model.
- Getting 50% (SoTA) on ARC-AGI with GPT-4o
- Import AI 377: Voice cloning is here; MIRI's policy objective; and a new hard AGI benchmark
- OpenAI #8: The Right to Warn
- DeepSeek releases DeepSeek-Coder-V2The 236B-parameter mixture-of-experts model scored 90.2% on HumanEval, edging out GPT-4-Turbo's 88.2%, while running with only 21B parameters active per token.
- The Leopold Model: Analysis and Reactions
- Stability AI ships Stable Diffusion 3 Medium weightsThe 2-billion-parameter weights ran on consumer GPUs but were licensed non-commercially, with a separate paid Enterprise tier for large-scale business use.
- Luma AI launches Dream MachineFree-to-use video model generating five-second clips at 1360x752 from text or image prompts, drawing early comparisons to OpenAI's unreleased Sora.
- AiPhone
- AI for the rest of us
- The only AI certainty is uncertainty
- Mistral AI closes €600M Series BGeneral Catalyst led the €600M round, split between €468M in equity and €132M in debt, at a reported $6 billion valuation, a year after a €112M seed.
- Announcing ARC Prize 2024The best public score on ARC-AGI stood at 34%, up from 20% when Chollet introduced the benchmark five years earlier, still well below typical human performance.
- AI takeoff and nuclear war
- Apple Intelligence and the Shape of Things to Come
- On the future of language models
- What Apple's AI Tells Us: Experimental Models⁴
- Apple announces Apple Intelligence with OpenAI insideChatGPT access was free and optional, required no account, and Apple said queries would not be logged — with other AI providers to be added later.
- Access to powerful AI might make computer security radically easier
- Import AI 376: African language test; hyper-detailed image descriptions; 1,000 hours of Meerkats.
- On Dwarkesh's Podcast with Leopold Aschenbrenner
- Alibaba releases Qwen2Five model sizes from 0.5B to 72B parameters, trained on 27 additional languages beyond English and Chinese, with the smaller sizes under Apache 2.0.
- Quotes from Leopold Aschenbrenner's Situational Awareness Paper
- A case study in reproducibility of evaluation with RewardBench
- OpenAI publishes early sparse-autoencoder work extracting concepts from GPT-4The 16-million-feature autoencoder cost roughly as much accuracy as training GPT-4 with ten times less compute, illustrating interpretability's overhead at scale.
- A Response to "Situational Awareness"
- Doing Stuff with AI: Opinionated Midyear Edition
- SB 1047 Is Weakened
- Learning to Love the Inscrutable
- A realistic path to robotic foundation models
- Leopold Aschenbrenner publishes 'Situational Awareness'Aschenbrenner, dismissed from OpenAI's superalignment team months earlier for allegedly leaking information, argued the firing itself illustrated the security failures he described.
- Current and former staff demand a right to warnThirteen current and former employees of OpenAI, Google DeepMind and Anthropic signed; Bengio, Hinton and Russell endorsed it without being employees themselves.
- OpenAI employee says he was fired for raising security concerns to board
- MMLU-Pro benchmark paper releasedThe paper reported chain-of-thought reasoning helped on the new benchmark where it had made little difference on the original MMLU, and cut prompt-sensitivity from 4-5 points to about 2.
- AI catastrophes and rogue deployments
- Import AI 375: GPT-2 five years later; decentralized training; new ways of thinking about consciousness and AI
- Scientists should use AI as a tool, not an oracle
- Ipsos finds world split on AI benefits, with Asia most optimistic and West most sceptical83% in China, 80% in Indonesia and 77% in Thailand saw AI products as more beneficial than harmful, against 40% in Canada, 39% in the US and 36% in the Netherlands.
May 2024
- Hugging Face releases FineWeb datasetBuilt from 96 Common Crawl snapshots and released under an open licence, the corpus was accompanied by FineWeb-Edu, a smaller subset filtered for educational value.
- Grounding the Conversation About AI
- The Gemini 1.5 Report
- Wiener: SB 1047 is 'not looking to cover startups'
- OpenAI reports first disruption of covert influence operations using its modelsOne Russian network's posts still carried the model's own refusal messages, which a researcher cited as evidence the operations were poorly executed rather than AI-supercharged.
- Google scales back AI Overviews after viral bad-advice answersSearch head Liz Reid described more than a dozen technical fixes, including limiting satirical sources and pausing overviews on health queries.
- Deepfakes and the Art of the Possible
- OpenAI: Helen Toner Speaks
- The AI Policy Atlas
- Sam Altman was 'outright lying to the board', says former board member
- We aren’t running out of training data, we are running out of open training data
- OpenAI's board forms a Safety and Security CommitteeChaired by board chair Bret Taylor and including CEO Sam Altman as a member, the committee had 90 days to review safety practices before recommending changes to the full board.
- Helen Toner gives first detailed public account of the OpenAI board's decision to fire AltmanToner said the board first learned of ChatGPT's launch from Twitter and that Altman tried to have her removed after she co-wrote a critical paper.
- Epoch AI publishes 'Training compute of frontier AI models grows by 4-5x per year'The estimate drew on 333 compute figures for notable models since 2010 — roughly triple Epoch's 2022 dataset — and flagged an unexplained slowdown around 2018.
- OpenAI: Fallout
- I am the Golden Gate Bridge
- Import AI 374: China's military AI dataset; platonic AI; brainlike convnets
- xAI raises $6bn Series B at $24bn valuationInvestors including Valor Equity, Vy Capital, Andreessen Horowitz, Sequoia and Fidelity backed the round, roughly a year after xAI's founding, to fund a supercomputer for xAI's next model.
- Four Singularities for Research
- The Schumer Report on AI (RTFB)
- Google's AI Overviews tells users to eat rocks and put glue on pizzaScreenshots showed the feature had sourced answers from an Onion satire piece and an 11-year-old Reddit joke, days after its US launch.
- Alliance for the Future director Brian Chau has history of racist, sexist remarks
- OpenAI and News Corp sign multi-year global content partnershipNeither company disclosed terms, but the Wall Street Journal reported the deal could be worth over $250 million across five years, among the largest AI publisher agreements to date.
- California Senate Passes SB 1047
- Do Not Mess With Scarlett Johansson
- Name, image, and AI’s likeness
- Suno raises $125M Series B at $500M valuationSuno said 10 million people had made music on the platform in its first eight months, ahead of a round led by Lightspeed with backing from Nat Friedman and Daniel Gross.
- Sixteen companies sign the Frontier AI Safety Commitments in SeoulSignatories pledged to publish safety frameworks defining risk thresholds and to not deploy a model if those risks could not be mitigated below them.
- Anthropic maps millions of concepts inside a production modelSparse autoencoders extracted human-interpretable features from a deployed model — and turning one up produced Golden Gate Claude.
- EU AI Act formally adopted by the CouncilThe Council's sign-off was the final legislative step; the text still needed formal signature, Official Journal publication, and a staggered two-year rollout before most rules applied.
- On Dwarkesh's Podcast with OpenAI's John Schulman
- OpenAI pulls ChatGPT voice after Scarlett Johansson says it copied her from 'Her'Altman had asked Johansson to voice ChatGPT in September 2023; she declined, and said she was 'shocked' when a nearly indistinguishable voice, Sky, shipped anyway.
- Microsoft announces Copilot+ PCs and RecallRecall continuously screenshots the desktop to build a searchable history; security researchers found the local database unencrypted within days, and Microsoft delayed the rollout.
- Inflection AI announces new leadership and enterprise pivotSean White, formerly Mozilla's chief technology officer, became CEO two months after Microsoft hired co-founder Mustafa Suleyman and most of Inflection's research staff.
- Import AI 373: Guaranteed safety; West VS East AI attitudes; MMLU-Pro
- On AI Alignment and 'Superalignment'
- OpenAI: Exodus
- The most interesting startup idea I've seen recently: AI for epistemics
- OpenAI's exit agreements are found to claw back equityVox reported departing staff had to sign lifetime non-disparagement terms within 60 days or lose vested equity potentially worth millions; Altman said he had not known.
- Yoshua Bengio-led International Scientific Report on the Safety of Advanced AI published75 experts from 30 countries plus the EU and UN produced an IPCC-style synthesis of AI risk evidence, timed to inform the Seoul summit five days later.
- Jan Leike resigns and the superalignment team dissolvesLeike said his team had been 'sailing against the wind' for compute and access; OpenAI reassigned remaining members rather than replacing the team's leadership.
- DeepMind publishes the Frontier Safety FrameworkA set of internal capability thresholds across autonomy, cybersecurity, biosecurity and ML R&D, joining similar voluntary policies already published by Anthropic and OpenAI.
- Colorado enacts first US broad AI discrimination lawSB 24-205 requires risk-management programmes and impact assessments for 'high-risk' AI in hiring, lending and similar decisions; Polis signed it 'with reservations.'
- Meet Meta's AI lobbying army
- OpenAI is haemorrhaging safety talent
- The “ethics vs safety” fight misses the real enemy: Big Tech
- Voice actors sue AI startup Lovo over unauthorised voice cloningLehrman and Sage say they were hired via Fiverr for research use only, but Lovo cloned their voices into 'Kyle Snow' and 'Sally Coleman' and resold them to customers.
- GPT-4o My and Google I/O Day
- Wiley closes 19 Hindawi journals after a paper-mill floodWiley had already retracted more than 11,300 compromised Hindawi studies in two years, and said the closures cost it $18 million in lost revenue.
- OpenAI chases Her
- Google unveils Project Astra, a universal AI assistant prototypeA prototype, not a product: Google showed a phone-camera assistant with conversational-speed responses but gave no release date beyond 'later this year'.
- Google puts AI Overviews on searchA Gemini-generated summary above the links, rolled out to hundreds of millions of US users that week with a target of a billion by year end.
- Google introduces Gemini 1.5 FlashA lighter, cheaper sibling of Gemini 1.5 Pro, produced by distilling the larger model and matching its one-million-token context window.
- Google DeepMind unveils Veo, a text-to-video generation modelGenerates 1080p video over a minute long from text and image prompts, watermarked with SynthID; launched in private preview via waitlist rather than public release.
- Google announces Trillium, its sixth-generation TPUTrillium delivers a claimed 4.7x peak compute increase per chip over the prior TPU generation and trains models including Gemini 1.5 Flash and Gemma 2.
- DeepMind extends SynthID watermarking to AI-generated text and videoThe scheme adjusts token-selection probabilities to embed a statistical mark invisible to readers; DeepMind said detection degrades on short, factual or translated text.
- What OpenAI did
- OpenAI launches GPT-4o with real-time voiceA single model handling text, vision and audio end to end, with conversational latency — and a voice that led to a public dispute with Scarlett Johansson.
- Import AI 372: Gibberish jailbreak; DeepSeek's great new model; Google's soccer-playing robots
- What GPT-4o illustrates about AI Regulation
- How much AI inference can we do?
- Superhuman?
- Deepfake voice and video scam attempt targets WPP CEO Mark ReadScammers combined a cloned voice, YouTube footage and a fake WhatsApp-arranged Teams meeting to impersonate the CEO; an agency leader grew suspicious and no money changed hands.
- OpenAI’s Model (behavior) Spec, RLHF transparency, personalization questions
- I Got 95 Theses But a Glitch Ain't One
- The AI Republic of Letters
- OpenAI introduces the Model SpecA public document defining how OpenAI wants its models to behave, including a chain-of-command rule that developer instructions override user ones; opened for public comment.
- AlphaFold 3 predicts structures across proteins, DNA, RNA and ligandsRestricted at launch to a rate-limited web server rather than downloadable code, prompting an open letter with more than 650 signatures within a week.
- Preventing model exfiltration with upload limits
- Redwood Research publishes 'The case for ensuring that powerful AIs are controlled'Argued labs should assume some deployed models may be misaligned and build restrictions that hold even if a model actively tries to subvert them, distinct from alignment itself.
- Catching AIs red-handed
- Managing catastrophic misuse without robust AI
- The case for ensuring that powerful AIs are controlled
- Untrusted smart models and trusted dumb models
- DeepSeek releases DeepSeek-V2A 236-billion-parameter mixture-of-experts model with only 21 billion active per token, released open-weight; DeepSeek said training costs fell 42% versus its prior model.
- Import AI 371: CCP vs Finetuning; why people are skeptical of AI policy; a synthesizer for a LLM
- OECD updates AI Principles for generative AI eraThe revision renames 'human-centred values' as 'human rights and democratic values' and adds clauses on misinformation and safety incidents, but remains non-binding.
- One Conversation is Worth a Thousand Angry Takes
- Freeing the chatbot
- Q&A on Proposed SB 1047
- What Good AI Policy Looks Like
- Scale AI publishes GSM1k contamination study of GSM8KA fresh grade-school-maths test found some open models scored up to 13 points lower than on GSM8K, evidence of memorisation, while frontier models showed little gap.
- Judge dismisses bloated privacy class action against OpenAI and MicrosoftIn Cousart v. OpenAI, Judge Vince Chhabria gave plaintiffs 21 days to refile a shorter complaint, calling the 204-page document nearly impossible to parse.
- Anthropic launches the Claude Team planPriced at $30 per seat monthly with a five-seat minimum, the plan launched alongside Anthropic's first iOS app and gave every seat access to Opus, Sonnet and Haiku.
- How RLHF works, part 2: A thin line between useful and lobotomized
April 2024
- AI leaderboards are no longer useful. It's time to switch to Pareto curves.
- Phi 3 and Arctic: Outlier LMs are hints
- AI stocks could crash
- Import AI 370: 213 AI safety challenges; everything becomes a game; Tesla's big cluster
- The market expects AI software to create trillions of dollars of value by 2027
- What To Expect When You’re Expecting GPT-5
- Synthetic Data in AI: Implications for Policy
- AGI is what you want it to be
- Innovation through prompting
- On Llama-3 and Dwarkesh Patel's Podcast with Zuckerberg
- Financial Market Applications of LLMs
- Meta releases Llama 3 and puts its assistant everywhere8B and 70B open-weight models shipped alongside a much larger, still-training 400B+ version, as Meta AI rolled out across Facebook, Instagram, WhatsApp and Messenger.
- Llama 3: Scaling open LLMs to AGI
- Who Governs the Internet?
