Google Antigravity Formalizes Weil Conjectures in 1.8M Lines

About 1.8 million lines of code. Exactly 28,800 subagent invocations. Zero human coauthors on the final Lean certificate. Those are the numbers behind Google's announcement this week that its Antigravity /teamwork multi-agent system completed a machine-verified formalization of Pierre Deligne's 1974 proof of the Weil conjectures. At the same time, the same system proved the Courtade-Kumar conjecture, an information-theory problem open since 2013, generating a 50-million-line machine-checkable proof that no human had ever produced before.

The report, published November 13 by Vahab Mirrokni, a Google Fellow and VP at Google Research, also covers molecular chemistry, distributed computing, software migration, and production scheduling at Google infrastructure scale. Each domain produced a headline number. The question the mathematical community will spend the next several months answering is whether those numbers replicate under independent review.

The scale of the math results alone is without precedent. The largest prior Lean formalization of a comparable theorem ran to roughly 13 million lines. The Courtade-Kumar formalization at 50 million lines is roughly four times that. The Weil conjectures formalization at 1.8 million lines may be smaller in raw size, but its mathematical depth, touching algebraic geometry, étale cohomology, and number theory simultaneously, makes it arguably the most technically demanding formal proof completed by an automated system to date.

How 28,800 Subagent Calls Turned Deligne's Proof Into 1.8 Million Lines of Lean

The Weil conjectures are a set of statements about the number of solutions to polynomial equations over finite fields, formulated by André Weil in 1949. They were considered so deep that proving them required inventing new mathematical machinery. Alexander Grothendieck spent most of the 1960s developing étale cohomology specifically to build the tools Deligne would eventually need. When Deligne finished the proof in 1974 and received the Fields Medal four years later, it marked one of the central achievements of 20th-century pure mathematics. The proof is not something most working mathematicians can read line by line, even those with PhDs in adjacent fields.

Formalizing a proof in Lean does not mean proving it again from scratch. It means translating every step into a machine-checkable format where a theorem prover verifies that each logical inference is valid. This is distinct from, and in many ways harder than, writing the original proof, because human proofs rely on shared background knowledge that must be made completely explicit for a machine. Every undefined term, every appeal to a lemma "left as an exercise," every informal diagram must be replaced with explicit formal content. For the Weil conjectures, that background knowledge includes decades of algebraic geometry that itself needed to be formalized first.

According to Google Research's official November 13 blog post, the LEAP (Lean Exploration and Proving) subsystem within /teamwork produced approximately 1.8 million lines of Lean 4 code across 28,800 invocations of specialized subagents. The Antigravity /teamwork architecture, described in Google's official platform blog, assigns separate roles to planner agents, worker agents, critic agents, and auditor agents. Critics review outputs without seeing the original agent's reasoning chain, which Google says prevents a failure mode where later agents inherit and amplify errors from earlier reasoning. After each step, automated Lean checking runs before the next subagent invocation begins, so errors surface immediately rather than compounding through thousands of subsequent steps.

The same research sprint also produced a Hilbert's tenth problem formalization spanning over 919,000 lines of Lean. Hilbert's tenth problem asks whether an algorithm can decide if a Diophantine equation has integer solutions. Yuri Matiyasevich proved in 1970 that no such algorithm exists, completing work started by Davis, Putnam, and Robinson over the prior two decades. That Google could formalize both Matiyasevich's result and Deligne's Weil proof within the same multi-week campaign suggests the approach generalizes across mathematical domains rather than depending on any single proof's structure.

The Courtade-Kumar Proof: Genuinely New Math, 50 Million Lines of Lean

The Courtade-Kumar conjecture is different from the Weil conjectures in one important way: it had never been proved before. Thomas Courtade and Michael Kumar proposed the conjecture in 2013. It addresses which Boolean functions maximize the mutual information between their output and a noisy version of their input, a question with direct implications for channel coding, cryptography, and information-theoretic lower bounds in theoretical machine learning. The problem is considered fundamental in information theory, and its resistance to proof over thirteen years was not for lack of attention.

