Results in mathematics and science where an AI system did a real share of the work. For each one: the problem, why it was hard, how it was solved step by step, what it changes, who did it, and how well it has been checked. Claims are kept apart from confirmed results.
53dossiers
14labs
9under review
as of
Under verification
These claims are recent and not yet independently checked. They may hold, shrink or fall. We update their status as specialists report.
Alpöge says that, working with Claude, he found a proof of the Riemann hypothesis, and that the method extends to the Generalized Riemann Hypothesis. He expects the Extended RH to follow with more time. The released Lean archive states Mathlib's `RiemannHypothesis` and GRH for Dirichlet L-functions (theorem names listed in the mirror README).
Status. As of 2026-11-21, Andrew Sutherland (MIT) has compiled the Lean code independently and reports a Lean kernel recheck that uses 'only the three standard axioms' (Scientific American). Kevin Buzzard says 'we still await a successful comparator run', which checks that the statement was not redefined. No human-readable proof exists yet, and no outside expert has checked the mathematics.
Full dossier: why, how, steps, sources
Lab
Anthropic Levent Alpöge is an Anthropic researcher; the work was done with Claude.
People
Levent Alpöge
AI systems
ClaudeLean 4 (v4.33.1) + Mathlib
In plain words
The Riemann hypothesis is a 167-year-old prediction about the hidden regularity of prime numbers, and much of modern number theory and cryptography theory is built on the assumption that it is true. An Anthropic mathematician says he and Claude have proved it, and a broader version too, and has published the proof only as computer-checkable Lean code. The code compiles for independent checkers, but experts are still confirming that it states exactly the right theorem. Even if it holds, nothing changes overnight for encryption or computing: results that already assumed RH would simply become unconditional.
Clay Millennium Prize, $1M; requires publication + 2 years + general acceptancePrize
Why it matters
The Riemann hypothesis (Riemann, 1859) is often called the most famous open problem in mathematics. It is a Clay Millennium Prize problem and the only one carried over from Hilbert's 1900 list. Thousands of results in number theory and computer science are proved assuming RH or GRH.
How it was done
Alpöge's note says he steered the work toward combining two ideas from a 'toy setting', a descendant of Stepanov's function-field argument with 'multiplicity', and that Claude then 'far outpaced' him. Example: Claude suggested chains of theta lifts of length √X to mimic multiplicity for a von Mangoldt sum of length X. Tao says the proof appears to use 'the full power of modern developments on the Langlands program and arithmetic geometry'. Alpöge released only the Lean formalization, because the human-readable argument would be an 'unreadable' document of 250,000+ pages.
Step by step
2026-10-06: 'Stepanov determinant' (first entry in the timeline Claude compiled for the note)
October 2026: CM-curve, theta-lift and Guinand–Weil steps (timeline in the note)
Full formalization in Lean 'without assumptions' with Claude
Announcement late Sunday night; Alpöge consults Anthropic CEO Dario Amodei about announcing
2026-11-20: 4-page note plus a 14-part Lean archive at alpo.ge
Same day: independent compile and kernel recheck by Sutherland; comparator run pending
Chronology
First step in Claude's timeline ('Stepanov determinant')
Switch to the von Mangoldt sum formulation (per note timeline)
Note and Lean archive released; Sutherland compiles; Tao blog post
Comparator run still reported pending
What it brings
If confirmed, it would be the first solution of the Riemann hypothesis and of a Clay problem through an essentially AI-produced, machine-checked proof. Conrey: 'I don't think any human was about to find this.' It also tests whether mathematicians will accept a proof that, for now, exists only in Lean.
What it does not show
No human-readable proof yet, and Alpöge says he has not mastered all the mathematics involved. The Extended RH is only expected, not claimed. Neither Clay nor Anthropic has made any statement of acceptance.
Human vs AI
Alpöge says he chose the strategy and toy model, and that Claude produced much of the higher-rank automorphic machinery, which he says he has not fully mastered. Tao describes the proof as generated 'using an internal AI model'.
Credit and priority
Alpöge names Claude as collaborator and thanks colleagues including Manjul Bhargava, Maksym Radziwiłł, Peter Sarnak, Jarred Sumner, David Dohan, Ben Michel, Ralph Furman, Eric Price and Cora Vinther. He also credits OpenAI's earlier results (an Ehrhart-related result and a Siegel-zero argument) with shaping his intuition. Tao frames the proof as using the automorphic, algebraic and arithmetic-geometry routes; in the comments he discusses Masullo's separate non-AI GRH manuscript without endorsing it.
Limits and debate
Not peer-reviewed; there is no written proof. A Lean compile shows only that the formal statement follows from the axioms. Experts must still confirm that the formal statement really is RH/GRH and that no definitions were changed; the comparator run is meant to check this. The archive is very large and the build is heavy, and a mirror README notes 4 files set maxHeartbeats 0 and 1 uses addDecl. The announcement came the day after Anthropic's stock-market debut; Alpöge calls the cynical reading 'false'. Ravi Vakil said a postdoc in his department independently submitted a separate, non-AI proof the same day. Clay rules require publication, two years and general acceptance.
Reactions
Brian Conrey (Riemann hypothesis specialist): Called it possibly the biggest thing ever in mathematics and seriously complicated; expects simpler proofs; had seen only the website remarks. source
Andrew Sutherland (MIT): Compiled the Lean proof independently and said 'the chance that anything is fishy here is approximately zero'. source
Kevin Buzzard (Imperial College London): The first question is what the theorem's statement in the Lean code is; still awaiting a successful comparator run. source
Terence Tao (Fields medallist, UCLA): Called it an enormous Lean-only proof that seems to use modern Langlands and arithmetic-geometry machinery; digesting it will take time. source
Hector Pasten (Pontifical Catholic University of Chile): If correct and understood, it completes a dictionary between different branches of mathematics. source
Carlo Pagano (Concordia University): Finds it very plausible that the new ideas will open new avenues in number theory. source
Nestor Guillen (Visiting professor, NYU): Praised Alpöge's openness about not fully understanding the proof, saying it shows AI working more like a tool. source
Fields it touches
Number theoryAutomorphic formsArithmetic geometryPrime distributionCryptography (conditional results that assume GRH)AI for mathematicsFormal verification
Claims smooth, compactly supported, divergence-free initial data on R³, and smooth periodic data on the torus T³, for which solutions of the 3D Navier–Stokes equations with no external force blow up in finite time. The authors say this disproves Fefferman's statements (A) and (B), the unforced existence-and-smoothness statements, and hence proves (C) and (D). Blow-up happens at rate (T−t)^(−1/4) in Ḣ¹, converging to a discretely self-similar profile.
Status. Posted 2026-11-20 and not peer-reviewed. Tao calls the paper 'technically human-readable' with a broadly clear strategy, but says it needs more work to become a professional proof. The Lean main theorems take the existence of the self-similar profile as a hypothesis. That profile comes from a numerical certificate, which is not yet released (only its hash is given) and is checked by interval arithmetic outside Lean. The repository's patched Lean package 'was not rebuilt' when assembled.
Full dossier: why, how, steps, sources
Lab
Anthropic Personal collaboration; the paper thanks Anthropic for supporting it.
With
New York University
People
Levent Alpöge, Tristan Buckmaster
AI systems
Claude (only Anthropic models, per the paper's AI statement)Lean 4 (v4.33.1) + MathlibInterval arithmetic (Arb via FLINT, 128-bit)
In plain words
Same equations as OpenAI's September claim, which describe water, air and blood flow, but now with no external pushing at all. The authors say they found a perfectly smooth starting flow that, left alone, develops infinite gradients in finite time. That would settle the deepest form of the Navier–Stokes question. Part of the proof rests on a numerical certificate that has not yet been released. For engineers nothing changes in practice: simulations already run at scales where this idealized blow-up is irrelevant.
Key figures
79 pagesPaper length
~600,000 lines in ~2,270 filesLean formalization
Lean v4.33.1 + MathlibToolchain
(T−t)^(−1/4) in Ḣ¹Blow-up rate
Interval arithmetic, 128-bit (Arb/FLINT) — not in Lean, not yet releasedCertificate check
Not disclosedCompute/tokens
Why it matters
This is the 'real' version of the question. OpenAI's September result needed a tuned external force, so (A)/(B) stayed open. Leray proposed looking for self-similar singular solutions in 1934, and exact self-similar ones were later ruled out (Nečas–Růžička–Šverák, Tsai), leaving discretely self-similar solutions as the next candidate.
How it was done
Two parts. (1) Theorem 1.1 shows that any smooth discretely self-similar profile with the stated decay yields blow-up on R³ and T³. (2) Theorem 7.1 shows such a profile exists near a numerically computed approximate one (a 'certificate') that satisfies explicit inequalities. Agents received a curated dossier of leads and guidance from the authors' earlier forced-Euler and forced hypodissipative Navier–Stokes manuscripts. A simpler mechanism first gave blow-up for a hyperdissipative variant, then a heavier computation produced the profile. The proof was done first in Lean; a draft was generated from the Lean proof, then rewritten with Claude under the authors' direction.
Step by step
Earlier unreleased work: forced Euler and forced hypodissipative NS manuscripts
Agents given a curated dossier; an 'uncontaminated experiment' produces unforced Euler and forced NS blow-up
Hyperdissipative unforced NS blow-up via a simpler mechanism
Heavier computation yields the approximate discretely self-similar profile (certificate)
Lean formalization (~600,000 lines) of the reduction (Theorems 1.1 and 7.1)
Draft generated from the Lean proof, rewritten with Claude; 79-page paper posted 2026-11-20
Chronology
Forced Euler/Boussinesq blow-ups (precursors)
Buckmaster mentions hypodissipative NS as a path to unforced results
Paper + Lean repository posted; Tao blog post
What it brings
If confirmed, this disproves global regularity for the actual unforced Navier–Stokes equations, the substance of the Millennium problem, with a combined human, AI, numerics and Lean workflow.
What it does not show
The Lean artifact alone does not establish the profile's existence; the certificate is promised for later. No claim of a peer-reviewed or Clay-accepted result.
Human vs AI
The authors set the strategy and curated leads; Claude agents did much of the search and proof writing, then the authors rewrote and simplified the text with Claude.
Credit and priority
Builds on Leray, Bradshaw–Tsai's open problems, Buckmaster–Coiculescu–Li on non-uniqueness, and the authors' unreleased computer-assisted work. It comes after OpenAI's forced NS (Sept 2026) and the September priority dispute. The acknowledgments thank Scott Armstrong for 'enslopification skills' used to draft the paper from the Lean proof, and thank Anthropic for supporting the collaboration.
Limits and debate
Key computer-assisted step not yet public: certificate data and its interval-arithmetic check are outside Lean and outside the repository. The Lean theorem is conditional on `PeriodicSimilarityProfileExists`. The formal statement has no blow-up rate, and paper Theorem 7.1's smoothness/closeness claims are not in any Lean statement file. Not peer-reviewed. Lean vs paper fidelity concerns of the 2610.08144 type apply here too.
Reactions
Terence Tao (Fields medallist, UCLA): Called the paper technically human-readable with a broadly clear strategy, but said it needs more work to become a professionally written proof. source
Fields it touches
PDE / fluid dynamicsComputer-assisted proofsTurbulence theoryAI for mathematicsFormal verification
Constructs odd, smooth, divergence-free periodic initial data on the 3D torus whose Euler solution blows up in finite time. Velocity stays bounded while vorticity at the origin tends to infinity, and the flow converges smoothly away from the origin.
Status. Tao lists it as the most 'digested' of the three 2026-11-20 results and calls it a 'human-readable proof'. The authors' Lean formalization claims one-to-one correspondence with the paper, uses only the standard axioms and contains no sorry. No independent build or expert review was found as of 2026-11-21.
Full dossier: why, how, steps, sources
Lab
Anthropic Paper thanks Anthropic for supporting the collaboration; Buckmaster (NYU), Coiculescu (NYU, Simons Junior Fellow).
With
New York University
People
Levent Alpöge, Tristan Buckmaster, Matei P. Coiculescu
AI systems
ClaudeCodexLean 4 (v4.34.0-rc2) + mathlib
In plain words
The Euler equations are Navier–Stokes without friction (viscosity). The authors show that, in a box with wrap-around edges, a smooth swirling flow can concentrate its spin infinitely at one point in finite time. It is a mathematical idealization and has no direct engineering consequence.
Key figures
55 pages (PDF)Paper length
propext, Classical.choice, Quot.sound only; no sorry (authors)Axioms
Why it matters
Fefferman's Clay description notes the regularity question is also open for Euler on R³ and T³. OpenAI's September proof covered R³ with compactly supported data; this paper addresses the periodic case with a multi-scale construction.
How it was done
Localized Craik–Criminale waves are used as building blocks in a multi-scale scheme following Córdoba and Martínez-Zoroa. Each step adds a higher-frequency oscillation near the origin that the preceding scales amplify. The first unforced proof was written by Anthropic models from the authors' forced-Euler and hypodissipative-NS drafts, then rewritten with Claude and Codex under human direction, with some parts replaced by the authors' own ideas.
Step by step
Unreleased second forced-Euler proof (Craik–Criminale mechanism)
Anthropic models write a first unforced proof from the drafts
Iterative rewriting with Claude and Codex; human ideas replace parts
Lean formalization mirroring every numbered result (spine.json ledger)
2026-11-20: 55-page manuscript + Lean public
Chronology
Manuscript frozen
Public repository created
Release alongside NS paper; Tao blog post
What it brings
A second unforced Euler blow-up, on the periodic domain, with a one-to-one paper-to-Lean mapping claimed by the authors.
What it does not show
Not a Navier–Stokes result, and not R³ (OpenAI's Euler paper covers R³).
Human vs AI
AI models wrote the first unforced proof; humans restructured, simplified and contributed ideas.
Credit and priority
The paper's 'Provenance' section says the authors had an unreleased Craik–Criminale-based forced-Euler proof before OpenAI's release, and points to Buckmaster's account of his communications with OpenAI. It builds on Córdoba–Martínez-Zoroa and Córdoba–Martínez-Zoroa–Zheng.
Limits and debate
The periodic Beale–Kato–Majda criterion cited in Corollary 8.3 is not formalized. The paper acknowledges the mechanism matches OpenAI's (Craik–Criminale amplification), so priority is entangled. Euler is not itself a Clay prize problem.
Reactions
Terence Tao (Fields medallist, UCLA): Lists it first, as the most digested of the three results: a human-readable proof of unforced Euler blow-up. source
Fields it touches
PDE / fluid dynamicsTurbulence theoryFormal verificationAI for mathematics
Signal's encryption protocol protects billions of chats, through Signal and variants in WhatsApp, Messenger and Google Messages. DeepMind's agents proved in Lean that the actual compiled machine code of libsignal, Signal's reference library, keeps key security promises for its post-quantum PQXDH handshake and Triple Ratchet: forward secrecy, authentication, post-compromise security and handshake secrecy.
Status. Proofs are checked by Lean's kernel, and a gatekeeper accepts a proof only if the comparator confirms it proves a held-out copy of the approved specification. The Lean development is public. The guarantees are only as good as about 26,000 lines of trusted specification and machine semantics, which humans reviewed. As of 2026-11-21 no independent third-party audit of those specifications has been reported.
Full dossier: why, how, steps, sources
Lab
Google DeepMind
With
Google, Signal (statement from its CTO; Signal Shot challenge co-organiser)
People
Swarat Chaudhuri, Pushmeet Kohli, Andrew Ferraiuolo, Anton Kovsharov, Tej Chajed, Sam Schlesinger, Karn Seth, Jonathan Katz, Charles Sutton, Ehren Kret
AI systems
AlphaProof CodeGemini 4 CarbonClaude Opus 5.5 (specification drafting/review)Lean 4Lean comparatorAArch64 formal semantics in Lean
In plain words
Signal's protocol is the maths that keeps your messages private even if someone later steals a key. 'Formally verifying' it means proving mathematically that the code really keeps those promises, so whole classes of attack are ruled out, not just the bugs someone thought to test for. DeepMind's AI agents produced those proofs for the actual compiled library in a month, work that used to take experts years. It does not mean Signal has zero bugs, and it says nothing about your phone, the apps, Signal's servers or group chats.
~1% of codebase (DeepMind blog)Flagged for human review
2 of 2Reintroduced historical bugs caught
15 in a Double Ratchet computational analysis (plus ProVerif query errors)Errors found in published analyses
seL4 verification: 20 person-yearsComparison
Why it matters
Earlier formal analyses of Signal covered abstract models, verification-friendly reimplementations, modified code or narrow properties. Verifying real code at this depth has historically cost years of expert labour: the seL4 microkernel took 20 person-years. The work is a milestone toward Signal Shot, the grand challenge launched in April 2026 by the Beneficial AI Foundation, Signal and Lean FRO.
How it was done
Humans and agents running on Gemini 4 Carbon and Claude Opus 5.5 turned published Signal security models into Lean specifications. Reviewer agents from a different model family critiqued each draft, and humans approved the final versions. Unmodified Rust code was compiled with the stock rustc compiler to AArch64 Linux machine code, which was loaded into Lean together with a formal semantics of AArch64. Teams of autonomous Gemini 4 Carbon prover agents then wrote the proofs; none of the authors wrote proof code. A gatekeeper agent accepted a proof only after Lean's kernel and the comparator checked it against the approved specification.
Step by step
Select properties and adapt existing symbolic and computational models of Signal to Lean, extending them to the post-quantum ratchet as deployed.
Iterate drafts between a spec-writer agent, cross-family reviewer agents and humans; humans approve and pin each specification.
Compile unmodified libsignal (v0.105.4 harness) with rustc to AArch64 and load the executable into Lean with a machine-code semantics.
Run teams of prover agents for days. The gatekeeper accepts only proofs that pass Lean's kernel and the comparator.
Validate the specifications by reintroducing two previously fixed libsignal bugs and proving in Lean that the buggy binaries violate the security theorems.
Publish the Lean development.
Chronology
Signal Shot challenge officially launched (Beneficial AI Foundation, Signal, Lean FRO)
DeepMind announces AlphaProof Code and the libsignal verification; paper and Lean artifacts released
Press coverage notes scope limits (one-to-one chats, own build, no new bugs found)
What it brings
This is the first proof that ties protocol-level security guarantees to the unmodified code of a widely deployed cryptographic implementation. The Rust compiler is not trusted: a compiler bug would make the proofs fail. Writing specifications against real code also exposed errors in the published literature: 15 gaps or bugs in a computational analysis of the Double Ratchet, plus mistakes in printed ProVerif queries. The authors argue that once proofs are cheap, the hard part becomes deciding what to prove.
What it does not show
This does not claim Signal is bug-free, and it found no new vulnerabilities in the parts checked (per The Next Web). It does not cover the Signal apps, servers, group messaging or key generation, and gives no post-quantum guarantee for the ratchet's healing. The AlphaProof Code framework itself has not been released; only the Lean artifacts have.
Human vs AI
A team of five worked for four weeks. Humans chose what to prove, guided and approved specifications, and reviewed the material agents flagged, about 1% of the codebase according to DeepMind's blog. All proof code was written by agents.
Credit and priority
The work builds on prior symbolic and computational Signal analyses (Cheval–Jacomme–Richards, Bhargavan et al., and others), whose two historical libsignal bugs were reused for validation. It responds to the Signal Shot challenge organised by the Beneficial AI Foundation (Max Tegmark), Lean FRO (Leonardo de Moura) and Signal (Rolfe Schmidt). DeepMind describes the paper's claims as firsts.
Limits and debate
The symbolic proofs assume perfect cryptography. The computational bounds give no guarantee against a quantum adversary. The ratchet theorems show the Triple Ratchet is no weaker than the Double Ratchet, not the post-quantum healing SPQR adds. The scenario is narrow: two honest parties, an authentic PKI, no multi-device state, no one-time Kyber prekeys, and side channels out of scope. The results cover the researchers' own build of libsignal-cli, not Signal's apps; per The Next Web, group messaging and key generation are not covered. Some environment assumptions are argued informally. Sources also differ on figures: the blog says 'over six million lines', the paper says about 6.5 million lines of proof. Some press noted that earlier verification work on Signal gave the project a head start.
Reactions
Ehren Kret (CTO, Signal): Calls AI agents translating live code into valid Lean theorems a major breakthrough that gives Signal greater confidence in its privacy guarantees. source
Swarat Chaudhuri (Google DeepMind (to Wired/The Register, via The Next Web)): Says there is 'a long distance to go' before scaling to all software and that the next frontier is specification, not proof. source
OpenAI: a deterministic, polynomial-space algorithm counts satisfying assignments of any k-CNF in time (n+m)^{O_k(1)}·2^{n/(log n)^{1/2+o(1)}}, which refutes ETH (3-SAT needs 2^{Ω(n)} time). Anthropic: a paper titled 'A 2^o(n) Time Algorithm for 3-SAT', credited to 'Claude, Anthropic'.