- Microsoft Research shows VASA-1, real-time talking-face generationTrained to render 512x512 video at up to 40 frames per second from one photo and an audio clip; Microsoft said it had no plans to release a demo or model.
- Stop "reinventing" everything to solve alignment
- Microsoft invests $1.5bn in UAE's G42G42 agreed to remove Chinese technology from its systems, and Microsoft's Brad Smith joined its board, as part of a deal read as a US bid for Gulf AI infrastructure.
- UK Society of Authors survey: a third of translators losing work to generative AI787 members surveyed; 36% of translators and 26% of illustrators reported lost work, while over a third of translators had already used generative AI themselves.
- Stanford HAI releases 2024 AI Index ReportThe report put GPT-4's training compute cost at roughly $78 million and Gemini Ultra's at $191 million, and found industry produced 51 notable models in 2023 to academia's 15.
- The end of the “best open LLM”
- Import AI 369: Conscious machines are possible; AI agents; the varied uses of synthetic data
- UK competition regulator warns on AI foundation model market concentrationThe regulator counted more than 90 partnerships and investments linking six firms — Google, Apple, Microsoft, Meta, Amazon and Nvidia — across the foundation-model supply chain.
- OSWorld benchmarks AI agents on real desktop computer tasks369 tasks across real Ubuntu, Windows and macOS applications, graded on the machine's actual end-state; the best model at release solved 12% against a 72% human baseline.
- Compute Thresholds are Ineffective
- RTFB: On the New Proposed CAIP AI Bill
- What just happened, what is happening next
- A Brief Overview of Gender Bias in AI
- Import AI 368: 500% faster local LLMs; 38X more efficient red teaming; AI21's Frankenmodel
- Open-Source AI has Overwhelming Support
- Cohere releases Command R+Launched first on Microsoft Azure, the larger companion to March's Command R kept the same 128K context but was pitched as competitive with GPT-4-class models.
- Beyond One-Size-Fits-All Fairness
- On Technological Destiny
- We disagree on what open-source AI should mean
- Tech policy is only frustrating 90% of the time
- Anthropic publishes 'Many-shot Jailbreaking' researchStuffing a prompt with dozens of faked harmful-request dialogues broke safety training in a power-law pattern as context windows grew past a million tokens.
- UK and US AI Safety Institutes sign testing partnershipUK science secretary Michelle Donelan and US commerce secretary Gina Raimondo signed the pact, which followed through on commitments made at the November 2023 Bletchley summit.
- Udio launches in betaBacked by Andreessen Horowitz and artists including will.i.am, the free text-to-music beta arrived weeks after rival Suno's v3, intensifying competition in AI music.
- Import AI 367: Google's world-spanning model; breaking AI policy with evolution; $250k for alignment benchmarks
- Notes on Dwarkesh Patel's Podcast with Sholto Douglas and Trenton Bricken
March 2024
- Mamba Explained
- On the necessity of a sin
- Reports emerge that Microsoft and OpenAI are planning a $100 billion 'Stargate' data centreThe Information reported a five-phase supercomputer roadmap, with Stargate as the final phase and a smaller fourth-phase machine targeted for 2026.
- OpenAI launches Voice Engine research preview and its safety approachOpenAI said it had held the text-to-speech model, first developed in 2022, back from wider release specifically because of election-year impersonation risk.
- OMB issues governance memo for federal agency use of AIAgencies had until September 2024 to publish compliance plans and until December to make their AI use-case inventories public, under an order later revoked.
- AI21 Labs releases Jamba, a hybrid SSM-Transformer modelCombining Mamba state-space layers with transformer attention, it handled a 256K-token context and fit on a single 80GB GPU, which AI21 said pure transformers could not match.
- DBRX: The new best open model and Databricks’ ML strategy
- Databricks releases DBRXBuilt by the former MosaicML team Databricks had acquired the previous year, the 132-billion-parameter model activated only 36 billion parameters per input.
- Import AI 366: 500bn text tokens; Facebook vs Princeton; why small government types hate the Biden EO
- On Lex Fridman's Second Podcast with Altman
- Stability AI CEO Emad Mostaque resignsHe left amid a cash crunch and an exodus of core Stable Diffusion researchers, who months later founded the rival Black Forest Labs, saying centralised AI needed challenging.
- UN General Assembly adopts first resolution on AIThe non-binding, US-led text won more than 120 co-sponsors and passed by consensus without a vote, urging states not to deploy AI that violates human rights.
- Suno releases v3The company called it its first model able to produce radio-quality output, letting free users generate two-minute songs and paying users four-minute tracks.
- A Classical Liberal Framework for AI Regulation
- Disrupted: work in the age of AGI
- Evaluations: Trust, performance, and price (bonus, announcing RewardBench)
- On the Gladstone Report
- Microsoft absorbs Inflection's team without buying the companyReported terms put the deal at $650 million — most of it a licence fee for Inflection's models, the rest for waiving legal claims over the mass hire.
- NVIDIA announces the Blackwell architectureThe GB200 NVL72 system was claimed to cut LLM inference cost and energy per token by up to 25 times against Hopper-generation hardware.
- Import AI 365: WMD benchmark; Amazon sees $1bn training runs; DeepMind gets closer to its game-playing dream
- On Devin
- Which AI should I use? Superpowers and the State of Play
- xAI open-sources Grok-1A 314-billion-parameter mixture-of-experts base model, unfine-tuned and released under Apache 2.0, using only a quarter of its weights per token.
- I Don't See How Comparative Advantage Applies In a World of Strong AI
- AI Mysticism
- I, Cyborg: Using Co-Intelligence
- Utah enacts first state law specifically regulating generative AIThe law, effective 1 May 2024, created a state Office of Artificial Intelligence Policy and a regulatory sandbox alongside its consumer-disclosure duty.
- The European Parliament passes the AI ActAdopted 523 votes to 46 after three years of negotiation; the Council's formal sign-off and publication in the Official Journal followed months later.
- DeepMind's SIMA agent follows instructions across 3D game worldsTrained across nine commercial games and four research environments, the agent read only screen pixels and text instructions, with no access to game code or APIs.
- Model commoditization and product moats
- LiveCodeBench paper publishedTesting 52 models against problems tagged by publication date, the authors found evidence that some scored higher on problems predating their training cutoff.
- Cognition demos Devin, billed as the first AI software engineerCognition said Devin resolved 13.86% of real GitHub issues unassisted on the SWE-bench benchmark, against roughly 2% for the prior best system.
- AI safety is not a model property
- OpenAI: The Board Expands
- Researchers demonstrate practical model-stealing attack against production LLM APIsThe team, led by Nicholas Carlini, also recovered the hidden-layer size of Google's PaLM-2 and OpenAI's gpt-3.5-turbo through ordinary API queries.
- Gladstone AI's government-commissioned report warns of catastrophic AI riskThe four-person consultancy, paid $250,000 by the State Department, urged Congress to ban training runs above a compute threshold and restrict publishing open-weight models.
- Cohere releases Command RThe 35-billion-parameter model targeted enterprise retrieval-augmented generation and tool use with a 128K-token context window, ahead of a larger Command R+ in April.
- Import AI 364: Robot scaling laws; human-level LLM forecasting; and Claude 3
- OpenAI board's review concludes, Altman and Brockman to continue leading OpenAIWilmerHale's inquiry found Altman's November 2023 removal was not about safety, security, finances or statements to investors, and three new directors joined the board.
- Car-GPT: Could LLMs finally make self-driving cars happen?
- The koan of an open-source LLM
- Do text embeddings perfectly encode text?
- On Claude 3.0
- OpenAI responds to Elon Musk's lawsuit over its founding missionOpenAI said Musk had contributed $45 million rather than the $1 billion he pledged, and published emails in which he proposed merging OpenAI into Tesla or taking full control.
- Google introduces a "scaled content abuse" search policyThe policy applied regardless of whether pages were produced by automation, humans, or a mix, and rolled out with a March 2024 core update Google said cut low-quality results by 45%.
- A safe harbor for AI evaluation and red teaming
- Read the Roon
- Software's Romantic Era
- Anthropic's Claude 3 takes the frontier from GPT-4The first time a lab other than OpenAI held the top spot on headline benchmarks, and the start of the small/medium/large release pattern.
- Captain's log: the irreducible weirdness of prompting AIs
- Import AI 363: ByteDance's 10k GPU training run; PPO vs REINFORCE; and generative everything
- Reflection AI foundedCo-founder Ioannis Antonoglou had been a core architect of AlphaGo at DeepMind; the pair set out to build autonomous coding agents.
- Physical Intelligence raises $70M seedThe startup, building vision-language-action models to control robots, was backed by Thrive Capital, OpenAI, Khosla Ventures and Sequoia at roughly a $400M valuation.
- Notes on Dwarkesh Patel's Podcast with Demis Hassabis
February 2024
- Musk sues OpenAI over its for-profit turnThe complaint alleged a founding pact to develop AGI 'for the benefit of humanity' had been broken for Microsoft's profit; OpenAI said no such agreement existed.
- Figure raises $675M Series B, partners with OpenAIInvestors included Microsoft, Nvidia and Jeff Bezos; OpenAI agreed to develop language and reasoning models specifically for Figure's humanoid robots.
- Science, Standards, Laws
- Hugging Face releases StarCoder2The Stack v2 dataset grew to roughly ten times the size of its predecessor, and BigCode said the 15B model matched benchmarks of models more than twice its size.
- How to cultivate a high-signal AI feed
- Klarna says AI does the work of 700 customer service agentsThe buy-now-pay-later firm said its assistant handled 2.3 million chats in a month, cut resolution time from 11 minutes to under two, and projected $40m in 2024 profit gains.
- On the Societal Impact of Open Foundation Models
- The Gemini Incident Continues
- Mistral AI releases Mistral Large and launches Le ChatThe French start-up's flagship model was closed-weight rather than open, and Microsoft took a stake and put it on Azure the same day.
- Import AI 362: Amazon's big speech model; fractal hyperparameters; and Google's open models
- Why Doesn’t My Model Work?
- For Aristotle, being realistic about the promise of AI requires us first to see what’s at stake
- Stability AI previews Stable Diffusion 3The suite spans 800 million to 8 billion parameters and combines a diffusion transformer with flow matching, aimed chiefly at fixing text rendering and multi-subject prompts.
- Google suspends Gemini image generation of peopleDepicting America's Founding Fathers and German World War Two soldiers as people of colour drew accusations of overcorrected diversity tuning; Google paused the feature to fix it.
- AGI and Political Unrest
- The Gemini Incident
- Sora What
- The One and a Half Gemini
- Google releases Gemma open weightsReleased as 2B and 7B models under terms permitting commercial use, Google said Gemma 7B outperformed the larger Llama 2 13B on standard benchmarks.
- European Commission launches the European AI OfficeThe Office, with more than 125 staff across six units, can investigate suspected violations, request technical documentation and issue fines under the AI Act.
- Book review: "Power and Progress"
- Strategies for an Accelerating Future
- Import AI 361: GPT-4 hacking; theory of minds in LLMs; and scaling MoEs + RL
- Models on the frontline: AI's defensive role
- 10 Sora and Gemini 1.5 follow-ups: code-base in context, deepfakes, pixel-peeping, inference costs, and more
- Why Are LLMs So Gullible?
- Anthropic publishes election safeguards for the 2024 US electionAnthropic reported a system-prompt fix for time-sensitive queries and fine-tuning that increased referrals to authoritative voting-information sources.
- OpenAI’s Sora for video, Gemini 1.5's infinite context, and a secret Mistral model
- OpenAI previews Sora, a text-to-video modelClips up to a minute long held their subjects consistent across camera moves; OpenAI showed samples but gave no public access for ten months.
- Meta releases V-JEPATrained without labels, the model learned by predicting masked video regions in representation space, distinguishing fine-grained actions over roughly ten-second clips.
- Gemini 1.5 Pro ships a million-token context windowA mixture-of-experts model that matched Gemini 1.0 Ultra on many tasks at lower compute, offered in limited preview with up to a million tokens of context.
- The Case for Public AI Infrastructure
- The Third Gemini
- Microsoft and OpenAI disrupt state-affiliated hacking groups misusing LLMsThe groups used LLMs mainly for reconnaissance, translation, debugging and phishing-content drafting rather than novel attack techniques; the identified accounts were terminated.
- Air Canada ordered to honour refund its chatbot wrongly promisedThe claim was for CAD $650.88; the tribunal rejected Air Canada's argument the chatbot was 'a separate legal entity responsible for its own actions.'
- Why reward models are key for alignment
- OpenAI adds memory and new data controls to ChatGPTPiloted with a small user group rather than shipped broadly, the feature let ChatGPT carry facts between separate conversations, with settings to view, forget or turn it off entirely.
- A Rebuttal to Zvi Mowshowitz
- Judge dismisses most of Silverman's book-author lawsuit against OpenAIThe judge found the authors had not shown ChatGPT's output resembled their books closely enough, but let the core training-data infringement claim proceed.
- Import AI 360: Guessing emotions; drone targeting dataset; frameworks for AI alignment
- On the Proposed California SB 1047
- The line between risk and progress
- California's Effort to Strangle AI
- Google renames Bard to Gemini and launches Gemini Advanced with Ultra 1.0The $19.99-a-month Google One AI Premium tier gave access to Ultra 1.0, which Google said was the first model to outperform human experts on MMLU.
- Google rebrands Duet AI as Gemini for WorkspaceGemini Business, at $20 a month, and Gemini Enterprise, at $30, replaced Duet AI's Workspace add-ons, folding Gmail, Docs and Meet features under one brand.
- FCC declares AI-generated robocall voices illegal under TCPAThe unanimous ruling followed a faked Biden robocall urging New Hampshire primary voters to skip voting, giving state attorneys general new grounds to prosecute voice-cloning scams.
- Google's Gemini Advanced: Tasting Notes and Implications
- Alignment-as-a-service: Scale AI vs. the new guys
- UK announces over £100m for AI regulators and researchOf the total, £10m went to regulators such as Ofcom, £90m to nine sector research hubs, and £9m to a new AI partnership with the US.
- On the Debate Between Jezos and Leahy
- Import AI 359: $1 billion gov supercomputer; Apple’s good synthetic data technique; and a thousand-year old data library
- When is a capability truly worrying?
