AI for science
AI as an instrument of scientific discovery — AlphaFold solving protein structure, Olympiad-level mathematics, and by 2026 original results including a conjecture counterexample and formally-verified proofs.
This thread follows AI used not as a product but as a scientific instrument. Its landmark is AlphaFold. DeepMind’s AlphaFold 2 solved protein-structure prediction at the CASP14 assessment to a degree much of the field had thought a decade away, and its publication in Nature with open code and a database made it a standard tool across biology within months.
The programme widened. AlphaFold 3 extended prediction across proteins, DNA, RNA and small molecules; AlphaProof and AlphaGeometry 2 reached silver-medal standard at the International Mathematical Olympiad. The scientific establishment ratified the shift when Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry for the protein work.
By 2026 the claim had moved from prediction to discovery. Anthropic reported that its Fable model had produced a counterexample to the Jacobian Conjecture, and OpenAI published ten formally-verified mathematical advances from an unreleased model. The open question the thread raises is where assistance ends and authorship begins — whether these systems are powerful instruments in human hands or are starting to do the science themselves.
Google Health reports AI matching radiologists on breast screening
A Nature paper claimed an AI system reduced false positives and false negatives in mammography against expert readers.
Models & capabilities · Benchmarks & progress
DeepMind publishes early AlphaFold protein structure work
Describes the CASP13-winning system, which used deep networks trained on genomic data to predict inter-residue distances rather than folding a structure directly.
Models & capabilities
The CORD-19 research dataset is released
Assembled with the White House OSTP, National Library of Medicine and Chan Zuckerberg Initiative, it opened with over 29,000 papers, most with full text.
Ideas & essays · Open weights & ecosystem
AlphaFold 2 solves protein structure prediction at CASP14
DeepMind's system predicted protein structures to roughly experimental accuracy, ending a fifty-year-old open problem in biology.
Models & capabilities · Benchmarks & progress
AlphaFold 2 is published in Nature and open-sourced
The 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.
Models & capabilities · Open weights & ecosystem
The AlphaFold Protein Structure Database opens
The 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.
Open weights & ecosystem
DeepMind controls a fusion plasma with reinforcement learning
A single network commanding all of a tokamak's control coils held plasma shapes on Switzerland's TCV reactor, including configurations conventional controllers struggle with.
Models & capabilities · Ideas & essays
AlphaFold's database expands to 200 million structures
The 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.
Open weights & ecosystem
AlphaTensor discovers new matrix multiplication algorithms
Found a 76-multiplication algorithm for a specific matrix size, improving on the best known method for the first time in over fifty years.
Models & capabilities
Meta pulls Galactica after three days
Its 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.
Models & capabilities · Culture & impact
AlphaDev discovers faster sorting algorithms, added to the C++ library
The 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.
Models & capabilities
AlphaMissense catalogues 71 million genetic variants for disease risk
Adapted from AlphaFold, the model classified 89% of all possible human missense variants as likely pathogenic or benign, versus 0.1% confirmed by human experts.
Models & capabilities
GraphCast beats conventional weather forecasting on speed and accuracy
Trained 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.
Models & capabilities
GNoME finds 2.2 million candidate new materials
DeepMind 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.
Models & capabilities
AlphaGeometry solves olympiad geometry problems near gold-medal level
The 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.
Models & capabilities · Benchmarks & progress
AlphaFold 3 predicts structures across proteins, DNA, RNA and ligands
Restricted at launch to a rate-limited web server rather than downloadable code, prompting an open letter with more than 650 signatures within a week.
Models & capabilities
AlphaProof and AlphaGeometry 2 reach silver-medal standard at the IMO
The 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.
Models & capabilities · Benchmarks & progress
DeepMind's AlphaProteo designs novel protein binders
Trained 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.
Models & capabilities
Hassabis and Jumper share the Nobel Prize in Chemistry
Half the prize went to Baker for computational protein design; the other half was split between Hassabis and Jumper for AlphaFold's structure prediction.
Culture & impact
DeepMind reduces quantum computing errors with AlphaQubit decoder
Trained 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.
Models & capabilities
Google Research launches an AI co-scientist to help generate hypotheses
A multi-agent Gemini 2.0 system with Generation, Reflection and Ranking agents proposed drug candidates later confirmed active in laboratory tests for two diseases.