Status. OpenAI says an independent working group led by Impagliazzo, Paturi and Williams will examine and explain its proof. OpenAI's Lean file proves a slightly weaker time bound than the paper's headline. OfficeChai reports that the Anthropic appendix says its Lean development (checked by Lean's kernel and the independent nanoda checker) was written and reviewed by AI models, with no person known to have checked it.
Full dossier: why, how, steps, sources
Lab
Multiple labs Claims circulated around OpenAI and Anthropic models.
People
Russell Impagliazzo, Ramamohan Paturi, Ryan Williams
AI systems
Unnamed internal OpenAI modelInternal Claude research modelLean 4
Why it matters
ETH, proposed by Impagliazzo and Paturi (with Zane), is a foundation of fine-grained complexity; hundreds of conditional lower bounds assume it.
How it was done
Anthropic, as quoted by Alpöge: 'The core algorithm was found on October 16 by an internal Claude research model with 14M output tokens and no human input.' Researchers then improved the presentation. OpenAI's result came from its internal model and was released at the review group's request so others could examine it.
Step by step
2026-10-16: Claude model finds core algorithm (Anthropic)
2026-10-25: OpenAI manuscript dated
2026-10-26: OpenAI release with Lean; review group announced
Hours later: Anthropic paper released
What it brings
If confirmed, it would overturn a central complexity-theory hypothesis. It has no practical effect: OfficeChai reports the bound beats brute force only beyond roughly 10^10,000 variables.
Human vs AI
Both labs describe the core discoveries as model-generated; humans handled presentation and review.
Limits and debate
Asymptotic only. OpenAI's Lean bound is weaker than its headline. Human review of both proofs is ongoing. Anthropic's half is reported here from press coverage and Levent Alpöge's post; the paper itself is not linked.
Claims that every rational Hodge class on a complex abelian variety with complex multiplication is a rational combination of algebraic cycle classes, in every dimension and codimension. Through Milne's theorems, the paper also claims the Tate conjecture for abelian varieties over finite fields and the Hodge standard conjecture for abelian varieties.
Status. OpenAI added a Lean formalization on 2026-10-26 and fixed its comparator namespaces on 2026-10-29. No independent expert verdict was found. Three related Hodge-type papers in the same release (split abelian eightfolds, Kuga–Satake, products of K3 surfaces) were withdrawn on 2026-10-07 over a sign error.
The Hodge conjecture is a Clay Millennium problem; CM abelian varieties are a central special case. Wikipedia describes the OpenAI result as a partial result toward Hodge.
How it was done
OpenAI says this result, like the zeta zero-free region, was an exception to its fixed procedure of about 3 hours of ChatGPT Pro compute per result. The Lean statement reuses the community HodgeConjecture project's formal statement for the comparator challenge.
Step by step
Manuscript released 2026-10-06
Related K3/eightfold papers withdrawn 2026-10-07
Lean formalization added 2026-10-26
Comparator fix 2026-10-29
What it brings
If correct, a major special case of a Millennium problem with strong arithmetic consequences.
Human vs AI
OpenAI says it was model-produced, with a non-standard procedure it does not detail.
Limits and debate
A special case, not the full Hodge conjecture. The Lean scope covers only the Hodge statement, not the Tate or standard-conjecture consequences. Related papers in the same family line were withdrawn.
Fields it touches
Algebraic geometryArithmetic geometry (Tate conjecture)AI for mathematics
Proves that the Riemann zeta function, every Dirichlet L-function, and every finite-order Hecke L-function over Q(√−3) have no zeros with real part greater than 7/8 (pole at s=1 excepted). By symmetry, all nontrivial zeta zeros therefore lie in the strip 1/8 ≤ Re(s) ≤ 7/8. A companion proof gets the weaker half-plane Re(s) > 11/12, and a related formalized result excludes Landau–Siegel zeros near 1 (1−β ≥ c/log q).
Status. Quanta reports that the 199-page 7/8 paper is formally verified in Lean and that experts 'largely agree' the result is correct. Their checking has centred on the shorter, human-edited 49-page 11/12 proof, which is not formalized. Peter Sarnak would not yet call it a theorem; no peer-reviewed publication found as of 2026-11-21.
The Riemann hypothesis says all the important zeros of the zeta function sit on the line Re(s)=1/2. Before this, mathematicians could only say they lie somewhere between 0 and 1, without even a slightly narrower band. OpenAI's model proved that none of them lies beyond 7/8. That is a big step for the theory of primes, but it is not the Riemann hypothesis and has no direct practical effect.
Key figures
Re(s) > 7/8 (formalized)Zero-free half-plane
Re(s) > 11/12 (not formalized)Alternate proof
199 pages7/8 paper length
49 pages11/12 paper length
721 manuscripts in 373 families (repo README); ~42% of top-line results formalizedRelease size
Why it matters
The Riemann hypothesis puts all zeros on Re(s)=1/2. Riemann showed in 1859 that they lie in the strip 0–1, and the edges were excluded in 1896, giving the prime number theorem. No fixed narrower strip (a 'quasi-RH') had been proved. Maksym Radziwiłł told Quanta that 'the vast majority of consequences of RH already follow from quasi-RH'.
How it was done
OpenAI says the zero-free-region work is one of two exceptions to its standard procedure (about 3 hours of ChatGPT Pro thinking compute per result). The 11/12 write-up was edited by a human for readability. Quanta reports that the key was finding the right family of functions, connected to metaplectic forms and a 2021 formula of Dunn and Radziwiłł. The 7/8 result is formalized in Lean, with comparator challenge files for zeta, Dirichlet and Hecke L-functions.
Step by step
Internal model produces the 7/8 proof (paper dated 2026-09-30)
Alternate 11/12 proof (2026-10-05), human-edited for readability
Lean formalization of the 7/8 bound for zeta, all Dirichlet L-functions and finite-order Hecke L-functions over Q(√−3)
Public release 2026-10-06 in the openai/math repository
Experts digest the 11/12 proof; follow-ups such as arXiv 2610.12234 improve the half-plane
Chronology
7/8 paper dated
11/12 alternate proof completed
OpenAI public release
Quanta reports broad expert agreement on correctness
What it brings
Lichtman called it 'the most significant AI-for-math achievement so far, by quite a margin' (Quanta, Oct 2026). Follow-up papers already build on the method.
What it does not show
Does not prove RH and does not place any zero on the critical line. OpenAI's model is unnamed and unreleased.
Human vs AI
The proof was produced by the model. Humans edited the 11/12 exposition, and outside mathematicians are now digesting it.
Credit and priority
Builds on Dunn–Radziwiłł (2021) and long-standing ideas Sarnak promoted. OpenAI's README says it defers 'to the math community completely on matters of credit, precedence, and concurrent work'.
Limits and debate
This is quasi-RH, not RH. It says nothing about how many zeros lie on the critical line. Radziwiłł called the papers 'horribly written'. The 11/12 proof is not formalized. Elsewhere in the same release, three unrelated Hodge-type papers were withdrawn on 2026-10-07 after a sign error; no correction to the quasi-RH paper is recorded.
Reactions
Jared Duker Lichtman (Mathematician, Stanford): Called it the most significant AI-for-math achievement so far, by a wide margin. source
Maksym Radziwiłł (Mathematician, NYU): Said the implications are staggering and that most consequences of RH follow from quasi-RH, but that the papers are horribly written. source
Peter Sarnak (Mathematician, Princeton): Is teaching a course on the proof but won't call it a theorem until humans check all the details. source
Alexander Dunn (Mathematician, Georgia Tech): Called the proof a surprising synthesis of well-known ideas. source
Will Sawin / Thomas Bloom (Mathematicians (Princeton / Manchester)): Both think full RH probably remains far off, though neither is certain. source
Fields it touches
Number theoryAnalytic number theoryDistribution of primesLandau–Siegel zerosAI for mathematics
Making 'green' hydrogen by splitting water in acid-based electrolysers relies on catalysts made of iridium or ruthenium, which are very scarce. Lila's AI-guided lab looked for alternatives and reports palladium-based oxide catalysts without iridium or ruthenium.
Status. arXiv preprint (Sept 2026), not peer-reviewed as of 21 Nov 2026; results come from the company's own campaign.
Full dossier: why, how, steps, sources
Lab
Lila Sciences
People
Ken J. Jenewein (first author), Rafael Gómez-Bombarelli (senior author)
Iridium supply is a recognised bottleneck for scaling electrolysers to gigawatt levels. Finding catalysts made from more available metals that are both active and durable has been a long-standing materials challenge.
How it was done
An AI system designed compositions, decided what to make and test next, and learned from each round, while robots synthesised and screened samples (over 90% automation; humans still moved samples between machines and checked safety). It evaluated 2,942 catalysts across 53 material systems and 26 elements. One lead candidate kept its performance within the target limit over 1,000 hours of operation in acid; other candidates degraded after roughly 200–470 hours.
Step by step
Define activity and stability goals
AI proposes compositions; robots synthesise and screen in batches
Machine-learning models update after each round
Leads advance to long-duration testing
Results posted as an arXiv preprint
What it brings
An example of an industrial 'autonomous science' lab producing a concrete materials lead, and of AI search going beyond what conventional design logic or off-the-shelf language models suggested, according to the authors.
Human vs AI
AI chose experiments in a closed loop; humans supervised, ensured safety and handled sample transfer.
Limits and debate
Preprint only. Lila says this is a materials discovery, not a commercial product, and that longer-term durability and performance in industrial formats are still being evaluated. Palladium is described as more available than iridium and ruthenium, not as abundant. The speed and human-time comparisons are the company's own.
Exhibits finite-time blow-up for the 3D incompressible Euler equations without any external force, from smooth, compactly supported, divergence-free initial data on R³.
Status. Released with the Navier–Stokes result and a Lean repository. Fan Zheng told Scientific American that OpenAI's proof follows the Córdoba–Martínez-Zoroa–Zheng framework while overcoming difficulties in removing the force; mathematicians were still comparing the Euler proofs (Oct 2026). No formal independent verification found.
Full dossier: why, how, steps, sources
Lab
OpenAI
AI systems
Unnamed internal OpenAI model (multi-agent)Lean (GPT-6 Astra for formalization)
Why it matters
Global regularity for 3D Euler (zero viscosity) was a famous open problem. Fefferman's Clay description says it is 'also open and very important', though it is not itself a prize problem.
How it was done
Nearly 100 agents worked about 50 hours on this 'easier' side problem. OpenAI's page says this result convinced the team that Navier–Stokes was the most promising target. The paper uses a localized-oscillation amplification mechanism iterated across scales.
Step by step
2026-09-01: multi-agent effort launched after rumours
Euler group (~100 agents, ~50 h) produces the regularity disproof
Formalization in Lean
2026-09-08: released with the NS paper
What it brings
Gave the first public claim of unforced Euler blow-up from smooth data on R³, and redirected the agent campaign to Navier–Stokes.
Human vs AI
Produced by OpenAI's agent system; the page describes no human mathematical contribution.
Credit and priority
Fan Zheng (quoted by Scientific American) said OpenAI's proof basically follows the framework of the Córdoba–Martínez-Zoroa–Zheng programme.
Limits and debate
OpenAI states its Euler result is unforced while Alpöge–Buckmaster's earlier one was forced. Alpöge, Buckmaster and Coiculescu later wrote that their unreleased second forced-Euler proof uses the 'same Craik–Criminale amplification mechanism as OpenAI's paper'. The Bastounis–Circelli–Hansen critique says the Euler Lean proof is likely to contain substantial mistranslations but did not examine it in detail.
Fields it touches
PDE / fluid dynamicsTurbulence theoryAI for mathematics
For every positive viscosity, constructs a solution of the 3D incompressible Navier–Stokes equations that starts at rest, is driven by a smooth force compactly supported in space and time, keeps bounded kinetic energy, and develops unbounded velocity in finite time. OpenAI says this establishes statement (C), and also (D), of the Clay formulation.
Status. Partial expert endorsement, not a full check. Charles Fefferman, who wrote the Clay problem description, told Quanta he was 'thrilled that the problem was solved'. Zhen Lei and Xiao Ren have published a readable reconstruction of the profile part and call it 'a major advance'. AFP (2026-11-20) reports that mathematicians say it appears correct but is not fully vetted. It is not peer-reviewed, Clay still lists Navier–Stokes as active, and OpenAI says it will not claim the prize. arXiv 2610.08144 argues the Lean code does not faithfully mirror the written proof.
Full dossier: why, how, steps, sources
Lab
OpenAI
People
Sébastien Bubeck
AI systems
Unnamed internal OpenAI model ('significantly more capable than GPT-6 Astra')~10,000-agent systemCodex (consolidation)GPT-6 Astra (Lean formalization)Lean
In plain words
The Navier–Stokes equations describe how water and air move and are used for aircraft design, weather forecasting and blood-flow studies. The open question was whether a smooth flow could ever 'blow up', reaching infinite speed at a point in finite time. OpenAI's AI agents built an example where it does, but only when the fluid is pushed by a carefully designed external force. It says nothing new about how real fluids behave in engineering practice, and the harder case with no force was left open.
Key figures
~10,000Concurrent agents (NS group)
~88 h (2026-09-01 → 09-05)Search time
+17 h (GPT-6 Astra), done 2026-09-06Lean formalization + verification
2.7 millionMessages (NS)
~130 billionOutput tokens (NS)
4.9 million messages, ~300 billion output tokensAll problems attempted
'millions of dollars' (Noam Brown); 'several million dollars' (Bubeck, per Quanta); $15 million reported by Scientific AmericanCost
166 pages (PDF)Paper length
$1M Clay; OpenAI will not claim itPrize
Why it matters
Navier–Stokes existence and smoothness has been a Clay Millennium problem since 2000. Fefferman's official text allows a disproof with a smooth external force f in (C)/(D), while (A)/(B), the existence statements, take f ≡ 0. A forced blow-up therefore settles the prize question as worded without settling whether unforced fluids can blow up.
How it was done
After hearing rumours on 2026-09-01, OpenAI launched groups of agents, each given a different variant of the problem (A/B for proof, C/D for disproof), with code execution and a cached copy of the internet. Codex consolidated insights across groups. The Navier–Stokes group involved about 10,000 concurrent agents and reached a resolution about 88 hours after launch. GPT-6 Astra formalized and verified it in Lean in 17 more hours. The construction is a self-similar vortex whose radial width shrinks faster than its axial length, with oscillations used to cancel the singular part of the residual force.
Step by step
2026-08-28: training of a new internal model begins
2026-09-01: rumours of Millennium results; agent groups launched on A/B and C/D variants
Euler side problem solved first (~100 agents, ~50 h)
2026-09-05: NS group (~10,000 agents) reaches resolution (~88 h after launch)
2026-09-06: Lean formalization and verification complete (+17 h, GPT-6 Astra)
2026-09-08: announcement, 166-page paper and Lean repo; 2026-09-10 investigation update
OpenAI updates 'Concurrent work' with investigation findings
Lei & Ren reconstruction (arXiv 2609.35406)
'Navier-Stokes lost in translation' critique (arXiv 2610.08144)
What it brings
First claimed resolution of a Clay Millennium problem by an AI company, and a showcase of massive multi-agent search plus autoformalization. Practical engineering impact is minimal, per experts such as Chris Combs.
What it does not show
Does not disprove (A)/(B), i.e. unforced smooth initial data with no force. Does not change engineering CFD practice. OpenAI does not claim the Millennium Prize.
Human vs AI
OpenAI describes the mathematics as produced by the agent system; humans set up the campaign and problem variants. Bubeck said OpenAI 'collectively had no research-level expertise in fluid dynamics'.
Credit and priority
OpenAI recognizes Alpöge and Buckmaster's priority on forced Euler and says it did not see their work before release. After an investigation, it says Buckmaster's Codex prompts 'could not have influenced the system in any way, including through training'. Buckmaster disputes this ('You are not going to be able to convince me that you did not get the key ideas from a leak'). Scientific American reports Bubeck offered Buckmaster lead authorship but asked that Alpöge's name be removed because of his affiliation with a competitor; a post quoted in Altman's 2026-09-08 thread denies ever asking for Alpöge to be removed from authorship of his own work. Both AI teams built on Córdoba and Martínez-Zoroa, whom Fefferman calls the 'heroes of the story'.
Limits and debate
Forced, not unforced: the fluid is stirred by a finely tuned external force, and the unforced statements (A)/(B) are not disproved by this result. Lean verification was OpenAI's own. Bastounis, Circelli and Hansen show mismatches between Lean and the written proof (derivative counts, a pressure-flux bound), but do not claim the theorem is false. Clay requires publication, two years and general acceptance. The result was surrounded by a priority dispute.
Reactions
Charles Fefferman (Princeton; author of the Clay problem description): Thrilled that the problem was solved; credits Córdoba and Martínez-Zoroa as the heroes. source
Diego Córdoba (Mathematician, Madrid): Said that ten years ago nobody believed Navier–Stokes had a singularity. source
Tristan Buckmaster (NYU): Says OpenAI's NS proof has new ideas but builds on his programme, and suspects a leak; OpenAI denies it. source
Fan Zheng (Mathematician, Spain): Says OpenAI's work basically follows the framework of the Córdoba–Martínez-Zoroa–Zheng programme. source
Zhen Lei and Xiao Ren (Mathematicians (arXiv 2609.35406)): Wrote a readable version of the profile construction and call OpenAI's work a major advance. source
Alexander Bastounis, Fabian Circelli, Anders C. Hansen (Authors of arXiv 2610.08144): The Lean proof does not correspond step by step to the natural-language proof, so Lean does not certify the written argument. source
Chris Combs (Professor of aerospace engineering): Posts saying 'we solved Navier–Stokes' are overblown: this is a very specific proof about a niche scenario. source
Noam Brown (OpenAI researcher): Acknowledged the result cost millions of dollars and compared it with the falling cost of earlier milestones. source
Sam Altman (OpenAI CEO): Said OpenAI initially believed the other team had also solved the problem and wanted a joint release. source
Fields it touches
PDE / fluid dynamicsTurbulence modellingAerospace and weather modelling (conceptual)AI for mathematicsFormal verification
Köthe's conjecture (1930) asserts that the sum of two nil left ideals of a ring is again nil. Working autonomously, a pre-release GPT-6 Astra disproved its matrix form (Krempa's version): it built a ring over a countable field with a nil two-sided ideal and a 2×2 matrix over that ideal that is not nilpotent, which also refutes the original conjecture.
Status. The Lean proof uses no 'sorry' and only standard axioms, checked against a trusted statement with Comparator. A generalised human-written version (arXiv 2609.07996) was reviewed by its second and third authors; an independent short construction followed (Greenfeld–King–Vendramin, arXiv 2609.15080).
Full dossier: why, how, steps, sources
Lab
OpenAI Found during an Epoch AI evaluation of an OpenAI model.
With
Epoch AI
People
Tom Adamczewski, Bernhard Böhmler, René Marczinzik
The conjecture had been open since Köthe posed it in 1930. Smoktunowicz's 2000 counterexample to Amitsur's conjecture had fuelled suspicion that Köthe's conjecture might be false.
How it was done
Epoch AI ran the model on 222 open statements with a $1,000 budget and 96 hours each; on this one it found and formalised a disproof in Lean with no human seeing or steering the search, at a cost of $432 and 26.1 hours. Humans then wrote a generalised paper that also answers a question of Rowen.
Step by step
Epoch AI places 222 open statements in a Lean harness
GPT-6 Astra searches autonomously ($1,000 / 96 h budget)
It produces a sorry-free Lean disproof of the matrix form
Comparator checks the theorem against the trusted statement
Adamczewski, Böhmler and Marczinzik publish a generalised human-written paper
What it brings
Among the most autonomous AI results on a classical problem to date, with a formal certificate; it prompted independent human constructions within a week.
Human vs AI
Fully autonomous discovery and formalisation; humans selected the statement set and wrote the exposition.
Limits and debate
The repository notes that the AI's informal proof account was not checked by a human mathematician (the formal proof was), and that the implication from the matrix form to the original conjecture is not formalised. Not peer reviewed as of 21 Nov 2026.
Fields it touches
Ring theory / noncommutative algebraFormal verification
Constructs solutions that blow up in finite time under a smooth external force for the incompressible porous media equation, the Boussinesq equations and the 3D incompressible Euler equations.
Status. Buckmaster says the proofs were Lean-verified, the first one on 2026-08-22. OpenAI publicly recognized their priority on forced Euler. No independent expert check of these papers was found as of 2026-11-21.
Full dossier: why, how, steps, sources
Lab
Anthropic Personal collaboration between Levent Alpöge (Anthropic) and Tristan Buckmaster (NYU), using Claude.
With
New York University
People
Levent Alpöge, Tristan Buckmaster
AI systems
ClaudeOpenAI Codex (GPT-5.6 Sol; Astra for write-ups and auditing)Lean
Why it matters
Pushes the Córdoba–Martínez-Zoroa programme, which had obtained blow-up with rough forcing, to smooth forcing, widely seen as a step toward Navier–Stokes.