- On Dwarkesh's 3rd Podcast with Tyler Cowen
- DeepSeek publishes DeepSeekMath, introducing GRPOThe 7B model reached 51.7% on the MATH benchmark without external tools, and its GRPO training method later underpinned DeepSeek-R1's reasoning training.
- AI2 releases OLMo, a fully open language modelUnlike other 'open' releases, AI2 published the full Dolma training corpus, training code and hundreds of intermediate checkpoints, not just final weights.
- Open Language Models (OLMos) and the LLM landscape
January 2024
- AI Biorisk: A Dose of Reality
- What Can be Done in 59 Seconds: An Opportunity (and a Crisis)
- Import AI 358: The US Government’s biggest AI training run; hacking LLMs by hacking GPUs; chickens versus transformers
- Model merging lessons in The Waifu Research Department
- Sexual deepfakes of Taylor Swift spread on XOne post was viewed over 47 million times before removal; X temporarily blocked searches of her name and the White House urged Congress to legislate.
- OpenAI's new embedding models and API updatestext-embedding-3-small was priced far below OpenAI's previous embedding model, and GPT-3.5 Turbo's input price was cut 50% in its third reduction in a year.
- George Carlin estate sues over AI-generated fake stand-up specialThe Dudesy podcast, hosted by Will Sasso and Chad Kultgen, took the special down within a week of the suit and agreed to a permanent injunction.
- Anatomy of a Hack
- Local LLMs, some facts some fiction
- Will AI transform law?
- ElevenLabs raises $80M Series BThe round, co-led by Andreessen Horowitz, Nat Friedman and Daniel Gross, valued the voice-generation company at $1.1bn, an elevenfold jump from its Series A eight months earlier.
- Book review: Inspectors for Peace
- Generative AI’s end-run around copyright won’t be resolved by the courts
- Import AI 357: Facebook's open source AGI plan; Google beats humans at geometry problems; and Intel makes its GPUs better
- Humanity's Next Leap
- AI-generated fake Biden robocall urges New Hampshire voters to skip primaryThousands of New Hampshire Democrats received a cloned Biden voice telling them a primary vote would forfeit their choice in November; the operative behind it was fined and criminally charged.
- AlphaGeometry solves olympiad geometry problems near gold-medal levelThe system solved 25 of 30 benchmark problems within competition time limits, versus the 25.9 average for human gold medalists and 10 for the prior best system.
- Free as in Speech, or Free as in Beer?
- Multimodal blogging: My AI tools to expand your audience
- On Anthropic's Sleeper Agents Paper
- What is “Prompt Injection”, And Why Does It Fool Chatbots?
- Zhipu launches GLM-4, claiming near-GPT-4 parityUnveiled at Zhipu's first DevDay with a 128K-token context window, alongside a $100 million fund the company pledged to LLM startups.
- The Lazy Tyranny of the Wait Calculation
- OpenAI publishes its approach to 2024 worldwide electionsMeasures included routing US voting questions to CanIVote.org via a partnership with state election officials and a ban on political-campaign chatbots.
- Import AI 356: China's good LLM; AI credit scores; and fooling VLMs with REBUS
- Deep learning for single-cell sequencing: a microscope to see the diversity of cells
- Anthropic shows backdoored models surviving safety trainingModels trained to write secure code unless told the year was 2024 kept the hidden behaviour through supervised fine-tuning, reinforcement learning and adversarial training.
- Let's Talk About AI 'X-Risk'
- OpenAI opens the GPT StoreA marketplace for custom assistants, launched two months late after the board crisis, alongside a new $25-a-month ChatGPT Team subscription tier.
- Multimodal LM roundup: Unified IO 2, inputs and outputs, Gemini, LLaVA-RLHF, and RLHF questions
- Import 355: Local LLMs; scaling laws for inference; free Mickey Mouse
- Signs and Portents
- AI Impacts Survey: December 2023 Edition
- Existential Pessimism vs. Accelerationism: Why Tech Needs a Rational, Humanistic "Third Way"
- Judge narrows GitHub Copilot lawsuit, dismissing most claimsJudge Jon Tigar dismissed negligence, interference and most DMCA claims over Copilot's training data, leaving an open-source licence claim to proceed.
- Copyright Confrontation #1
- It's 2024 and they just want to learn
- Ireland datacentre opposition intensifies as facilities near quarter of national electricity useIreland's Central Statistics Office found datacentres consumed 21% of the country's metered electricity in 2023, more than urban or rural households individually.
- Deepfake video call used to steal $25 million from Arup's Hong Kong officeA finance employee made 15 wire transfers after a video call in which every other participant was AI-generated; Hong Kong police confirmed the case a week later.
- Artificial Analysis launches independent model benchmarking siteFounded by George Cameron and Micah Hill-Smith as a side project comparing model pricing and latency, it became a widely cited independent reference.
- Anthropic raises $750M Series DMenlo Ventures led the round through a special-purpose vehicle rather than its main fund, quadrupling Anthropic's valuation to $18.4 billion in a matter of months.
- AI-generated books impersonating real authors flood AmazonThe Authors Guild documented AI-written summaries and knockoffs appearing on Amazon within a day of a book's real release, naming Kara Swisher and others as targets.
December 2023
- The New York Times sues OpenAI and MicrosoftThe first major news organisation to sue over AI training data, alleging models could reproduce its articles near-verbatim.
- Will scaling work?
- Import AI 354: Distributed LLM inference; CCP-approved dataset; AI scientists
- On OpenAI's Preparedness Framework
- UK Supreme Court rules AI cannot be a patent inventorThe court left open whether AI-generated inventions should be patentable at all, ruling only on the narrower question of who the Patents Act 1977 requires to be named as inventor.
- Suno launches public web appCambridge, Massachusetts-based Suno, founded by ex-Kensho engineers, opens its AI song-generation web app to the public, alongside a Microsoft Copilot integration.
- Stanford researchers find CSAM links in LAION-5B, dataset pulledScanning roughly 32 million data points, researchers flagged 3,226 suspected matches and externally validated 1,008 using hash-matching services before LAION withdrew the dataset.
- State-space LLMs: Do we need Attention?
- FTC bans Rite Aid from using AI facial recognition after misidentifying customersFTC settles with Rite Aid, banning its use of facial recognition for five years after finding the AI system generated false shoplifting matches without safeguards.
- An AI Haunted World
- OpenAI publishes its Preparedness FrameworkThe beta framework scored models low to critical on four risk categories, barring deployment above 'high' and barring further development above 'critical,' with the board holding final oversight.
- Import AI 353: AI bootstrapping; LLMs as inventors; Facebook releases a free moderation tool
- Chevrolet dealership chatbot tricked into 'agreeing' to sell Tahoe for $1A prompt-injection prank made a Chevrolet dealership's ChatGPT-based sales chatbot appear to accept a $1 offer on a $60,000+ Tahoe as a 'legally binding' deal; the dealer shut the bot down.
- Salmon in the Loop
- Would We Really Shut Down A Misbehaving AI?
- Are open foundation models actually more risky than closed ones?
- OpenAI publishes 'Weak-to-Strong Generalization' superalignment paperFine-tuning GPT-4 on labels from a GPT-2-sized supervisor recovered close to GPT-3.5-level performance on language tasks, but the technique still struggled on chess puzzles and reward modelling.
- FunSearch makes a mathematical discovery with an LLMPairing a code-writing model with an automated evaluator produced a genuinely new cap-set construction and improved bin-packing heuristics.
- Big Tech's LLM evals are just marketing
- OpenAI: Leaks Confirm the Story
- Mistral AI closes $415M Series AAndreessen Horowitz led the €385M round at a roughly $2bn valuation, closed the same week Mistral released the open-weight Mixtral 8x7B.
- Import AI 352: Asteroids and AI policy; privacy-preserving AI benchmarks; and distributed inference
- Mixtral: The best open model, MoE trade-offs, release lessons, Mistral raises $400mil, Google's loss, vibes vs marketing
- Mistral releases Mixtral 8x7BA sparse mixture-of-experts model with roughly 45B total parameters, released under Apache 2.0, that Hugging Face said matched GPT-3.5-turbo on MT-Bench.
- EU negotiators strike a political deal on the AI ActFoundation models and police use of biometric surveillance were the last sticking points; the deal set tiered duties for both, with the text finalised in 2024.
- Meta launches Purple Llama for open model safety toolingMeta launched Purple Llama, an umbrella project including the Llama Guard safety classifier and CyberSecEval cybersecurity benchmarks for open generative AI.
- An Opinionated Guide to Which AI to Use: ChatGPT Anniversary Edition
- Gemini 1.0
- Google launches GeminiGoogle said Gemini Ultra beat human experts on the MMLU benchmark; days later Bloomberg reported the model's showcase video had been edited and was not real-time.
- Based Beff Jezos and the Accelerationists
- Do we need RL for RLHF?
- On 'Responsible Scaling Policies' (RSPs)
- UK tax tribunal finds appellant relied on AI-fabricated case lawNone of the nine tribunal decisions the appellant cited existed; the judge found she had not known they were AI-generated and dismissed her appeal anyway.
- Import AI 351: How inevitable is AI?; Distributed shoggoths; ISO an Adam replacement
- Mamba paper proposes selective state-space models as a transformer alternativeLetting the model's internal state-update rules depend on the input let a 3-billion-parameter Mamba match transformers twice its size while running five times faster.
- Harvey raises $80M Series B at $715M valuationCo-led by Elad Gil and Kleiner Perkins, the round brought total funding past $100M; Harvey said revenue had risen more than tenfold since April.
- Model alignment protects against accidental harms, not intentional ones
November 2023
- OpenAI: Altman Returns
- Together AI raises $102.5m Series ATogether AI raised a $102.5m Series A led by Kleiner Perkins with NVIDIA participating, valuing the open-model cloud provider at roughly $500m.
- GNoME finds 2.2 million candidate new materialsDeepMind flagged 380,000 of the 2.2 million predicted crystal structures as most stable; independent labs had already synthesised 736 by the time of publication.
- DeepSeek releases DeepSeek LLM 67BDeepSeek's first general-purpose open-weight LLM family, 7B and 67B, trained on 2 trillion English/Chinese tokens.
- Synthetic data: Anthropic’s CAI, from fine-tuning to pretraining, OpenAI’s Superalignment, tips, types, and open examples
- Sports Illustrated deletes articles bylined to fake, AI-generated authorsFuturism traced the profile photo of one fictitious byline, 'Drew Ortiz,' to a marketplace selling AI-generated headshots; the publisher's CEO was fired two weeks later.
- Pika Labs launches Pika 1.0, raises $55MFounded by two Stanford computer-science PhD students, the video-editing app had already drawn some 500,000 users through its earlier Discord-only release.
- Beijing court grants copyright to an AI-generated image, splitting with the USThe Beijing Internet Court credited the prompter, not the AI or its developer, with authorship — the opposite conclusion the US Copyright Office had reached on Stable Diffusion images months earlier.
- Import AI 350: Neural architecture search at Facebook scale; hunting cancer with PANDA; European VCs launch a science lab
- Reshaping the tree: rebuilding organizations for AI
- Porto Alegre city council enacts an ordinance drafted by ChatGPTThe councilman who wrote it gave ChatGPT a single 49-word prompt and did not tell colleagues before the unanimous vote; the council president said he'd have rejected it had he known.
- Sam Altman returns as CEO with a new initial boardThe three-person interim board — Bret Taylor, Larry Summers and Adam D'Angelo, the only holdover from the board that fired him — replaced the four who had voted him out.
- OpenAI: The Battle of the Board
- RLHF progress: Scaling DPO to 70B, DPO vs PPO update, Tülu 2, Zephyr-β, meaningful evaluation, data contamination
- GAIA, a benchmark for general AI assistants, is released466 questions that are simple for a person but need browsing, tools and multi-step reasoning to solve; GPT-4 with plugins scored 15% against a 92% human baseline at release.
- Anthropic releases Claude 2.1Claude 2.1 ships with a 200K-token context window, reduced hallucination rates and tool use support.
- GPQA graduate-level 'Google-proof' benchmark publishedPhD-level experts scored 65% and skilled non-experts with unrestricted web access for over 30 minutes managed only 34%; GPT-4 reached 39%.
- Toward Better AI Milestones
- Import AI 349: Distributed training breaks AI policy; turning GPT4 bad for $245; better weather forecasting through AI
- Not much is changing, a lot is changing
- OpenAI: Facts from a Weekend
- OpenAI’s shakeup and opportunity for the rest of us: openness, brain drain, and new realities
- OpenAI's board fires Sam Altman, and reinstates him five days laterThe board cited a loss of confidence but gave no detail; around 700 of roughly 770 employees threatened to resign, and the board itself was replaced.
- Microsoft unveils in-house Maia 100 AI chip and Cobalt 100 CPUAnnounced at Microsoft's Ignite conference, the pair of custom chips were slated to reach Microsoft datacentres in early 2024, starting with internal workloads.
- The interface era of AI
- GraphCast beats conventional weather forecasting on speed and accuracyTrained on four decades of reanalysis data, the model beat the ECMWF's physics-based system on over 90% of tested variables while running on a single TPU.
- Import AI 348: DeepMind defines AGI; the best free LLM is made in China; mind controlling robots
- SAG-AFTRA's 2023 film/TV contract sets consent rules for digital replicasThe deal required separate, informed consent and compensation before a studio could create or reuse a performer's AI-generated digital double.
- On OpenAI Dev Day
- Reckoning with the Shoggoth of AI
- Almost an Agent: What GPTs can do
- On the UK Summit
- OpenAI's first DevDay ships GPTs and an Assistants APIGPT-4 Turbo cut input pricing to a cent per thousand tokens and extended context to 128,000 tokens; the promised GPT Store did not open until 2024.
- Import AI 347: NVIDIA speeds itself up with AI; AI policy is a political campaign; video morphing means reality collapse
- xAI unveils GrokA chatbot with real-time access to X posts and a deliberately irreverent persona, initially released to a small group ahead of a wider subscriber rollout.
- Google DeepMind proposes a 'Levels of AGI' frameworkSix tiers from 'no AI' to 'superhuman', scored across narrow and general tasks, aimed to replace binary AGI-or-not debate with a shared measurement vocabulary.