Models & capabilities
Isomorphic Labs raises $600 million to advance AI-designed drugs toward clinical trials
Thrive 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.
Money & business
DeepMind's AlphaEvolve pairs Gemini with automated evaluators to discover algorithms
Not 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.
Models & capabilities
Google launches Weather Lab with an experimental AI cyclone model
The experimental model produces 50 possible storm-path outcomes roughly a week ahead and was developed with feedback from the US National Hurricane Center.
Models & capabilities
DeepMind launches AlphaGenome for predicting genome regulatory activity
The 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.
Models & capabilities
Google DeepMind marks five years of AlphaFold's impact on biology
DeepMind 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.
Models & capabilities
DeepMind's Genie 3 generates navigable, real-time interactive worlds
The 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.
Models & capabilities
Trump launches Genesis Mission to accelerate AI-driven science
The 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.
Government & policy · Compute & infrastructure
OpenAI introduces FrontierScience benchmark
GPT-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.
Benchmarks & progress
OpenAI launches Prism
Built 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.
Models & capabilities
Shanghai AI Laboratory open-sources Intern-S1-Pro, a 1-trillion-parameter scientific model
Only 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.
Open weights & ecosystem · Models & capabilities
Google upgrades Gemini 3 Deep Think to V2
Google 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.
Models & capabilities
Anthropic tests Claude on BioMysteryBench
On 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.
Benchmarks & progress
OpenAI model credited with disproving the Erdős unit distance conjecture
An internal general-purpose reasoning model found an infinite family of constructions beating a bound mathematicians had assumed near-optimal since 1946.
Benchmarks & progress · Models & capabilities · Ideas & essays
Google DeepMind's Co-Scientist reaches Nature publication as a multi-agent research tool
The 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.
Models & capabilities
Google DeepMind's AlphaProof Nexus solves nine open Erdős problems
The 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.
Models & capabilities · Ideas & essays
Anthropic's Fable model produces counterexample to the Jacobian Conjecture
Harvard mathematician Levent Alpöge said Claude Fable 5 found the three-variable counterexample in an evening; it disproves the conjecture from three dimensions upward.
Benchmarks & progress · Models & capabilities · Ideas & essays
Anthropic reports Claude finding novel cryptographic weaknesses
Anthropic'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.
Security & misuse · Benchmarks & progress
Epoch AI expands FrontierMath to 50 unsolved research problems
Unlike 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.
Benchmarks & progress
OpenAI publishes ten formally-verified math advances from unreleased Astra model
OpenAI 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.
Benchmarks & progress · Models & capabilities · Ideas & essays
DeepMind's WeatherNext model improves cyclone forecasting accuracy
WeatherNext gives roughly an extra day of predictive accuracy on cyclone track, intensity and wind structure versus prior forecasting models, per a Nature paper.
Models & capabilities
Inherent releases Faraday, an AI agent trained to reproduce research papers
The 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.
Models & capabilities · Benchmarks & progress
Anthropic reports Claude designed working protein binders
Wet-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.
Models & capabilities · Ideas & essays
ASI-Bench measures how far agents are from autonomous science
Withdrawing 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.
Benchmarks & progress
Benchmark finds AI agents complete about a fifth of full science workflows
The 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.
Benchmarks & progress
Anthropic expands free and discounted Claude access for scientists
Premium 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.
Culture & impact · Money & business
Anthropic previews a standard letting AI agents operate lab equipment
Anthropic 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.
Models & capabilities · Open weights & ecosystem
Google DeepMind's Co-Scientist AI is validated in physical lab experiments
The 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.
Models & capabilities · Benchmarks & progress
Anthropic says Claude formalised Fermat's Last Theorem
The full Lean proof ran to about 13 million lines and 29,500 intermediate theorems — more than five times the size of Mathlib — built in roughly 11 days.
Models & capabilities · Benchmarks & progress
OpenAI says an internal model resolved a Navier–Stokes problem, amid a credit dispute
An unreleased model, run as ~10,000 agents, produced a Lean-verified finite-time blow-up proof for a forced fluid; OpenAI declined the prize amid a credit dispute with NYU and Anthropic.
Models & capabilities · Benchmarks & progress
Anthropic says Claude found a new CRISPR-like enzyme system
About 950 Claude agents spent 21 hours searching genomic data for unusual reverse transcriptases; human scientists in Anthropic's new wet lab then ran first experiments, reported in an unreviewed preprint.
