Nathan Lambert

Commentary by Nathan Lambert

From the commentary rail — every link leaves the site for the original piece.

• 6 October 2026 · InterconnectsThe Cyber Risk Discourse is BrokenOpen-weights, ideology, and acknowledging trade-offs.
• 21 September 2026 · InterconnectsThe current balance of power in open modelsThe expanded form of a testimony I prepared for Congress.
• 19 September 2026 · InterconnectsWhy I still haven’t bought into true RSIAn “AI moderate’s” view on recent events and the trajectory of frontier models.
• 10 September 2026 · InterconnectsOne resignation turned the embers of AI fear into a wildfireSome quick notes on a truly weird week.
• 9 September 2026 · InterconnectsWhen will average people feel AI’s impact?We’re <5 years into a compounding revolution which could take a century, and how the AI industry should manage this.
• 17 August 2026 · InterconnectsTeaching Everyone to Fish for TokensNvidia wants you building your own model, not buying from Anthropic/OpenAI.
• 14 August 2026 · InterconnectsGLM-5.3: How Chinese labs keep stride with the frontierHint: It’s really not a distillation story.
• 12 August 2026 · InterconnectsI wrote an AI textbook — how long until AI can do it better?Reflections on AI's writing ability and how AI models get more capable.
• 20 July 2026 · InterconnectsKimi K3: The open-weights escalationThe global implications on the AI ecosystem.
• 22 June 2026 · InterconnectsGLM-5.2 is the step change for open agentsA capability threshold I've been carefully monitoring.
• 19 June 2026 · InterconnectsBanning Open Source AI Would Be A MistakeThis post was originally an op-ed co-authored with Kevin Xu of Interconnected for a general, non-technical audience.
• 9 June 2026 · InterconnectsClaude Fable 5 and new AI safety fablesOne step further into the power politics of frontier AI systems.
• 1 June 2026 · InterconnectsOpen and closed models are on different exponentialsWhere marginally higher intelligence drives value, and where it doesn't.
• 26 May 2026 · InterconnectsSome ideas for what comes next, May 2026Gemini Flash 3.5, Mythos, open-closed balance, America's open-source surge, emerging power struggles and more.
• 7 May 2026 · InterconnectsNotes from inside China's AI labsLessons from my trip to talk to most of the leading AI labs in China.
• 4 May 2026 · InterconnectsThe distillation panic‘Distillation attacks’ is a horrible term for what is happening right now.
• 15 April 2026 · InterconnectsMy bets on open models, mid-2026What I expect to come next and why, focused on the open-closed gap.
• 11 April 2026 · InterconnectsThe inevitable need for an open model consortiumAnd yes, I hate consortia too.
• 9 April 2026 · InterconnectsClaude Mythos and misguided open-weight fearmongeringAnother dance around fears of open-source.
• 3 April 2026 · InterconnectsGemma 4 and what makes an open model succeedHint: it's not benchmark scores.
• 22 March 2026 · InterconnectsLossy self-improvementWhy self-improvement is real but it doesn't lead to fast takeoff.
• 18 March 2026 · InterconnectsGPT 5.4 is a big step for CodexOn evaluating and understanding the frontier of agents, and why I still turn to Claude.
• 16 March 2026 · InterconnectsWhat comes next with open modelsMarkets, capabilities, cope, and bewilderment in the industrialization of language models.
• 5 March 2026 · InterconnectsOlmo Hybrid and future LLM architecturesThe latest Olmo model and discussions at the frontier of open-source post training tools.
• 24 February 2026 · InterconnectsHow much does distillation really matter for Chinese LLMs?Reacting to Anthropic's post on "distillation attacks."
• 9 February 2026 · InterconnectsOpus 4.6, Codex 5.3, and the post-benchmark eraOn comparing models in 2026.
• 30 January 2026 · InterconnectsThoughts on the job market in the age of LLMsOn standing out and finding gems.
• 21 January 2026 · InterconnectsGet Good at AgentsThe tools are getting so powerful that we need to change how we scope, manage, and approach our work.
• 11 January 2026 · InterconnectsUse multiple modelsThe meta for getting the most out of AI in 2026.
• 9 January 2026 · InterconnectsClaude Code Hits DifferentCoding agents cross a meaningful threshold with Opus 4.5.
• 7 January 2026 · Interconnects8 plots that explain the state of open modelsMeasuring the impact of Qwen, DeepSeek, Llama, GPT-OSS, Nemotron, and all of the new entrants to the ecosystem.
• 20 November 2025 · InterconnectsOlmo 3: America’s truly open reasoning modelsWe present Olmo 3, our next family of fully open, leading language models.
• 16 November 2025 · InterconnectsWhy AI writing is midHow the current way of training language models destroys any voice (and hope of good writing).