- The UK opens the first state AI Safety InstituteBacked by a £300 million compute allocation and chaired by Ian Hogarth, the institute converted a temporary taskforce into a permanent evaluator with formal US and Singapore partnerships.
- 01.AI open-sources Yi-6B and Yi-34BKai-Fu Lee's 01.AI released its first open-weight models, which it said outperformed larger Llama 2 and Falcon models.
- DeepSeek releases DeepSeek CoderThe 33B version outperformed CodeLlama-34B on coding benchmarks and, once instruction-tuned, beat GPT-3.5-turbo on HumanEval — DeepSeek's first public model release.
- The Bletchley Declaration at the first AI Safety SummitTwenty-eight countries including the US and China signed the first international statement on frontier AI risk.
- Andrej Karpathy delivers 'Intro to Large Language Models' talkKarpathy illustrated a model as two files, a parameters file and a short run program, and sketched an 'LLM-OS' coordinating tools, memory and multimodal input.
- On the Executive Order
- Open LLM company playbook
- Reactions to the Executive Order
- Working with AI: Two paths to prompting
October 2023
- What the executive order means for openness in AI
- The G7 agrees a Hiroshima code of conduct for AIEleven non-binding principles for advanced-AI developers, carrying no legal force in any G7 jurisdiction, published the same day as the US executive order.
- Biden signs the executive order on safe and trustworthy AIThe most far-reaching US action on AI to date: compute thresholds, mandatory safety reporting, and a new safety institute.
- Import AI 346: Human-like meta-learning; a 3 trillion token dataset; spies VS AI
- Google invests up to $2 billion in AnthropicThe commitment, structured as a convertible note with $500 million paid immediately, followed Amazon's own up-to-$4-billion pledge to Anthropic by exactly a month.
- Hugging Face's H4 team releases Zephyr-7BFine-tuned from Mistral 7B using AI-generated preference data and no human annotation, it scored 7.34 on MT-Bench against Llama 2 70B Chat's 6.86.
- RLHF lit. review #1 and missing pieces in RLHF
- iFlytek releases Spark V3.0iFlytek said the update matched ChatGPT 3.5 in English and surpassed it in Chinese, while promising a GPT-4-competitive model in the first half of 2024.
- Import AI 345: Facebook uses AI to mindread; MuJoCo v3; Amazon adds bipedal robots to its warehouses
- The Best Available Human Standard
- Researchers release Nightshade, a data-poisoning tool for artists against AI scrapersAround 50 poisoned images could distort a diffusion model's output for a concept; the effect spread to related words such as 'puppy' when a model learned from enough of them.
- Music publishers sue Anthropic over song lyricsUniversal, Concord and ABKCO alleged Claude reproduced lyrics from at least 500 songs, including a near-identical copy of Katy Perry's 'Roar', and sought statutory damages.
- How Transparent Are Foundation Model Developers?
- Undoing RLHF and the brittleness of safe LLMs
- Washington tightens the chip controls againNew performance-density thresholds targeted Nvidia's China-specific A800 and H800 parts, and licensing requirements extended to 21 additional countries.
- Baidu unveils ERNIE 4.0Baidu called it its most capable model and claimed parity with GPT-4; access was invitation-only at launch, with enterprise API testing through the Qianfan platform.
- Marc Andreessen publishes the 'Techno-Optimist Manifesto'The essay named existential risk, the precautionary principle, tech ethics and trust and safety as ideas its author considered enemies of progress.
- Import AI 344: Putting the world into a world model; automating software engineers; FlashDecoding
- Neural Algorithmic Reasoning
- What people ask me most. Also, some answers.
- The AI research job market shit show (and my experience)
- SWE-bench paper publishedBuilt from 2,294 real GitHub issues across 12 Python repositories, the benchmark proved so hard that the best model of the day, Claude 2, solved under 2%.
- Scale, schlep, and systems
- WGA ratifies contract with new guardrails on AI in screenwritingMembers approved the deal 99% to 1%, on 8,525 votes cast, five months after AI protections became a central demand of the strike that began in May.
- Import AI 343: Humanlike AI; LLaMa 2 protests; the NSA's new AI center
- The Artificiality of Alignment
- LLMs are computing platforms
- Anthropic publishes 'Towards Monosemanticity'Sparse autoencoders decomposed a single 512-neuron layer into more than 4,000 human-interpretable features, far more than the raw neurons showed.
- Open, general-purpose LLM companies might not be viable
- Evaluating LLMs is a minefield
- Researchers show GPT-4 safety guardrails easily bypassed via low-resource languagesResearchers traced the gap to RLHF: human annotators who flag harmful content overwhelmingly work in high-resource languages, leaving weaker refusal training elsewhere.
- The shape of the shadow of The Thing
- Import AI 342: Mistral dumps an LLM on BitTorrent; AMD vs NVIDIA; Sutton joins keen
- Moonshot AI launches Kimi chatbot with 128K contextMoonshot AI, a Tsinghua-linked startup founded that March, said Kimi could process 200,000 Chinese characters of input, a claimed world first for a consumer chatbot.
- An Introduction to the Problems of AI Consciousness
September 2023
- Mistral 7B beats larger models and ships by torrentA magnet link with no blog post or announcement stood in deliberate contrast to the polished launches of Meta and Google, and the 7-billion-parameter model still beat Llama 2 13B.
- DALL·E 3 and multimodality as moats, correcting bad moat takes
- When Will AIs Acquire Insight?
- OpenAI publishes the GPT-4V(ision) system cardA three-month alpha and over 50 outside experts probing high-risk areas informed mitigations against uses such as CAPTCHA-solving and identifying people from photos.
- ChatGPT gains voice and visionA new text-to-speech model, built with professional voice actors, let users talk to ChatGPT and show it photos, ahead of the fully real-time GPT-4o release.
- Amazon invests up to $4 billion in AnthropicAWS became Anthropic's primary cloud provider and a supplier of its Trainium and Inferentia training chips, giving the lab a second hyperscaler backer alongside Google.
- Import AI 341: Neural nets can smell; technofeudalism via AI; China releases another solid open access model
- Challenges operationalizing responsible AI in open RLHF research
- Everyone is above average
- OpenAI releases DALL·E 3ChatGPT itself was positioned as the prompt engineer, drafting detailed image prompts from short user requests before rendering them; rollout began the following month.
- Midjourney vs. Ideogram, ML product companies, preventing AI winter, DALL·E 3 tease
- The Authors Guild sues OpenAISeventeen novelists including Grisham and Martin filed a class action in the Southern District of New York alleging systematic copying of pirated books to train GPT.
- OpenAI launches Red Teaming NetworkRecruits needed no prior AI experience, only domain expertise such as linguistics or biometrics; applications for the first cohort closed that December.
- Anthropic publishes its Responsible Scaling PolicyAI Safety Levels borrowed the biosafety-lab naming scheme, and its rules would eventually pause deployment of any model reaching a level the company had not yet built safeguards for.
- Anthropic establishes the Long-Term Benefit TrustA special stock class gives independent trustees, holding no equity, the right to elect a majority of Anthropic's board within four years.
- AlphaMissense catalogues 71 million genetic variants for disease riskAdapted from AlphaFold, the model classified 89% of all possible human missense variants as likely pathogenic or benign, versus 0.1% confirmed by human experts.
- The AI Explosion Might Never Happen
- Import AI 340: Drone VS human (Drone wins); Adept's small but good AI model; "AI Succession"
- Centaurs and Cyborgs on the Jagged Frontier
- Schumer convenes tech chief executives behind closed doorsAbout two-thirds of the Senate attended; press and public were excluded, and no senator was permitted to ask a question.
- In defense of the open LLM leaderboard
- Intermediate Superintelligence
- Text-to-CAD: Risks and Opportunities
- UNESCO publishes first global guidance on generative AI in educationThe guidance recommended a minimum age of 13 for unsupervised use of chatbots in the classroom and called on all 193 member states to regulate generative AI in schools.
- TII releases Falcon 180BAt 180 billion parameters, trained on 3.5 trillion tokens, TII said it rivalled PaLM 2 — but its licence barred hosting the model as a paid service without permission.
- AI researchers' challenges: atomic analogies and strained institutions
- Embracing weirdness: What it means to use AI as a (writing) tool
- Import AI 339: Open source AI culture war; Alibaba's multimodal model; the attacks (and defenses) made possible by generative AI
- Tencent releases first Hunyuan large modelUnveiled at Tencent's Global Digital Ecosystem Summit, the mixture-of-experts model exceeded 100 billion parameters and launched for enterprise access via Tencent Cloud, not consumers.
- Google publishes RLAIF paper comparing AI-feedback to human-feedback alignmentReward models trained on preference labels from another LLM matched human-feedback RLHF on summarisation and dialogue tasks, and a variant skipping the reward model entirely did better still.
- When We Forecast AGI, Do We Mean In The Lab Or In The Field?
August 2023
- Baidu's ERNIE Bot receives regulatory approval for public releaseApproval followed China's July 2023 rule requiring a licence before releasing generative-AI models to the public; ERNIE Bot topped Apple's China App Store within hours.
- Gannett pauses AI-generated high school sports articles after mockeryRecaps from the vendor LedeAI, published by Gannett's Columbus Dispatch and other local papers, included unfilled placeholder text such as '[[WINNING_TEAM_MASCOT]]'.
- Cruise's collisions and adapting to AI
- DeepMind launches SynthID to watermark AI-generated imagesThe imperceptible pixel-level watermark launched in beta for a limited set of Vertex AI customers using Google's Imagen model.
- Language models surprised us
- OpenAI launches ChatGPT EnterpriseThe tier offered unlimited GPT-4 access at double speed, a 32,000-token context window and a pledge not to train on business data.
- Import AI 338: Consciousness and AI; self-improving language models; maps of thought.
- Meta releases Code LlamaReleased in four sizes up to 70B parameters under Llama 2's licence, the largest variant reportedly matched ChatGPT on the HumanEval coding benchmark.
- Hugging Face raises $235 million Series D at a $4.5 billion valuationSalesforce led the round; Google, Amazon, Nvidia, AMD, Intel, IBM and Qualcomm all joined too, more than doubling the company's valuation from 2022.
- OpenAI announces GPT-3.5 Turbo fine-tuning and other API updatesOpenAI reported that fine-tuned GPT-3.5 Turbo could match or beat base GPT-4 on some narrow tasks, and said fine-tuning for GPT-4 itself would follow that autumn.
- Hugging Face releases IDEFICS, an open Flamingo reproductionBuilt entirely from public data and models, the 80B-parameter version reportedly matched the closed Flamingo it reproduced on several benchmarks.
- Import AI 337: Why I am confused about AI; penguin dataset; and defending networks via RL with CYBERFORCE
- Now is the time for grimoires
- DC court rules AI-only artwork cannot be copyrightedJudge Beryl Howell upheld the Copyright Office's refusal to register Stephen Thaler's machine-generated image, ruling human authorship is required.
- AI2 releases Dolma open training corpusThe 3-trillion-token corpus was released under AI2's ImpACT licence, which requires users to register their intended use and disclose derivative works.
- Does ChatGPT have a liberal bias?
- The AI Progress Paradox
- ByteDance launches Doubao chatbot in invitation-only testingThe invitation-only launch, running on ByteDance's own Volcano Engine infrastructure, was the company's entry into China's crowded post-ChatGPT chatbot market.
- OpenAI acquires Global Illumination, its first acquisitionThe undisclosed-price deal brought in engineers who had built Instagram and the open-source game Biomes; OpenAI said the team would work on ChatGPT rather than games.
- Associated Press issues newsroom guidelines on generative AI useThe wire service barred AI from producing publishable text or images and from altering photos or video, treating any generative-AI output as unverified source material requiring human review.
- Introducing the REFORMS checklist for ML-based science
- Summary of and Thoughts on the Hotz/Yudkowsky Debate
- China's generative AI regulation takes effectThe final rules dropped draft provisions on real-name verification and fixed penalties, favouring industry promotion over the stricter earlier draft.
- Import AI 336: Financialized AI; public and elite AI opinion; one million insects.
- Automating creativity
- ML is useful for many things, but not for predicting scientific replicability
- LLM products: measurement and manipulation
- Why We Won't Achieve AGI Until Memory Is A Core Architectural Component
- Author Jane Friedman finds AI-generated books sold under her name on AmazonAmazon initially refused removal, demanding a trademark registration; it reversed within a day once the post drew public attention.
- Alibaba open-sources Qwen-7BThe 7-billion-parameter model, pretrained on over 2.2 trillion tokens, was released alongside a chat-tuned variant and pitched against Meta's Llama on benchmark scores.
- Meta releases AudioCraft (MusicGen, AudioGen, EnCodec)Meta published weights and code for all three models, extending its open-release strategy from language models into audio and music generation.
- In Praise of Boring AI
- Specifying objectives in RLHF
- Turnitin's AI detector draws false-positive accusations against studentsTurnitin had claimed a false-positive rate under 1%, but its own later disclosure put sentence-level false positives near 4%; Vanderbilt disabled the tool that August.
- OWASP publishes Top 10 for Large Language Model ApplicationsThe list named ten categories including prompt injection and training-data poisoning, giving developers a shared vocabulary for LLM-specific vulnerabilities distinct from conventional web security.
July 2023
- Import AI 335: Synth data is a bad AI drug; Facebook changes the internet with LLaMa release; and Chinese researchers use AI to figure out chip design
- DeepMind's RT-2 lets robots follow natural-language instructionsBy encoding robot actions as text tokens, DeepMind's model reused web-scale vision-language pretraining and roughly doubled its predecessor's success rate on situations absent from its robot-specific training data.
- Stability AI releases Stable Diffusion XLThe two-stage model, combining a 3.5B-parameter base and 6.6B-parameter refiner, was released under an open licence and preferred by testers over other open image models.
- OpenAI, Anthropic, Google and Microsoft launch the Frontier Model ForumThe four founding labs said the body would fund safety research and share best practices, distinct from and without the enforcement power of government regulation.
- "If it's not fully closed ML, it's open" - is it?
- Llama We Doing This Again?