How it was done
Starting from the Córdoba and Martínez-Zoroa construction, the authors used LLMs heavily to extend it. Buckmaster says the first LLM-generated proof was 'the most horrendous' he had read and that they verified it in Lean on 2026-08-22.
Step by step
About a year of slower progress (literature, IPM with smooth forcing)
2026-08-15: blow-up with smooth forcing for Boussinesq and Euler
2026-08-22: Lean verification
2026-09-03: rumours circulate; Buckmaster emails an OpenAI mathematician
2026-09-07: public release of three results
Chronology
Smooth-forcing blow-up for Boussinesq and Euler obtained
Verified in Lean
Buckmaster writes to OpenAI amid rumours
Public release
What it brings
First smooth-forcing blow-up for 3D Euler, and the trigger (through rumours) for OpenAI's multi-agent Navier–Stokes campaign.
Human vs AI
Humans chose the programme and directed the work; LLMs wrote much of the proof, which was then Lean-checked.
Credit and priority
Buckmaster credits the programme's basic idea to Diego Córdoba and Luis Martínez-Zoroa and says Martínez-Zoroa 'deserves a Fields Medal'. OpenAI's page says: 'We recognize the priority of their work on forced Euler.'
Limits and debate
Forced, not unforced. Buckmaster himself says the Euler write-up 'can only be described as AI slop'. A hypodissipative Navier–Stokes result was mentioned but not released at the time.
Fields it touches
PDE / fluid dynamicsAI for mathematicsFormal verification
Fermat's Last Theorem says that a^n + b^n = c^n has no solutions in positive whole numbers when n is 3 or more. Andrew Wiles (with Richard Taylor) proved it in the 1990s on paper. Claude agents translated that proof, end to end, into the Lean proof assistant, so a computer now checks every logical step down to Lean's basic axioms.
Status. September artifact: public repo; builds only if the proof depends on exactly Lean's three standard axioms; comparator confirms the statement matches Mathlib's FermatLastTheorem; independent kernel nanoda accepted it; Kevin Buzzard compiled it and ran the comparator himself. November 30-hour run: reported only by Anthropic (launch post + engineering post); the engineering post links no repository for that proof, so it has not been independently checked as of 2026-11-21.
Full dossier: why, how, steps, sources
Lab
Anthropic
With
Columbia University (Prove2Me platform, Tianyi Peng's group), Imperial College London (FLT project code reused)
People
Tianyi Peng, Kevin Buzzard, Barry Zhang, Elle Najt, Keller Scholl, Lance Martin
AI systems
Claude (internal research model roughly comparable to Claude Fable 5.1)Claude Fable 5.2 (early pre-release build)Prove2MeLean 4Mathlibleanprover/comparatornanoda (independent Rust kernel)
In plain words
Fermat's Last Theorem is a statement about whole numbers that was famously hard to prove; Wiles's 1990s proof runs to well over a hundred pages and leans on decades of other work. 'Formalizing' it means rewriting every step in a programming language (Lean) so that a small, trusted computer program can check the whole argument mechanically. Claude agents did that translation almost alone. Nothing changes in everyday life or in what mathematicians believe: experts were already near-certain the proof was right. What changes is the evidence that AI can now translate big, real mathematics into machine-checked form quickly.
Key figures
~11 days (Aug 7–17, 2026)Sept run duration
~13 million lines of Lean (>5x Mathlib)Sept proof size
29,511Theorems in Sept proof
~6 billionSept output tokens
60,475Lean modules built
5 h 32 min at 96 jobsFull build time
1,000 (1 lead + 999 helpers)Nov agents
28.7 h ("30 hours" in launch post)Nov time to FLT
~1/3 the output tokens, proof ~1/7 as longNov efficiency vs Sept
>1 million, none blockedNov tool calls monitored
26% of the formalization in 18 hOpus 5.5 same setup
Why it matters
FLT is one of the most famous theorems in mathematics, and its proof rests on a vast amount of modern number theory (Frey, Serre, Ribet, Wiles, Taylor–Wiles). Formalizing it was treated as a multi-year human project: Kevin Buzzard's EPSRC-funded Imperial College effort began in 2024, with £1M over 5 years. Buzzard had promised to reduce FLT to results known in the 1980s, not to finish the whole proof.
How it was done
In September, dozens of Claude agents worked largely autonomously for about 11 days on Prove2Me, a Columbia platform that keeps a graph of theorem statements and separates statements from proofs. They followed the 1995 Darmon–Diamond–Taylor exposition of Wiles's argument and reused pieces of the Imperial FLT project, flt-regular and Mathlib. Humans wrote no mathematics and no Lean beyond the one-line goal statement. Tianyi Peng gave occasional high-level priorities, such as 'Jacobian as a scheme sounds high priority'. In November, an early Fable 5.2 build ran as 1 lead plus 999 helper agents, each in a sandbox with no network access, coordinating through Git and a shared Postgres database. Without Prove2Me or a researcher steering, they proved the final theorem at hour 28.7.
Step by step
2024: Kevin Buzzard's Imperial FLT formalization project begins (EPSRC, £1M over 5 years), building on Mathlib.
Aug 7, 2026: Claude provers start on the Prove2Me platform; the target is the one-line Lean statement of FLT.
Aug 7–17: agents formalize the chain of steps: Taylor–Wiles primes, Ribet's level-lowering, Mazur's results, Langlands–Tunnell, modularity lifting. A false lemma that passed one review was caught by another.
Aug 17: the full proof checks in Lean (~13 million lines, 29,511 theorems), then goes through external checking: from-scratch build, comparator, nanoda.
Sep 4: Anthropic publishes the repo and a timeline write-up; Buzzard independently compiles it and runs the comparator.
Nov 9: Anthropic reports that 1,000 early Fable 5.2 agents, given 60 target statements, a one-sentence route sketch, definitions and 47 lemmas from the September proof, reached FLT at hour 28.7 (54/60 statements by hour 30).
Full FLT proof checks in Lean (end of ~11-day run)
Anthropic publishes repo + write-up; Buzzard's 'Anthropic has beaten me to it' post
Fable 5.2 / Mythos 5.2 launch: 1,000 early Fable 5.2 agents redo the formalization in ~30 h (FLT at hour 28.7)
Anthropic amends the launch post's 'Mathematical formalization' subsection and adds a footnote on inputs given to the model
What it brings
This is the first complete computer-checked proof of FLT, years ahead of the human-led schedule. Buzzard says AI autoformalization artefacts are 'now robust enough to be built upon'. The November run suggests that a large swarm of agents can organize itself on a big formal-proof project in about a day, with real-time safety monitoring across more than a million tool calls. It is a large-scale stress test of agent coordination, not a new theorem.
What it does not show
No new mathematics and no new proof strategy. The November run did not start from scratch: it reused September lemmas and statement lists, and its artifact has not been independently verified. Neither run produces a human-readable exposition, and Buzzard doubts Anthropic will build one. This is not the modern proof route that Buzzard's grant targets.
Human vs AI
September: AI wrote all the Lean, and one researcher gave occasional priority nudges and encouragement. November: humans set up the task and sandboxes, restarted the run from its last save near hour 18, and reviewed classifier blocks (none was triggered). The harness sent the lead agent 19 automatic time reminders.
Credit and priority
The mathematics belongs to Frey, Serre, Ribet, Wiles and Taylor–Wiles, in the Darmon–Diamond–Taylor exposition. The Lean code builds on Mathlib, the Imperial College FLT project led by Buzzard and flt-regular: ATTRIBUTION.md lists 106 files from the latter two and 23 that reproduce Mathlib text. Prove2Me is by Chen, Marwaha, Lu, Yuen and Peng (arXiv 2608.28433). Buzzard notes the formalization covers primes p >= 17, with the remaining cases handled by the existing formalization for odd regular primes.
Limits and debate
This verifies known mathematics; it is not a new result. The September proof is about 13 million lines, more than 5x Mathlib. Anthropic itself says it is 'likely much longer than it needs to be' and 'written to be checked rather than read'. No tool can check that each intermediate theorem means what its machine-generated name suggests. The proof follows the older 1995 exposition, not the modern route Buzzard is formalizing. The November run had a head start: 47 proved lemmas and the 60-statement milestone list both came from the September work. Anthropic calls it a 'toy experiment' on a model that may not have finished pre-release alignment assessment. Its proof has no public repo, and the launch post's formalization text was amended on Nov 11 with a footnote on what the model was given.
Reactions
Kevin Buzzard (Imperial College London, leads the EPSRC FLT formalization project): Headlined his post 'Anthropic has beaten me to it'; verified the build and comparator himself, but says mathematically the work 'tells us essentially nothing', since the proof was already near-certain. source
Kevin Buzzard (Imperial College London): Sees the process as the real news: if an AI swarm can formalize thousands of pages in eleven days, formalizing modern research could soon be quick. He checked for cheating and found nothing. source
Kevin Buzzard (quoted by Anthropic): Says the work proves FLT with no assumptions other than the axioms of mathematics and that such artefacts are now robust enough to build on. source
Fields it touches
Number theoryFormal verification / interactive theorem provingAutoformalizationProof-assistant libraries (Mathlib)Refereeing and error-checking of the mathematical literatureMulti-agent AI coordination and safety
A complex structure lets a space be described locally by complex coordinates, like the 2-sphere viewed as the Riemann sphere. Among spheres, only S² and S⁶ admit almost complex structures, and whether S⁶ admits a genuine (integrable) one was Hopf's question. Alpöge announced 'Yup': a compact complex threefold diffeomorphic to S⁶, built by compactifying the (3,4,∞) triangle-group family of complex 2-tori.
Status. Philip Engel (UIC) wrote a self-contained proof of the construction (arXiv:2609.38442, 29 Sep 2026). Robert Bryant (Duke) spoke of an 'emerging consensus that the construction is plausible' (SciAm, 24 Sep), and Jeff Viaclovsky has published further families of complex structures on S⁶. Lean formalizations exist (Boris Alexeev; Dean Cureton, standard axioms only), but their completeness is reported inconsistently: Wikipedia says 'parts', SciAm says 'step by step'. Not peer-reviewed.
Full dossier: why, how, steps, sources
Lab
Anthropic Levent Alpöge is an Anthropic researcher; the work was done with Claude.
People
Levent Alpöge
AI systems
Claude (internal version; Alpöge says Opus 5 wrote the 100+ page PDF)Lean 4 (independent formalizations)
In plain words
Mathematicians classify shapes by which kinds of coordinates can be laid over them. A 2-dimensional sphere can be described with complex numbers, which is why it is so central in physics and engineering math. For 79 years nobody knew whether the 6-dimensional sphere could be too. Alpöge and Claude built such a description, and other mathematicians have since rewritten and extended it. It has no practical consequence; it is a landmark in pure geometry.
The question has been open since 1947 (Hopf). Alpöge notes it is 'a problem with a lot of claims either way in the literature that fail due to subtleties' (a reply in his thread points to a non-existence paper by Atiyah). Alpöge's PDF explains why an earlier argument [CDP20] does not apply.
How it was done
Per Alpöge, 'claude really contains multitudes', and he 'got opus 5 to write every detail down in a certainly far too long 100+ page pdf'. Wikipedia and SciAm describe it as generated with an internal Claude version. Alpöge also posted two pages meant to be enough to replicate the key finite computation: π₁ vanishes and the homology is right, which suffices since there are no exotic 6-spheres. Mathematicians then reconstructed and formalized the proof: Engel's pedagogical account (27 Aug notes, 29 Sep arXiv), and Lean projects by Alexeev and Cureton.
Step by step
Alpöge works on the problem with an internal Claude model.
Construction: (3,4,∞) triangle-group family of complex 2-tori, completed at three special points (Kodaira logarithmic transforms; Mumford toric degeneration at the cusp).
23 Aug 2026: X announcement with a two-page certificate and the 108-page PDF written by Claude.
27 Aug: Engel posts expository notes; Lean formalization efforts by Alexeev and Cureton follow.
Late Sep–Oct 2026: Engel's arXiv proof and Viaclovsky's new families on arXiv.
Chronology
X announcement and alpo.ge/s6.pdf (108 pp.)
Engel posts expository notes
Scientific American coverage
Viaclovsky two-parameter family (arXiv:2609.33785)
Engel self-contained proof (arXiv:2609.38442)
Viaclovsky 'Combs and complex structures on the six-sphere' (arXiv:2610.10451)
What it brings
If correct, it answers a 79-year-old question and opens a new class of geometric objects. Engel: 'There is a new geometric object in the world that we now know about and can explore.' Follow-ups already show large families of complex structures on S⁶.
What it does not show
It is not a peer-reviewed result yet. It does not say the octonionic almost complex structure is integrable (it is not). It has no applied consequence.
Human vs AI
Alpöge directed the work and supplied the short certificate. Claude produced the detailed manuscript (Engel: 'produced by Claude under the direction of Levent Alpöge'). Humans re-derived and formalized it afterwards.
Credit and priority
Alpöge and Claude. Lean work is by Boris Alexeev (an OpenAI researcher, per SciAm) and by Dean Cureton. The human exposition and verification is Engel's. The construction relies on classical tools (Kodaira, Mumford degenerations, triangle groups).
Limits and debate
The original 108-page document was described as 'nigh-indecipherable' (SciAm). How much guidance Alpöge gave is not disclosed, a point Ilka Agricola (Marburg) criticised. The completeness of the Lean formalizations is reported inconsistently. There is no peer-reviewed publication yet.
Reactions
Robert L. Bryant (Duke University): 'Seems to be an emerging consensus that the construction is plausible.' source
Philip Engel (University of Illinois Chicago): Wrote a self-contained proof; 'a new geometric object in the world'. source
Ilka Agricola (University of Marburg): Criticised the lack of detail on prompts and human guidance as typical of tech announcements. source
Jeff Viaclovsky (mathematician, arXiv:2609.33785): Built a two-parameter family of complex structures on S⁶, not biholomorphic to the Alpöge–Claude examples. source
The rank of an elliptic curve y² = x³ + Ax + B over the rationals measures how many independent rational solutions generate all others. Whether ranks are unbounded is open. Alpöge and cryptographer Ava Howell, working with an internal Claude model, posted a curve of rank at least 30 and, three days later, one of rank at least 31.
Status. There is no paper. The lower bounds rank ≥ 30 and ≥ 31 rest on explicit independent rational points posted on the ICARM leaderboard (curves #273 and #302), and ICARM marks #302 verified on 2026-09-25. The exact ranks (30, 31) are conditional on BSD and GRH. A rank ≥ 32 curve by Elkies, Klagsbrun and Li was submitted on 12 Nov 2026, so the record no longer stands.
Full dossier: why, how, steps, sources
Lab
Anthropic Levent Alpöge is an Anthropic researcher; the work was done with Claude.
With
NSF Institute for Computer-Aided Reasoning in Mathematics (ICARM) leaderboard
People
Levent Alpöge, Ava Howell
AI systems
Claude (internal Anthropic model)
In plain words
Elliptic curves are the equations behind much modern cryptography, though this result has no direct security implications. Their 'rank' counts how many independent rational solutions you need to generate all the others. Big ranks are extremely rare, and finding a bigger one is a decades-long sport. With Claude's help, Alpöge and Howell broke the record twice in four days. Human experts then beat it again in November.
Key figures
rank 29 (Elkies–Klagsbrun, 2024)Previous record
rank ≥ 30 (20 Aug) and ≥ 31 (23 Aug 2026)New records
3 daysTime between records
~18 yearsRank 28 → 29 took
rank ≥ 32, Elkies–Klagsbrun–Li, submitted 12 Nov 2026Current record (as of 21 Nov)
Why it matters
Going from rank 28 (Elkies, 2006) to 29 (Elkies–Klagsbrun, 2024) took about 18 years. The problem was listed on Epoch AI's FrontierMath open-problem list.
How it was done
Scientific American reports they used an internal Claude variant with a 'relatively simple prompt'. Leaderboard commentary describes curve #302 as a specialization of a parametrized family, with witness points supplied for independent checking. Humans chose the direction and checked results; secondary coverage describes Claude as a fast search collaborator. Compute was not disclosed.
Step by step
Alpöge and Howell set Claude on searching families of high-rank curves.
20 Aug 2026: curve #273 (rank ≥ 30) submitted to the ICARM leaderboard with 30 witness points.
23 Aug 2026: curve #302 (rank ≥ 31) submitted with 31 points, noted 'BSD + GRH certified to rank 31'.
25 Sep 2026: ICARM marks #302 verified.
12 Nov 2026: Elkies–Klagsbrun–Li curve #6782 (rank ≥ 32) supersedes the record.
Chronology
Rank ≥ 30 curve #273 submitted
Rank ≥ 31 curve #302 submitted
Scientific American coverage
ICARM verifies #302
Rank ≥ 32 curve #6782 by Elkies, Klagsbrun, Li
What it brings
It set two rank records in four days after an 18-year plateau and triggered a burst of human follow-ups: other rank-31 curves (#846, #7331) and then rank ≥ 32. It shows AI-assisted search can move long-stagnant computational records.
What it does not show
It does not prove unbounded ranks or any conjecture. The exact rank is not proven unconditionally. It has no cryptographic impact.
Human vs AI
Credit on the registry is 'Claude, Levent Alpöge, and Ava Howell'. The detailed division of labour is not published.
Credit and priority
Credit goes to Claude, Alpöge and Howell (ICARM). The search methods descend from the Elkies–Klagsbrun line of work. The record has since passed to Elkies, Klagsbrun and Li.
Limits and debate
Only lower bounds are unconditional; the exact rank depends on BSD and GRH. Curve #273's original submission silently dropped one witness point because of a leaderboard parser bug that was later fixed. The record was surpassed by a human team on 12 Nov 2026. There is no write-up of the method.
Reactions
Epoch AI (FrontierMath open-problems list): Marked the problem 'solved (human + AI)', first solved with an internal Anthropic Claude model. source
Fields it touches
Number theoryArithmetic geometryCryptography (elliptic curves)Computational number theory
An umbilic point on a surface is a point where it curves equally in every direction. The Carathéodory conjecture says every smooth closed convex surface in 3D space has at least two. The related Loewner conjecture bounds the 'index' of an isolated umbilic by 1. Alpöge posted explicit formulas giving, for each k, an umbilic of index 1 + k/2, and a C∞ (infinitely smooth) convex body with exactly one umbilic, of index 2. Both conjectures fail in the C∞ setting.
Status. Alpöge's construction exists only as an X post and a privately circulated PDF; we found no public paper and no Lean formalization (one X user said they were starting one). Differential geometers Naoya Ando and Masaaki Umehara (arXiv:2610.34036, 25 Oct 2026) published a self-contained, simpler variant that gives the same conclusion. That preprint is not yet peer-reviewed.
Full dossier: why, how, steps, sources
Lab
Anthropic Levent Alpöge is an Anthropic researcher; the work was done with Claude.
People
Levent Alpöge, John-Paul Smith
AI systems
Claude
In plain words
Think of the points on a ball where it is curved the same way in all directions. On a perfect sphere every point is like that. For any egg-shaped convex surface, people believed there must always be at least two such special points, a claim from about 1924. Alpöge, a colleague and Claude wrote down a perfectly smooth convex shape with only one. This is pure geometry with no direct practical impact.
Key figures
~1924 (attributed to Carathéodory)Conjecture age
exactly 1, index 2Umbilics on the counterexample
1 + k/2 for every positive integer kIndex family
Why it matters
The conjecture is attributed to Carathéodory around 1924 (via Hamburger). Its history is full of disputed proofs: Hamburger for analytic surfaces (1940–41), Bol, Klotz, Ivanov (2002), and Guilfoyle–Klingenberg, who announced a proof for C^{3,α} surfaces in 2008 and claimed the global case in 2025. Alpöge cited this 'complicated literature situation' as the reason to look at it.
How it was done
Alpöge discussed the problem with @AchimWar, asked whether it was OK to 'look at with a model', and then 'me, John-Paul Smith, and of course claude checked' the construction. A PDF write-up circulated privately for a week before the 19 August post. The post gives a trigonometric seed f and functions g_k built from it via a support-function parametrization. Alpöge noted that the constants (1/4, 100, 10^10) 'might need some kicking'. Ando and Umehara then gave a simpler variant with a trigonometric seed.
Step by step
Alpöge and @AchimWar discuss favourite problems; Carathéodory comes up.
Alpöge explores it with Claude; Alpöge, John-Paul Smith and Claude check a construction.
A PDF write-up circulates privately for about a week.
19 Aug 2026: explicit formulas posted on X as a 'form certifiable by models'.
25 Oct 2026: Ando–Umehara arXiv:2610.34036 give a simpler self-contained construction (exactly one umbilic, index 2).
Chronology
Approx.: Alpöge discusses problem with @AchimWar ('a week ago')
Explicit counterexample posted on X
Ando–Umehara simpler construction on arXiv
What it brings
It settles the smooth case of a century-old problem negatively and also disproves the Loewner index bound in C∞. It sharpens the open question to the real-analytic category, where Hamburger-type index bounds apply.