Models & capabilities · Ideas & essays
Anthropic says Claude computed a nine-loop physics amplitude
Working largely unsupervised over several days, the model extended a calculation of particle-scattering formulas one loop past the previous 2023 record, for an estimated $1,000-2,000 of compute.
Models & capabilities · Ideas & essays
OpenAI publishes hundreds of AI-written proofs of open maths problems
722 unreviewed manuscripts in 372 result families, including claimed proofs of the Unique Games Conjecture and a zero-free strip for the zeta function; 235 families come with Lean checks.
Models & capabilities · Ideas & essays · Benchmarks & progress
AI companies pledge $2.4bn in compute and tools to the White House's science push
Nvidia pledged $1bn, AMD $500m, OpenAI $200m, Anthropic and Google $150m each; the credits go to research projects at more than 15 federal agencies in the Genesis Mission.
Government & policy · Compute & infrastructure
GPT-6 Astra solves 49 of 50 unpublished theoretical-physics problems on a new benchmark
Physicist-built problems the authors call 'turn-the-crank' research tasks; Claude Opus 5.5 solved 46, Gemini 3.8 Flash 34 and the open-weight DeepSeek-V4.1-Flash 27.
Benchmarks & progress · Ideas & essays
Chinese mathematicians form a 1,000-strong alliance on AI and mathematics
Led by Fields medallist Shing-Tung Yau at Tsinghua, eight institutions plan specialised maths models, open infrastructure for AI-assisted proof, interpretability work and a journal.
Ideas & essays
Anthropic says Claude Mythos 5 improved on the best published quantum circuits
Against the smallest previous elliptic-curve design, one circuit used 12% more qubits but 17 times fewer gates; the post cites no outside expert review.
Models & capabilities · Ideas & essays
AI-designed bacteriophages are given to a patient for the first time, Radical Numerics says
The company says its cocktail eliminated the patient's resistant bacterial isolate in the laboratory before administration, but is withholding the patient's clinical course and outcome.
Ideas & essays
Microsoft researchers say AI agents helped resolve the 31-year-old KLS conjecture
The proof was posted to arXiv on 1 October and completed on 4 October; Microsoft's post says agents explored over 100 approaches but does not name the systems.
Ideas & essays · Models & capabilities
Axiom Math releases a computer-checked proof of the classification of finite simple groups
AI agents wrote about 43.1 million lines of Lean in 24 days for the roughly 15,000-page proof, and found 34 noteworthy deviations from the published record.
Models & capabilities · Ideas & essays
OpenAI and Anthropic each say AI disproved a central conjecture in computer science
Both posted Lean-checked subexponential 3-SAT algorithms; OpenAI also claimed Kakeya in all dimensions and an unreleased 1,147-page Yang-Mills proof. No completed independent review of any had been reported by 2 November.
Models & capabilities · Ideas & essays
Anthropic reports the first 3D structure of a prokaryotic vault particle, found with Claude
Claude Science searched about 15 billion predicted proteins; a phage-derived vault of 96 protein copies was solved by cryo-EM, with a first model in 34 hours. Function remains unknown.
Ideas & essays
Axiom Math releases a Lean formalisation of the Stacks Project
About 16 million lines of Lean cover the roughly 7,500-page algebraic-geometry reference; Axiom says its agents found dozens of statements in the original needing correction.
Ideas & essays
FrontierPhysics benchmark finds AI agents fall short on physics research
56 tasks written by 42 physicists from their own research; 22 models in 26 configurations passed at most 46 of 168 rollouts, with rigour and expert intuition the main weaknesses.
Benchmarks & progress · Ideas & essays
Anthropic analyses 207,000 research sessions to see how scientists use Claude
Data analysis is about a third of Claude sessions and nearly half of Claude Code's; a modelled full-automation scenario puts computer-science output up to 15-fold.
Ideas & essays · Labs & people
DeepMind and Google Quantum AI report AI-discovered quantum programs verified in Lean
Agent teams filled in Lean specifications; the abstract reports gains over prior baselines of 290x for quantum arithmetic and 11x, 8x and 2.5x for fusion, photochemistry and drug-metabolism simulations.
Ideas & essays