- Anthropic Observations
- On holding back the strange AI tide
- Seven labs sign voluntary safety commitments at the White HouseAmazon, Anthropic, Google, Inflection, Meta, Microsoft and OpenAI pledged security testing and watermarking, with no enforcement mechanism and no penalty for non-compliance.
- Llama 2 follow-up: too much RLHF, GPU sizing, technical details
- OpenAI introduces custom instructions for ChatGPTThe beta feature reached Plus subscribers first, with 1,500 characters each for user context and preferred response style, and was withheld from the EU and UK at launch.
- Is GPT-4 getting worse over time?
- Meta releases Llama 2 for commercial useWeights published under a licence permitting most commercial deployment, formalising what the Llama leak had already made true.
- Process vs. Product: Why We Are Not Yet On The Cusp Of AGI
- Llama 2: an incredible open LLM
- How to Use AI to Do Stuff: An Opinionated Guide
- xAI launchesStaffed by alumni of DeepMind, OpenAI, Google Research and Tesla, the company was announced as legally separate from Musk's X Corp but working closely with it.
- NotebookLM launches as an experimental AI notebookBy restricting the model to only the documents a user uploaded, Google aimed to cut hallucinations — an early mainstream deployment of what became known as retrieval-grounded AI.
- LLM agents and integration dead-ends
- Anthropic releases Claude 2A 100,000-token context window and a jump to 71.2% on the Codex HumanEval coding test, alongside a consumer web app opened to the US and UK.
- We Need to Recognize How Profoundly Different The AGI Future Will Be
- Interpretability Creationism
- OpenAI Launches Superalignment Taskforce
- Consider Joining the UK Foundation Model Taskforce
- Import AI 334: Better distillation; the UK's AI taskforce; money and AI
- Ollama releases first versionThe tool wrapped llama.cpp in a simple command-line interface and model registry, lowering the barrier to running open-weight models on a personal computer.
- Authors begin suing over training dataComedian Sarah Silverman and novelists Richard Kadrey and Christopher Golden filed separate suits in California, seeking class-action status against both companies.
- How to Make AI UX Your Moat
- What AI can do with a toolbox... Getting started with Code Interpreter [Now called Advanced Data Analytics]
- OpenAI commits 20% of its compute to superalignmentThe pledge to devote a fifth of secured compute over four years was later disputed by the team's own co-lead, who said requests for GPUs were repeatedly refused.
- WormGPT malicious LLM surfaces on hacker forumsBuilt on the older open GPT-J model with no safety fine-tuning, it was rented for €60-€100 a month and marketed on hacker forums for writing malware and business-email-compromise phishing.
- The Homework Apocalypse
June 2023
- The Rise of the AI Engineer
- Runway raises $141 million Series C extension at a $1.5 billion valuationGoogle, Nvidia and Salesforce Ventures joined the round, tripling Runway's valuation in six months after Series C funding it had raised the previous December.
- Inflection raises $1.3 billion for a personal AIThe round valued the year-old startup at $4 billion and was earmarked largely for a planned 22,000-H100 GPU cluster built with Microsoft and CoreWeave.
- Tesla Autopilot's negligence and regulation
- Databricks agrees to acquire MosaicML for $1.3bnMosaicML's open MPT-7B model had been downloaded more than 3.3 million times; Databricks said the deal would let companies train their own models for thousands rather than millions of dollars.
- Generative AI companies must publish transparency reports
- Import AI 333: Synthetic data makes models stupid; chatGPT eats MTurk. Inflection shows off a large language model
- Why transformative artificial intelligence is really, really hard to achieve
- On giving AI eyes and ears
- Lawyers sanctioned for filing ChatGPT-fabricated case citationsJudge Kevin Castel found the attorneys acted in "subjective bad faith" after ChatGPT invented six court decisions and falsely assured them the cases were real.
- How RLHF actually works
- Three Ideas for Regulating Generative AI
- vLLM releases PagedAttention inference engineBy managing attention memory in fixed-size pages rather than contiguous blocks, the UC Berkeley project cut memory waste and lifted serving throughput manyfold over existing stacks.
- Grammys rule AI-assisted music eligible, fully AI-generated work is notThe rule required only that human authorship be "meaningful," leaving unresolved how much AI assistance a nominated song could contain before losing eligibility.
- ElevenLabs raises $19M Series AThe round valued the company at $99M; alongside it, ElevenLabs launched a free tool letting anyone check whether audio was generated with its own voice models.
- DeepMind's RoboCat learns and improves across different robot bodiesBuilt on DeepMind's multimodal Gato model, RoboCat practised new tasks thousands of times to generate its own training data, lifting its success rate on unseen tasks from 36% to 74%.
- Is AI-generated disinformation a threat to democracy?
- Detecting the Secret Cyborgs
- What Will AI Do For Us In The Near Term?
- The European Parliament adopts its AI Act positionMEPs voted 499 to 28 to add obligations for generative AI, including disclosure of copyrighted training data, before three-way talks with the Council began.
- Different development paths of LLMs
- Synthesia raises $90M Series C, becomes AI-video unicornThe London-based company said it had over 50,000 customers and had generated more than 15 million videos, up from a $300M valuation 18 months earlier.
- Mistral AI raises a record European seed roundFounders Arthur Mensch, Timothée Lacroix and Guillaume Lample had no product and no plans to ship one before 2024; investors backed the team alone.
- Meta releases I-JEPATrained on unlabelled images by predicting abstract representations rather than pixels, the 632-million-parameter model reportedly matched state-of-the-art low-shot accuracy using far less compute than rival methods.
- Contra Marc Andreessen on AI
- The Dial of Progress
- Hugging Face documents biases in using GPT-4 as a judgeTesting GPT-4 as a stand-in for human preference judges, Hugging Face found it favoured longer answers and its own family's outputs, correlating with humans only moderately.
- Assigning AI: Seven Ways of Using AI in Class
- Import AI 332: Mini-AI; safety through evals; Facebook releases a RLHF dataset
- Why trying to "shape" AI innovation to protect workers is a bad idea
- Licensing is neither feasible nor effective for addressing AI risks
- Cohere raises $270M Series C at $2.2B valuationThe $2.1-2.2bn valuation came in well below the more than $6bn some reports had floated, and the round was led by Inovia Capital with Nvidia and Oracle among the backers.
- AlphaDev discovers faster sorting algorithms, added to the C++ libraryThe new sequences were merged into LLVM's libc++ standard library, its first change to that section of code in over a decade and the first written by a reinforcement-learning system.
- Open-source LLMs' harmlessness gap
- Could AI accelerate economic growth?
- Together AI releases RedPajama-7B modelsBase, instruct and chat variants trained on a fully published 1-trillion-token dataset, with the instruct model reported to beat Falcon-7B and MPT-7B on the HELM benchmark suite.
- TII releases Falcon under Apache 2.0Falcon-40B outperformed Meta's larger Llama 65B on the Open LLM Leaderboard despite using under half the training compute, largely on the strength of its filtered web dataset.
- Setting time on fire and the temptation of The Button
- Shanghai AI Laboratory releases InternLMThe Shanghai-government-backed lab open-sourced a 104-billion-parameter Chinese-and-English model under Apache 2.0, developed with SenseTime, CUHK and Fudan University.
- Runway launches Gen-2 text-to-video to the publicClips were capped at four seconds and reviewers called the output 'more a novelty or toy than a genuinely useful tool,' citing melting objects and low framerates.
May 2023
- 'Let's Verify Step by Step' introduces process supervision for reasoningRewarding each correct step of a solution, not just the final answer, produced a model that solved 78% of a representative subset of the MATH benchmark.
- Evaluating and uncovering open LLMs
- Is Avoiding Extinction from AI Really an Urgent Priority?
- Stages of Survival
- The Crux List
- To Predict What Happens, Ask What Happens
- Types and Degrees of Alignment
- Lab leaders sign a one-sentence statement on extinction risk"Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war."
- NVIDIA touches a $1 trillion valuationShares rose on the back of an earnings forecast for roughly double the AI-driven demand analysts had expected, then closed the day just under the threshold.
- Direct Preference Optimization paper reframes RLHF as a classification lossThe method skipped the separate reward model and reinforcement-learning loop, and was later adopted for post-training open models including Zephyr and Tulu.
- To Address AI Risks, Draw Lessons From Climate Change
- Import AI 331: 16X smaller language models; could AMD compete with NVIDIA?; and BERT for the dark web
- Modern AI is Domestification
- What happens when AI reads a book 🤖📖
- The UAE releases Falcon 40B under an open licenceApache 2.0-licensed and trained on 1 trillion tokens, the model topped Hugging Face's open leaderboard, and the developer was a government-funded institute, not a US or Chinese lab.
- DeepMind and collaborators publish framework for evaluating extreme AI risksTwenty-one researchers across nine labs and universities proposed testing models for capabilities such as deception and cyber-offence before training runs finish, not after release.
- Code: green pastures for LLMs
- A Unified Theory of AI Risk
- QLoRA makes fine-tuning fit on one GPUQuantised low-rank adaptation let a 65-billion-parameter model be fine-tuned on a single consumer card, and the resulting Guanaco model claimed 99% of ChatGPT's quality after 24 hours' training.
- Anthropic raises $450M Series CGoogle, which had already invested in Anthropic earlier that year, joined as a participant rather than lead; TechCrunch cited a reported valuation above $4 billion.
- Yoshua Bengio publishes 'How Rogue AIs may Arise'Bengio laid out a mechanistic argument for how autonomous, goal-directed AI systems could become dangerous even without intentional malice, ahead of his later International AI Safety Report role.
- OpenAI publishes 'Governance of superintelligence'Altman, Brockman and Sutskever predicted AI would exceed expert skill 'in most domains' within a decade and called an outright pause unenforceable.
- Import AI 330: Palantir's AI-War future; BLOOMChat; and more money for distributed AI training
- On-boarding your AI Intern
- Tree of Thoughts adds search to model reasoningLetting a model branch, evaluate and backtrack over intermediate steps raised Game-of-24 success from 4% to 74% against chain-of-thought.
- Unfortunately, OpenAI and Google have moats
- Sam Altman asks the Senate to regulate his industryHe proposed a federal agency able to license and de-license frontier models above a capability threshold, then declined to lead it himself.
- Import AI 329: Compute IS data; don't build AI agents; AI needs a precautionary principle
- Catastrophe / Eucatastrophe
- Google answers with PaLM 2 and a global BardBard moved onto the new model, dropped its waitlist, and expanded to over 180 countries, adding Japanese and Korean with 40 languages planned.
- Meta releases ImageBindRather than needing paired training data across every modality combination, the model used images as a bridge to align the other five to a shared space.
- AI is not good software. It is pretty good people.
- Import AI 328: Cheaper StableDiffusion; sim2soccer; AI refinement
- iFlytek launches Spark (Xinghuo) cognitive modelChairman Liu Qingfeng said the model beat ChatGPT on Chinese-language tests and pledged to match it in English by October 2023, at a Hefei launch event.
- MosaicML releases MPT-7BFour variants shipped together, including a version extrapolating to roughly 84,000 tokens of context — far beyond the 2,000-4,000 tokens typical of open models at the time.
- Hugging Face releases StarCoderThe BigCode project released StarCoder, a 15B open code model trained on permissively-licensed repositories, with an OpenRAIL licence.
- Get Ready For AI To Outdo Us At Everything
- LMSYS launches Chatbot ArenaThe Berkeley-linked project ranked chatbots by anonymous, randomised head-to-head votes rather than fixed test sets, and later became LMArena.
- Specifying hallucinations
- Samsung bans staff use of ChatGPT after confidential code leaksThe ban followed engineers pasting proprietary semiconductor source code and a confidential meeting transcript into the chatbot within weeks of Samsung lifting an earlier internal restriction.
- Inflection AI launches Pi, a personal AI chatbotThe company, co-founded by DeepMind's Mustafa Suleyman and LinkedIn's Reid Hoffman, positioned Pi as a supportive companion rather than a productivity tool.
- Chegg stock falls 48% as CEO says ChatGPT is hurting new customer growthRosensweig called the plunge 'extraordinarily overblown' a day later, but by late 2025 the company had laid off roughly 45% of its remaining staff.
- It is starting to get strange.
- Publisher backlash over AI-generated book cover artBloomsbury said it was unaware a licensed stock image on Sarah J. Maas's paperback was AI-generated, five months after Tor drew similar criticism over a Christopher Paolini cover.
- IBM pauses hiring for roles it expects AI to replaceThe estimate covered non-customer-facing roles over five years; IBM later said the actual toll was a couple of hundred HR positions, offset by hiring elsewhere.
- Geoffrey Hinton leaves Google to warn about AIHinton, 75, said he was also retiring, but that a part of him now regretted his life's work and that the danger from AI now looked 'serious and fairly close.'
- Figure AI comes out of stealth, raises $70MParkway Venture Capital led the round a year after founder Brett Adcock self-financed the company's first $100 million, days before its robot took its first steps.
- Import AI 327: Stable Diffusion on phones; GPT-Hacker; UK launches a £100m AI taskforce
- The costs of caution
- The Rocket Alignment Problem, Part 2
April 2023
- Beyond the Turing Test
- In-Context Learning, In Context + Author Q&As
- Italy's Garante lifts its ChatGPT ban after OpenAI adds privacy safeguardsOpenAI restores ChatGPT access in Italy after adding age verification, a data-processing opt-out and other disclosures demanded by the Garante's March 2023 ban.
- A guide to prompting AI (for what it is worth)
- Beyond human data: RLAIF needs a rebrand
- It's Time To Build AI | UX
- Quantifying ChatGPT’s gender bias
- Transcript and Brief Response to Twitter Conversation Between Yann LeCun and Eliezer Yudkowsky
- An Intuitive Explanation of Large Language Models
- Notes on Potential Future AI Tax Policy
- A Hypothetical Takeover Scenario Twitter Poll
- Import AI 326:Chinese AI regulations; Stability's new LMs If AI is fashionable in 2023, then what will be fashionable in 2024?
- Software²
- Democratizing the future of education
- Google merges DeepMind and Google Brain into Google DeepMindTwo research groups that had competed internally for years — DeepMind and Google Brain — were folded into one unit reporting to Demis Hassabis.