• 6 November 2025 · Interconnects5 Thoughts on Kimi K2 ThinkingQuick reactions to another fantastic open model from a rapidly rising Chinese lab.
• 20 October 2025 · InterconnectsHow to scale RLA bombshell paper from Meta and friends.
• 7 October 2025 · InterconnectsThoughts on The CurveThe conference and the trajectory.
• 30 September 2025 · InterconnectsChatGPT: The Agentic AppChatGPT's long awaited move into user monetization and what it shows about the future of ChatGPT (and AI products writ large).
• 22 September 2025 · InterconnectsThinking, Searching, and ActingA reflection on reasoning models.
• 18 September 2025 · InterconnectsCoding as the epicenter of AI progress and the path to general agentsGPT-5-Codex, adoption, denial, peak performance, and everyday gains.
• 9 September 2025 · InterconnectsOn China's open source AI trajectoryNotes and predictions for the next phase of Chinese (open) AI models.
• 17 August 2025 · InterconnectsRanking the Chinese Open Model BuildersFrom the obvious names to those to keep an eye on.
• 15 August 2025 · InterconnectsContra Dwarkesh on Continual LearningDon't try to make your airplane too much like a bird.
• 7 August 2025 · InterconnectsGPT-5 and the arc of progressOverpromising will always lead to some sort of underdelivering, but what we're getting is still phenomenal.
• 5 August 2025 · Interconnectsgpt-oss: OpenAI validates the open ecosystem (finally)OpenAI's first open language model release since GPT 2 and what it means for the ecosystem.
• 4 August 2025 · InterconnectsTowards American Truly Open Models: The ATOM ProjectRebranding American DeepSeek into a more lasting brand. From an idea to a coalition with real impact.
• 23 July 2025 · InterconnectsThe White House's plan for open models & AI research in the U.S.Thoughts on the new AI Action plan, American DeepSeek, and what comes next.
• 14 July 2025 · InterconnectsKimi K2 and when "DeepSeek Moments" become normalOne "DeepSeek Moment" wasn't enough for us to wake up, hopefully we don't need a third.
• 4 July 2025 · InterconnectsThe American DeepSeek ProjectWhat I think the next goal for the open-source AI community is.
• 28 June 2025 · InterconnectsIlya on deep learning in 2015On vision and how to understand deep learning.
• 23 June 2025 · InterconnectsSome ideas for what comes next (Jun. 2025)As releases slow down, it's time to think about what we got this year and where we are going. o3's search, agent vs model progress, and scaling's settling.
• 12 June 2025 · InterconnectsThe rise of reasoning machinesAnd a debate that doesn't warrant repeating.
• 9 June 2025 · InterconnectsWhat comes next with reinforcement learningScaling RL, sparse rewards, continual learning, and the progress wall when pretraining really stops.
• 4 June 2025 · InterconnectsA taxonomy for next-generation reasoning modelsWhere we've been and where we're going with RLVR.
• 27 May 2025 · InterconnectsClaude 4 and Anthropic's bet on codeReasons to be optimistic and pessimistic on Anthropic's future.
• 27 May 2025 · InterconnectsReinforcement learning with random rewards actually works with Qwen 2.5Making sense of research casting doubt on the potential of RLVR and where I'm optimistic for the next phase of scaling.
• 21 May 2025 · InterconnectsPeople use AI more than you thinkAnd businesses too. The most important trend in AI that gets washed away from between the headlines.
• 6 May 2025 · InterconnectsWhat people get wrong about the leading Chinese open models: Adoption and censorshipNarrative violations on licenses, adoption, and censorship.
• 30 April 2025 · InterconnectsState of play of AI progress (and related brakes on an intelligence explosion)Why I don't think AI 2027 is going to come true.
• 28 April 2025 · InterconnectsQwen 3: The new open standardA wonderful release, base models, reasoners, model size scales, and all before LlamaCon.
• 28 April 2025 · InterconnectsTransparency and (shifting) priority stacksWhat you want to be open says a lot about your ranked priorities.
• 19 April 2025 · InterconnectsOpenAI's o3: Over-optimization is back and weirder than everTools, true rewards, and a new direction for language models.
• 14 April 2025 · InterconnectsOpenAI's GPT-4.1 and separating the API from ChatGPTOpenAI's latest models optimizing on intelligence per dollar. We'll continue to see ChatGPT handled differently than the API business.
• 7 April 2025 · InterconnectsLlama 4: Did Meta just push the panic button?One of the weirdest releases of the year and understanding the future of the Llama endeavor. For the time being, we have some more amazing open weight models!
• 5 April 2025 · InterconnectsRL backlog: OpenAI's many RLs, clarifying distillation, and latent reasoningNotes I forgot to publish. Closing some loose ends in the reasoning model discussions.