The Google Research team, working with academic collaborators at CUHK, USC, and CMU, used /teamwork's Long Proof pattern to develop a proof of the conjecture in full generality. Gemini models calculated analytical bounds and produced computer-verifiable certificates. The result is a Lean 4 formalization of approximately 50 million lines, according to the Google Research post. It represents the first machine-verified proof of this conjecture anywhere, with human collaborators validating the mathematical strategy while the agentic system handled the formal verification work.

This distinction between formalization and new proof matters for how the result should be interpreted. When an AI system formalizes an existing human proof, the correctness of the underlying mathematics is not in question, only the correctness of the translation. When an AI system contributes to proving an open conjecture for the first time, the mathematics itself is new. Independent verification by the mathematical community becomes the threshold the work must clear, not just Lean's automated checker.

The research sprint also advanced several other open problems in theoretical computer science. The /teamwork system produced a new deterministic algorithm for Boolean matrix multiplication that achieves the speedup target from a STOC 2024 open question, a problem where combinatorial methods had produced only logarithmic improvements over fifty years. Online inverse linear optimization now has a polynomial-time solution, down from exponential. An O(n log n) integer multiplication algorithm was produced that matches GMP, the standard reference implementation, bit-for-bit up to one billion bits. These results collectively represent a volume of theoretical computer science output that would be unusual for an entire research group in a single year, let alone a single multi-week campaign.

Thirty Chemistry Agents at 95%: The NMR Molecular Structure Problem

Not every result in the Google Research announcement involves formal proof. In molecular chemistry, the team applied a 30-agent self-correcting loop to the problem of identifying molecular structures from nuclear magnetic resonance spectroscopy data. NMR structure determination is a standard step in pharmaceutical research, materials science, and natural product chemistry. It requires interpreting spectral patterns that encode the atomic arrangement of a molecule. Experienced chemists do it well, but the process is time-consuming, requires deep domain expertise, and creates a persistent bottleneck in high-throughput synthesis campaigns.

The 30-agent system reached 95% top-1 exact match and 100% top-3 accuracy on complex benchmark molecules, while achieving nearly double the performance of a single-agent baseline on the same task, per the LavX reporting of the Google Research post. The agents include specialized subagents for different spectral analysis steps, with a physics-informed simulator that generates and validates structural candidates against measured spectra. When one agent's structural proposal fails the physics check, the loop self-corrects rather than propagating the error forward.

The implications for drug discovery pipelines are real. Pharmaceutical research campaigns in lead optimization routinely generate hundreds of novel compounds per week. Each requires structural confirmation before the program can advance. At 95% first-choice accuracy with automated physics checking, an agentic NMR system could cut the time between synthesis and characterization, particularly for compounds whose structures are not derivable from simple analogy to known scaffolds. The benchmark molecules tested in this evaluation are complex by the standard of the benchmark, but whether they represent the full difficulty distribution of real-world pharmaceutical discovery compounds is a gap the Google Research post does not address.

In a separate result, the same general framework was applied to review 578 submissions to the STOC 2027 academic conference, more than a third of all submissions, through a Deep Review system. In a survey of 70 respondents who opted in, 90% rated the feedback as mostly or almost entirely factual, 90% said they would use the system again, and critically, 14% reported finding technical gaps that took over an hour to fix. That last figure is the most consequential for the peer review conversation. AI-found errors that would have required hours of reviewer time to catch represent a genuine addition to the review process, distinct from faster summarization of a paper's own claims.

What Machine-Checked Proof Certifies and What It Leaves Open

Lean verification certifies exactly one thing: that the formal statements in the Lean file follow from each other by valid inference rules. Critics argue that it does not automatically certify that those formal statements accurately represent the mathematical claim the original conjecture made. That translation step, from informal mathematical language into Lean's formal language, still requires a human expert to audit. If the Lean formalization correctly proves a subtly different statement than the conjecture as originally posed, the checker still passes, and a reader without deep Lean expertise could easily miss the distinction. This is not a theoretical concern: a 2026 independent review found that one Lean formalization of an OpenAI-claimed mathematics result did not match the accompanying informal proof, even though the Lean file checked out.

