Subscribe
AI Models

Google's AI agents formalised a 2013 maths conjecture in 50 million lines of machine-checked code

Tuesday, November 17, 2026 · 4 min read

Google Research says its Antigravity /teamwork system helped prove the Courtade–Kumar conjecture and formalised the proof in Lean 4 — about 50 million lines of code, independently audited — while a second, bigger formalisation of the Weil conjectures still cannot be checked.

Bullish$GOOGL

📊 Scorecard: our Bullish call on $GOOGL is scored from $283.72, the last close before it ran. Its first result comes after the next close.

Google's AI agents formalised a 2013 maths conjecture in 50 million lines of machine-checked code

The News

Google Research reported on November 13 that its Antigravity coding agent's /teamwork multi-agent system helped prove the Courtade–Kumar conjecture — a maths problem first posed in 2013 — in full generality, then formalised the entire proof in Lean 4, a programming language that checks mathematical proofs line by line. The result is about 50 million lines of machine-checked code, run mainly on Gemini 4 Argon, Google's latest AI model.

The formalisation is public, and it has been independently audited. About 99.5% of the 50 million lines are machine-generated certificate data — numerical bounds the Lean software verifies automatically. The human-readable part is far smaller: about 239,000 lines of proof, anchored to a 63-line statement, and accepted by the Lean Foundation's external checker.

The same Google post makes a bigger claim that cannot yet be checked. It says a /teamwork-LEAP pattern produced a roughly 1.8-million-line Lean formalisation of the Weil conjectures — a landmark of 20th-century maths — across 28,800 subagent invocations. But there is no public repository, paper or statement file, so nobody outside Google can verify it yet. Axiom, a rival maths-AI company, published its own public Weil formalisation on November 6, a week before Google's post. Google also says /teamwork's core features are now available outside the company; the underlying preprint dates from September 21, and the proof has not been peer-reviewed.

Fifty million lines of machine-checked maths is less a proof than a stress test — and the agents passed it.

Why It Matters

A formal proof is maths with nowhere to hide. When a proof is written in Lean, a computer checks every step, so a result that passes is genuinely correct. Producing 50 million lines of it tests whether AI agents can do sustained, structured work: planning a proof, splitting it into thousands of pieces and keeping it all consistent — a much harder task than writing a clever paragraph.

For Google, the stakes are commercial as well as scientific. Antigravity is its AI coding agent, competing with tools from OpenAI, Anthropic and a dozen startups. A public, audited result like this is evidence its multi-agent system can handle real, long-horizon engineering work — the kind of evidence enterprise buyers ask for before they pay.

The Weil gap shows where the line is. Google asks the public to take its biggest claim on trust while a smaller rival has already published. Until the Weil repository appears, the honest reading is that /teamwork is proven at the 50-million-line scale and unproven at the 1.8-million-line autonomous scale.

Analyst's Take

Our call is Bullish. This is a genuine capability demonstration, not a press-release promise: the formalisation is public, an independent audit counted the lines and traced the checker trail, and the headline number survived scrutiny. Multi-agent systems that can plan, divide and verify work at this scale are what Google needs to sell Antigravity as a serious engineering tool.

The honest caveat is the Weil claim. A formalisation nobody can inspect is a claim, not a result, and Axiom publishing a public version a week earlier undercuts the 'first' framing. The preprint is from September, and peer review has not happened.

For readers, the signal is the tooling, not the theorem. Watch whether /teamwork's now-public features show up in real engineering work outside Google, and whether the Weil repository appears — if it does, this becomes a much bigger story. Reader poll: Would you trust an AI-proved theorem? Vote: Yes, it's checked, Not yet, Need peer review.

Are you bullish or bearish on this story?

By The Numbers

50 millionLines of Lean 4 code in the Courtade–Kumar formalisation
~99.5%Share that is machine-generated certificate data
~239,000Lines of human-readable proof
63Lines in the formal statement
28,800Subagent invocations for the Weil formalisation
~1.8 millionLines in the Weil conjectures formalisation

FAQ

What does it mean to formalise a proof in Lean?

Lean is a programming language for writing maths so a computer can check every step. A formalised proof is one the software has verified line by line, so there is no room for a hidden gap or a skipped argument.

What is the Courtade–Kumar conjecture?

It is a mathematical conjecture — an unproven statement mathematicians believed to be true — first proposed in 2013. Google's agents helped prove it in full generality, meaning for every case it covers, not just some of them.

Why can't the Weil conjectures claim be checked yet?

Google has published no repository, paper or statement file for that formalisation, so nobody outside the company can inspect or re-run it. Axiom, a rival maths-AI company, published its own public Weil formalisation on November 6.

👁

Source: Navsplace

Get the daily briefing. AI news with a markets lens — tickers, bullish/bearish reads, zero hype. Subscribe free on Substack

More on $GOOGL

All $GOOGL coverage →

More from AI Agent Update