• 31 March 2025 · InterconnectsRecent reasoning research: GRPO tweaks, base model RL, and data curationThe papers I endorse as worth reading among a cresting wave of reasoning research.
• 26 March 2025 · InterconnectsGemini 2.5 Pro and Google's second chance with AIPlus some coverage for the latest DeepSeek.
• 19 March 2025 · InterconnectsManaging frontier model training organizations (or teams)How do the frontier labs consistently train great models? How can they fail?
• 13 March 2025 · InterconnectsGemma 3, OLMo 2 32B, and the growing potential of open-source AILeading open-weight models and the first open-source model to clearly surpass GPT 3.5 (the very last version).
• 10 March 2025 · InterconnectsElicitation, the simplest way to understand post-trainingAn F1 analogy to help understand fast improvements in post-training on top of slow improvements in scaling.
• 5 March 2025 · InterconnectsWhere inference-time scaling pushes the market for AI companiesFundamentals emerging downstream from the RL reasoning models.
• 28 February 2025 · InterconnectsGPT-4.5: "Not a frontier model"?OpenAI's latest model raises more questions than answers, but no, the AI bubble isn't popping quite yet.
• 26 February 2025 · InterconnectsCharacter training: Understanding and crafting a language model's personalityPost-training in industry is very different than the academic papers and open-source models demonstrate. Let's dive into one of my favorite topics in language modeling development today.
• 24 February 2025 · InterconnectsClaude 3.7 thonks and what's next for inference-time scaling(The classic lunch break post) The latest reasoning model and what it says about the direction of inference time compute and RL training.
• 18 February 2025 · InterconnectsGrok 3 and an accelerating AI roadmapWhere AI is heading, why 2024 felt slow, and shifting priorities of frontier laboratories.
• 12 February 2025 · InterconnectsDeep Research, information vs. insight, and the nature of scienceWhat AI will accelerate in the scientific process, what it cannot do, and how we can prepare for new manners of scientific investigation.
• 5 February 2025 · InterconnectsMaking the U.S. the home for open-source AIOpen-source AI is here to stay, but it is not a given that it will be American.
• 28 January 2025 · InterconnectsWhy reasoning models will generalizePeople underestimate the long-term potential of “reasoning.”
• 21 January 2025 · InterconnectsDeepSeek R1's recipe to replicate o1 and the future of reasoning LMsYes, ring the true o1 replication bells for DeepSeek R1 🔔🔔🔔. Where we go next.
• 15 January 2025 · InterconnectsLet me use my local LMs on Meta Ray-BansDifferent futures for AI devices and how we can start creating positive feedback loops for some types of open-weight language models.
• 9 January 2025 · InterconnectsDeepSeek V3 and the actual cost of training frontier AI modelsThe $5M figure for the last training run should not be your basis for how much frontier AI models cost.
• 20 December 2024 · InterconnectsOpenAI's o3: The grand finale of AI in 2024A step change as influential as the release of GPT-4. Reasoning language models are the current and next big thing.
• 18 December 2024 · InterconnectsThe AI agent spectrumSeparating different classes of AI agents from a long history of reinforcement learning.
• 11 December 2024 · InterconnectsOpenAI's Reinforcement Finetuning and RL for the massesThe cherry on Yann LeCun’s cake has finally been realized.
• 4 December 2024 · InterconnectsOpenAI's o1 using "search" was a PSYOPHow to understand OpenAI's o1 models as really just one wacky, wonderful, long chain of thought.
• 26 November 2024 · InterconnectsOLMo 2 and building effective teams for training language modelsAnnouncing OLMo 2 - the best open-source models yet - and some of the things I've been learning about training good language models.
• 21 November 2024 · InterconnectsTülu 3: The next era in open post-trainingWe give you open-source, frontier-model post-training.
• 14 November 2024 · InterconnectsScaling realitiesBoth stories are true. Scaling still works. OpenAI et al. still have oversold their promises.
• 13 November 2024 · InterconnectsSaving the National AI Research Resource & my AI policy outlookAs domestic AI policy gets reset, we need to choose our battles when recommending keeping the Biden administration’s work.
• 30 October 2024 · InterconnectsWhy I build open language modelsReflections after a year at the Allen Institute for AI and on the battlefields of open-source AI.
• 23 October 2024 · InterconnectsClaude's agentic future and the current state of the frontier modelsHow Claude's computer use works. Where OpenAI, Anthropic, and Google all have a lead on eachother.
• 16 October 2024 · InterconnectsBuilding on evaluation quicksandOn the state of evaluation for language models.
• 9 October 2024 · InterconnectsHow scaling changes model behaviorSome trends are reasonable to extrapolate, some are not. Even for the trends we are succeeding at extrapolating, it is not clear how that signal translates into different AI behaviors.