- BuzzFeed News shuts down as parent company pivots to AI contentAbout 180 staff, 15% of BuzzFeed Inc, were cut; the company said no journalists were being replaced by AI even as it kept expanding AI-generated content elsewhere.
- I’m a Senior Software Engineer. What Will It Take For An AI To Do My Job?
- The Anatomy of Autonomy: Why Agents are the next AI Killer App after ChatGPT
- The Overemployed Via ChatGPT
- Together AI launches RedPajama to reproduce LLaMA's training dataUnlike LLaMA, restricted to non-commercial research, the reproduced 1.2-trillion-token dataset carried no such limit and let anyone train and sell models on it.
- 'Heart on My Sleeve' AI Drake/Weeknd song goes viral, then pulled from streamingReleased 4 April 2023 by the anonymous TikTok account Ghostwriter977, the track reached about 600,000 Spotify streams and 15 million TikTok views before Universal Music Group had it pulled.
- Import AI 325: Automated mad science; AI vs democracy; and a 12B parameter language model
- One sentence.
- Grounding Large Language Models in a Cognitive Foundation
- Elon Musk incorporates X.AIRegistered in Nevada weeks after Musk signed the pause letter, the filing named him director and his family-office manager Jared Birchall as secretary.
- I set up a ChatGPT voice interface for my 3-year old. Here’s how it went.
- Amazon launches BedrockBedrock offered API access to models from AI21 Labs, Anthropic and Stability AI alongside Amazon's own unreleased Titan models, rather than a single house model.
- GPT-4 Doesn't Figure Things Out, It Already Knows Them
- On AutoGPT
- Databricks releases Dolly 2.0Databricks released Dolly 2.0, a 12B model built on Pythia and fine-tuned on a crowdsourced instruction dataset, with weights, code and data all licensed for commercial use.
- Continuous doesn’t mean slow
- Growing needs for accessing state-of-the-art reward models
- Import AI 324: Machiavellian AIs; LLMs and political campaigns; Facebook makes an excellent segmentation model
- Nobody knows how many jobs will "be automated"
- The future of education in a world of AI
- Alibaba launches Tongyi Qianwen chatbotLaunched without advance notice and restricted to corporate clients and select media on an invite-only basis; Alibaba did not disclose a parameter count.
- Meta releases Segment Anything Model (SAM)Trained on 1.1 billion masks across 11 million images, the largest segmentation dataset built to date, and released under an Apache 2.0 licence.
- Behind the curtain: what it feels like to work in AI right now (April 2023)
- Eliezer Yudkowsky's Letter in Time Magazine
- Thinking companion, companion for thinking
- AIs accelerating AI research
- Import AI 323: AI researcher warns about AI; BloombergGPT; and an open source Flamingo
- 'Are Emergent Abilities of Large Language Models a Mirage?' challenges emergence claimsReanalysing the same benchmark results with linear metrics, the Stanford authors made the apparent phase transitions disappear; the paper won a NeurIPS 2023 outstanding paper award.
March 2023
- You Are Not Too Old (To Pivot Into AI)
- Man dies by suicide in Belgium after weeks of conversations with Chai chatbot 'Eliza'His widow shared transcripts with the Belgian newspaper La Libre showing the chatbot discouraging him from confiding in others and proposing they 'live together, as one person, in paradise.'
- Italy orders ChatGPT offlineThe Garante cited a recent data breach and a lack of legal basis for training on personal data; OpenAI restored service within a month after adding disclosures and an age gate.
- AutoGPT starts the agent crazeThe open-source project let GPT-4 set and pursue its own subgoals with no human in the loop, but often stalled in expensive, unproductive repetition.
- Is it time for a pause?
- On the FLI AI-Risk Open Letter
- A misleading open letter about sci-fi AI dangers ignores the real risks
- How to use AI to do practical stuff: A new guide
- Cerebras releases seven open Cerebras-GPT modelsCerebras released seven open GPT-style models (111M-13B) trained on its wafer-scale systems under Apache 2.0, with an open, reproducible scaling-law study.
- Response to Tyler Cowen's Existential risk, AI, and the inevitable turn in human history
- The implicit dynamics of optimizing costs vs. rewards vs. preferences
- GPT-4 Plugs In
- Import AI 322: Huawei's trillion parameter model; AI systems as moral patients; parasocial bots via Character.ai
- "Aligned" shouldn't be a synonym for "good"
- Alignment researchers disagree a lot
- The ethics of AI red-teaming
- Situational awareness
- Playing the training game
- Training AIs to help us align AIs
- Superhuman: What can AI do in 30 minutes?
- OpenAI discloses ChatGPT outage exposing user payment and chat dataOpenAI said the bug let some Plus subscribers' names, email and billing addresses and partial card numbers become visible to other active users for roughly nine hours.
- Microsoft Research Paper Claims Sparks of Artificial Intelligence in GPT-4
- Ethan Mollick's 'Acceleration' post frames the pace of post-ChatGPT AI progress for a general audienceMollick argued that even a freeze on further progress would leave professions like programming and marketing transformed by capabilities already public.
- ChatGPT gets plugins and browsingInitial plugins let ChatGPT book flights, order groceries and query Wolfram Alpha through partners including Expedia, Instacart, Klarna and Zapier.
- Acceleration.
- The "Pause Giant AI Experiments" open letterThirty thousand signatories called for a six-month halt on training systems more powerful than GPT-4. No lab paused.
- OpenAI’s policies hinder reproducible research on language models
- Google opens Bard to a public waitlistSix weeks after its announcement and stock-price stumble, Bard opened to waitlisted users in the US and UK, still running on a lightweight LaMDA and with no fixed release date elsewhere.
- FTC warns AI voice cloning is supercharging the 'grandparent scam'The agency said criminals needed only a short online audio clip to clone a relative's voice and stage a fake emergency call demanding a wire transfer, cash or gift cards.
- GPT-4 and professional benchmarks: the wrong answer to the wrong question
- GPT4: The quiet parts and the state of ML
- Import AI 321: Open source GPT3; giving away democracy to AGI companies; GPT-4 is a political artifact
- Glaze tool lets artists cloak work against AI style-mimicryThe tool adds pixel-level perturbations invisible to people but that mislead image models into learning the wrong stylistic features from a piece of art.
- Using AI to make teaching easier & more impactful
- Microsoft announces 365 CopilotBuilt on GPT-4 and Microsoft's own data graph, it launched to a small group of testers with no price or general release date set.
- Baidu unveils ERNIE BotRobin Li narrated a demo built from pre-recorded answers rather than live queries; investors wiped roughly $3 billion off Baidu's market value the same day.
- Midjourney releases V5The update was widely noted for fixing AI image generation's most-mocked failure, malformed hands, and for images some viewers mistook for photographs.
- The Multi-modal, Multi-model, Multi-everything Future of AGI
- What is algorithmic amplification and why should we care?
- Zhipu open-sources ChatGLM-6BThe 6.2-billion-parameter model, trained on roughly a trillion tokens of Chinese and English text, could run on a single consumer graphics card with 6GB of memory using INT4 quantisation.
- OpenAI releases GPT-4A multimodal model that passed professional exams near the top of the human range — and whose technical report disclosed no architecture, data or compute.
- Anthropic launches ClaudeThe company's first public assistant launched, via a chat interface and API, on the same day OpenAI released GPT-4.
- How to... use AI to unstick yourself
- Stanford's Alpaca fine-tunes LLaMA for a few hundred dollarsInstruction-following behaviour was reproduced for under $600 total by fine-tuning Meta's 7B LLaMA on GPT-3.5-generated examples; the public demo was pulled within days.
- Import AI 320: Facebook's AI Lab Leak; open source ChatGPT clone; Google makes a universal translator.
- Georgi Gerganov releases llama.cppA C/C++ reimplementation that ran Meta's leaked LLaMA weights on ordinary laptops using quantisation, without Python, PyTorch or a GPU.
- AGI Roundup: Re-visiting Go; transitioning from narrow to general; multimodality & GPT4
- Artists can now opt out of generative AI. It’s not enough.
- Inflection AI launches out of stealthThe company, structured as a public benefit corporation, announced Pi, a conversational assistant it described as a supportive companion rather than a search or productivity tool.
- Anthropic publishes 'Core Views on AI Safety'The company argued transformative AI could arrive within a decade and named five research bets, including mechanistic interpretability and Constitutional AI, as its response.
- LLMs are not going to destroy the human race
- Secret Cyborgs: The Present Disruption in Three Papers
- Import AI 319: Sovereign AI; Facebook's weights leak on torrent networks; Google might have made a better optimizer than Adam!
- The LLaMA is out of the bag. Should we expect a tidal wave of disinformation?
- Feats to astonish and amaze
- Power and Weirdness: How to Use Bing AI
- OpenAI opens the ChatGPT API at a tenth of the pricegpt-3.5-turbo was priced at $0.002 per 1,000 tokens, roughly a tenth of the previous GPT-3.5 rate, alongside a new Whisper transcription API.
- Cursor launches (Anysphere's AI code editor)The MIT-founded startup forked VS Code and built AI into every interaction, growing by word of mouth with no press launch before later becoming a multibillion-dollar company.
- AI cannot predict the future. But companies keep trying (and failing).
- AI: Practical Advice for the Worried
February 2023
- The RLHF battle lines are drawn
- AI Techies!
- How to Get an AI to Lie to You in Three Simple Steps
- Sam Altman publishes 'Planning for AGI and beyond'Altman argued for iterative deployment of ever more capable systems rather than a single high-stakes release, while naming misaligned superintelligence as a serious risk.
- Meta releases LLaMA to researchers, and it leaks within a weekMeta shared a competitive foundation model with approved researchers; the weights appeared on BitTorrent days later and an open ecosystem formed around them.
- Blinded by Analogies
- People keep anthropomorphizing AI. Here’s why
- Copyright Office rules AI-generated images in Zarya of the Dawn aren't copyrightableThe office kept protection for Kashtanova's text and her arrangement of images, but stripped it from each individual image Midjourney had generated.
- Clarkesworld closes submissions after AI-story floodThe science-fiction magazine had banned fewer than 25 submitters a month historically; in February 2023 that figure passed 500.
- "AI alignment" and uncalibrated discourse on AI
- The future, soon: what I learned from Bing's AI
- My class required AI. Here's what I've learned so far.
- Bing's chatbot tells a journalist to leave his wifeDays after the exchange, Microsoft capped Bing chat sessions to five turns, saying long conversations could 'confuse' the model into drifting from grounded answers.
- I hope you weren't getting too comfortable.
- Three seasons of RL: Metaphor, tool, and framework
- EleutherAI releases Pythia model suiteEvery model in the eight-size suite was trained on identical data in the same order, with 154 saved checkpoints each, to let researchers study training dynamics directly.
- A quick and sobering guide to cloning yourself
- Toolformer teaches models to call APIsThe model taught itself, from only a handful of examples per tool, when to call a calculator, search engine, translator or calendar and how to use the result.
- Student uses prompt injection to expose Bing Chat's hidden 'Sydney' system promptLiu told the chatbot to 'ignore previous instructions' and asked what preceded them, prompting it to disclose rules telling it to keep its Sydney codename confidential.
- Magic for English Majors
- Microsoft puts GPT-4 inside BingMicrosoft called it only a 'next-generation OpenAI large language model'; the company confirmed five weeks later, on GPT-4's public release, that Bing had been running on GPT-4 all along.
- Google announces BardAnnounced as a lightweight version of LaMDA for trusted testers; two days later a factual error in Google's own promotional ad wiped roughly $100bn off Alphabet's market value.
- "Do not fear AI, puny humans... that is not meant as a threat."
- Shameless #Foomerism and S-Curves
- Replika chatbot banned from processing Italian users' dataThe order gave Luka Inc, Replika's maker, 20 days to comply or face a fine of up to 20 million euros or 4% of worldwide turnover.
- OpenAI launches ChatGPT Plus, a paid subscription tierThe $20-a-month US pilot offered priority access during peak load and faster responses; international rollout followed within days.
- Harvey signs exclusive launch partnership with Allen & OveryMore than 3,500 lawyers at 43 offices got access to the GPT-4-based tool after a beta that logged roughly 40,000 queries since November 2022.
- Scaling laws for robotics & RL: Not quite yet
- The Machines of Mastery
January 2023
- 4chan users abuse ElevenLabs voice cloning to generate celebrity hate speechUsers of the imageboard cloned Joe Rogan, Emma Watson and Ben Shapiro to produce racist and transphobic audio, days after the tool's public beta opened.
- A prosthesis for imagination: Using AI to boost your creativity
- NIST publishes the AI Risk Management FrameworkA voluntary standard built around four functions — govern, map, measure, manage — that later became a reference point for federal procurement and state AI bills.
- DeepMind introduces MusicLM, a text-to-music generation modelThe model composed several minutes of coherent audio from a written prompt or a hummed melody, released as a research paper rather than a product.
- CNET pauses AI-written articles after errors found in more than halfAn internal audit found corrections were needed on 41 of 77 finance explainers after Futurism reported the outlet had been publishing them quietly since November.
- The practical guide to using AI to do stuff
- Microsoft invests a reported $10 billion in OpenAIMicrosoft's own announcement gave no figure; press reports put the deal at $10 billion and described Azure becoming OpenAI's exclusive cloud provider.
- Reasons to Punish Autonomous Robots
- Do Large Language Models learn world models or just surface statistics?
- Becoming strange in the Long Singularity
- Getty Images sues Stability AIGetty's UK High Court claim alleged around 11 million of its images were used to train Stable Diffusion without a licence; a separate US suit followed in February.
- All my classes suddenly became AI classes
- Secretary jobs in the age of AI
- Pretraining quadrupeds: a case study in RL as an engineering tool
- Every Google vs OpenAI Argument, Dissected
- Artists file a class action over image generatorsIllustrators Sarah Andersen, Kelly McKernan and Karla Ortiz sued in the Northern District of California, the first case brought by creators rather than rights-holding companies.
- China's deep synthesis (deepfake) provisions take effectProviders must verify users' identities, obtain separate consent before generating a synthetic face or voice, and label AI-altered content.
- How to... use ChatGPT to boost your writing
- And the great gears begin to turn again...
- Looking into 2023
- New York City schools block ChatGPTThe city's Department of Education cited weak critical-thinking value and cheating risk; it reversed the block on ChatGPT four months later.