What it does not show
It does not cover real-analytic surfaces, where the classical question is believed true. No peer-reviewed paper or formal proof of Alpöge's own version has been published.
Human vs AI
Alpöge credits Claude as a collaborator alongside John-Paul Smith. The division of labour is not described.
Credit and priority
Credit goes to Alpöge, John-Paul Smith and Claude. @AchimWar raised the problem. Ghomi and Howard (2012) had earlier built weaker C¹ counterexamples. Ando–Umehara's construction is a human simplification of Alpöge's.
Limits and debate
The result is about C∞ surfaces. The real-analytic case is not disproved by this construction, and the conjecture is believed (via index bounds) for analytic surfaces. A smooth counterexample is in direct tension with Guilfoyle–Klingenberg's claimed C^{3,α} proof. We found no public response from them. Alpöge's own write-up is not public, and he flagged that some constants might need adjustment.
Reactions
Naoya Ando & Masaaki Umehara (differential geometers, arXiv:2610.34036): Gave a simpler self-contained variant of the construction. source
A Hadamard matrix is a square grid of +1 and −1 entries whose rows are pairwise orthogonal. The Hadamard conjecture says one exists for every order divisible by 4. On 12 August 2026 Alpöge posted a 23,828-character string of + and − characters, plus a sed decoder. It encodes Hadamard matrices at all twelve orders below 2000 where none had been known: 668, 716, 892, 1132, 1244, 1388, 1436, 1676, 1772, 1916, 1948 and 1964.
Status. The matrices are their own proof: checking means computing HHᵀ = nI. Egan, Kotsireas, London and Ó Catháin (arXiv:2610.27175, Oct 2026) independently verified and analysed all twelve, and Ion Nechita checked order 668 with a short script. The discoverers have not published a paper or their method.
Full dossier: why, how, steps, sources
Lab
Anthropic Levent Alpöge is an Anthropic researcher; the work was done with Claude.
People
Levent Alpöge, Philippe Voinov, Saul Reynolds-Haertle
AI systems
Claude (internal Anthropic model, per Epoch AI)
In plain words
Hadamard matrices are grids of +1 and −1 whose rows are perfectly 'uncorrelated'. They are used in error-correcting codes, signal processing and experiment design. Mathematicians believe one exists for every size that is a multiple of 4, but for some sizes nobody had ever found one, and 668 was the smallest such gap. The team found one for 668 and for the eleven other missing sizes under 2000. The general conjecture remains unproved.
Key figures
12 (all open orders < 2000)New orders
23,828 ± charactersAnnouncement length
2005Order 668 open since
2396 (Egan et al.)Smallest unknown order now
95 (Egan et al.)Unknown orders ≤ 10,000 now
16 below 10,000 (Egan et al., with Claude Opus 5.5)Further orders found in follow-up
Why it matters
Order 668 had been the smallest unresolved case of the Hadamard conjecture since 2005, when order 428 was settled (Wikipedia; X replies). It was listed as an open problem on Epoch AI's FrontierMath page. Earlier AI benchmark attempts on it had failed (HorizonMath, per a reply to the announcement).
How it was done
Alpöge described it as 'weekend fun' with @tehwalris (Philippe Voinov), Saul Reynolds-Haertle 'and of course claude', and said there was 'a search (and an actual idea, which i don't claim!!)'. The method was not disclosed. Egan et al. reverse-analysed the matrices. Each is a Goethals–Seidel array over an abelian group: nine over cyclic groups (five unbordered from prime-order circulants, four with a width-4 border from Legendre quadruples) and three over non-cyclic groups with wider borders. The b = 0 generators are unusually spectrally flat, which suggests a restricted search.
Step by step
The human–AI team (Alpöge, Voinov, Reynolds-Haertle, Claude) combines a structural idea with a computer search.
Matrices are found at all 12 open orders below 2000, including 668.
12 Aug 2026: they are posted on X as a ±-string plus an obfuscated sed decoder.
Mathematicians decode and check HHᵀ = nI (Nechita, 13 Aug).
Oct 2026: Egan et al. verify all twelve, identify Goethals–Seidel structure, and use Claude Opus 5.5 to help find 16 further orders below 10,000.
Chronology
Alpöge posts the ±-string and decoder on X
Nechita blog post verifying order 668
Egan, Kotsireas, London, Ó Catháin arXiv:2610.27175 (v2 3 Nov)
What it brings
For the first time a Hadamard matrix is known for every order divisible by 4 below 2000 (MathWorld). The smallest unknown order jumps from 668 to 2396. The result also prompted follow-up human work that extended the method.
What it does not show
It does not prove the Hadamard conjecture. No construction method or theorem was published by the discoverers. It does not settle any order above 2000.
Human vs AI
Alpöge says the 'actual idea' was not his, and Epoch AI notes humans may have contributed core ideas. The split between Voinov, Reynolds-Haertle and Claude is not public. Epoch has 'provisionally' marked the problem as solved by AI and will revise this if humans contributed significantly.
Credit and priority
Credit goes to Alpöge, Philippe Voinov and Saul Reynolds-Haertle together with Claude. Alpöge says he 'only claim[s] bad suggestions'. Epoch AI's 'solved by AI' label is explicitly provisional. The Goethals–Seidel and Legendre-pair machinery the matrices use is classical human work.
Limits and debate
No write-up of the method has been published. The announcement format (a raw ± string) drew criticism, including a public request from @littmath for an explanation. This is computer-assisted construction, not a step toward proving the Hadamard conjecture in general.
Reactions
Ion Nechita (CNRS): Verified the order-668 matrix with a ~20-line script on his phone and guessed at a model-plus-search recipe. source
Ronan Egan, Ilias Kotsireas, Stephen London, Pádraig Ó Catháin (authors of arXiv:2610.27175): Independently verified all twelve matrices, characterised them as Goethals–Seidel arrays and extended the list. source
@littmath (mathematician on X): Asked how it was found and whether it was an autonomous AI construction; Alpöge replied that a search plus a human idea was involved. source
Epoch AI (FrontierMath open-problems list): Marked the problem solved, with provisional AI attribution. source
Proves that at least 67.2% of the nontrivial zeros of the Riemann zeta function lie on the critical line Re(s)=1/2, i.e. satisfy the Riemann hypothesis. The previous proven lower bound cited by Anthropic was 41.6%.
Status. Anthropic reports a Lean formalization that passes the standard comparator tool, internal validation by Anthropic mathematicians Levent Alpöge and Ralph Furman, and a short-notice look by outside experts Brian Conrey and Dan Goldston. No journal publication found as of 2026-11-21.
Full dossier: why, how, steps, sources
Lab
Anthropic
People
Jarred Sumner, Eric Easley, Levent Alpöge, Ralph Furman, Brian Conrey (reviewer), Dan Goldston (reviewer)
AI systems
Unreleased research version of ClaudeClaude CodeLean (comparator)
In plain words
The Riemann zeta function encodes how prime numbers are spread out; its 'zeros' are special points, and the Riemann hypothesis says they all sit on one vertical line. Mathematicians can prove that some share of them do. This result raises the proven share from about 42% to about 67%. It does not settle the hypothesis and has no direct practical application.
Key figures
67.2% of zeros on the critical lineNew lower bound
41.6%Previous bound (as cited)
~31 million (two Claude Code sessions)Output tokens
~60Subagents
~2,400Shell commands
54arXiv papers checked for novelty
Why it matters
The Riemann hypothesis (Riemann, 1859) says all nontrivial zeros lie on the critical line; proving that a positive and growing fraction do is the classical 'partial' route. The fraction had moved slowly through the work of a series of mathematicians; Claude's argument builds on work by Aryan, by Baluyot–Goldston–Suriajaya–Turnage-Butterbaugh, and on a 2000 paper of Bombieri.
How it was done
A non-mathematician staff member, Jarred Sumner, asked Claude to 'take a real stab' at the full Riemann hypothesis; about 650 first-round ideas failed. In a second push of about a day and a half, Claude coordinated roughly 60 subagents that ran about 2,400 shell commands, wrote hundreds of Python scripts, reviewed proofs, hunted counterexamples and downloaded 54 arXiv papers to check novelty. The improved bound emerged as a by-product. Claude then produced a Lean formalization with staff member Eric Easley.
Step by step
Prompt to attack the full Riemann hypothesis (first ~650 ideas fail)
Second ~1.5-day multi-agent session in Claude Code (~60 subagents)
Bound improved from 41.6% to 67.2%; subagents check proof and novelty against 54 arXiv papers
Lean formalization with Eric Easley; passes comparator
Validation by Anthropic mathematicians and short-notice review by Conrey and Goldston
Revised paper posted 2026-08-13
Chronology
Anthropic announces the 67.2% bound
Revised paper posted
What it brings
A large jump in a decades-old partial result, obtained mostly autonomously in about 31 million output tokens. It was an early public example of an AI system producing a formally verified, research-level number-theory result.
What it does not show
No proof of the Riemann hypothesis; no claim that the method extends to one.
Human vs AI
Anthropic describes the human prompting as mostly encouragement ('keep going'); humans formalized with Claude and validated the paper afterwards.
Credit and priority
Builds on Aryan; Baluyot, Goldston, Suriajaya and Turnage-Butterbaugh; and Bombieri (2000). Human contributors named by Anthropic: Jarred Sumner (prompting), Eric Easley (formalization), Levent Alpöge and Ralph Furman (validation).
Limits and debate
Not a proof of the Riemann hypothesis; Anthropic wrote that it does not expect these techniques to lead to one. Made with an unreleased model, so outsiders cannot reproduce the run.
Reactions
Brian Conrey and Dan Goldston (External experts in the zeta function): Examined the paper on short notice, per Anthropic. source
Fields it touches
Number theoryAnalytic number theoryDistribution of prime numbersAI for mathematics
Sendov's conjecture says that if a polynomial of degree n ≥ 2 has all its roots in the unit disk, then every root has a critical point (a root of the derivative) within distance 1. Lech Mazur posted a computer-assisted proof for all n ≥ 2, generated with GPT-5.6 Pro; Tao showed the argument also proves the stronger Phelps–Rodriguez conjecture.
Status. Mazur's Lean 4 development proves the exact statement (about 90,000 lines per Tao); Tao produced a streamlined proof and a roughly 15,000-line Lean formalisation. The mid-range of degrees relies on exact computer certificates; independent specialist review of the exposition was still invited.
Full dossier: why, how, steps, sources
Lab
Independent researchers Lech Mazur, using an OpenAI model.
People
Lech Mazur, Terence Tao
AI systems
GPT-5.6 ProLean 4
Why it matters
Previously it was known only for small degrees (n ≤ 8, Brown–Xiang) and for sufficiently large n (Tao's earlier work), with no explicit threshold. Tao judged the AI-generated proof remarkably elementary.
How it was done
Mazur designed an orchestration and verification workflow in which GPT-5.6 Pro explored, developed and audited the proof. The finite range of degrees is certified by exact rational arithmetic, while large degrees use elementary estimates. Tao spent several days, with heavy AI assistance, simplifying the argument and highlighting its key polar inequality, which he could not find in prior literature.
Step by step
Mazur runs GPT-5.6 Pro in a custom orchestration and verification workflow
The model develops a reduction to two scalar propositions plus exact certificates
A separate Lean 4 formalisation proves the exact theorem
Mazur posts the paper (5 August 2026)
Tao publishes a digest, a streamlined proof and a shorter Lean formalisation (12 August 2026)
What it brings
Resolves a well-known conjecture in the geometry of polynomials, plus the Phelps–Rodriguez strengthening, with machine-checked proofs.
Human vs AI
Per Mazur's disclosure, AI played a substantial role in discovering and deriving the mathematics; Mazur curated, prompted, reconciled outputs and takes responsibility.
Limits and debate
Mazur states the Lean proof is not a line-by-line verification of his manuscript and does not verify the certificate programs. Earlier human claims of proofs of Sendov's conjecture had failed. Not peer reviewed as of 21 Nov 2026.
Fields it touches
Complex analysisGeometry of polynomialsComputer-assisted proof
OpenAI announced ten results produced by an internal version of Astra, including the first construction of a non-sofic group, a disproof of a Connes rigidity conjecture, polynomial-factor hardness of approximating the closest vector problem, results on Ehrhart's volume conjecture, and resolutions of Erdős problems 146, 180 and 183.
Status. OpenAI published Lean certificates and takes responsibility for correctness; no independent peer review was found. The dispute concerns novelty and attribution: experts showed key steps relied on uncited prior work, and OpenAI updated its wording.
Full dossier: why, how, steps, sources
Lab
OpenAI
People
Andreas Thom, Francesco Fournier-Facio, Stephen Miller
AI systems
Internal version of Astra (OpenAI's next major model)Lean
Why it matters
Whether non-sofic groups exist had been a central open question in group theory. OpenAI said the solutions cost roughly $2,000 at API rates.
How it was done
OpenAI says the arguments were generated by its system, while staff used the same model to help prepare a nearly 250-page manuscript and Lean formalisations, published with 'reasoning walkthroughs'.
Step by step
During internal development, Astra is run on open problems
It produces arguments for ten problems
OpenAI staff, with the model, prepare manuscripts and Lean certificates
Release on 1 August 2026 with a GitHub repository of proofs
Experts identify uncited prior work; OpenAI revises its language
What it brings
If the results hold, they would be the largest single batch of AI-generated research mathematics. The release intensified debate on attribution standards for AI-generated proofs.
Human vs AI
OpenAI attributes the mathematical arguments to the model; humans prepared manuscripts.
Limits and debate
Stephen Miller said the sphere-packing proof relies on a 2016 argument he co-wrote and called this 'research misconduct'. The non-sofic step combined ideas from 2016 and 2019 papers (one co-authored by Andreas Thom, who called the result 'creative and at the same time elementary'). The Ehrhart result's scope (OpenAI's page versus other summaries describing only an upper bound) was not independently confirmed as of 21 Nov 2026.
Fields it touches
Group theoryOperator algebrasDiscrete geometryComplexity theoryCombinatoricsPost-quantum cryptography (lattice problems)Coding theoryQuantum information
The Jacobian conjecture (1939) says that if a polynomial map from n-dimensional complex space to itself has a Jacobian determinant that is a nonzero constant, so that it is locally invertible everywhere, then it has a polynomial inverse. Alpöge posted an explicit map C³→C³ whose Jacobian determinant is the constant −2 but which sends several distinct points to the same point, so it cannot be inverted. Adding identity coordinates extends this to every dimension n ≥ 3. The 2-dimensional case is still open.
Status. The explicit map was checked in Lean 4 by Paul Lezeau and others: PR #4474 to Google DeepMind's Formal Conjectures repo was opened 20 July and merged 26 July 2026. It was also checked in Isabelle/HOL (Archive of Formal Proofs, 20 July 2026). The identities are short enough to check by hand. We re-checked them with sympy (det = −2, and the published collisions hold). There is no peer-reviewed paper by Alpöge.
Full dossier: why, how, steps, sources
Lab
Anthropic Levent Alpöge is an Anthropic researcher; the work was done with Claude.
People
Levent Alpöge, Akhil Mathew (suggested the problem)
AI systems
Claude Fable 5Lean 4 (independent check)Isabelle/HOL (independent check)
In plain words
A polynomial map is a recipe that takes three numbers and returns three numbers using only multiplication and addition. The conjecture said that if such a recipe can be undone locally around every point, it can be undone everywhere, like a reversible code. Alpöge posted a recipe that passes the local test everywhere yet sends three different inputs to the same output, so no undo exists. Nothing changes for engineering or cryptography tomorrow. The result settles a pure-math question.
Key figures
posed 1939 (Keller)Conjecture age
constant −2Jacobian determinant
7, 6, 4 (computed from the posted map)Polynomial degrees of the 3 components
>20 million (Fortune, 21 July)Views of the announcement
20 July → 26 July 2026Lean PR open→merged
Why it matters
Ott-Heinrich Keller posed the conjecture in 1939 (Fortune). It became one of the best-known open problems in algebraic geometry. Alpöge called it 'the canonical crank graveyard' because of its long record of false proofs. Tao wrote that the example is 'extremely quick' to verify but 'appears like a massive miracle', because a naive brute-force search looks hopeless.
How it was done
Akhil Mathew asked Alpöge about the problem. Alpöge credits 'my other close friend fable' (Claude Fable 5) with the work that produced the map. The exact prompts and the model's output have not been published (The Conversation, 22 July). The result was released as a three-line formula plus Wolfram Alpha screenshots of the determinant and of two colliding points. Humans then checked and explained it: Tao derived it from multiplication of a linear and a quadratic form, Arno van den Essen gave an 'elementary' route to the same map, and Remy van Dobben de Bruyn gave a coordinate-free geometric explanation.
Step by step
Akhil Mathew suggests the Jacobian conjecture to Alpöge.
Claude Fable 5 produces an explicit degree-(7,6,4) polynomial map C³→C³ with constant Jacobian −2.
Alpöge posts the map on X on 20 July 2026, with Wolfram Alpha checks of the determinant and of colliding points.
Formalizations follow within a day: an Isabelle/HOL entry in the Archive of Formal Proofs on 20 July, and a Lean PR to Formal Conjectures (opened 20 July, merged 26 July).
Tao's 'digestion' (21 July) and later arXiv papers (van den Essen; van Dobben de Bruyn) explain the construction in human terms.
Chronology
Alpöge posts the map on X ('hello there the jacobian conjecture is false')
Isabelle/HOL verification published in the Archive of Formal Proofs; Lean PR #4474 opened
Terence Tao posts 'A digestion of the Jacobian conjecture counterexample'
Lean disproof merged into Google DeepMind Formal Conjectures
van Dobben de Bruyn arXiv:2608.27341 (coordinate-free explanation)
van den Essen arXiv:2609.17795 (elementary derivation)
What it brings
An 87-year-old conjecture was settled negatively in dimensions ≥ 3 by a certificate anyone can check in seconds. Alpöge agreed in replies that related conjectures, such as Dixmier, then also fail in general. The episode became a reference case for AI-found counterexamples that humans later explain and formalize.
What it does not show
It does not resolve the 2-dimensional Jacobian conjecture. It does not come with a human-readable proof or a published paper by the discoverers. It has no practical or engineering consequence.
Human vs AI
A human chose the problem (Mathew) and a human posted and checked the result (Alpöge). Alpöge credits Claude Fable 5 with finding the map. How the work was split in detail has not been disclosed.
Credit and priority
The idea of trying the problem came from Akhil Mathew (Fortune; Alpöge's post). Alpöge credits 'fable' (Claude Fable 5). Hacker News commenters questioned the wording that Fable 'produced' the map, because the post does not say how it was found. Later explanations are human work: Tao, van den Essen, van Dobben de Bruyn.
Limits and debate
This is a counterexample, not a theory: Akhil Mathew told Fortune it can be checked but 'lacks an explanation', and the explanations came later from humans. The 2-variable case, the classical core of the conjecture, is untouched. Some coverage describes the map over ℝ³; the post states it over ℂ³ (it has integer coefficients, so it works over both). There is no transcript of the AI session.
Reactions
Kevin Buzzard (Imperial College London): Called it 'a big day' and 'very exciting'; noted machines are still poor at posing good questions. source
Akhil Mathew (University of Chicago): Correct and checkable but without explanation; 'a very rapid and very unsettling change', especially for junior mathematicians. source
Terence Tao (UCLA, Fields medallist): Gave a conceptual derivation; quick to verify yet 'appears like a massive miracle'. source
Melissa Lee (Monash University): The difficulty lay in searching the space of maps; the method had not been made public. source
How many pairs of points at distance exactly 1 can n points in the plane have? Erdős conjectured the maximum grows like n^(1+o(1)), barely faster than n. The OpenAI model constructed, for infinitely many n, configurations with at least n^(1+δ) unit distances for a fixed δ > 0, disproving the conjecture.
Status. Checked by nine external mathematicians, who posted a shorter, verified and generalised exposition (arXiv 2605.20695). Not yet journal-published; Gowers said he would have recommended acceptance to the Annals without hesitation.
Full dossier: why, how, steps, sources
Lab
OpenAI
With
Nine external mathematicians (co-authors of arXiv 2605.20695)
People
Noga Alon, Thomas F. Bloom, Timothy Gowers, Daniel Litt, Will Sawin, Arul Shankar, Jacob Tsimerman, Victor Wang, Melanie Matchett Wood, Sébastien Bubeck, Mark Sellke, Mehtaab Sawhney
Put n dots on a sheet of paper: how many pairs can be exactly 1 cm apart? Since 1946 the best arrangements were essentially square grids, and Erdős believed nothing could do substantially better. OpenAI's model found a cleverer arrangement, built from advanced number theory, that beats this by a fixed power of n, so the belief was wrong. It has no direct engineering use; it matters because experts accept that an AI settled a famous question largely by itself.