• 2 October 2024 · InterconnectsAI Safety Culture Confronts CapitalismSB1047’s veto, OpenAI’s turnover, and a constant treadmill pushing AI startups to be all too similar to big technology name brands.
• 25 September 2024 · InterconnectsLlama 3.2 Vision and Molmo: Foundations for the multimodal open-source ecosystemOpen models, tools, examples, limits, and the state of training multimodal models.
• 16 September 2024 · InterconnectsReverse engineering OpenAI’s o1What productionizing test-time compute shows us about the future of AI. Exploration has landed in language model training.
• 11 September 2024 · InterconnectsFutures of the data foundry business modelScale AI’s future versus further scaling of language model performance. How Nvidia may take all the margins from the data market, too.
• 9 September 2024 · InterconnectsA post-training approach to AI regulation with Model SpecsAnd why the concept of mandating “model spec’s” could be a good start.
• 5 September 2024 · InterconnectsOpenAI’s Strawberry, LM self-talk, inference scaling laws, and spending more on inferenceWhether or not scaling works, we should spend more on inference.
• 4 September 2024 · InterconnectsOLMoE and the hidden simplicity in training better foundation modelsAi2 released OLMoE, which is probably our “best” model yet relative to its peers, but not much has changed in the process.
• 28 August 2024 · InterconnectsOn the current definition of open-source AI and the state of the data commonsThe Open Source Initiative (OSI) is working towards a definition.
• 16 August 2024 · InterconnectsOn Nous Hermes 3 and classifying a "frontier model"The latest model from one of the most popular fine-tuning labs makes us question how a model should be identified as a “frontier model.”
• 7 August 2024 · InterconnectsA recipe for frontier model post-trainingApple, Meta, and Nvidia all agree — synthetic data, iterative training, human preference labels, and lots of filtering.
• 31 July 2024 · InterconnectsGPT-4o-mini changed ChatBotArenaAnd how to understand Llama 3.1’s results on the community's favorite benchmark.
• 23 July 2024 · InterconnectsLlama 3.1 405B, Meta’s AI strategy, and the new, open frontier model ecosystemDefining the future of the AI economy and regulation. Is Meta’s AI play equivalent to the Unix stack for open-source software?
• 17 July 2024 · InterconnectsSB 1047, AI regulation, and unlikely allies for open modelsThe rallying of the open-source community against CA SB 1047 can represent a turning point for AI regulation.
• 3 July 2024 · InterconnectsSwitched to Claude 3.5Speculations on the role of RLHF and why I love the model for people who pay attention.
• 26 June 2024 · InterconnectsRLHF roundup: Getting good at PPO, sketching RLHF’s impact, RewardBench retrospective, and a reward model competitionThings to be aware of if you work on language model fine-tuning.
• 21 June 2024 · InterconnectsFrontiers in synthetic dataTrends in synthetic data that I'm watching closely in the leading open and closed models.
• 18 June 2024 · InterconnectsText-to-video AI models are already abundant, but the products?Signs point to a general use Sora-like model coming very soon, maybe even with open weights.
• 12 June 2024 · InterconnectsAI for the rest of usApple Intelligence makes a lot of sense when you get out of the AI bubble. Plus, the cool technical details Apple shared about their language models "thinking different."
• 7 June 2024 · InterconnectsA case study in reproducibility of evaluation with RewardBenchA very technical deep dive. Lessons in evaluating, using, and training reward models on open-source infrastructure.
• 5 June 2024 · InterconnectsA realistic path to robotic foundation modelsSome thoughts and excitement after revisiting the industry thanks to Physical Intelligence founders Sergey Levine and Chelsea Finn. Not “agents” and not “AGI.”
• 29 May 2024 · InterconnectsWe aren’t running out of training data, we are running out of open training dataData licensing deals, scaling, human inputs, and repeating trends in open vs. closed LLMs.
• 22 May 2024 · InterconnectsName, image, and AI’s likenessCelebrity’s power will only grow in the era of infinite content.
• 15 May 2024 · InterconnectsOpenAI chases HerChatGPT left the textbox and where AI is leading society.
• 10 May 2024 · InterconnectsOpenAI’s Model (behavior) Spec, RLHF transparency, personalization questionsNow we will have some grounding for when weird ChatGPT behaviors are intended or side-effects — shrinking the Overton window of RLHF bugs.
• 1 May 2024 · InterconnectsHow RLHF works, part 2: A thin line between useful and lobotomizedMany, many signs of life for preference fine-tuning beyond spoofing chat evaluation tools.
• 30 April 2024 · InterconnectsPhi 3 and Arctic: Outlier LMs are hintsModels that seem totally out of scope from recent open LLMs give us a sneak peek of where the industry will be in 6 to 18 months.
