
Google Antigravity teamwork claims a 2013 conjecture proof and a 1.8 million line Lean formalization
By Himanshu Sakre·
On November 13, 2026, Google Research said multi-agent teams built on Gemini 4 helped prove an information theory conjecture and formalize the Weil conjectures in Lean. The proof is on arXiv; the rest is Google's own account.
Google Research says its Antigravity /teamwork system, a set of cooperating AI agents running on Gemini 4 models, has helped crack a 13-year-old information theory problem and build one of the largest machine-checked formalizations on record. The post went up on November 13, 2026. The headline numbers are big: 50 million lines of Lean for one proof, 1.8 million for another. The caveats are in how much of it anyone outside Google has checked.
What Google says it did
The post, written by Google Fellow Vahab Mirrokni, says the team used multi-agent systems "to prove information theory's 2013 Courtade-Kumar conjecture in full generality." It adds that the team "formalized the entire proof in Lean 4, producing 50 million lines of code." Lean is a proof assistant: it accepts a proof only if every logical step type-checks.
A second result concerns the Weil conjectures, a landmark set of statements in number theory. The post says the system "fully autonomously produced an end-to-end, machine-checked formalization of the Weil conjectures," and puts the size at about 1.8 million lines of Lean produced across 28,800 subagent invocations. It also reports about 919,000 lines of Lean for a formalization of a recent proof of Hilbert's tenth problem over general rings of integers.
The post then lists engineering claims: a 170,000-line Rust port of the libgav1 video decoder that it says runs 20 to 30 percent faster than the C++ original, custom TPU kernels, and a change to Google's Borg scheduler that it says cut TPU preemption by 24 percent on some hardware. Those figures come from Google's own post and we have not seen them reproduced independently.
The part you can check today
The Courtade-Kumar result is the one with a public paper. The preprint, arXiv:2609.24931, is titled "A Proof of the Most Informative Boolean Function Conjecture." It was submitted on September 21, 2026 and revised on September 24. Its seven authors are Zijie Chen, Amin Gohari, Adel Javanmard, Honghao Lin, Vahab Mirrokni, Chandra Nair and David P. Woodruff. The abstract describes the proof as computer-assisted and says "The entire proof, including all numerical certificates, has been formally verified in Lean end-to-end."
“All of this work is done in collaboration with human experts, and we are making these tools available externally.”
The conjecture itself comes from Thomas Courtade and Gowtham Kumar, who posted it to arXiv on February 11, 2013 and presented it at ISIT 2013 in Istanbul. It says that if you send a string of fair coin flips through a noisy channel, no single yes-or-no function of the input can carry more than 1 minus the binary entropy of the noise level about the output. The 2013 paper noted that the conjecture was unproven. A plain-language walk through the statement is in our explainer on the conjecture.
The arXiv page lists no journal publication or peer review. That does not make the proof wrong. It means mathematicians have not yet had their turn, and a Lean certificate answers a narrower question than people assume.

What a Lean check does and does not settle
A Lean proof guarantees that the stated theorem follows from the stated axioms. It does not guarantee that the stated theorem is the theorem humans care about. Someone has to read the formal statement and confirm it matches the original conjecture. Google's post stresses that the work was done with human experts, but it does not say how the formal statements were audited.
That is why "fully autonomous" deserves a close read. In the Weil case it describes how the proof code was produced. It does not describe the whole pipeline, including who chose the formal statement and who audited it. We covered the same question for Google DeepMind's AlphaProof code signal work, and our explainer on what a Lean check does not prove sets out the limits.
Size is also a poor proxy for difficulty. Fifty million lines of generated Lean is a measure of how much machinery the agents produced, not of how hard the idea was. The proof's core, per the abstract, is a differential-equation method that the authors describe as a limiting form of auxiliary-receiver techniques from network information theory. That is a human-legible idea, and the line count says little about it.
Access, and what to watch
Google says the tools are available now. External users can run /teamwork and /boost with Gemini 4 models in Antigravity, and Deep Review gives users two free calls a week. The company has opened applications for 10,000 academic researchers to receive six months of free access to its highest usage tier. A formal proof pattern and a rewrite pattern are listed as coming soon, with tutorials promised in the coming weeks.
Google's earlier Antigravity Teamwork post, dated August 31, 2026 and built on Gemini 3.7 Flash, claimed seven open problems and named four of them. It reported that only one, Knuth's Cycles Conjecture, was verified in Lean. The November post moves the claim from a handful of results to formal verification at far larger scale, which is a real change in ambition even if the evidence is still mostly self-reported.
Our take
The Courtade-Kumar paper is the one to take seriously, because it is public, signed by named researchers and checked in a proof assistant. Treat the Weil formalization and the engineering numbers as promising, vendor-reported results until someone outside Google reproduces them. We will watch for three things: an independent audit of the formal theorem statements, mathematicians' verdicts on the preprint, and whether the 10,000-researcher program produces results Google did not curate.
Frequently asked questions
What is Google Antigravity teamwork?
It is a multi-agent system in Google's Antigravity tool where teams of AI agents critique and iterate on each other's work. Per Google Research's November 13, 2026 post, it runs on Gemini 4 models and works alongside human experts. Users can run /teamwork and /boost in Antigravity now.
Did AI prove the Courtade-Kumar conjecture?
Google says its multi-agent systems helped prove it in full generality. The arXiv paper 2609.24931 is by seven human authors and calls the proof computer-assisted, with Lean verification claimed. It has no listed peer review yet, so the result is not independently confirmed.
What did Google formalize in Lean?
Google says it produced 50 million lines of Lean 4 for the Courtade-Kumar proof and about 1.8 million lines for the Weil conjectures across 28,800 subagent invocations. It also cites about 919,000 lines for a Hilbert's tenth problem result. All figures are from Google's own post.
Does a Lean proof mean the result is correct?
A Lean proof means the formal statement follows from the axioms. Humans still need to confirm the formal statement matches the original problem, and Google's post does not describe that audit.
Can researchers use Antigravity teamwork for free?
Google opened applications for 10,000 academic researchers to get six months of free access to its highest usage tier. Deep Review gives two free calls per week. Other patterns, such as formal proof and rewrite, are listed as coming soon.
Sources
What each one is, and whose it is.
- 1
Advancing science and engineering with Antigravity's teamwork feature, Google Research (November 13, 2026)
Vendor announcement - 2
A Proof of the Most Informative Boolean Function Conjecture, arXiv (September 21, 2026)
PaperIndependent of the vendorNot peer reviewed, preprint - 3
Which Boolean Functions are Most Informative?, arXiv (February 11, 2013)
PaperIndependent of the vendorNot peer reviewed, preprint - 4
Antigravity Teamwork multi-agent, Google (August 31, 2026)
Vendor announcement