Key figures
≈ 6.24·10^-38Exponent δ (explicit proof)
0.014Exponent δ (Sawin refinement)
9 mathematiciansExternal checkers / co-authors
O(n^(4/3))Previous best upper bound
Why it matters
The unit distance problem is one of the best-known questions in discrete geometry; Erdős's square-grid-based construction had not been beaten, and the best known upper bound O(n^(4/3)), which still stands, dates to Spencer, Szemerédi and Trotter. OpenAI called it the first time a prominent open problem in a central subfield was solved autonomously by AI.
How it was done
While being evaluated on a collection of Erdős problems, the model produced the counterexample in one shot; it was then refined with Codex. The construction replaces Erdős's Gaussian integers with number fields having richer symmetries, using infinite class field towers and Golod–Shafarevich theory, tools well known to number theorists but not previously connected to this geometry problem.
Step by step
An internal general-purpose reasoning model is evaluated on Erdős problems
It generates a disproof with no human intervention
The argument is refined with Codex
OpenAI researchers ask Daniel Litt and others to check it
Nine mathematicians write a simplified, generalised exposition (20 May 2026)
Chronology
OpenAI announcement and companion paper by nine mathematicians (arXiv 2605.20695)
What it brings
OpenAI described it as the first autonomous AI solution of a prominent open problem in a central subfield, and the external mathematicians' assessments (Alon, Bloom, Gowers) back its significance. Will Sawin's refinement improves the exponent to δ = 0.014.
What it does not show
Does not determine the true growth rate of unit distances; the O(n^(4/3)) upper bound is untouched. No practical application claimed.
Human vs AI
The core mathematics was AI-generated; humans verified, simplified and contextualised it.
Credit and priority
Ideas retrospectively attributable to Ellenberg–Venkatesh, Golod–Shafarevich and Hajir–Maire–Ramakrishna; the exposition and simplification are by the nine external mathematicians.
Limits and debate
The original exponent is tiny (about 6.24·10^-38 from the explicit proof) before Sawin's refinement. The ideas can in retrospect be traced to Ellenberg–Venkatesh, Golod–Shafarevich and Hajir–Maire–Ramakrishna. Bloom gives only a 'moderated yes' on whether it deepens understanding of discrete geometry. Gowers notes counterexamples can come from 'trying lots of things'.
Reactions
Noga Alon (Mathematician, co-author of the companion paper): Called it an outstanding achievement settling a long-standing open problem. source
Timothy Gowers (Fields medallist, Cambridge): As an Annals referee he would have recommended acceptance without hesitation, while noting counterexamples are easier for search-based systems than proofs. source
Thomas Bloom (Maintainer of erdosproblems.com): Found it surprising and impressive for an AI to solve a problem of this calibre. source
Melanie Matchett Wood (Mathematician, co-author of the companion paper): Warned that mathematicians urgently need to plan how to keep their work rigorous and correct. source
Fields it touches
Discrete geometryCombinatoricsAlgebraic number theoryIncidence geometryClass field theory
A primitive set is a set of integers in which no element divides another. Problem #1196, a 1966 conjecture of Erdős, Sárközy and Szemerédi, concerns primitive sets of large numbers. An amateur, Liam Price, obtained a proof from a single prompt to GPT-5.4 Pro and posted it on erdosproblems.com.
Status. The proof was checked by Tao and others and developed in an 8-author preprint (arXiv 2605.00301, 1 May 2026) labelled preliminary. Tao's wiki records a formalisation via Gauss on 16 April 2026.
Full dossier: why, how, steps, sources
Lab
Independent researchers Liam Price using an OpenAI model; checked by an expert team including Terence Tao.
People
Liam Price, Terence Tao, Jared Duker Lichtman, Boris Alexeev, Kevin Barreto, Yanyang Li, Jibran Iqbal Shah, Quanyu Tang, Nat Sothanaphan
AI systems
GPT-5.4 ProGPT-5.4 ThinkingGauss (formalisation)
Why it matters
The AI's argument used a formula well known in neighbouring areas that no one had applied to this question. Distilled by experts, it became a new method that also settled #1217 and gave a short proof of the Erdős primitive set conjecture.
How it was done
GPT-5.4 Pro produced a raw proof that experts found hard to follow. Lichtman, Tao and co-authors extracted its key idea: a Markov-chain method on the divisibility poset with von Mangoldt weights to bound Erdős sums of primitive sets.
Step by step
Liam Price prompts GPT-5.4 Pro with the problem
He posts the model's proof on erdosproblems.com (13 April 2026)
Experts check and simplify the argument
The idea is generalised into the 'von Mangoldt chains' method
An 8-author preprint proves #1196, #1217 and related results
What it brings
One of the clearest cases where an AI output contributed a reusable new technique rather than a one-off solution.
Human vs AI
The core idea came from the AI; experts turned a poor raw write-up into a proof and generalised it. Tao's wiki lists both an AI-standalone solution and a human–AI collaborative one.
Limits and debate
Lichtman said the raw output was 'actually quite poor'. Tao said the problem may have been easier than expected and the long-term significance is still open. The preprint is preliminary.
Fields it touches
Number theoryCombinatoricsProbability (Markov chains)
Gemini was used to evaluate 700 Erdős problems marked 'open'. Of 13 problems examined in depth, 5 got apparently novel autonomous solutions and 8 were matched to solutions already in the literature. A later formal-proof agent autonomously resolved 9 of 353 open Erdős problems in Lean and proved 44 of 492 OEIS conjectures, at a few hundred dollars per problem.
Status. Both are arXiv preprints. In the first, human experts judged correctness and novelty; in the second, proofs are Lean-checked.
Full dossier: why, how, steps, sources
Lab
Google DeepMind
People
Tony Feng, Thang Luong, Quoc V. Le, George Tsoukalas, Swarat Chaudhuri, Pushmeet Kohli, Thomas Hubert
AI systems
GeminiLean-based proof-search agents
Why it matters
These are systematic sweeps rather than cherry-picked anecdotes, and they quantify both success rates and how often 'new' AI solutions turn out to be known.
How it was done
In the January study, AI-driven natural-language verification narrowed candidates and human experts judged correctness and novelty. In the May study, agents alternated LLM generation with Lean verification; a basic agent reproduced the Erdős successes but cost more on the hardest problems.
Step by step
Screen hundreds of 'open' Erdős problems with Gemini
Filter candidate solutions with AI verification
Human experts check correctness and search the literature
Run Lean-based proof-search agents on 353 Erdős problems and 492 OEIS conjectures
Report counts and costs
What it brings
The authors concluded that the open status of many solved problems reflected obscurity rather than difficulty; the formal pipeline shows open problems can be attacked with guaranteed-correct output at low cost.
Human vs AI
Semi-autonomous in the first study (human verification); autonomous with Lean checking in the second.
Limits and debate
Low success rates (about 2.5% of Erdős problems and 9% of OEIS conjectures in the formal study). The first paper warns of 'subconscious plagiarism' and revised its autonomous count from 6 to 5 after one solution was reclassified as an independent rediscovery.
Erdős Problem #728 asks about divisibility among factorials. The result shows that, for any constants 0 < C1 < C2 and 0 < ε < 1/2, there are infinitely many triples (a, b, n), with a and b between εn and (1−ε)n, such that a!·b! divides n!·(a+b−n)! and a+b−n lies between C1·log n and C2·log n.
Status. Aristotle produced a Lean proof; Nat Sothanaphan's arXiv writeup (2601.07421, v5 of 26 January 2026) translates it into ordinary mathematics. Tao's wiki records it as a full solution, with similar methods found in a paper by Pomerance.
Full dossier: why, how, steps, sources
Lab
Independent researchers Kevin Barreto, using OpenAI (GPT-5.2) and Harmonic (Aristotle) tools.
With
erdosproblems.com community
People
Kevin Barreto, Nat Sothanaphan, Terence Tao
AI systems
GPT-5.2 ProAristotle (Harmonic)Lean
In plain words
Paul Erdős left behind more than a thousand problems, from very hard to forgotten curiosities. #728 asks a precise question about when products of factorials divide other such products. An OpenAI model produced the argument, and a second AI (Harmonic's Aristotle) translated it into a form a computer can check line by line. It is a real but modest result; its importance is that AI produced it with little human help.
Key figures
v1 12 Jan 2026 → v5 26 Jan 2026Writeup versions
#729 solved 8–10 Jan 2026Follow-on problem
Why it matters
Tao called it a milestone: the first Erdős problem solved more or less autonomously by AI in the spirit of the problem, with the result not replicated in existing literature as far as anyone knew.
How it was done
Kevin Barreto ran GPT-5.2 Pro to produce a proof, after feedback on an initial attempt; Harmonic's Aristotle then produced a Lean formalisation. The proof reduces the question to a binomial-coefficient divisibility and analyses it prime by prime with Kummer's theorem and a counting argument. AI tools were also used to rapidly write and rewrite human-readable expositions.
Step by step
The erdosproblems.com community reconstructs the intended statement of the misformulated problem
GPT-5.2 Pro generates a proof; feedback on a first attempt leads to a revised run
Aristotle formalises the proof in Lean
Literature search finds related work by Pomerance using similar methods
Nat Sothanaphan posts an arXiv writeup of the Lean proof
Chronology
Related Pomerance literature located (5–6 Jan)
Full solution with Lean proof logged (Aristotle + GPT-5.2 Pro)
Tao describes it as solved more or less autonomously
Sothanaphan posts arXiv writeup
What it brings
Shifted the Erdős-problems conversation from 'AI found it in the literature' to genuine AI-generated solutions with machine-checked proofs; a follow-on problem (#729) was solved days later.
What it does not show
Not a famous or central open problem; not proof that AI can solve hard Erdős problems in general.
Human vs AI
Largely AI: humans chose the problem, clarified the statement, gave feedback on a first attempt and wrote up the result.
Credit and priority
Pomerance's earlier work contains similar results by similar methods (logged by Tao's wiki as found on 5–6 January 2026). Credit for formalisation goes to Harmonic's Aristotle; the run was operated by Kevin Barreto.
Limits and debate
The problem as stated by Erdős was misformulated, and the intended version was only reconstructed months earlier, which helps explain the lack of prior work. Similar results by similar methods existed (Pomerance). Tao's wiki stresses selection bias and that obscurity, not difficulty, often explains why such problems are open.
Reactions
Terence Tao (Fields medallist, UCLA): Called it a milestone and a genuine increase in capability, consistent with AI using existing methods, and found the AI's rapid rewriting of expositions even more interesting. source
Fields it touches
Number theoryFormal verificationAI-written exposition
OpenAI and academic collaborators documented cases where GPT-5 contributed to research. In mathematics these included a density estimate that Sawhney and Sellke turned into a full proof of Erdős Problem #848, a sharper step-size bound for a gradient-descent theorem, and a lower bound for clique-avoiding codes.
Status. A curated collection of case studies co-authored with external researchers; individual results were checked by the human authors (e.g. Bubeck checked the gradient-descent bound by hand). Not a systematic evaluation.
Full dossier: why, how, steps, sources
Lab
OpenAI
With
Vanderbilt, UC Berkeley, Columbia, Oxford, Cambridge, Lawrence Livermore National Laboratory, The Jackson Laboratory
People
Mehtaab Sawhney, Mark Sellke, Sébastien Bubeck, Timothy Gowers, Kevin Weil
AI systems
GPT-5GPT-5 Pro
Why it matters
It was the first broad, lab-published account of a frontier LLM acting as a research collaborator across fields, and it followed directly on the October 2025 Erdős controversy.
How it was done
Experts posed research questions to GPT-5 and iterated; the model suggested ideas, constructions or proof sketches that humans checked and completed.
Step by step
Researchers bring open questions from their own fields
GPT-5 proposes ideas, estimates or proof outlines
Humans verify, correct and complete the arguments
Cases, including failures, are written up in a joint paper
What it brings
Framed GPT-5 as strongest where structure is explicit and feedback loops are fast, such as mathematics and theoretical computer science.
Human vs AI
Human-led; the model contributed key ideas in some cases but required expert oversight throughout.
Limits and debate
Curated examples, not a random sample. In the clique-codes case the model reproduced a result from an earlier paper without citing it until asked. The paper notes hallucinated citations and proofs. Gowers found it useful as a fast critic but not yet at co-authorship level.
Given an open-ended research goal and a dataset, Kosmos runs for up to 12 hours, cycling through data analysis, literature search and hypothesis generation, and writes a fully cited report.
Status. Preprint with an evaluation by independent scientists (79.4% of sampled statements judged accurate); not peer-reviewed as of 21 Nov 2026. An independent test in radiation biology found mixed results.
Full dossier: why, how, steps, sources
Lab
Edison Scientific Commercial spinout of FutureHouse.
With
Academic collaborators (beta users)
People
Ludovico Mitchener (first author), Sam Rodriques, Andrew D. White
AI systems
Kosmos
Key figures
up to 12 hoursRun length
~42,000 linesCode per run
~1,500Papers read per run
79.4% (of 102 sampled)Statements judged accurate
$200 per runPrice at launch
Why it matters
Earlier AI agents lost coherence after a limited number of steps. Kosmos uses a shared 'world model' so that its analysis and literature agents can pursue one objective across about 200 agent runs.
How it was done
A data-analysis agent and a literature agent share a structured memory of findings. A typical run executes about 42,000 lines of code and reads about 1,500 papers. The team highlights seven discoveries: three reproduce findings from preprinted or unpublished manuscripts Kosmos had not seen, and four are presented as new contributions (for example in perovskite solar cells, neuroscience and statistical genetics).
Step by step
Scientist provides an objective and dataset
Parallel cycles of analysis and literature search update a shared world model
Hypotheses are generated and tested in code
After up to ~20 cycles, a cited report is produced
Independent experts audit sampled statements
What it brings
Pushed autonomous data-analysis agents to much longer, traceable runs; offered commercially at $200 per run at launch.
Human vs AI
Humans choose the objective and dataset and must vet the report; Kosmos does the analysis and literature synthesis autonomously.
Limits and debate
Accuracy was 79.4% overall but lower for synthesis statements (57.9%). Edison itself says Kosmos 'often goes down rabbit holes or chases statistically significant yet scientifically irrelevant findings'. The '6 months of work in a day' figure is an estimate from seven beta users. An independent radiation-biology test (Nusrat & Nusrat, arXiv Nov 2025) found one well-supported discovery, one uncertain result and one false hypothesis out of three.
Fields it touches
AI for scienceData-driven discoveryMetabolomicsMaterials scienceNeuroscienceStatistical genetics
Analysis & PDECombinatoricsGeometry & topologyNumber theory
A Fields medallist and collaborators applied AlphaEvolve to 67 problems across analysis, combinatorics, geometry and number theory. It matched the best known results in most cases, improved several, and in some cases generalised finite results into formulas valid for all inputs.
Status. arXiv preprint 2511.02864 (v3, 22 December 2025) with a public repository of problems, prompts, verifiers and evolved programs; not peer reviewed as of 21 Nov 2026.
Full dossier: why, how, steps, sources
Lab
Google DeepMind
People
Bogdan Georgiev, Javier Gómez-Serrano, Terence Tao, Adam Zsolt Wagner
AI systems
AlphaEvolveGemini Deep ThinkAlphaProof
Why it matters
It is the most systematic expert-run test of an AI discovery tool, documenting failures and exploits as well as successes.
How it was done
Experts wrote exploit-resistant scoring functions, sometimes disguising well-known problems so the model would not just recall answers. AlphaEvolve evolved code producing candidate constructions. In the finite-field Kakeya case it slightly improved the best 3D construction; humans then proved the construction with Deep Think and formalised it with AlphaProof.
Step by step
Select 67 problems and write verifiers (using exact or interval arithmetic)
Run AlphaEvolve to evolve construction programs
Inspect outputs for verifier exploits and memorised extremisers
Prove selected constructions (Deep Think) and formalise them (AlphaProof)
Publish problems, prompts and code openly
What it brings
Showed AI search can cheaply sweep many problems and variants, while also documenting limits: one-sided bounds only, weak results in analytic number theory, and a tendency to game imperfect verifiers.
Human vs AI
Experts designed verifiers, chose problems and proved results; AlphaEvolve performed the search.
Limits and debate
No counterexamples to major conjectures (e.g. Sidorenko, Sendov, Crouzeix) were found. Tao warns that results can reflect verifier exploits and that AI output should not be used without independent verification.
An OpenAI vice-president posted that GPT-5 had found solutions to 10 previously unsolved Erdős problems and made progress on 11 others. In fact the model had found existing published solutions that the site's maintainer had not known about.
Status. The claim of new solutions was withdrawn: GPT-5 had located existing papers that solved problems listed as 'open' on erdosproblems.com. The literature-search value itself was acknowledged.
Full dossier: why, how, steps, sources
Lab
OpenAI
With
erdosproblems.com
People
Kevin Weil, Sébastien Bubeck, Thomas Bloom, Demis Hassabis, Yann LeCun
AI systems
GPT-5GPT-5 Pro
Why it matters
The episode set the tone for how AI claims on Erdős problems are judged: 'open' on the database can mean only that the maintainer is unaware of a solution, so apparent AI breakthroughs must be checked against the literature.
How it was done
Researchers used GPT-5 Pro as a literature-search engine on problems marked open. It surfaced old references containing solutions, which was then misdescribed publicly as the model solving the problems.
Step by step
GPT-5 is used to search the literature on 'open' Erdős problems
It finds papers with existing solutions
Kevin Weil posts that GPT-5 'found solutions' to 10 unsolved problems
Thomas Bloom calls this 'a dramatic misrepresentation'
The posts are deleted; Bubeck acknowledges only solutions in the literature were found
What it brings
Led to more careful categorisation of AI contributions, for example Tao's wiki separating 'AI standalone', 'AI alongside literature' and 'literature search' entries.
Human vs AI
The AI performed literature search; the mischaracterisation was human.
Limits and debate
This entry is a caveat for the field: several later AI 'solutions' were also found to have precedents in the literature (see the DeepMind Erdős study, which matched 8 of 13 candidate solutions to existing literature).
Fields it touches
Number theoryCombinatoricsLiterature searchResearch integrity
C2S-Scale is a 27-billion-parameter model that treats gene-activity data from single cells like text. Researchers used it to search for existing drugs that might make 'cold' tumours — ones the immune system ignores — more visible to immune cells, but only in a specific immune context.
Status. Described in a bioRxiv preprint (not peer-reviewed as of 21 Nov 2026) with laboratory validation in human cell models only; the authors call it 'an early first step'.
Full dossier: why, how, steps, sources
Lab
Google DeepMind Google DeepMind with Google Research.
With
Yale University
AI systems
Cell2Sentence-Scale 27B (C2S-Scale), built on Gemma
Key figures
27 billion parametersModel size
>4,000Drugs screened in silico
>50 millionSingle-cell profiles in training corpus
Why it matters
Many tumours do not respond to immunotherapy because the immune system fails to notice them. A model that can propose context-specific drug effects could help prioritise which combinations to test.
How it was done
The model ran a virtual screen of more than 4,000 drugs under two simulated conditions and ranked candidates expected to work only in the immune-relevant one. It highlighted silmitasertib, which Google says had not previously been reported in the literature for this effect. Yale and Google researchers then tested the prediction in human cell models of a cell type the model had not seen in training, and observed the predicted effect in combination with a low dose of an immune signalling molecule.
Step by step
Train a large model on single-cell gene-activity data rendered as text
Run a two-context virtual screen of 4,000+ drugs
Rank context-specific candidates
Test the top prediction in human cell models
Release preprint, model and code
What it brings
An example of a biology foundation model generating a non-obvious, testable hypothesis that held up in a first lab test; the model was released openly for researchers.
Human vs AI
The model proposed and ranked candidates; scientists designed the screen, chose what to test and did the lab work.
Limits and debate
Cell-model results only: no animal or human data. The authors stress that further preclinical and clinical validation is needed before any therapeutic relevance. The claim of novelty is the developers'.
Finds new families of 'unstable' self-similar blow-up solutions for the incompressible porous media (IPM) equation and the Boussinesq equations, the latter being closely related to 3D Euler with a boundary. Unstable means the solution blows up only from initial data tuned with extreme precision.
Status. A numerical discovery by a team that includes leading singularity experts (Buckmaster, Gómez-Serrano). It is not a rigorous proof and was not presented as one; residuals reach near machine precision, which the authors say is enough for future computer-assisted proofs. No journal version had appeared as of 21 Nov 2026.
Full dossier: why, how, steps, sources
Lab
Google DeepMind
With
Brown University, New York University, Stanford University
Earlier numerics mostly found stable singularities. Unstable ones are expected to matter for the open boundary-free 3D Euler and Navier–Stokes cases. The authors also report an empirical formula linking blow-up rate to the order of instability.
How it was done
PINNs were trained to minimize the residual of the self-similar equations, using a high-precision Gauss-Newton optimizer and careful architectures. For some solutions the accuracy approaches double-precision GPU limits.
Step by step
Formulate self-similar ansatz for IPM and Boussinesq
Train PINNs with second-order optimization to near machine precision
Identify multiple unstable profiles and fit an empirical λ-vs-instability law
Publish as arXiv 2509.14185 with a view to computer-assisted proofs
What it brings
Showed that machine learning can find delicate unstable singularities that traditional numerics missed, a precursor step for the 2026 computer-assisted and AI-assisted blow-up programmes.