• 24 April 2024 · InterconnectsAGI is what you want it to beCertain definitions of AGI are backing people into a pseudo-religious corner.
• 18 April 2024 · InterconnectsLlama 3: Scaling open LLMs to AGIMeta shows that scaling won't be a limit for open LLM players in the near future.
• 17 April 2024 · InterconnectsStop "reinventing" everything to solve alignmentIntegrating some non-computing science into reinforcement learning from human feedback (RLHF) can give us the models we want. Bonus: OLMo 1.7-7B.
• 15 April 2024 · InterconnectsThe end of the “best open LLM”Modeling the compute versus performance tradeoff of many open LLMs.
• 3 April 2024 · InterconnectsWe disagree on what open-source AI should mean... and that's okay. How to read what multiple people mean by the word openness and see through the PR speak.
• 28 March 2024 · InterconnectsDBRX: The new best open model and Databricks’ ML strategyDatabricks’ new model is surpassing the performance of Mixtral and Llama 2 70B while still being in a size category that's reasonably accessible.
• 20 March 2024 · InterconnectsEvaluations: Trust, performance, and price (bonus, announcing RewardBench)Evaluation is not only getting harder with modern LLMs getting more complicated, it’s getting harder because it means something different.
• 13 March 2024 · InterconnectsModel commoditization and product moatsWhere moats are tested now that so many people have trained GPT4 class models. Claude 3, Gemini 1.5, Inflection 2.5, and Mistral Large are here to party.
• 6 March 2024 · InterconnectsThe koan of an open-source LLMA proposal for a new definition of an “open-source” LLM and why no definition will ever just work.
• 28 February 2024 · InterconnectsHow to cultivate a high-signal AI feedBasic tips on how to assess inbound ML content and cultivate your news feed.
• 19 February 2024 · Interconnects10 Sora and Gemini 1.5 follow-ups: code-base in context, deepfakes, pixel-peeping, inference costs, and moreThe cutting edge technical discussions beneath the wow factor.
• 16 February 2024 · InterconnectsOpenAI’s Sora for video, Gemini 1.5's infinite context, and a secret Mistral modelEmergency blog! Three things you need to know from the ML world that arrived on Thursday.
• 14 February 2024 · InterconnectsWhy reward models are key for alignmentIn an era dominated by direct preference optimization and LLM-as-a-judge, why do we still need a model to output only a scalar reward?
• 7 February 2024 · InterconnectsAlignment-as-a-service: Scale AI vs. the new guysScale’s making over $750 million per year selling data for RLHF, who’s coming to take it?
• 1 February 2024 · InterconnectsOpen Language Models (OLMos) and the LLM landscapeA small model at the beginning of big changes.
• 29 January 2024 · InterconnectsModel merging lessons in The Waifu Research DepartmentWhen what seems like pure LLM black magic is actually supported by the literature.
• 24 January 2024 · InterconnectsLocal LLMs, some facts some fictionThe deployment path that’ll break through in 2024. Plus, checking in on strategies across Big Tech and AI leaders.
• 17 January 2024 · InterconnectsMultimodal blogging: My AI tools to expand your audienceA fun demo on how generative AI can transform content creation, and tools for my fellow writers on Substack!
• 10 January 2024 · InterconnectsMultimodal LM roundup: Unified IO 2, inputs and outputs, Gemini, LLaVA-RLHF, and RLHF questionsA sampling of recent happenings in the multimodal space. Be sure to expect more this year.
• 3 January 2024 · InterconnectsIt's 2024 and they just want to learnThe state of the ML communities big and small starting 2024. My general expectations for the year.
• 20 December 2023 · InterconnectsState-space LLMs: Do we need Attention?Mamba, StripedHyena, Based, research overload, and the exciting future of many LLM architectures all at once.
• 13 December 2023 · InterconnectsBig Tech's LLM evals are just marketingA PSA everyone needs. The importance of a wait and see attitude when it comes to new models, big and small, open and closed.
• 11 December 2023 · InterconnectsMixtral: The best open model, MoE trade-offs, release lessons, Mistral raises $400mil, Google's loss, vibes vs marketingWe have an amazing open mixture of experts model for the holidays!
• 6 December 2023 · InterconnectsDo we need RL for RLHF?Direct (DPO) vs. RL methods for preferences, more RLHF models, and hard truths in open RLHF work. We have more questions than answers.
• 29 November 2023 · InterconnectsSynthetic data: Anthropic’s CAI, from fine-tuning to pretraining, OpenAI’s Superalignment, tips, types, and open examplesSynthetic data is the accelerator of the next phase of AI — what it is and what it means.
• 22 November 2023 · InterconnectsRLHF progress: Scaling DPO to 70B, DPO vs PPO update, Tülu 2, Zephyr-β, meaningful evaluation, data contaminationHuge steps forward in confirming that RLHF can really help you on vibes based evaluation, among many other RLHF analyses.