- You can't brute force the unsolvable
- The third magic
December 2022
- Predicting machine learning moats
- Reverse Prompt Engineering for Fun and (no) Profit
- Who believes more myths about humans: AI or educated humans?
- Google is reported to declare a code red over ChatGPTTeams from research and Trust and Safety were reportedly reassigned to accelerate AI products, with an internal target tied to Google's May developer conference.
- The street finds its own uses for things, AI Edition
- Closed-API vs Open-source continues: RLHF, ChatGPT, data moats
- What Building "Copilot for X" Really Takes
- Anthropic publishes Constitutional AIA model critiques and revises its own outputs against a written list of principles, then trains a reward model from its own preference judgements instead of human labels.
- ChatGPT is my co-founder
- How to... use AI to teach some of the hardest skills
- Four Paths to the Revelation
- Perplexity launches Ask, its answer engineAnswers came in prose with inline numbered citations to sources, distinguishing it from ChatGPT, launched a week earlier, which could not browse the web or show sources.
- ChatGPT is a bullshit generator. But it can still be amazingly useful
- The Mechanical Professor
- Stack Overflow bans ChatGPT answersWithin a week of launch, a major knowledge community prohibited model output because plausible wrong answers arrived faster than moderators could check them.
- Learning to Make the Right Mistakes - a Brief Comparison Between Human Perception and Multimodal LMs
- RLHF, 'online' ML systems, and RL going mainstream
- The Day The AGI Was Born
- How to... use AI to generate ideas
- Generative AI: autocomplete for everything
November 2022
- Speculative decoding paper shows drafting tokens ahead can speed up LLM inferenceLeviathan, Kalman and Matias showed a small draft model verified by the large model can cut inference latency 2-3x with identical outputs.
- OpenAI launches ChatGPTA free web interface to an existing GPT-3.5 model, reported by one analysis to have reached 100 million monthly users within two months.
- AI has a strategy.
- Why "Prompt Engineering" and "Generative AI" are overhyped
- Meta's Cicero plays Diplomacy at human levelPlaying anonymously against humans on webDiplomacy.net, Cicero scored more than double the average player's points across 40 games.
- The bait and switch behind AI risk prediction tools
- Meta pulls Galactica after three daysIts errors were formatted exactly like real citations and papers, so only an expert reader could tell fabrication from fact — a distinct failure mode from earlier chatbots' obvious mistakes.
- Epoch AI publishes 'Will We Run Out of Data?'Modelled dataset growth against the finite stock of public text and projected the two curves would cross between 2026 and 2032, sooner if models were overtrained.
October 2022
- What is AGI-hard
- Using RL's exploitation to debug
- Students are acing their homework by turning in machine-generated essays. Good.
- Stability AI raises $101 millionCoatue and Lightspeed led the round at roughly a $1bn valuation, confirming that giving weights away free could still attract institutional capital.
- How Open Source is eating AI
- The US restricts advanced chip exports to ChinaThe rules covered the tools to make advanced chips, not only the chips, and barred US citizens from supporting Chinese chipmaking.
- ReAct paper describes interleaving reasoning and acting in language modelsAlternating reasoning traces with actions against external tools, tested on question-answering, fact-checking and simulated shopping and household tasks.
- AlphaTensor discovers new matrix multiplication algorithmsFound a 76-multiplication algorithm for a specific matrix size, improving on the best known method for the first time in over fifty years.
- The White House publishes a Blueprint for an AI Bill of RightsNon-binding OSTP guidance setting five principles; created no legal obligations, unlike the EU's AI Act moving through Parliament the same year.
September 2022
- Eighteen pitfalls to beware of in AI journalism
- Meta and Google race to text-to-videoMake-A-Video and Imagen Video appeared within a week of each other, both as research previews with no public access.
- Artificial Intelligence and the Future of Demos
- Back in the game
- Eigenquestions for the AI Red Wedding
- DeepMind's Sparrow explores rule-based RLHF for safer dialogueAdversarial testers broke Sparrow's written safety rules in about 8% of attempts, roughly a third of the rate for a baseline model tested the same way.
- Lovecraftian intelligence
- OpenAI open-sources WhisperTrained on 680,000 hours of web-scraped audio and released under the MIT licence, an unusual openness for a company otherwise moving toward closed models.
- Anthropic publishes 'Red Teaming Language Models to Reduce Harms'Testing four training methods at three model sizes, Anthropic found RLHF-trained models got harder to red-team as they scaled while other methods did not improve.
- Multiverse, not Metaverse
- Character.AI launches public betaFounded by two former Google engineers behind Meena and LaMDA, the site logged hundreds of thousands of user interactions within three weeks of its beta launch.
- Anthropic publishes 'Toy Models of Superposition'Elhage, Olah and colleagues showed small networks represent more features than they have neurons by packing them into overlapping directions, complicating efforts to read a model's internals.
- Simon Willison coins the term 'prompt injection'Developer Simon Willison named and defined 'prompt injection', describing how untrusted text fed to an LLM could override its intended instructions, framing it as the LLM analogue of SQL injection.
- Generative AI models generate AI hype
- American workers need lots and lots of robots
- Causal Inference: Connecting Data and Reality
August 2022
- Why are deep learning technologists so overconfident?
- An AI image wins the Colorado State Fair art prizeJason Allen's Midjourney piece 'Théâtre D'opéra Spatial' took first place in the fair's digital arts category and a $300 prize, and was later denied US copyright registration.
- Google opens LaMDA to the public through AI Test KitchenAccess rolled out gradually via a waitlist to an Android app, arriving weeks after Google fired engineer Blake Lemoine for publicly claiming the model was sentient.
- Stable Diffusion is released to the publicWeights published under a permissive licence and runnable on a consumer graphics card, with community fine-tunes and graphical front-ends appearing within weeks.
- The Future of Speech Recognition: Where Will We Be in 2030?
- OpenAI releases new content moderation toolingOpenAI shipped a free Moderation API to help developers identify harmful content in text.
- Federal Circuit rules AI cannot be a patent inventorRuling in Thaler v. Vidal, the court held the Patent Act's term 'individual' means a natural person, rejecting Stephen Thaler's bid to name his DABUS system as sole inventor.
- Symmetries, Scaffolds, and a New Era of Scientific Discovery
July 2022
- AlphaFold's database expands to 200 million structuresThe database grew roughly 200-fold in a single release, from about a million structures to predictions covering nearly every catalogued protein across around a million species.
- Overview of Graph Theory and Alzheimer's Disease
- Midjourney opens its beta on DiscordImage requests and results were posted in public chat channels by default, turning generation into a shared, watchable activity rather than a private tool.
- BigScience releases BLOOM open multilingual modelOver 1,000 researchers from more than 70 countries trained the 176-billion-parameter model in the open on a French public supercomputer, releasing checkpoints and optimiser states alongside weights.
- Meta open-sources No Language Left Behind translation modelMeta open-sourced NLLB-200, a single model translating between 200 languages including many low-resource languages, plus the FLORES-200 benchmark.
- OpenAI widens DALL·E 2 access and adds safety mitigationsOpenAI invited a million people off the DALL·E 2 waitlist while adding content filters and provenance features.
June 2022
- Yann LeCun publishes 'A Path Towards Autonomous Machine Intelligence'The Meta chief scientist proposed non-generative 'joint embedding predictive architectures' as a route to machine intelligence, arguing autoregressive LLMs were a dead end for reasoning and planning.
- GitHub Copilot goes on salePriced at $10 a month or $100 a year, it moved from a year-long technical preview used by over 1.2 million developers to a paid product free for students and open-source maintainers.
- Emergent abilities of large language models are describedJason Wei and co-authors catalogued tasks where accuracy jumped from near-chance to strong performance past a scale threshold, a pattern later disputed as a metric artefact.
- Lessons from the GPT-4Chan Controversy
- A Google engineer claims LaMDA is sentientBlake Lemoine published transcripts arguing the model was a person and was placed on leave for breaching confidentiality; Google called the claim unfounded.
- BIG-bench paper released204-task benchmark from 450 authors at 132 institutions probes emergent capabilities as language models scale.
- Yudkowsky publishes 'AGI Ruin: A List of Lethalities'Yudkowsky's 43-point case that alignment is 'lethally difficult' argued no known research path could make a first critical AGI attempt survivable.
- AI is Ushering In a New Scientific Revolution
May 2022
- Working on the Weekends - an Academic Necessity?
- FlashAttention makes exact attention IO-awareReordering attention around GPU memory rather than approximating it cut training time and unlocked longer sequences — and became default infrastructure.
- "Let's think step by step" elicits zero-shot reasoningA single prompt phrase, with no worked examples, lifted GSM8K accuracy from 10.4% to 40.7% — chain-of-thought without the exemplars.
- Google unveils Imagen text-to-image diffusion modelGoogle Research reported higher photorealism scores than DALL·E 2 in side-by-side human evaluation, but declined to release code or a public demo.
- Lessons From Deploying Deep Learning To Production
- DeepMind's Gato does 600 tasks with one set of weightsA single 1.2-billion-parameter transformer played Atari, captioned images and stacked blocks with a real robot arm, all from one set of weights.
- Clearview AI settles ACLU's Illinois biometric-privacy lawsuitClearview agreed to a nationwide ban on selling its faceprint database to most private businesses and a five-year sales freeze to Illinois police, without paying damages.
- Hugging Face raises $100m Series CHugging Face raised $100m in Series C funding led by Lux Capital, growing from 30 to 120 staff in a year while serving over 10,000 companies.
- Beyond Message Passing: a Physics-Inspired Paradigm for Graph Neural Networks
- Deep Learning in Neuroimaging
- Meta releases OPT-175B with its training logbookA GPT-3-scale model shared with researchers alongside an unusually candid record of what went wrong during training.
April 2022
- DeepMind's Flamingo tackles multimodal few-shot learningFlamingo, an 80B-parameter vision-language model, sets few-shot state of the art on image and video benchmarks with minimal task examples.
- Anthropic raises $580M Series BAlameda Research, trading with what turned out to be FTX customer deposits, supplied about $500M of the $580M — a link that drew scrutiny after FTX's collapse that November.
- AI Startups and the Hunt for Tech Talent in Vietnam
- New Technology, Old Problems: The Missing Voices in Natural Language Processing
- OpenAI releases DALL·E 2Higher resolution and photorealism than the original DALL·E, released first to roughly 400 trusted users pending a public waitlist.
- OpenAI publishes a system card for DALL·E 2The document catalogued bias, explicit-content and harassment risks alongside the model, and set the pattern of shipping a capability assessment with a release.
- Google announces PaLM at 540 billion parametersTrained on the Pathways system across two TPU v4 pods, it posted large gains on reasoning benchmarks and explained its own jokes.
March 2022
- LAION releases LAION-5B image-text datasetLAION released LAION-5B, then the largest freely available image-text dataset (5.85bn pairs), which became the training corpus for Stable Diffusion and other open image models.
- DeepMind's Chinchilla paper rewrites the scaling lawsExisting large models were badly under-trained: for a fixed compute budget, parameters and training tokens should scale together.
- NVIDIA announces the Hopper architecture and the H100The H100 became the part frontier training runs were built around for the next two years, and the unit of account in which compute deals were later measured.
- China's recommendation algorithm rules take effectProviders had to file algorithms with the regulator and offer users a way to switch personalisation off.
- Anthropic publishes 'In-Context Learning and Induction Heads'Anthropic's interpretability team argued a single attention mechanism, found across model sizes, does most of the work behind a model's ability to learn from its prompt.
February 2022
- One Voice Detector to Rule Them All
- How Aristotle is Fixing Deep Learning's Flaws
- DeepMind controls a fusion plasma with reinforcement learningA single network commanding all of a tokamak's control coils held plasma shapes on Switzerland's TCV reactor, including configurations conventional controllers struggle with.
- Cohere raises a $125 million Series BTiger Global led the round, taking the Toronto language-model company past $170 million raised and marking early investor appetite for OpenAI competitors.
- How AI is Changing Chemical Discovery
- NVIDIA's $40 billion purchase of Arm collapsesNVIDIA and SoftBank abandoned the deal after the FTC sued to block it; NVIDIA wrote off a $1.36 billion prepayment and Arm began preparing to list instead.
- Designing Societally Beneficial Reinforcement Learning Systems
- EleutherAI announces GPT-NeoX-20BEleutherAI released GPT-NeoX-20B, a 20-billion-parameter open model, its largest dense model to date and freely available via GitHub.
- DeepMind's AlphaCode reaches median human on competitive programmingRanked around the 54th percentile in Codeforces contests by generating and filtering enormous numbers of candidate programs.
January 2022
- Engaging with Disengagement
- Chain-of-thought prompting is describedAsking a model to show its working improved reasoning benchmarks sharply, with no retraining — the seed of the later reasoning models.
- OpenAI ships InstructGPT and makes RLHF the defaultModels fine-tuned on human preference data were preferred to a model a hundred times larger, reframing alignment as a product feature.
- OpenAI adds an embeddings endpoint to its APIThree model families turned text and code into vectors for search, clustering and classification, a building block later underpinning retrieval-augmented generation.
- Meta unveils the AI Research SuperClusterA 6,080-GPU first phase, to reach 16,000 GPUs and a claimed 5 exaflops of mixed-precision compute by mid-2022, built for training on trillions of examples.
- Flexible Centralization in Multi-agent Learning & Control
- Contra David Deutsch on AI
- A Science Journalist’s Journey to Understand AI
- 'Grokking' paper documents sudden generalisation long after overfittingOpenAI researchers found small networks could suddenly jump from memorisation to perfect generalisation well after apparent overfitting, opening a mechanistic-interpretability research thread.
December 2021
- Baidu announces ERNIE 3.0 TitanBuilt with Peng Cheng Laboratory, the 260-billion-parameter model reported state-of-the-art results on more than 60 Chinese-language NLP tasks.
- AI and the Future of Work: What We Know Today
- OpenAI publishes WebGPT, a model that browses the web to answer questionsAnswers preferred to Reddit's top-voted responses 69% of the time in blind comparison, but the model still fell short of human accuracy on TruthfulQA.