Human vs AI
The neural networks are numerical solvers designed and steered by the human team; no autonomous AI reasoning was involved.
Limits and debate
Numerical evidence, not a proof. Covers IPM, Boussinesq and Euler with boundary, not the boundary-free Euler or Navier–Stokes problems.
Fields it touches
PDE / fluid dynamicsComputer-assisted proofsTurbulence and singularity theoryScientific machine learning
Gauss turns human mathematics into computer-checked Lean proofs. It produced about 25,000 lines (about 1,100 theorems and definitions) formalising the strong prime number theorem in three weeks, then formalised Maryna Viazovska's Fields-Medal proofs that the E8 and Leech lattices give the densest sphere packings in dimensions 8 and 24.
Status. Outputs are Lean developments checked by the Lean kernel. The sphere-packing work was done in collaboration with the human-led project that had already built a blueprint for dimension 8.
Full dossier: why, how, steps, sources
Lab
Math Inc.
With
EPFL, Formalising Sphere Packing in Lean project
People
Jesse Han, Maryna Viazovska, Sidharth Hariharan, Terence Tao, Alex Kontorovich
AI systems
Gauss (autoformalization agent)Lean 4
Why it matters
Formalising deep theorems by hand takes experts months or years. Tao and Kontorovich had been leading a human strong-PNT formalisation effort, and the sphere-packing project had been running since March 2024.
How it was done
Gauss interleaves natural-language reasoning with formal Lean work, using tools and the Lean compiler. For the strong PNT, most statements and proofs were generated by Gauss with targeted human scaffolding and review of key lemmas. For dimension 8 it completed the formalisation in five days from the human blueprint and fixed a typo in the published paper; for dimension 24 it worked without a blueprint for about two weeks, producing more than 200,000 lines.
Step by step
Humans or a project blueprint provide the target theorem and structure
Gauss drafts statements and proofs in Lean, iterating against the compiler
Humans review key lemmas and strategy
The finished development compiles in Lean, certifying the proof
Results are merged into or released alongside community projects
What it brings
Showed that agents can compress years of formalisation into days or weeks, a precondition for machine-checking new AI-generated mathematics. A January 2026 version of Gauss reproduced the three-week PNT result in two to three days, per IEEE Spectrum.
Human vs AI
Mixed: human-built blueprints and scaffolding plus agent-written formal proofs; both teams describe dimension 24 as a human–AI collaboration.
Limits and debate
Formalisation verifies known theorems; it does not discover new mathematics. Line counts for dimension 24 vary between sources (about 200,000 per Math Inc and EPFL). The dimension-8 work overlapped with a blueprint the human team had already built.
Fields it touches
Formal verificationAnalytic number theoryDiscrete geometryProof assistantsMathematical libraries (Mathlib)
Two general-purpose language models each solved 5 of the 6 problems of the 2025 IMO, scoring 35 out of 42, a gold-medal score. Both worked in natural language from the official problem statements, within the two 4.5-hour sessions, without tools or internet access.
Status. DeepMind's solutions were graded and certified by IMO coordinators (the IMO reviewed answers only, not the system). OpenAI did not enter officially; its proofs were graded by three former IMO medalists and published on GitHub.
Full dossier: why, how, steps, sources
Lab
Multiple labs Google DeepMind (Gemini Deep Think) and OpenAI (experimental reasoning model).
With
International Mathematical Olympiad (coordinators, for DeepMind)
People
Gregor Dolinar, Alexander Wei
AI systems
Gemini Deep Think (advanced version)OpenAI experimental reasoning model
In plain words
Every year, the best high-school mathematicians sit two 4.5-hour exams with three proof problems each. In 2025, two AI systems sat the same paper and wrote answers in ordinary mathematical prose, scoring 35 of 42 points, enough for gold. Google's score was confirmed by the official graders; OpenAI's was graded by three former medallists. It shows strong reasoning, but olympiad problems are designed to have neat solutions, unlike open research questions.
Key figures
35/42Score (both)
5 of 6Problems solved
2 × 4.5 hTime limit
67 of 630Human gold medallists
Why it matters
One year after AlphaProof needed manual Lean translation and days of compute, general LLMs wrote full natural-language proofs at gold level under contest conditions. Gold went to 67 of 630 contestants in 2025.
How it was done
DeepMind used an advanced Gemini Deep Think with 'parallel thinking' (exploring and combining several solution paths), new reinforcement-learning techniques, a curated corpus of high-quality solutions and general hints on IMO problems. OpenAI said its result came from general-purpose reinforcement learning and test-time compute scaling rather than task-specific methods.
Step by step
Models receive the official IMO 2025 problem statements
Each writes natural-language proofs within the 2 × 4.5-hour limit, no tools
DeepMind: IMO coordinators grade with student criteria and certify 35/42
OpenAI: three former IMO medalists grade independently, reaching unanimous scores of 35/42
OpenAI publishes its proofs on GitHub
Chronology
OpenAI's Alexander Wei announces a 35/42 gold-level score
DeepMind announces IMO-certified 35/42 for Gemini Deep Think
What it brings
Marked the shift from specialised formal systems to general LLM reasoning as the frontier of AI mathematics, and preceded the 2025–2026 wave of research-level results.
What it does not show
Not research mathematics; no claim the models can solve open problems; no tools or formal verification involved.
Human vs AI
Fully autonomous problem solving; humans only graded.
Credit and priority
Only DeepMind's result went through official IMO grading; OpenAI's grading was organised by OpenAI.
Limits and debate
OpenAI's result was not certified by the IMO. DeepMind's model was given a curated solution corpus and general hints. Competition problems have known solutions, unlike research problems.
Reactions
Gregor Dolinar (President of the IMO): Confirmed that Google DeepMind reached the milestone with 35 of 42 points. source
Simon Willison (Software developer and commentator): Called OpenAI's result notable while noting the proofs' slightly strange format. source
Rentosertib is an experimental drug for idiopathic pulmonary fibrosis (IPF), a progressive scarring of the lungs. Insilico reports that both the disease target and the molecule were found with its generative-AI platforms, and the drug has now been tested in patients.
Status. Phase 2a trial (71 patients, 12 weeks) peer-reviewed in Nature Medicine (2025). A 320-patient, 52-week Phase III trial began in China in 2026. The drug is investigational and not approved by any regulator; the FDA granted it orphan-drug designation for IPF in February 2023.
Full dossier: why, how, steps, sources
Lab
Insilico Medicine
With
Peking Union Medical College Hospital (Phase III lead site), Shanghai Pulmonary Hospital
People
Alex Zhavoronkov (Insilico CEO, corresponding author), Zuojun Xu (Phase III leading principal investigator)
Idiopathic pulmonary fibrosis slowly scars the lungs and has no cure. Insilico used AI to choose a new biological target and to design a pill against it. In a small 12-week trial the drug looked reasonably safe and lung capacity improved slightly in the highest-dose group while it declined on placebo. That is encouraging, but it is an early signal: a large Phase III trial started in 2026 and must succeed before the drug could be approved.
Key figures
71 (4 arms)Phase 2a patients
12 weeksPhase 2a duration
+98.4 mL (95% CI 10.9–185.9) vs −20.3 mLLung-capacity (FVC) change, highest-dose arm vs placebo
~320 patients, 52 weeks, 47 centresPhase III enrolment target
Why it matters
IPF has no therapies that reverse its course. Few AI-discovered or AI-designed drugs had reached human trials, so rentosertib became a closely watched test of whether AI can shorten drug discovery and produce drugs that work in people.
How it was done
At a high level, Insilico's PandaOmics software analysed biological data to propose a target not previously linked to fibrosis, and its Chemistry42 generative-chemistry platform designed and optimised a molecule against it; another tool was used to help plan trials. Humans then carried the candidate through standard preclinical and clinical development. The Phase 2a trial was randomised, double-blind and placebo-controlled; its main goal was safety, with lung function as a secondary measure.
Step by step
AI-assisted target identification (PandaOmics)
Generative design and optimisation of a candidate molecule (Chemistry42)
Preclinical development and earlier clinical testing
Phase 2a: 71 patients, 12 weeks, published in Nature Medicine (2025)
Phase III (GENESIS-IPF-3): ~320 patients, 52 weeks, 47 centres in China (started 2026)
Chronology
FDA orphan-drug designation for IPF (month only)
Phase 2a results published in Nature Medicine
Phase III trial announced
First Phase III patients enrolled (announced)
What it brings
The authors describe it as the first phase 2a trial of an AI-generated drug aimed at a target that was itself discovered using generative AI, making it a reference case for AI drug discovery.
What it does not show
Not an approved medicine; not proof of efficacy (Phase 2a was designed for safety); not evidence that AI removes the need for standard clinical trials.
Human vs AI
AI tools proposed the target and generated candidate molecules; chemists, biologists and clinicians selected, tested and developed the drug through regulated trials.
Credit and priority
Developed by Insilico Medicine; clinical work led by investigators at Chinese hospitals including Peking Union Medical College Hospital.
Limits and debate
The Phase 2a trial was small and short; its primary endpoint was safety. The lung-function improvement in the highest-dose group had a wide confidence interval and was a secondary measure, not confirmed efficacy. Adverse events leading to discontinuation included liver toxicity and diarrhoea. Phase III results are needed before any approval; the late-stage trial is being run in China.
Reactions
Alex Zhavoronkov (CEO, Insilico Medicine): Wrote that whether AI can bring meaningful, sustained disruption to drug development had remained unanswered, framing the trial as a test case. source
Zuojun Xu (Phase III leading principal investigator, Peking Union Medical College Hospital): Said the AI-identified target had never previously been linked to fibrosis. source
Fields it touches
Drug discoveryRespiratory medicineFibrotic diseasesAgeing-related disease
Robin chains together AI agents for literature review, candidate evaluation and data analysis so that the system proposes hypotheses, chooses experiments, interprets the results and updates its ideas — with humans doing the physical lab work.
Status. Published in Nature 655, 497–505 (2026). The drug candidate was validated only in laboratory cell-culture experiments; no clinical testing for this use has been reported.
Full dossier: why, how, steps, sources
Lab
FutureHouse
People
Ali E. Ghareeb (first author), Sam Rodriques (senior author)
Dry age-related macular degeneration is the major cause of blindness in the developed world, and treatment options are limited. More broadly, Robin was presented as one of the first systems to automate both hypothesis generation and data analysis in a lab-in-the-loop cycle.
How it was done
The literature agents generated and ranked therapeutic strategies and candidate compounds; human scientists ran the chosen cell-culture experiments; the analysis agent interpreted the data and proposed follow-ups. Robin identified ripasudil, a drug already used clinically for glaucoma, as a candidate never previously proposed for dry AMD, and its activity was confirmed in vitro. FutureHouse states that all hypotheses, experimental choices, analyses and main-text figures were produced by Robin.
Step by step
Literature agents propose and rank therapeutic strategies
Human scientists run the selected cell-culture assays
Data-analysis agent interprets results
Robin proposes a follow-up experiment and analyses it
Results written up (about 2.5 months from start to submission) and later peer-reviewed in Nature
What it brings
A concrete demonstration of a semi-autonomous 'lab-in-the-loop' AI system producing a testable drug-repurposing lead, published alongside Google's Co-Scientist in the same Nature issue.
Human vs AI
AI generated hypotheses, chose experiments and analysed data; humans carried out all physical experiments and oversaw the project.
Limits and debate
Evidence is preclinical (cell culture), which is far from showing benefit in patients. Repurposing a known drug is a narrower task than discovering a new one. The announcement did not list explicit limitations.
Fields it touches
BiomedicineAI for scienceOphthalmologyDrug repurposing
AlphaEvolve is an evolutionary coding agent that uses Gemini models to write and improve programs. It found a way to multiply two 4×4 complex-valued matrices with 48 scalar multiplications, improving on Strassen's 1969 algorithm in that setting, and found a configuration of 593 spheres touching a central sphere in 11 dimensions, a new lower bound for the kissing number.
Status. Announced with a technical white paper; outputs are explicit constructions verified by code. Tao and Gómez-Serrano are acknowledged for contributions and advice; their own large-scale study followed in November 2025.
Full dossier: why, how, steps, sources
Lab
Google DeepMind
People
Terence Tao, Javier Gómez-Serrano
AI systems
AlphaEvolveGemini FlashGemini Pro
In plain words
Many maths problems ask for the best possible arrangement of something, such as how many equal balls can touch one ball in 11-dimensional space. AlphaEvolve writes and rewrites computer programs that build such arrangements and keeps the ones that score best. It found a slightly better 11-dimensional arrangement (593 balls) and a quicker recipe for multiplying 4×4 tables of complex numbers. These are records, not proofs that the answers are optimal, and the practical speed-up for ordinary computing is limited.
It extended the FunSearch idea from small functions to whole codebases and from one problem to dozens. Applied to more than 50 open problems, it matched the best known results in about 75% and improved them in about 20%.
How it was done
An ensemble of Gemini Flash (breadth of ideas) and Gemini Pro (depth) proposes code changes; automated evaluators score the results; the best programs are kept and mutated further. The output is a program that constructs the mathematical object, which can be checked independently.
Step by step
Humans encode a problem as code plus an automatic evaluator
Gemini Flash and Pro propose program modifications
Evaluators run and score each candidate
An evolutionary database keeps and recombines the best programs
Researchers verify the resulting constructions
Chronology
DeepMind announces AlphaEvolve
Georgiev, Gómez-Serrano, Tao and Wagner post their 67-problem study
What it brings
Beyond mathematics, a scheduling heuristic it found recovers on average 0.7% of Google's worldwide compute resources. It became the basis for the Tao–Gómez-Serrano–Georgiev–Wagner study and later Erdős and Ramsey-number work.
What it does not show
No proofs of optimality; no resolution of the kissing problem or of the matrix-multiplication exponent.
Human vs AI
Humans choose problems and write evaluators; the search for improved constructions is automated.
Credit and priority
Builds directly on FunSearch and on prior human records; Terence Tao and Javier Gómez-Serrano are acknowledged for contributions and advice.
Limits and debate
Improvements are typically small numerical gains on bounds, and the system gives one-sided bounds (constructions), not proofs of optimality. The 48-multiplication result is for complex-valued matrices.
Reactions
Terence Tao (Fields medallist, UCLA): Its main advantage is breadth and scale, but domain experts are needed to design exploit-resistant verifiers and to interpret outputs. source
Fields it touches
AlgorithmsCombinatoricsDiscrete geometryAnalysisData-centre schedulingChip designAI training efficiency
The AI Scientist is a pipeline of foundation models that comes up with machine-learning research ideas, writes and runs the code, analyses results, writes the whole paper and reviews it.
Status. The system is described in Nature (March 2026). The AI-written workshop paper passed first-round review at an ICLR 2025 workshop and was withdrawn by prior agreement; independent evaluators found serious quality problems in the earlier version.
Full dossier: why, how, steps, sources
Lab
Sakana AI
With
University of British Columbia, University of Oxford
People
Chris Lu, Cong Lu, Robert Tjarko Lange, Jeff Clune, Jakob Foerster
AI systems
The AI ScientistThe AI Scientist-v2
In plain words
Sakana built software that writes a machine-learning research paper from idea to final PDF. One of three such papers got good enough scores at a workshop — a smaller, less selective venue attached to a major conference — and was then withdrawn as agreed in advance. It shows AI can produce work that passes a first look by reviewers, not that it can do top-tier research on its own.
Key figures
3 (of 43 workshop submissions)AI papers submitted
1 (scores 6, 7, 6)Passed review
~70%Workshop acceptance rate
42%Experiments failing from code errors (independent test of v1)
Why it matters
Automating the full research cycle has been a long-standing ambition of AI. Passing real peer review, even at a workshop, was a symbolic test of whether AI-generated research could meet human standards.
How it was done
With the cooperation of ICLR leadership and workshop organisers and university ethics approval, Sakana submitted three papers generated end to end by The AI Scientist-v2 to the ICLR 2025 'I Can't Believe It's Not Better' workshop; reviewers knew 3 of 43 submissions might be AI-generated. One received scores of 6, 7 and 6, above the acceptance threshold; the other two did not. Humans chose the broad topic and which three papers to submit.
Step by step
AI Scientist v1 released (Aug 2024)
v2 generates candidate papers end to end
Humans pick three to submit to an ICLR 2025 workshop under an agreed protocol
One paper passes first-round review (average 6.33)
Paper withdrawn before publication as agreed
System paper published in Nature (March 2026)
Chronology
Original AI Scientist released (month only)
Workshop peer-review result announced
Technical report and open-source v2 released
Paper published in Nature
What it brings
A milestone in automated research and a prompt for the community to decide how venues should handle AI-generated manuscripts.
What it does not show
Not acceptance at a main conference track; not a published paper (it was withdrawn); not research in fields beyond machine learning.
Human vs AI
AI generated ideas, code, experiments, text and figures; humans chose the topic, selected submissions and handled the ethics arrangement.
Credit and priority
Developed by Sakana AI with academic co-authors at UBC (Jeff Clune's group) and Oxford (Jakob Foerster's group).
Limits and debate
Workshop acceptance rates (about 60–70%; the Nature paper cites 70%) are far higher than main conference tracks, and Sakana's own review judged none of the three papers good enough for the main track. The AI made citation errors. An independent evaluation of the first version (Beel, Kan & Baumgart, 2025) found 42% of experiments failed due to coding errors, poor novelty checks and some hallucinated numbers, likening output to a rushed undergraduate paper. The Nature paper itself warns of risks such as overloading review systems.
Reactions
Joeran Beel, Min-Yen Kan and Moritz Baumgart (Independent evaluators (arXiv, Feb 2025)): Found critical shortcomings but called the system a leap forward in research automation, with unprecedented speed and cost. source
Davide Castelvecchi (Nature news): Covered the Nature paper as the first peer-reviewed account of how to build an AI scientist. source
Fields it touches
Machine learning researchAI for scienceScientific publishingPeer review
A team of cooperating AI agents built on Google's Gemini model reads literature and a scientist's stated goal, then generates, debates, ranks and refines research hypotheses for human scientists to test in the lab.
Status. System paper peer-reviewed in Nature (May 2026); the bacterial gene-transfer work appeared in Cell (2025). Validations are laboratory (cell-line and organoid) experiments, not clinical results.
Full dossier: why, how, steps, sources
Lab
Google DeepMind Google Research with Google DeepMind.
With
Imperial College London and the Fleming Initiative, Houston Methodist Hospital, Sequome, Stanford University
People
Juraj Gottweis, Vivek Natarajan, José R. Penadés (Imperial College London)
AI systems
AI co-scientist (Co-Scientist), built on Gemini 2.0 / Gemini
In plain words
Scientists spend a lot of time deciding what experiment to try next. Google built a team of AI agents that reads the literature and proposes ranked ideas. Some of those ideas held up when real labs tested them in cells and mini-organs. That is promising, but no treatment came out of it, and humans still chose the questions and ran every experiment.
Key figures
6 plus a supervisorSpecialised agent roles
15 (subset of 11 for novelty/impact)Expert-curated research goals in blog evaluation
about two daysImperial case: time to leading hypothesis
Why it matters
Choosing which hypotheses to test is one of the slowest, most expertise-bound parts of research. A tool that reliably proposes good, novel ideas could speed up discovery, but claims that AI generates genuinely new scientific ideas need experimental proof.
How it was done
Specialised agents (generation, reflection, ranking via a tournament, evolution, proximity and meta-review, coordinated by a supervisor) iteratively improve hypotheses, with quality rising as more computing time is spent. Partner labs tested selected outputs: proposed drug-repurposing candidates for acute myeloid leukaemia reduced tumour-cell viability in cell lines; proposed targets for liver fibrosis showed activity in human liver organoids; and, given a question from Penadés's group at Imperial, it independently proposed the explanation for a bacterial gene-transfer puzzle that the group had found but not yet published.
Step by step
Scientist states a research goal in natural language
Agents generate and critique candidate hypotheses
A tournament ranks hypotheses; top ones are refined
Experts review the shortlist
Partner labs test selected hypotheses experimentally
Peer-reviewed publication in Nature (2026)
Chronology
System announced with preprint
IEEE Spectrum reports on the Imperial result after its publication in Cell
Peer-reviewed paper published in Nature
What it brings
One of the first AI hypothesis-generation systems with laboratory follow-up across several biomedical problems, published in Nature alongside FutureHouse's Robin.
What it does not show
Not a new approved drug or clinical result; not a fully autonomous scientist; lab validations were in cell lines and organoids.
Human vs AI
Scientists set the question, filter the AI's suggestions and do all experiments; the AI proposes and ranks ideas.
Credit and priority
The bacterial gene-transfer finding was made experimentally by Penadés's group before the AI was asked; the AI's proposal matched it. Lab validations were carried out by partner institutions.