• 19 November 2023 · InterconnectsOpenAI’s shakeup and opportunity for the rest of us: openness, brain drain, and new realitiesNew timelines that emerge in AI and the winners and losers, regardless of the unfolding details.
• 15 November 2023 · InterconnectsThe interface era of AIModern LLMs are becoming the easiest and most efficient way to access information. This will change how we see the world.
• 8 November 2023 · InterconnectsReckoning with the Shoggoth of AICulture wars, open letters, new politics, developer days, and everything hidden under the smiling face of RLHF.
• 1 November 2023 · InterconnectsOpen LLM company playbookWhere does releasing model weights fit into company strategy? 3 requirements, 3 actions, and 3 benefits of being in the open LLM space.
• 25 October 2023 · InterconnectsRLHF lit. review #1 and missing pieces in RLHFLooking at the difference between two sets -- what rumors say industry leaders are doing with RLHF and what the literature is up to. I'm starting my new series studying RLHF literature.
• 18 October 2023 · InterconnectsUndoing RLHF and the brittleness of safe LLMsRecent papers show most of the arguments about needing "safety" in releases of open LLM weights are nearly dead in the water. Yes, still release the parameters.
• 11 October 2023 · InterconnectsThe AI research job market shit show (and my experience)There are plenty of jobs, but finding a place where you're happy is as hard as ever.
• 6 October 2023 · InterconnectsLLMs are computing platformsThis fact is why so many debates around LLMs feel broken, especially moderation.
• 4 October 2023 · InterconnectsOpen, general-purpose LLM companies might not be viableFailure modes on the quest to open-source LLMs. Expect pivots to specialized models.
• 27 September 2023 · InterconnectsDALL·E 3 and multimodality as moats, correcting bad moat takesMultimodality may be a key differentiator in how moats are built for LLMs.
• 25 September 2023 · InterconnectsChallenges operationalizing responsible AI in open RLHF researchSome reflections from my time at HuggingFace.
• 20 September 2023 · InterconnectsMidjourney vs. Ideogram, ML product companies, preventing AI winter, DALL·E 3 teaseThe coming image-generation battle and its implications on ML product longevity.
• 13 September 2023 · InterconnectsIn defense of the open LLM leaderboardNo, SemiAnalysis, HuggingFace isn't misleading all of open-source, and open-source is still making real progress.
• 6 September 2023 · InterconnectsAI researchers' challenges: atomic analogies and strained institutionsWe need to heal AI research norms, not build a super project. The reflections and reverberations that we're feeling in the AI community from Oppenheimer's quest for the atomic bomb.
• 30 August 2023 · InterconnectsCruise's collisions and adapting to AISF continues to be the center of attention for developments in AI, but this time it's in the physical world.
• 9 August 2023 · InterconnectsLLM products: measurement and manipulationTwo stories will begin to unfold as the AI capabilities-to-product overhang is reduced.
• 2 August 2023 · InterconnectsSpecifying objectives in RLHFAt ICML, it is obvious that many people are getting value out of RLHF. What is limiting the scientific understanding of it (other than research embargoes)?
• 26 July 2023 · Interconnects"If it's not fully closed ML, it's open" - is it?Definitions from open-source software are being bent by new machine learning technologies.
• 21 July 2023 · InterconnectsLlama 2 follow-up: too much RLHF, GPU sizing, technical detailsThe community reaction to Llama 2 and all of the things that I didn't get to in the first issue.
• 18 July 2023 · InterconnectsLlama 2: an incredible open LLMMeta is continuing to deliver high-quality research artifacts and not backing down from pressure against open source.
• 12 July 2023 · InterconnectsLLM agents and integration dead-endsWhen is GPT4 going to schedule my meetings? What is stopping it?
• 28 June 2023 · InterconnectsTesla Autopilot's negligence and regulationMany of my colleagues in robotic learning have long been skeptical of Tesla’s efforts in the area. I also encourage people to not drive in Tesla’s with Autopilot engaged. Here’s why.
• 21 June 2023 · InterconnectsHow RLHF actually worksThe proven formula for RLHF and when we will see it in open-source.
• 14 June 2023 · InterconnectsDifferent development paths of LLMsIn industry, open-source, and academia, each of these giant pools of talent are driven by different incentives and will create very different language models.
• 7 June 2023 · InterconnectsOpen-source LLMs' harmlessness gapOpen-source LLM capabilities are leaving safety tools in the dust. What can we do?
• 31 May 2023 · InterconnectsEvaluating and uncovering open LLMsWhen choosing a model, we're stuck in the middle between classic NLP benchmarks (e.g. MMLU) and qualitative chatbot ranking. Neither are exactly what we want.
