🇺🇸 English 🇰🇷 한국어 🇯🇵 日本語
Home › AI & Machine Learning
AI & Machine Learning

Google AI Agents Formalize Weil Conjectures

Google Research said on Nov. 13 that /teamwork-LEAP, a multi-agent workflow in its Antigravity platform paired with Gemini 4 Argon, autonomously produced a machine-checked Lean formalization of the Weil conjectures spanning about 1.8 million lines across 28,800 subagent invocations. The post also claims a 919,000-line Hilbert's tenth problem formalization and opens free access applications for 10,000 researchers.

Google says a team of AI agents has turned one of 20th-century mathematics' landmark results into code a computer can check line by line. In a Google Research post dated Nov. 13, Google Fellow and VP Vahab Mirrokni wrote that /teamwork-LEAP, a multi-agent workflow inside the Antigravity platform, "fully autonomously produced an end-to-end, machine-checked formalization of the Weil conjectures" in about 1.8 million lines of Lean.

Mathematician Pierre Deligne smiling, photographed indoors
Pierre Deligne, whose 1974 proof completed the Weil conjectures, at the Institute for Advanced Study in Princeton in March 2005 (file photo). 사진: User:Eecc / 위키미디어 공용(퍼블릭 도메인)

Key takeaways

  • Google's /teamwork-LEAP produced a roughly 1.8-million-line Lean formalization of the Weil conjectures across 28,800 subagent invocations.
  • The same system wrote more than 919,000 lines of Lean for the first formalization of a recent proof of Hilbert's tenth problem over general rings of integers.
  • Google is opening applications for 10,000 academic researchers to get six months of free access to its highest Antigravity usage tier.

What are the Weil conjectures?

The Weil conjectures are results about counting the solutions of polynomial equations over finite fields. Their final piece fell in 1974, when Belgian mathematician Pierre Deligne completed the proof. The arguments draw on a large body of modern algebraic geometry, which is why a full formal version is a heavy lift.

Formalizing a proof in Lean 4 means rewriting it as code that the Lean kernel verifies step by step. Google's post notes the caveat that matters most: once human mathematicians confirm the formal theorem statement matches the original problem, "the proof only compiles if every logical step holds."

How Google's AI agents formalized the proof

LEAP is a framework developed with colleagues in Google Cloud AI and paired with the Gemini 4 Argon model. It works from informal, human-written proofs. Google said the Weil formalization took 28,800 subagent invocations.

The broader /teamwork design rests on two ideas. Agents explore in parallel, and dedicated "falsifiers" such as the Lean 4 proof checker, unmodified test suites or simulators decide which candidates actually work. Context is also split up on purpose: each specialized agent receives only the files and tools it needs for one step, while critics and auditors review results from clean contexts.

Other math results in the post

Google listed several more formal-verification claims. The system wrote over 919,000 lines of Lean for Hilbert's tenth problem over general rings of integers and independently reproduced a formalization of a recent proof of finite-time blowup for the 3D Navier–Stokes equations with smooth forcing.

It also produced what Google calls the first formalizations of the Birkhoff conjecture, open since 1927, and the 3D Bochner–Riesz conjecture, where it found and fixed a gap in a recently released informal proof. Separately, researchers from Google, CUHK, USC and CMU proved the 2013 Courtade–Kumar conjecture and formalized the proof in 50 million lines of Lean 4.

Context: proofs you can audit

Large AI math claims have drawn scrutiny this year, as with OpenAI's 10,000-agent Navier-Stokes attempt. Machine-checked proofs shift the question from "is the reasoning right?" to "is the statement right?" That is a narrower problem for human reviewers. However, the post links the Weil result to background pages rather than to a public repository, so outside mathematicians cannot yet inspect the 1.8 million lines themselves.

The post also reports engineering results, including about 800,000 lines of safe Rust for Google's Zircon microkernel and a Rust port of the libgav1 video decoder that runs 20% to 30% faster than the C++ original.

What's next

Google says external users can already run Gemini 4 models with /teamwork and /boost in Antigravity. Its formal proof pattern is slated to arrive in upcoming updates, with tutorials on the Antigravity blog in the coming weeks. Publication of the Weil formalization's source would be the clearest next test.

FAQ

Did Google's AI prove the Weil conjectures? No. Deligne finished the proof in 1974. Google's agents translated existing mathematics into a machine-checked Lean formalization.

Can researchers check the Weil formalization? Not yet from the post itself, which does not link a public code repository for it.

Who can get free access? Google opened applications for 10,000 academic researchers to receive six months of free access to its highest usage tier.

References

#Google Research #Antigravity #teamwork #LEAP #Weil conjectures #Lean 4 #formal verification #Gemini 4 Argon #Pierre Deligne #AI math #multi-agent AI #Hilbert's tenth problem
S
Seung Jung
AI Newsway · Developer

Editor responsible for AI Newsway. Writes about AI, developer tools, cloud and productivity technology.

More by this reporter ›
Copyright © AI Newsway All rights reserved. Unauthorized reproduction, redistribution, and AI training prohibited.
Share
f X K

Comments 0

Comments are currently disabled.

Be the first to comment.