Limits and debate
Google's own blog notes its internal quality score is not based on independent ground truth and its expert evaluation used a small sample (15 research goals). Lab results are early-stage and do not show benefit in patients. In the Imperial case the AI started from a focused question and the group's data and papers; commentators debated how much was new insight versus recombining prior knowledge.
Reactions
José R. Penadés (Microbiologist, Imperial College London): Was so surprised by the matching hypothesis that he asked whether Google had accessed his computer, and found even the incorrect suggestions well reasoned. source
Kriti Gaur (Elucidata (life-sciences analytics)): Argued that until it shows original, verifiable insight it 'remains a powerful assistant, but certainly not a co-scientist'. source
Kyle Swanson (Computer scientist, Stanford University): Positive but noted that users still need to be somewhat expert to get the most from such systems. source
Fields it touches
BiomedicineAI for scienceDrug repurposingMicrobiologyHepatology
MatterGen is a generative AI model that proposes new crystal structures directly, and can be steered toward desired chemistry, symmetry or properties such as stiffness or magnetism — rather than screening a fixed list of candidates.
Status. Published in Nature (Jan 2025). A 2026 Materials Horizons paper by Mikkel Juelsholt argues the one experimentally made example is a compound reported decades earlier and present in MatterGen's training data. No public author response as of 21 Nov 2026.
Full dossier: why, how, steps, sources
Lab
Microsoft Research Microsoft Research AI for Science.
With
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences
Designing a material with a target property usually means testing huge numbers of candidates; generating good candidates on demand could shortcut that search for batteries, catalysts and more.
How it was done
A diffusion model (the same family as AI image generators) was trained on about 608,000 stable materials from the Materials Project and Alexandria databases, and fine-tuned for property targets. The authors report its structures are more than twice as likely to be new and stable as those of earlier generative models. As a proof of concept, a team in Shenzhen synthesised one proposal, TaCr2O6, aimed at a stiffness (bulk modulus) of 200 GPa; it measured 169 GPa, within 20% of target.
Step by step
Train diffusion model on ~608,000 known stable structures
Fine-tune toward property constraints
Screen outputs with the MatterSim simulator and novelty checks
Synthesise one candidate and measure its property
Release code under MIT licence with training data
What it brings
Demonstrated property-conditioned generation of crystals at scale and released an open model, while the follow-up critique sharpened the question of how to check AI-proposed 'novelty'.
Human vs AI
The model proposes structures; humans selected, synthesised and measured the single test compound.
Limits and debate
Microsoft itself noted the synthesised structure matched the proposal 'with the caveat of compositional disorder' between Ta and Cr. Juelsholt's crystallographic analysis concludes the made material is identical to a previously reported Ta/Cr oxide described in the early 1970s and included in the training set, echoing the disorder problem raised about GNoME. Only one experimental example was reported.
At the 2024 International Mathematical Olympiad, DeepMind's systems solved 4 of 6 problems for 28 of 42 points, the top of the silver range (gold required 29). AlphaProof solved two algebra problems and one number-theory problem, including the hardest problem of the contest, solved by only five human contestants; AlphaGeometry 2 solved the geometry problem.
Status. AlphaProof's proofs are Lean-checked; IMO 2024 solutions were scored by Timothy Gowers and Joseph Myers under IMO rules; the method was published in Nature on 12 November 2025.
Full dossier: why, how, steps, sources
Lab
Google DeepMind
People
Thomas Hubert, Rishi Mehta, Laurent Sartran, Timothy Gowers, Joseph Myers
AI systems
AlphaProofAlphaGeometry 2GeminiLean
In plain words
The IMO is the world championship of high-school mathematics, with six very hard proof problems. DeepMind's systems scored like a silver medallist, but with two important differences from students: humans first rewrote the questions in a computer-checkable language, and the AI sometimes worked for days. The upside is that every AlphaProof answer is checked by software, so it cannot be a convincing-sounding but wrong proof. It changes nothing in everyday life; it is a milestone in machine reasoning.
Key figures
28/42 (4 of 6 problems)Score
29 pointsGold threshold
~80 million auto-formalisedTraining statements
up to 3 daysLongest solve
19 seconds after formalisationAlphaGeometry 2 on P4
Why it matters
The IMO is a long-standing benchmark of creative mathematical reasoning. AlphaProof was the first AI system to reach medal level, and every one of its proofs is machine-verified.
How it was done
AlphaProof couples a pre-trained language model with AlphaZero-style reinforcement learning inside the Lean proof assistant. A fine-tuned Gemini model auto-formalised natural-language problems, yielding about 80 million formal statements for training. For the hardest problems it used 'test-time RL', generating and learning from variants of the target problem during the search.
Step by step
Auto-formalise millions of natural-language problems into Lean with a fine-tuned Gemini model
Train the prover by reinforcement learning on these statements
At IMO 2024, humans manually translate the official problems into Lean
Run proof search with test-time RL, taking from minutes up to three days per problem
Lean checks the proofs; Gowers and Myers score them
Chronology
DeepMind announces IMO 2024 silver-standard result
AlphaProof paper published in Nature
What it brings
Demonstrated that RL on formal mathematics can reach elite competition level with guaranteed-correct output; AlphaProof was later combined with AlphaEvolve and Deep Think in research workflows.
What it does not show
Not a gold medal, not achieved within contest time, not end-to-end from natural language, and no progress on combinatorics problems.
Human vs AI
Humans translated the problems into Lean; the proofs were found autonomously.
Credit and priority
Problem formalisation was done by humans; scoring was done by external mathematicians Timothy Gowers and Joseph Myers.
Limits and debate
Not under competition conditions: some problems took up to three days, versus two 4.5-hour sessions for students. Both combinatorics problems went unsolved. The pipeline was not end-to-end, since formal statements were prepared manually.
Reactions
Timothy Gowers (Fields medallist): Scored the AI's solutions together with Joseph Myers using IMO point-awarding rules. source
AlphaFold 3 extends structure prediction from single proteins to complexes: proteins together with DNA, RNA, small drug-like molecules (ligands) and some chemical modifications.
Status. Published in Nature (May 2024). Initially released without code, which drew criticism; code and weights were released for academic use in November 2024 and the paper was corrected accordingly.
Full dossier: why, how, steps, sources
Lab
Google DeepMind
With
Isomorphic Labs
People
Josh Abramson (first author)
AI systems
AlphaFold 3AlphaFold Server
Why it matters
Biology happens when molecules interact, so the shape of a complex is often more useful than the shape of one protein. Predicting how small molecules sit on proteins had been the domain of physics-based docking tools.
How it was done
DeepMind and Isomorphic Labs trained a new model that predicts the joint 3D structure of mixed sets of molecules. The authors report at least a 50% improvement over existing methods for protein interactions with other molecule types, and on the PoseBusters benchmark it outperformed traditional physics-based docking without needing structural input. A free AlphaFold Server was opened for non-commercial research.
Step by step
Model developed jointly by Google DeepMind and Isomorphic Labs
Paper published in Nature with pseudocode; access via the AlphaFold Server
Researchers and reviewers criticise the absence of code
Code and weights released for academic use (Nov 2024); paper corrected
Chronology
Paper and AlphaFold Server released
Code and weights released for academic use
What it brings
Broadened AI structure prediction to the kinds of interactions most relevant to drug design and gene regulation, and sparked a debate about openness standards for high-impact AI papers.
Human vs AI
Humans designed and trained the model and validated it on benchmarks; the model makes the predictions.
Limits and debate
Performance figures are the developers' own benchmark results. Access was restricted at launch (non-commercial server, daily limits), and academic-only terms still apply to the released weights. Nature's editor cited biosecurity and ethical considerations in its code-availability decision.
Reactions
Roland Dunbrack (Reviewer of the paper, Fox Chase Cancer Center): Called the lack of code access 'a big step backwards', as reported by Retraction Watch. source
AlphaGeometry solved 25 of 30 olympiad geometry problems (IMO-AG-30, drawn from 2000–2022) within competition time limits. The previous best computer method (Wu's method) solved 10; an average IMO gold medallist solved 25.9 of the same problems.
Status. Published in Nature on 17 January 2024 (author correction 23 February 2024); code and model open-sourced.
Full dossier: why, how, steps, sources
Lab
Google DeepMind
With
New York University
People
Trieu H. Trinh, Yuhuai Wu, Quoc V. Le, He He, Thang Luong
AI systems
AlphaGeometry (neural language model + symbolic deduction engine)
Why it matters
Olympiad geometry was hard for AI because few human-written formal proofs exist to learn from. AlphaGeometry showed this data bottleneck could be bypassed with synthetic data.
How it was done
A language model suggests auxiliary constructions (extra points or lines), and a symbolic engine performs rigorous deductions from them. The model was trained on 100 million synthetic theorem–proof examples, about nine million involving added constructions, with no human demonstrations. Proofs are checked by the symbolic engine.
Step by step
Generate 100 million synthetic geometry theorems and proofs by symbolic means
Train a language model to propose auxiliary constructions
At solve time, alternate between the symbolic engine and model suggestions
Output a machine-checked proof
What it brings
Its successor, AlphaGeometry 2, solved the geometry problem at IMO 2024 as part of DeepMind's silver-medal result.
Human vs AI
Humans designed the system and data pipeline; problem solving is autonomous once a problem is formalised.
Limits and debate
Covers Euclidean plane geometry only (no geometric inequalities or combinatorial geometry), roughly a third of a typical olympiad. Problems must be translated into its formal language.
Fields it touches
Euclidean geometryAutomated theorem provingNeuro-symbolic AISynthetic training data
A cap set is a collection of points in a high-dimensional grid (coordinates 0, 1, 2) in which no three points lie on a line. FunSearch found a cap set of 512 points in dimension 8, larger than previously known, and improved the best lower bound on how fast the largest cap sets grow (the 'capacity') from 2.2180 to 2.2202.
Status. Published in Nature (Romera-Paredes et al., 14 December 2023), with Jordan Ellenberg (UW–Madison) as co-author. Constructions are explicit and machine-checkable.
Full dossier: why, how, steps, sources
Lab
Google DeepMind
With
University of Wisconsin–Madison, University of Lyon
People
Bernardino Romera-Paredes, Alhussein Fawzi, Jordan Ellenberg, Pushmeet Kohli
AI systems
FunSearchPaLM 2
In plain words
Imagine a 3×3×3×... grid in 8 dimensions and the challenge of picking as many points as possible with no three in a straight line. Mathematicians know the best answer only in low dimensions. FunSearch did not prove a theorem; it found a bigger valid example (512 points) and a better recipe for building large examples in very high dimensions. It has no immediate practical application; its significance is that an AI produced new, checkable mathematics.
Key figures
512 pointsCap set in dimension 8
2.2180 → 2.2202Capacity lower bound
PaLM 2Base LLM
Why it matters
The cap set problem is a famous question in extremal combinatorics; DeepMind called the gain the largest increase in cap-set size in 20 years. It was the first time an LLM-based system produced a new, verifiable discovery on an established open problem.
How it was done
Instead of producing answers directly, a pre-trained LLM wrote small computer programs that generate candidate solutions. An automated evaluator scored each program and rejected incorrect ones; the best programs were fed back as examples, so the population of programs evolved. Because the output is a program, humans can read how the construction works.
Step by step
Humans specify the problem and an automatic scoring function
The LLM (PaLM 2 in the original work) proposes programs that build candidate cap sets
The evaluator runs and scores programs, discarding wrong or weak ones
Top programs are returned to a pool and used to prompt further rounds (evolutionary loop)
Researchers inspect the best programs and extract the mathematical construction
Chronology
Paper received by Nature
Paper accepted
Nature publication and DeepMind announcement
What it brings
Established the 'LLM + evaluator + evolution' recipe that later became AlphaEvolve. The same system also found better online bin-packing heuristics.
What it does not show
No proof about the true asymptotic size of cap sets; the bound is still far from the upper bound. Not a general-purpose theorem prover.
Human vs AI
Humans posed the problem and wrote the evaluator; the LLM-driven search generated the constructions; mathematicians interpreted them.
Credit and priority
Builds on decades of human cap-set constructions; Jordan Ellenberg is a co-author and collaborator. The system's search depends on a human-written evaluator.
Limits and debate
The new capacity bound remains far from the known upper bound; this is incremental progress on bounds, not a resolution of the cap set problem.
Reactions
Jordan Ellenberg (Mathematician, University of Wisconsin–Madison): Collaborated on the work and co-authored the Nature paper. source
Gravitational-wave detectors such as LIGO are giant laser interferometers. Urania searched the vast space of possible layouts of mirrors, lasers and other optics for designs that would be more sensitive than the best human-made blueprints.
Status. Peer-reviewed in Physical Review X (2025). The designs are computed blueprints; no report was found of any being built.
Full dossier: why, how, steps, sources
Lab
Max Planck Institute Max Planck Institute for the Science of Light.
With
California Institute of Technology (LIGO)
People
Mario Krenn, Yehonathan Drori, Rana X Adhikari
AI systems
Urania
Why it matters
Detector sensitivity sets how much of the Universe can be observed for black-hole and neutron-star mergers. Existing designs come from decades of expert work, and the space of unconventional layouts was largely unexplored.
How it was done
The team reframed detector design as a continuous optimisation problem solved with machine-learning-inspired methods. The algorithm proposed unorthodox layouts, which physicists then simplified and analysed for months to understand. They report designs that outperform current next-generation proposals under realistic constraints — increasing the potentially observable volume of the Universe by up to 50-fold — and published 50 top designs as an open 'Detector Zoo'.
Step by step
Encode all candidate optical components and noise models
Optimise layouts with gradient-based, ML-inspired search
Filter for realistic experimental constraints
Physicists interpret the winning designs
Publish the best 50 in a public Detector Zoo
What it brings
An example of AI producing experimental designs that human experts initially found baffling but then learned from. Adhikari told Quanta one insight would have given 'something like 10 or 15% better LIGO sensitivity all along'.
Human vs AI
The algorithm generated and optimised the designs; physicists defined constraints, cleaned up and interpreted the outputs.
Limits and debate
Performance gains are simulated; none of the designs had been built, according to the sources read. Quanta reports the team suspects the AI exploited noise-reduction principles identified decades earlier by Russian physicists but never pursued experimentally — so the 'new physics' partly rediscovers known theory.
DeepMind trained graph neural networks to propose new inorganic crystal structures and to predict whether they would be stable, reporting 2.2 million new crystals, of which about 380,000 were judged most stable.
Status. Peer-reviewed in Nature (2023), but solid-state chemists Anthony Cheetham and Ram Seshadri (Chemistry of Materials, 2024) found scant evidence that the predicted compounds are simultaneously novel, credible and useful. A separate critique (PRX Energy, 2024) challenged the companion A-Lab synthesis claims; the A-Lab team disputed it.
Full dossier: why, how, steps, sources
Lab
Google DeepMind
With
Lawrence Berkeley National Laboratory (A-Lab, companion paper)
AI systems
GNoME (Graph Networks for Materials Exploration)
In plain words
Engineers are always looking for new solid materials, for example for better batteries. DeepMind's AI produced a huge list of hypothetical crystals that its calculations said should be stable. Chemists who looked closely found that many are minor variants of things already known, and that being stable on a computer is not the same as being makeable or useful. The list is a big resource, but it did not hand us millions of new usable materials.
Key figures
2.2 millionPredicted new crystals
~380,000Most stable subset
736Independently made by outside labs (per DeepMind)
41 of 58A-Lab syntheses claimed (companion paper)
10 of 384,870 entriesRandom sample checked by critics
Why it matters
New crystalline materials underpin batteries, solar cells, chips and magnets, and finding them has traditionally relied on slow, experience-guided lab work. A reliable way to screen enormous numbers of hypothetical compounds would accelerate that search.
How it was done
The AI generated candidate structures and predicted their stability; promising candidates were checked with standard quantum-mechanical (DFT) calculations, and the results were fed back to improve the model in repeated rounds. DeepMind said external researchers had independently made 736 of the structures in concurrent work, and a companion Berkeley Lab paper reported an autonomous robotic lab synthesising 41 new materials (out of 58 targets).
Step by step
Graph neural networks propose candidate crystals
Stability predicted by the model, then checked with DFT calculations
Active-learning loop retrains the model on new results
~380,000 most-stable predictions contributed to the open Materials Project database
Companion A-Lab paper attempts robotic synthesis of selected targets
Chronology
GNoME and A-Lab papers published in Nature
PRX Energy critique of A-Lab published (vol. 3, 011002; date approximate)
Cheetham & Seshadri perspective published online in Chemistry of Materials (month approximate)
What it brings
Greatly enlarged public databases of computed candidate structures and became a flagship claim for AI-driven materials discovery, while also prompting a debate about what counts as a 'new material'.
What it does not show
Not 2.2 million synthesised or tested materials; stability was computed at effectively zero temperature, not demonstrated in the lab for the vast majority; no specific new battery or superconductor was shown.
Human vs AI
Humans built the model and the DFT verification pipeline; the AI generated and filtered candidates at scale. Experimental confirmation was limited to a small fraction.
Credit and priority
Built on decades of public crystal data (the Materials Project, ICSD) and DFT methods. Critics stressed the need to incorporate domain expertise in synthesis and crystallography.
Limits and debate
Cheetham and Seshadri argue that many entries are ordered arrangements of metal atoms that would in reality be disordered (making them versions of known compounds), that a large fraction of the 384,870 compositions adopt structures already in the ICSD database, and that the predictions should be called crystalline compounds rather than 'materials' since utility is not shown. The '800 years of knowledge' framing is DeepMind's.
Reactions
Anthony K. Cheetham and Ram Seshadri (Materials chemists, UC Santa Barbara (Cheetham also National University of Singapore)): Found 'scant evidence for compounds that fulfill the trifecta of novelty, credibility, and utility', while saying the methods hold promise. source
Robert Palgrave and Leslie Schoop (Chemists, co-authors of the PRX Energy critique of A-Lab): Reported systematic errors in how the robotic lab identified its products; Chemistry World summarised their conclusion as no new materials discovered. source
Gerbrand Ceder (A-Lab lead, Berkeley Lab): Provided additional data that he argued confirms the A-Lab synthesised the reported compounds. source
GraphCast is a machine-learning model that produces 10-day global weather forecasts; GenCast produces probabilistic 15-day forecasts as an ensemble of 50 or more possible futures.
Status. Both peer-reviewed (Science 2023; Nature 2024). Accuracy comparisons are the authors' evaluations against ECMWF systems; ECMWF has run GraphCast experimentally.
Full dossier: why, how, steps, sources
Lab
Google DeepMind
AI systems
GraphCastGenCast
Key figures
>90% of 1,380GraphCast: variables beating HRES
<1 minute on one TPU v4GraphCast: 10-day forecast time
97.2% of 1,320GenCast: targets beating ENS
~8 minutes on one TPU v5GenCast: 15-day forecast time
0.25° (~28 km)Resolution
Why it matters
Weather forecasting has relied for decades on numerical models that solve physics equations on supercomputers, with ECMWF's systems as the global benchmark. Faster and more accurate forecasts, especially of extremes, translate directly into lives and money saved.
How it was done
Both models learned from about four decades of ECMWF's ERA5 reanalysis (a reconstruction of past global weather). GraphCast was better than ECMWF's deterministic HRES on more than 90% of 1,380 test variables and lead times, and produces a 10-day forecast in under a minute on one TPU v4 machine. GenCast, a diffusion-style ensemble model, beat ECMWF's ENS ensemble on 97.2% of 1,320 targets and generates a 15-day ensemble member in about 8 minutes on one TPU v5.
Step by step
Train on ERA5 reanalysis (~40 years)
GraphCast: deterministic 10-day model at 0.25° resolution, published in Science (2023)
Comparisons against ECMWF HRES, including cyclone tracks
GenCast: ensemble model trained on data to 2018 and tested on 2019, published in Nature (2024)
Code and weights released openly
What it brings
Showed that learned models can match or beat leading operational physics-based forecasts on most benchmarks at a fraction of the compute; ECMWF has run GraphCast forecasts as an experiment.
Human vs AI
Humans designed the models and evaluation; the models learned atmospheric dynamics from data and produce forecasts autonomously once trained.
Limits and debate
The models depend on physics-based reanalysis and analyses for training and starting conditions, so they complement rather than replace traditional systems. Reported figures are the developers' own benchmarks; single-storm successes (e.g. Hurricane Lee) are anecdotes.
Fields it touches
MeteorologyEarth system scienceDisaster preparednessRenewable energy forecastingClimate science
AlphaDev searched directly at the level of assembly instructions for faster ways to sort very short lists (e.g. 3–5 numbers), which are building blocks of everyday sorting code.
Status. Published in Nature 618, 257–263 (2023); the routines were accepted into the LLVM libc++ standard library and a hashing routine into Abseil. A follow-up note by Cassio Neri found shorter programs for one case the paper said could not be shortened.