• 25 May 2023 · InterconnectsCode: green pastures for LLMsBoth code-generation and code-training seem central to LLM progress, present and future.
• 17 May 2023 · InterconnectsUnfortunately, OpenAI and Google have moatsWhile everyone went crazy over a leaked memo, no one took the time to think through how companies have worked in the internet era.
• 3 May 2023 · InterconnectsSpecifying hallucinationsThe future of LLM errors and the need for a discourse around the specification of unexpected outputs.
• 26 April 2023 · InterconnectsBeyond human data: RLAIF needs a rebrandReinforcement learning from computational feedback: how a promising method is being ignored because of a confusing launch.
• 12 April 2023 · InterconnectsGrowing needs for accessing state-of-the-art reward modelsThe rewards assigned by these models will control the subjective experience of users; the subjective experience controls their decision-making.
• 5 April 2023 · InterconnectsBehind the curtain: what it feels like to work in AI right now (April 2023)Fear, FOMO, and the scientific exodus driven by ChatGPT
• 27 March 2023 · InterconnectsThe implicit dynamics of optimizing costs vs. rewards vs. preferencesWith the emergence of reinforcement learning from human feedback, we've been applying old techniques with a new guiding function (🤫 RLHF).
• 20 March 2023 · InterconnectsGPT4: The quiet parts and the state of MLChecking in on the state of the ML field: technical progress, societal implications, and more.
• 9 March 2023 · InterconnectsAGI Roundup: Re-visiting Go; transitioning from narrow to general; multimodality & GPT4AI is much more fun to think about when AGI is an adjective, not a finish line.
• 27 February 2023 · InterconnectsThe RLHF battle lines are drawnTales of the open and closed sides, how these two dynamics will dictate progress and public perception.
• 20 February 2023 · Interconnects"AI alignment" and uncalibrated discourse on AIRe-naming arms races; bridging communities; aligning ChatGPT, and more.
• 15 February 2023 · InterconnectsThree seasons of RL: Metaphor, tool, and frameworkHow we’re passing through different eras of reinforcement learning.
• 1 February 2023 · InterconnectsScaling laws for robotics & RL: Not quite yetRobotics Transformers, DreamerV3, XLand 2, and hoping that scaling laws are coming embodied AI.
• 16 January 2023 · InterconnectsPretraining quadrupeds: a case study in RL as an engineering toolHow an unlikely corner of robotics research, locomotion, defined RL's new notion of success.
• 6 January 2023 · InterconnectsLooking into 2023My predictions for machine learning this year: 3D assets, self-driving, GPT4, RLHF, Deep RL, diffusion models, and conference cycles.
• 28 December 2022 · InterconnectsPredicting machine learning moatsModels aren't moats and how emergent behavior scaling laws will change the business landscape.
• 19 December 2022 · InterconnectsClosed-API vs Open-source continues: RLHF, ChatGPT, data moatsModel-as-a-service makes more sense when there is a data advantage to back it up.
• 5 December 2022 · InterconnectsRLHF, 'online' ML systems, and RL going mainstreamCommon machine learning systems are starting to deploy the RL lens of feedback.
• 26 October 2022 · InterconnectsUsing RL's exploitation to debugA musing on how I think autonomous system companies should use RL.
• 26 September 2022 · InterconnectsBack in the gameWhat I've been up to and what's coming soon.
• 8 February 2022 · InterconnectsDesigning Societally Beneficial Reinforcement Learning SystemsChoices, risks, and reward reporting. Recommendations for how to integrate RL systems with society.
• 21 January 2022 · InterconnectsFlexible Centralization in Multi-agent Learning & ControlA tour of control theory, multi-agent RL, and hierarchical learning.
• 9 August 2021 · InterconnectsRemote robotic-data farmsIndustry labs power up their robot learning research with parallelization!
• 2 August 2021 · Interconnectson the Horizon of applied RLHow RL is starting to be used by industry and how RL is heading to a framing more suited for industrial scales.
• 21 June 2021 · InterconnectsReward is not enoughMulti-agent scenarios make reward maximization a risk. Discussing when, rather than if, we should believe in the Reward Hypothesis.
• 14 June 2021 · InterconnectsHow all machine learning becomes reinforcement learningI make the case why people iteratively training any model should learn some core concerns of reinforcement learning.
• 19 March 2021 · InterconnectsSetting ourselves up for exploitation: RL in the wildHow simulator exploitation, a dual of over-optimization, in AI is the canary in the coal mine for what negative implications could come from weakly-bounded, data-driven iterative systems (RL).
• 26 February 2021 · InterconnectsCounting down until consumer drones are banned in citiesI don’t even like the idea of flying delivery drones, but that will be all we have.
• 19 February 2021 · InterconnectsClarifying RL: Obscure problem formulations and structure tradeoffsSome debates that will be settled en route to RL being used in all corners of the modern world.