The bear case for the Courtade-Kumar result specifically is that the mathematical community has not yet independently reviewed either the proof strategy or the Lean formalization. All figures in the Google Research post are self-reported by Google and its collaborators. Interesting Engineering's coverage notes that Google linked to no preprint on arXiv, no public Lean repository, and no independent verification at the time of the announcement. The Lean community's standard for a claimed formalization is that the repository is public, that an experienced Lean developer can build it without modification, and that the formal statement matches expert consensus on the original result. None of those conditions have yet been confirmed by external parties.

This is not unusual for a major lab announcement. The practical timeline for formal verification community acceptance of a large Lean proof runs from weeks, for results under a few hundred thousand lines, to months for a 50-million-line proof that takes weeks of reviewer time just to navigate. The Weil conjectures formalization, at 1.8 million lines, is the more tractable case for external review, since it formalizes an established proof whose steps are known. The Courtade-Kumar proof requires reviewers to also evaluate whether the proof strategy is mathematically sound and novel, independent of Lean's automated checking.

The 10,000-Researcher Program and the Competitive Stakes

Google's decision to offer 10,000 academic researchers six months of free access to Antigravity's highest usage tier is simultaneously a service announcement and a network effects play. Researchers who build workflows on Antigravity's /teamwork, /boost, and Deep Review tools become invested in the platform's continued availability. The feedback from a large external user base is itself training signal and evaluation data for the next iteration of the system. Coverage of the access program notes that Google has not published the selection criteria or the application deadline, which means the program functions partly as a signaling mechanism to the academic community as well as a service.

The competitive landscape for AI math and science tools is active. Axiom Math, a startup, completed a Lean formalization of the Classification of Finite Simple Groups in October 2026, producing approximately 43 million lines of Lean code across 24 days of agent operation. OpenAI has pursued AI-generated mathematics papers, though independent review found mismatch issues in at least one Lean formalization. Anthropic has not announced a formal proof initiative. DeepMind's AlphaProof system targets competition-style mathematics problems rather than research-frontier conjectures, and operates by training proof-search models rather than by using general-purpose agentic workflows. The technical approaches are different enough that direct comparison is premature.

What Google is doing with /teamwork is qualitatively different from competition math benchmarking. The targets are open-frontier research conjectures, the workflow is explicitly collaborative between human researchers and AI agents, and Lean formalization functions as a verification output rather than as the proof method itself. That combination, if it scales and if the results replicate, creates a new category for AI in research: not a tool that assists humans in doing math, but a system that contributes to mathematical knowledge in ways that require the same external validation any human proof does.

Three Verification Deadlines to Watch Over the Next 90 Days

The most important signal to track in the next 30 days is whether a public Lean repository for the Courtade-Kumar proof appears on GitHub or the Lean4 community Zulip. Without a public repository, the mathematical community cannot begin the independent review that would convert a Google announcement into an accepted mathematical result. If the repository becomes public within the next month, the timelines for formal acceptance compress by weeks. If it stays private, that itself becomes a data point about the confidence Google has in external scrutiny.

The second marker is STOC 2027 paper outcomes. The Deep Review system processed 578 submissions. Decisions from the program committee should be announced in early 2027. If papers that received Deep Review feedback and acted on substantive corrections show higher acceptance rates than the baseline, 10 or more percentage points above the overall conference rate, that provides the first real-world signal about whether AI peer review assistance translates into better science selection, not just faster pre-submission checking. That data point will matter as other venues decide whether to adopt similar systems.

The third marker is the response from other frontier AI labs. Google has now claimed the largest Lean formal proof in history by a margin of roughly 7 million lines over Axiom Math's Classification of Finite Simple Groups result, and done it with a general-purpose agentic system rather than a specialized mathematics model. OpenAI, Anthropic, and xAI will face competitive pressure to demonstrate comparable or superior formal proof capabilities. Watch for model releases or blog posts in the December 2026 and January 2027 window specifically citing formal mathematics benchmarks. Those announcements will be competing with the Google results for the same pool of academic researchers now being offered free access to Antigravity.

When a 50-million-line machine-checked proof of a 13-year-old open conjecture still needs months of independent review to be accepted, the AI is not the bottleneck anymore.

---