Full dossier: why, how, steps, sources
Lab
Google DeepMind
People
Daniel J. Mankowitz, Andrea Michi
AI systems
AlphaDev (based on AlphaZero)
Why it matters
Sorting and hashing run constantly in nearly all software, so even small speedups to core library routines multiply across vast numbers of computers.
How it was done
DeepMind turned program writing into a single-player 'assembly game' in which the agent adds one instruction per move and is rewarded for correct, fast code, using reinforcement learning built on AlphaZero. The resulting short-sequence sorts were reported up to 70% faster for short sequences and about 1.7% faster for sequences over 250,000 elements, and a hashing routine 30% faster for 9–16-byte inputs.
Step by step
Frame algorithm discovery as an assembly-instruction game
Train an AlphaZero-style RL agent with rewards for correctness and speed
Extract the best small sorting and hashing routines
Submit them to LLVM libc++ and Abseil, where they were accepted
What it brings
One of the first cases of AI-discovered code being merged into a widely used standard library; DeepMind estimates the hashing routine runs trillions of times a day.
Human vs AI
Engineers framed the game, verified correctness and integrated the routines; the agent found the instruction sequences.
Limits and debate
Gains are concentrated on very short inputs; the effect on large sorts is small. MIT Technology Review noted it did not beat the best human four-item sort. Critics argued some speedup reflected a missing branchless implementation in the library, and Neri (arXiv, July 2023) presented two counterexamples to the claim that the 17-instruction 3-element sort was shortest possible, plus a faster C/C++ version.
Matrix multiplication underlies most scientific computing and AI. AlphaTensor searched for recipes that multiply matrices with fewer scalar multiplications. For 4×4 matrices in arithmetic modulo 2 it found a method using 47 multiplications instead of the 49 obtained by applying Strassen's algorithm twice; for a 4×5 by 5×5 product it used 76 multiplications versus the previous 80.
Status. Published in Nature on 5 October 2022; algorithms are checkable by direct verification of the tensor decompositions.
Full dossier: why, how, steps, sources
Lab
Google DeepMind
People
Alhussein Fawzi, Matej Balog, Bernardino Romera-Paredes, Demis Hassabis, Pushmeet Kohli
Strassen showed in 1969 that the schoolbook method is not optimal, and finding better small-case algorithms has been a long-standing challenge. DeepMind described it as the first improvement over Strassen's two-level algorithm in a finite field since its discovery 50 years earlier.
How it was done
The task was turned into a single-player game: decompose a 3D tensor into as few rank-one pieces as possible, each piece being one multiplication. An AlphaZero-style agent learned to play it by reinforcement learning, starting with no knowledge of existing algorithms, in a search space DeepMind put at 30 orders of magnitude larger than Go's move space. It found up to thousands of algorithms per matrix size, and some were tuned to specific hardware.
Step by step
Recast matrix multiplication as tensor decomposition (a game)
Train an AlphaZero-based agent with reinforcement learning
Search for low-rank decompositions for many matrix sizes
Verify each decomposition exactly
Optimise selected algorithms for Nvidia V100 and TPU v2 hardware
What it brings
Showed that RL can discover provably correct algorithms. Hardware-tuned variants multiplied large matrices 10–20% faster than commonly used algorithms on the same hardware, per DeepMind.
Human vs AI
Humans framed the game and the verification; the agent found the decompositions autonomously.
Limits and debate
The 47-multiplication 4×4 result holds only in characteristic-2 (mod 2) arithmetic, not over the real numbers. In 2025 AlphaEvolve found a 48-multiplication method for complex-valued 4×4 matrices, a setting where the 47-multiplication result does not apply.
Fields it touches
Algorithms / algebraic complexityNumerical linear algebraGPU/TPU computingMachine learning infrastructure
A tokamak confines super-hot plasma with magnetic fields from many coils, which must be adjusted thousands of times per second. A single neural network learned to drive all 19 of the TCV tokamak's magnetic coils directly from sensor readings and to hold the plasma in a range of requested shapes.
Status. Published in Nature 602, 414–419 (2022); demonstrated on the real TCV tokamak in Lausanne. Follow-up work on practical RL magnetic control continued.
Full dossier: why, how, steps, sources
Lab
Google DeepMind DeepMind at the time.
With
EPFL Swiss Plasma Center
AI systems
Deep reinforcement learning controller
Key figures
19Magnetic coils controlled by one network
up to ~3 sTCV plasma duration
Why it matters
Plasma shape and stability control is central to fusion research and traditionally uses a set of separate, hand-engineered controllers. Learning a controller from a target specification could make it faster to explore new plasma configurations.
How it was done
The controller was trained by reinforcement learning entirely inside an EPFL plasma simulator, then deployed on the real machine. It produced conventional and elongated shapes, an ITER-like shape, negative triangularity and 'snowflake' configurations, and sustained 'droplets' — two separate plasmas in the vessel at once, not previously done on TCV.
Step by step
Specify target plasma shapes and behaviours
Train an RL agent in EPFL's plasma simulator
Deploy the learned controller on the TCV tokamak
Validate a range of shapes on real plasma shots (each lasting up to ~3 s)
What it brings
Described in commentary as the first use of reinforcement learning for feedback control on a tokamak; a demonstration of sim-to-real AI control in a demanding physics experiment.
Human vs AI
Physicists defined goals and the simulator; the AI learned the control policy, which then ran autonomously during experiments.
Limits and debate
TCV is a small research tokamak whose plasmas last about three seconds; the work concerns shaping and control, not fusion energy gain. Experiment time is limited and machine conditions vary day to day.
Mathematicians used machine learning to spot hidden relationships in large datasets of mathematical objects. In knot theory it revealed that a knot's algebraic signature is tied to its geometry through a new quantity, the 'natural slope'. In representation theory it pointed to structures that led to a new algorithm bearing on the combinatorial invariance conjecture for Kazhdan–Lusztig polynomials.
Status. Published in Nature (Davies et al., 2021). The knot-theory relation was proved by the human mathematicians; the representation-theory algorithm was checked computationally on more than 3 million examples.
Full dossier: why, how, steps, sources
Lab
Google DeepMind
With
University of Oxford, University of Sydney
People
Alex Davies, Marc Lackenby, András Juhász, Geordie Williamson
AI systems
Supervised neural networks with attribution (saliency) analysis
Why it matters
The combinatorial invariance conjecture, linking Bruhat-interval graphs to Kazhdan–Lusztig polynomials, had been open for roughly 40 years. The work was an early demonstration that AI could help professional mathematicians form conjectures, not just compute.
How it was done
Researchers trained supervised models to predict one invariant from others; when a model succeeded, attribution techniques showed which input features mattered. Those hints steered Lackenby toward the natural slope, after which the humans proved the exact relationship. For Kazhdan–Lusztig polynomials, the models highlighted broken dihedral intervals and extremal reflections, which helped Williamson propose a new algorithm.
Step by step
Mathematicians hypothesise that a relation exists between two families of invariants
A neural network is trained to predict one from the other; success signals a real pattern
Attribution methods identify which features drive the prediction
Humans turn these hints into precise conjectures
Humans prove the knot-theory theorem; the representation-theory algorithm is verified on over 3 million cases
What it brings
A template for 'AI-guided intuition': the model points, the mathematician proves. It was published in Nature and influenced later AI-for-math work.
Human vs AI
The AI only detected patterns and highlighted features; every conjecture and proof was formulated by humans.
Limits and debate
The representation-theory result is an algorithm and supporting evidence, not a proof of the full combinatorial invariance conjecture.
Before a chip is manufactured, its large blocks ('macros') must be placed on the silicon so that wiring, power use and speed are good — a step called floorplanning. AlphaChip treats this as a game, placing components one at a time on a grid and being rewarded for the final layout quality.
Status. Peer-reviewed in Nature (2021) with a 2024 addendum after Nature opened an investigation and second review (editor's note, Sept 2023). Independent evaluations (Cheng, Kahng et al.; Markov's meta-analysis in Communications of the ACM, Nov 2024) report it lags simulated annealing and commercial tools; Google's rebuttal says those studies did not run the method as described.
Full dossier: why, how, steps, sources
Lab
Google DeepMind Started at Google Brain, later Google DeepMind.
With
MediaTek (external adopter)
People
Azalia Mirhoseini, Anna Goldie, Jeff Dean (co-author of 2024 rebuttal), Igor L. Markov (critic)
AI systems
AlphaChip (Circuit Training)
Why it matters
Floorplanning is a hard combinatorial problem that can take human experts weeks or months per chip, and better layouts translate into faster, more efficient processors.
How it was done
Google trained a reinforcement-learning agent with a graph neural network that improves as it is pre-trained on more chip blocks. Google says it produces 'superhuman or comparable' layouts in hours and has been used in every TPU generation since 2020, in its Axion datacenter CPUs, and was extended by MediaTek. Open-source code and, in 2024, a pre-trained checkpoint were released.
Step by step
2020 preprint; 2021 Nature paper with open-source Circuit Training code
Layouts used in Google TPU generations
ISPD 2023 study by Cheng, Kahng et al. finds weaker results than baselines
Nature editor's note (Sept 2023) and second peer review
Addendum, checkpoint release and 'AlphaChip' name (Sept 2024)
Markov meta-analysis in CACM and Google rebuttal 'That Chip Has Sailed' (Nov 2024)
What it brings
Google credits it with triggering a wave of AI-for-chip-design research and with real deployment in its own hardware; it is also a leading case study in reproducibility disputes over industrial AI claims.
Human vs AI
Engineers set objectives and integrate the layouts into full chip design flows; the agent proposes macro placements.
Limits and debate
Google did not release the TPU design blocks used in the paper, which critics say rules out full external reproduction. Markov's meta-analysis argues Google RL lags human designers, simulated annealing and commercial software while being slower; Google responds that critics skipped pre-training, used far less compute and unrepresentative test cases. The replication Google cites was by another internal Google team.
Proteins are long chains of amino acids that fold into precise 3D shapes, and the shape determines what the protein does. AlphaFold2 predicts that 3D shape from the amino-acid sequence alone, often at near-experimental accuracy, even when no similar structure is known.
Status. Blind-tested in the CASP14 community assessment, published in Nature (2021) and recognised by the 2024 Nobel Prize in Chemistry (half to Hassabis and Jumper). Widely reproduced and used.
Full dossier: why, how, steps, sources
Lab
Google DeepMind DeepMind at the time.
People
John Jumper, Demis Hassabis, David Baker (co-laureate, University of Washington, for protein design)
AI systems
AlphaFold2
In plain words
Almost everything a cell does is done by proteins, and to understand or target a protein you usually need its 3D shape. Getting that shape experimentally requires painstaking lab work. AlphaFold2 gives a good estimate from the sequence alone. It speeds up research enormously, but it does not replace experiments, and it did not by itself produce new medicines.
Key figures
92.4 overall; 87.0 on hardest targetsCASP14 median GDT score (0–100)
virtually all ~200 million knownProteins with predicted structures
>2 million in 190 countriesUsers (per Nobel committee, 2024)
87 protein domainsCASP14 test set
11 million SEKNobel prize amount (shared)
Why it matters
Researchers had tried to predict protein structures from amino-acid sequences since the 1970s, and the problem was notoriously difficult; CASP chair John Moult said the field had been stuck on it for nearly 50 years. The CASP blind competition, founded in 1994 by Moult and Krzysztof Fidelis, had tracked slow progress for decades.
How it was done
A deep neural network was trained on the experimentally determined structures in public databases together with evolutionary information from related sequences. Entered blind in CASP14 under the team name AlphaFold2, its structures were far more accurate than competing methods. DeepMind and EMBL-EBI then used it to predict structures for nearly all catalogued proteins and released them openly.
Step by step
Earlier AlphaFold version competes at CASP13 (2018)
A completely redesigned network (AlphaFold2) is entered blind in CASP14, May–July 2020
Methods paper and code published in Nature, July 2021
Predicted structures for virtually all ~200 million known proteins made available
Nobel Prize in Chemistry awarded to Hassabis and Jumper, October 2024
Chronology
CASP14 results announced by DeepMind
Nature methods paper published online
Nobel Prize in Chemistry announced
What it brings
Structure prediction became a routine computational step for biologists worldwide, used by more than two million people in 190 countries according to the Nobel committee. It is the most widely cited example of AI producing a foundational scientific tool.
What it does not show
Not a solution to how proteins physically fold over time; not an experimental measurement; not, on its own, a new drug.
Human vs AI
Humans designed the architecture, curated training data and ran the assessments; the trained model then produces structure predictions automatically for any sequence.
Credit and priority
The Nobel was split: half to David Baker for computational protein design, half to Demis Hassabis and John Jumper for structure prediction. AlphaFold built on decades of public structural data (the Protein Data Bank) and on the CASP community assessment.
Limits and debate
Predictions are models, not measurements: accuracy varies (the model reports its own confidence), and it is weaker on flexible or disordered regions, on how proteins move, and on some complexes. It does not by itself explain the physical folding process.
Reactions
Royal Swedish Academy of Sciences (Nobel Committee for Chemistry): Said AlphaFold2 has been used to predict the structure of virtually all 200 million known proteins and by more than two million people in 190 countries. source
John Moult (Co-founder and chair of CASP): Said the field had been stuck on how proteins fold for nearly 50 years (quoted by DeepMind at the CASP14 announcement). source
Three studies from the same research network used machine learning to find antibiotic candidates against drug-resistant bacteria: first by screening libraries of existing molecules (halicin, abaucin), then by having generative AI invent entirely new candidate molecules (2025).
Status. All three results are peer-reviewed (Cell 2020; Nature Chemical Biology 2023; Cell 2025). All evidence is from lab-dish and mouse studies; according to the sources read, none of these compounds had entered human trials.
Full dossier: why, how, steps, sources
Lab
MIT James Collins lab, with Regina Barzilay for halicin.
With
McMaster University (Jonathan Stokes lab), Broad Institute, Phare Bio (nonprofit developer, 2025 compounds)
People
James J. Collins, Regina Barzilay, Jonathan M. Stokes, Gary Liu, Aarti Krishnan
AI systems
Deep neural networks for antibacterial-activity predictionGenerative AI models for de novo design
In plain words
Bacteria are becoming resistant to the antibiotics we have, and new ones are rare. MIT and McMaster researchers taught AI models to predict which molecules kill bacteria, then used them to sift through millions of possibilities in days; later they had AI dream up brand-new molecules. Several candidates cleared infections in mice. None has yet been shown to work in people, so these are promising leads, not new medicines on pharmacy shelves.
Key figures
>100 million in ~3 days (plus ~6,000 repurposing library)Halicin: molecules screened in silico
<2 hoursAbaucin: screening time for 6,680 compounds
240 / 9Abaucin: hits tested / antibiotics found
>36 million2025: compounds generated
22 / 62025: MRSA compounds synthesised / active
Why it matters
Drug-resistant infections are a serious threat to modern medicine. A. baumannii, the target of abaucin, is on the World Health Organization's list of the most dangerous antibiotic-resistant bacteria, and the 2025 work targeted drug-resistant gonorrhoea and multidrug-resistant Staphylococcus aureus (MRSA).
How it was done
For halicin, a model trained on about 2,500 molecules screened a roughly 6,000-compound drug-repurposing library and then more than 100 million molecules in about three days, ranking candidates that lab tests then confirmed. For abaucin, the team tested about 7,500 compounds in the lab, trained a model on the results, and used it to screen 6,680 unseen compounds in under two hours; 240 top hits were tested, yielding nine antibiotics. In 2025, generative AI proposed more than 36 million hypothetical compounds, which were filtered computationally; a few dozen were synthesised and tested, and two leads worked in mouse infection models of drug-resistant gonorrhoea and MRSA.
Step by step
Assemble lab data on which molecules inhibit bacterial growth
Train a model to predict antibacterial activity
Use the model to rank large libraries (or, in 2025, generate new molecules)
Test top-ranked candidates in lab dishes
Test leads in mouse infection models
Hand promising leads to developers for optimisation
Chronology
Halicin study published in Cell
Abaucin study published in Nature Chemical Biology
Generative antibiotic-design study published in Cell
What it brings
Helped establish machine learning as a practical tool for antibiotic discovery, moving from re-screening known molecules to designing new ones, and drew attention to narrow-spectrum candidates that may slow resistance.
What it does not show
Not approved drugs; not tested in humans; not proof that AI-found antibiotics will avoid resistance in clinical use.
Human vs AI
Scientists generated the training data, chose what to synthesise and ran all lab and animal experiments; the models screened or generated candidates and ranked them.
Credit and priority
All three studies came out of James Collins's group at MIT and its alumni (Jonathan Stokes, now at McMaster), with the Broad Institute's Drug Repurposing Hub supplying a key screening library. Development of the 2025 leads passed to the nonprofit Phare Bio.
Limits and debate
No human efficacy data. In the 2025 study synthesis was a bottleneck: of 80 gonorrhoea candidates selected, only two could be made. Halicin had previously been studied as a possible diabetes drug, so it was a repurposing find rather than a new molecule. Lab and mouse success often does not translate to approved medicines.
Reactions
James J. Collins (Bioengineer, MIT): Cautioned that the models are 'doing quite well at designing compounds that can attack in a Petri dish' — a narrower claim than clinical success (as cited by MIT News from the Financial Times). source
Jonathan Stokes (Biochemist, McMaster University): Said AI lets researchers rapidly explore vast regions of chemical space. source
Fields it touches
MicrobiologyDrug discoveryAntimicrobial resistancePublic health
Not discoveries yet: partnerships, tools and programmes that could produce the next dossier. Each links to live coverage on AI Signal.
Isomorphic Labspartnership
Isomorphic Labs and MSD (Merck) strategic research collaboration
Isomorphic Labs will apply its AI drug design engine (IsoDDE), together with selected MSD datasets, to design small-molecule candidates against an undisclosed oncology target; MSD handles preclinical and clinical development, manufacturing and commercialisation. Isomorphic receives an upfront payment plus possible milestones and tiered royalties (amounts undisclosed), and MSD joins Isomorphic's announced partners, alongside Lilly, Novartis and Johnson & Johnson.
Why watch. This becomes a discovery dossier if a candidate from the deal is publicly named as reaching the clinic, with data showing what the AI design step contributed. Demis Hassabis framed it as a 'proof point' for Isomorphic's approach.
Anthropic acquires Mithril Medicine, a one-person AI-run biotech
Anthropic disclosed on Oct 27 that it had acquired Mithril Medicine, whose only full-time employee, founder Cody Tranbarger, used Claude to coordinate contracts, logistics and experiments with 11 external labs on two rare childhood muscular dystrophies; one programme reached promising mouse results in under 18 months. Nature News (Nov 11, 2026) asked whether this lean 'one-person biotech' model is drug discovery's future, and noted that many human scientists contributed to the work.
Why watch. Everything so far is preclinical and self-reported; Anthropic says human trials could come as soon as late 2028 if the candidates hold up. Peer-reviewed data or clinical entry would turn this into a dossier.
OpenAI commits models and funding to DOE's Genesis Mission science campaigns
OpenAI announced support for the U.S. Department of Energy's Genesis Mission: $4M of Codex access for about 2,000 researchers at national labs and universities, $3M of API support for two large scientific campaigns, and access to GPT-Rosalind for selected national-lab biology projects. The named campaigns are high-temperature superconductor research and 'An Atlas of the Machine-Accessible Frontier'.
Why watch. A published, checkable result from the superconductor campaign or the 'Atlas' study would turn this into a discovery dossier.
Signal Shot: grand challenge to formally verify the Signal protocol and app in Lean
The Beneficial AI Foundation launched Signal Shot with Signal and Lean FRO to machine-check the full Signal protocol stack and app in Lean, and to build tools for continuous verification as the software evolves. Steering co-chairs are Max Tegmark and Leonardo de Moura, and Signal research engineer Rolfe Schmidt is an organising co-chair.
Why watch. DeepMind's AlphaProof Code (Nov 10) is described as a milestone toward this challenge. Verification of the Signal apps, group messaging or continuous re-verification would be the next dossier-level steps.
DARPA expMath: 13 university teams start building AI co-authors for mathematics
DARPA's Exponentiating Mathematics (expMath) programme, managed by Patrick Shafto, held its kickoff in March 2026 with 11 technology teams (Caltech, CMU, Harvard, Johns Hopkins, MIT, NYU, Princeton, Stanford, UC Berkeley, UCLA, Maryland) and 2 evaluation teams (Cornell, UC Irvine). Its stated aim is AI 'co-author' systems that propose and prove useful abstractions; UCLA's three-year, $5M Team ALPHA includes Terence Tao.
Why watch. This becomes a dossier when an expMath team publishes a checked new result (for example a formalized new theorem) credited to its AI system, or when the TA2 teams publish evaluations against professional mathematics.
Each dossier is written from primary sources: the paper or preprint, the lab's own post, the Lean repository when there is one, and named specialists' public comments. The status reflects the state of public verification on the date shown, not our own judgement of whether a proof is correct. A formal proof checks that the argument proves the statement written in Lean; whether that statement is the famous problem is a separate human question, and we flag it when it is debated. Corrections are welcome via our methodology page.