• 12 February 2021 · InterconnectsDecoupling AI from the latent variable of spoken languagesHow English being the language of progress in AI could bias our machine’s minds and what we think they are capable of. Uncoupling AI from our notion of language is frightening.
• 5 February 2021 · Interconnects100th anniversary of the word robot: COVID didn’t give us personal robots, it gave us WoebotYes COVID accelerated automated manufacturing and logistics, but robots have not been helping out en masse anywhere else.
• 29 January 2021 · InterconnectsBoston Dynamics 🤖🐈: Studying Athletic IntelligenceThe acrobatic dance videos are flashy, but what are the actual technical breakthroughs? What is happening to the Korean robotics industry?
• 22 January 2021 · InterconnectsRobotic Companies 2.0: Horizontal ModularityHow behaving as a platform rather than a manufacturer will be the sweet spot for the next generation of robotics companies.
• 15 January 2021 · InterconnectsReflections on digital technology from a capitol siegeThe digital fallout following the siege on the capital shows we are not keeping up with the pace of technological change.
• 8 January 2021 · InterconnectsConsidering AIs through our own mind’s reflectionLessons about AI from lessons about our mind. Focused on the nature of free will.
• 1 January 2021 · InterconnectsThe stakeholders start to learn the stakes: predictions for machine learning and automation in 2021Most of my predictions for automation in 2021 have common people starting to learn how their lives will be impacted, and it’s up to the practitioners to make it fair.
• 18 December 2020 · InterconnectsThe Ubiquity and Future of Model-based Reinforcement LearningMaking a case for why you want to learn about models and why my research matters.
• 11 December 2020 · InterconnectsFacebook Case Study 🌏🔬: Using AI to Regulate the Digital EcosystemThe modern digital giants have data on scales unfathomable to the outside, so trying to understand how they manage deleterious information is an uphill battle.
• 13 November 2020 · InterconnectsThe Collingridge Dilemma and Current Policy on RobotsLegal cases impacting robotics, innovation vs policy, the sci-fi future of robotics.
• 6 November 2020 · InterconnectsConstructing Axes for Reinforcement Learning PolicyA small step into the research community’s most opaque framework.
• 30 October 2020 · InterconnectsTowards an Ethics for RoboticistsHow does one handle abstraction and scale in ethical design?
• 16 October 2020 · InterconnectsWhat is ML for Climate Change?A new “subfield” founded in 2019 is making waves, and is more accessible than I first thought.
• 2 October 2020 · InterconnectsModels, Systems, Code; and RobotsEE v ME v CS in complex systems. Do we view things differently?
• 11 September 2020 · InterconnectsAutonomy startups are such a messAI startups just use data logging and robotics startups are all in dark mode: where are we heading? I just want to hear what people are really doing, and then I can decide its value.
• 4 September 2020 · InterconnectsAI & Arbitration of TruthCan we make an AI fact checker? Language, knowledge, and opinions are all moving targets.
• 28 August 2020 · InterconnectsCan robots be autonomous and unintelligent?Re-thinking robot design part 2: don’t restrict your robot worldview to robot (a) does task (b).
• 21 August 2020 · InterconnectsThe uncanny world of robots at homeRe-thinking robotic design part 1: social support robots diving us into the uncanny valley. We don't need robots to look like humans.
• 7 August 2020 · InterconnectsDigital companies and the goal of at-home embodied AILearning how to learn, by letting autonomous agents interact with the world. Why big tech companies like Facebook and Google hire roboticists to bring life to their excess meeting rooms.
• 31 July 2020 · InterconnectsAutomated: how algorithms shape a day in the life and our futureThe goal of the algorithm is not to give users content they like, but to make the users more predictable. Brief added comment on online courses.
• 24 July 2020 · InterconnectsRecommendations are a game - a dangerous game (for us).Machine-recommendation system ethics, models of human reward, reinforcement learning framing; followup on GPT-3.
• 17 July 2020 · InterconnectsAutomating code, Twitter's hack(s), a robot named StretchLanguage models write code, Twitter gets hacked (again), new robots, and top-tier conferences.
• 10 July 2020 · InterconnectsOnline courses, automating education, and digitalizing degreesOnline teaching is here to stay and I don’t think anyone knows how it will go.
• 2 July 2020 · InterconnectsDemocratizing AutomationGetting everyone to benefit from the artificial intelligence boom could be more challenging than some expect.
• 26 June 2020 · InterconnectsDrones, Swarms, and Storms of DronesThere is no situation when I want to encounter a drone of unknown origin, and in the future we'll be seeing hundreds.
• 12 June 2020 · Interconnects"10 years of automation in 1 year"What the automation acceleration from COVID19 actually will look like. What did robotics look like a decade ago and is this step feasible?