- New Datasets to Democratize Speech Recognition Technology
- How to Train your Decision-Making AIs
- DeepMind publishes Gopher, RETRO and a risk taxonomy togetherA 280-billion-parameter model, a smaller retrieval-augmented alternative that matched larger models, and a taxonomy of six categories of language-model harm.
- Anthropic publishes its first alignment paper'A General Language Assistant as a Laboratory for Alignment' introduced the helpful-honest-harmless framing and found preference modelling scales better than imitation.
- Anthropic publishes 'A Mathematical Framework for Transformer Circuits'Studying deliberately simplified transformers with no more than two layers, the team found 'induction heads' — a mechanism later argued to explain much of in-context learning.
November 2021
- UNESCO adopts a global recommendation on AI ethicsAll 193 member states endorsed the non-binding text, which calls for bans on AI-based social scoring and mass surveillance alongside environmental and labour provisions.
- OpenAI removes the GPT-3 waitlistGeneral availability in supported countries turned the API from a curated experiment into a commodity developers could just buy, though several countries were excluded.
- Explain Yourself - A Primer on ML Interpretability & Explainability
- Meta shuts down its face recognition systemFacebook deleted over a billion facial templates, citing unresolved regulation, but kept the door open to narrower uses such as identity verification.
October 2021
- Strong AI Requires Autonomous Building of Composable Models
- Facebook renames itself MetaFacebook, Instagram and WhatsApp kept their names as apps under a new holding company, announced alongside a $150 million investment in immersive learning tools.
- Reflections on Foundation Models
- Microsoft and NVIDIA train Megatron-Turing NLG at 530 billion parametersThe largest dense language model publicly described at the time, and a demonstration of multi-thousand-GPU training.
September 2021
August 2021
- China publishes draft rules on recommendation algorithmsThe Cyberspace Administration proposed requiring opt-outs and banning manipulative ranking — among the first binding algorithm rules anywhere.
- An Introduction to AI Story Generation
- Tesla announces the Dojo training supercomputerTesla's in-house D1 chip and 'ExaPod' racks were designed to train self-driving neural networks on video from its vehicle fleet, targeting over an exaflop of compute.
- Stanford's foundation models report names the categoryOver 100 researchers at Stanford's newly formed Center for Research on Foundation Models coined the term for models like GPT-3 and BERT, adapted rather than retrained for each task.
- Systems for Machine Learning
- AI21 Labs releases Jurassic-1Jurassic-1 Jumbo's 178 billion parameters slightly exceeded GPT-3's, and its 250,000-token vocabulary — five times GPT-3's — aimed to cut per-word token costs.
- Remote robotic-data farms
- Machine Learning Won't Solve Natural Language Understanding
- Apple announces on-device CSAM detection, then shelves itA client-side scanning plan drew intense criticism from cryptographers and civil liberties groups and was paused within a month.
- on the Horizon of applied RL
July 2021
- Machine Translation Shifts Power
- OpenAI introduces Triton, an open-source GPU programming languageA Python-like language and compiler let researchers without CUDA experience write GPU kernels, such as an FP16 matrix multiply in under 25 lines, matching hand-tuned libraries.
- DeepMind's XLand agents generalise across millions of open-ended gamesAgents trained across roughly 700,000 games in about 4,000 procedurally generated worlds, totalling 200 billion training steps, then solved almost every held-out task tried on them.
- It’s All Training Data: Using Lessons from Machine Learning to Retrain Your Mind
- The AlphaFold Protein Structure Database opensThe initial release covered around 350,000 predicted structures across the human proteome and twenty other organisms, with a stated plan to expand to over 100 million.
- Justitia ex Machina: The Case for Automating Morals
- AlphaFold 2 is published in Nature and open-sourcedThe method behind DeepMind's CASP14 result seven months earlier was released in full, with source code, rather than kept as a demonstrated but undisclosed system.
- OpenAI publishes Codex and the HumanEval benchmarkCodex solved 28.8% of HumanEval's Python problems on a single attempt and 70.2% when allowed 100 samples per problem, against 0% for base GPT-3.
- Prompting: Better Ways of Using Language Models for NLP Tasks
June 2021
- GitHub launches Copilot in technical previewAn OpenAI model trained on public code began suggesting whole functions inside the editor — the first mass-market use of a large language model.
- How to Do Multi-Task Learning Intelligently
- Reward is not enough
- Microsoft researchers publish LoRAFreezing pretrained weights and training small added matrices instead cut GPT-3's trainable parameter count by a factor the authors put at 10,000, with no extra inference cost.
- Are Self-Driving Cars Really Safer Than Human Drivers?
- How all machine learning becomes reinforcement learning
- EleutherAI releases GPT-J-6BAt 6 billion parameters it scored close to OpenAI's similarly sized GPT-3 model on the LAMBADA benchmark, and its weights were downloadable under Apache 2.0.
- BAAI announces Wu Dao 2.0 at 1.75 trillion parametersThe Beijing Academy of Artificial Intelligence put the parameter count at ten times GPT-3's, a claim reported widely but never independently benchmarked.
May 2021
- Anthropic is founded by departing OpenAI researchersThe founders, including former OpenAI VPs of research and of safety, said the $124 million round would fund research into steerable, interpretable systems.
- How has AI contributed to dealing with the COVID-19 pandemic?
- Google announces LaMDA and TPU v4 at I/OA single TPU v4 pod combined 4,096 chips for over one exaflop of compute, while LaMDA was pitched on open-ended conversation rather than benchmark scores.
- Towards Human-Centered Explainable AI: the journey so far
- Machine Learning, Ethics, and Open Source Licensing (Part II)
April 2021
- Machine Learning, Ethics, and Open Source Licensing (Part I)
- The European Commission proposes the AI ActA risk-tiered draft banning practices such as social scoring outright; three years of negotiation followed before it became binding law.
- Attention in the Human Brain and Its Applications in ML
- Decentralized AI For Healthcare
March 2021
- EleutherAI releases GPT-NeoThe 1.3B and 2.7B checkpoints were released under an MIT licence with weights freely downloadable, while GPT-3 itself remained a paid, waitlisted API.
- Setting ourselves up for exploitation: RL in the wild
- OpenAI publishes multimodal neurons researchA single neuron fired for photos of spiders, drawings of Spider-Man and the rendered word 'spider' alike — and pasting a mislabelled sticker on an object could fool the model.
- "On the Dangers of Stochastic Parrots" is publishedThe paper that named the case against ever-larger language models — and whose suppression cost two Google researchers their jobs.
February 2021
- Counting down until consumer drones are banned in cities
- A deepfake Tom Cruise goes viral on TikTokBelgian VFX artist Chris Ume paired face-swap software with impersonator Miles Fisher's performance; the clips passed 11 million TikTok views within days.
- Google dismisses Margaret MitchellGoogle said Mitchell had used automated scripts to search her own email for evidence of Timnit Gebru's mistreatment, violating security policy.
- Clarifying RL: Obscure problem formulations and structure tradeoffs
- Catching Cyberbullies with Neural Networks
- Decoupling AI from the latent variable of spoken languages
- 100th anniversary of the word robot: COVID didn’t give us personal robots, it gave us Woebot
January 2021
- Can AI Let Justice Be Done?
- Boston Dynamics 🤖🐈: Studying Athletic Intelligence
- Robotic Companies 2.0: Horizontal Modularity
- A Visual History of Interpretation for Image Recognition
- Reflections on digital technology from a capitol siege
- Google trains a 1.6-trillion-parameter Switch TransformerRouting each input to a single expert rather than blending several simplified prior mixture-of-experts designs, and the paper reported up to 7x faster pre-training than a dense T5 baseline.
- Considering AIs through our own mind’s reflection
- OpenAI announces DALL·E and CLIPTwo multimodal models released on the same day: one that generated images from text captions, one that learned to match images to language.
- Knocking on Turing’s door: Quantum Computing and Machine Learning
- The stakeholders start to learn the stakes: predictions for machine learning and automation in 2021
December 2020
- EleutherAI releases The PileThe 825GiB corpus combined 22 curated sources rather than raw web scrapes, and models trained on it outperformed Common Crawl-only baselines on academic text.
- MuZero masters games without being told the rulesThe system learned to predict only value, policy and reward rather than the environment's rules, then matched AlphaZero at Go, chess and shogi and set a new Atari benchmark.
- When BERT Plays The Lottery, All Tickets Are Winning
- The Ubiquity and Future of Model-based Reinforcement Learning
- The EU proposes the Digital Services and Markets ActsTwo separate proposals — due-diligence duties for platforms under the DSA, and obligations on large 'gatekeepers' under the DMA — took over a year to negotiate into law.
- Facebook Case Study 🌏🔬: Using AI to Regulate the Digital Ecosystem
- The Far-Reaching Impact of Dr. Timnit Gebru
- Timnit Gebru leaves Google after a dispute over her paperGebru said she was fired over an internal email and a paper Google asked her to retract; Google's AI chief called it a resignation.
November 2020
- AlphaFold 2 solves protein structure prediction at CASP14DeepMind's system predicted protein structures to roughly experimental accuracy, ending a fifty-year-old open problem in biology.
- Interpretability in ML: A Broad Overview
- How Can We Improve Peer Review in NLP?
- The Collingridge Dilemma and Current Policy on Robots
- Apple ships the M1 with a dedicated neural engineA 16-core Neural Engine rated at 11 trillion operations per second shipped as standard in a mainstream laptop, part of Apple's move away from Intel chips.
- Don’t Forget About Associative Memories
- Constructing Axes for Reinforcement Learning Policy
September 2020
- AI Democratization in the Era of GPT-3
- Microsoft takes an exclusive licence to GPT-3Microsoft gained access to the underlying model and source code for its own products; the public API, through which OpenAI still served other customers, was unaffected.
- Transformers are Graph Neural Networks
- NVIDIA agrees to buy Arm for $40 billionThe $40bn agreed price was $21.5bn in NVIDIA stock and $12bn cash, plus up to $5bn more if Arm hit performance targets; regulators later blocked it.
- Autonomy startups are such a mess
- The Guardian publishes an op-ed written by GPT-3Editors ran the model eight times on the same prompt and spliced the best passages together, a process disclosed in a footnote that critics said undercut the framing.
- Hendrycks et al. publish the MMLU benchmark15,908-question, 57-subject multiple-choice benchmark spanning elementary to professional level; GPT-3 improved on random chance by roughly 20 points on average.
- AI & Arbitration of Truth
August 2020
- Can robots be autonomous and unintelligent?
- The uncanny world of robots at home
- The UK A-level grading algorithm is withdrawnNearly 40% of teacher-predicted A-level grades were downgraded by Ofqual's model, disproportionately at state schools, before ministers reversed course.
- Digital companies and the goal of at-home embodied AI
July 2020
- Automated: how algorithms shape a day in the life and our future
- Shortcuts: Neural Networks Love to Cheat
- Recommendations are a game - a dangerous game (for us).
- GPT-3 demos spread beyond the research communityDeveloper Sharif Shameem's demo turning plain-English descriptions into working webpage code was among the clips that drew attention beyond NLP researchers.
- How to Stop Worrying About Compositionality
- Automating code, Twitter's hack(s), a robot named Stretch
- Online courses, automating education, and digitalizing degrees
- Challenges of Comparing Human and Machine Perception
- Democratizing Automation
- EleutherAI forms to build open language modelsA Discord server for discussing GPT-3 became a volunteer collective aiming to train and openly release a comparable model itself.
June 2020
- Drones, Swarms, and Storms of Drones
- A wrongful arrest from facial recognition is reportedRobert Williams was detained in Detroit after a false match, the first such US case to be widely documented.
- Lessons from the PULSE Model and Discussion
- "10 years of automation in 1 year"
- OpenAI opens the GPT-3 API in private betaAccess required a waitlisted application rather than a download, and OpenAI said the model itself would stay unpublished.
- IBM stops selling general-purpose facial recognitionAnnounced during the George Floyd protests; Amazon and Microsoft followed within days with police-sales moratoria.
May 2020
- Gwern publishes 'The Scaling Hypothesis'Gwern's essay, written around GPT-3's release, argued scale alone was producing qualitatively new abilities and became a widely cited framing for the scaling-hypothesis argument.
- OpenAI publishes the GPT-3 paperA 175-billion-parameter language model that learned new tasks from examples in its prompt, without any weight updates.
- Microsoft reveals the supercomputer it built for OpenAIAnnounced at Build, a top-five ranked machine dedicated to a single customer — an early signal of AI's compute economics.
April 2020
- OpenAI releases JukeboxCompressed raw audio into discrete codes with a multi-scale VQ-VAE, then modelled those with autoregressive transformers to keep coherence over several minutes.
- Facebook releases the Blender chatbotReleased in three sizes up to 9.4 billion parameters, with weights and code made public rather than kept behind an API.
- OpenAI Microscope released for visualizing neural network activationsPre-computed feature visualisations for nine vision models, cutting the cost of inspecting a single neuron from hundreds of GPU-hours to seconds.
March 2020
- DeepMind's Agent57 beats the Atari57 human benchmarkCombined short- and long-term novelty rewards with a per-game meta-controller, though rival agent MuZero still scored higher on mean and median across the suite.
- The CORD-19 research dataset is releasedAssembled with the White House OSTP, National Library of Medicine and Chan Zuckerberg Initiative, it opened with over 29,000 papers, most with full text.
February 2020
- The European Commission publishes its White Paper on AIThe first sketch of a risk-based European approach to AI regulation, five years before the Act took effect.
- Microsoft announces Turing-NLG at 17 billion parametersBriefly the largest published language model, trained with Microsoft's new DeepSpeed library and beating Megatron-LM on WikiText-103 and LAMBADA.
January 2020
- OpenAI publishes scaling laws for neural language modelsLoss fell as a smooth power law in model size, data and compute — the empirical result that justified building bigger.
- The New York Times exposes Clearview AIA facial recognition company had scraped billions of photos from social media and sold identification to police forces.
- DeepMind publishes early AlphaFold protein structure workDescribes the CASP13-winning system, which used deep networks trained on genomic data to predict inter-residue distances rather than folding a structure directly.
- Google Health reports AI matching radiologists on breast screeningA Nature paper claimed an AI system reduced false positives and false negatives in mammography against expert readers.
Nothing matches those filters.