Field intelligence for AI-first professionalsVol. II · Nº 195 · Monday, November 16, 2026
← Phantom Notes
Nº 190AI AgentsNovember 13, 20264 min read

Google's Agents Wrote 50 Million Lines of Lean to Prove a 2013 Conjecture. The Headline "Firsts" Are Formalizations of Proofs Humans Already Wrote, and the Free Top Tier Is 10,000 Applications With No Stated Criteria.

Google Research reported on November 13 what Antigravity's /teamwork multi-agent workflows and LEAP, built with Google Cloud AI and paired with Gemini 4 Argon, have done in mathematics: a machine-checked Lean formalization of the Weil conjectures in about 1.8 million lines across 28,800 subagent invocations, more than 919,000 lines formalizing a recent proof of Hilbert's tenth problem over general rings of integers, first formalizations of the Birkhoff conjecture and 3D Bochner-Riesz, and a full proof of the 2013 Courtade-Kumar conjecture by a Google, CUHK, USC, and CMU team, formalized in 50 million lines of Lean 4. Every result is Google's account, done with human experts by Google's own statement; most firsts are formalizations of existing proofs; the formal proof pattern is not yet in Antigravity; and the 10,000 free six-month top-tier seats for academics are by application with no criteria published. We read the post.

By T.W. Ghost


Google Research published today, November 13, 2026, a set of mathematics results produced with Antigravity's /teamwork multi-agent workflows and LEAP, a system built with Google Cloud AI and paired with Gemini 4 Argon. The list: an autonomous, machine-checked Lean formalization of the Weil conjectures, about 1.8 million lines across 28,800 subagent invocations; more than 919,000 lines formalizing a recent proof of Hilbert's tenth problem over general rings of integers; the first formalizations of the Birkhoff conjecture from 1927 and of 3D Bochner-Riesz, where the agents found and fixed a gap in a recent informal proof; and a full proof, in general, of the 2013 Courtade-Kumar conjecture by a team from Google, CUHK, USC, and CMU, formalized in 50 million lines of Lean 4. Google also claims infrastructure wins from the same agents, memory savings it estimates above one pebibyte and 24% less TPU preemption on some hardware.

Read the verbs. Formalized, formalized, formalized, proved. The Weil conjectures, Hilbert's tenth over rings of integers, Birkhoff, and Bochner-Riesz are existing human proofs translated into machine-checked Lean, which is real work and is not discovery. The Courtade-Kumar result is presented as a proof, and Google's post says that "all of this work is done in collaboration with human experts," so the agents are not the authors. Line counts measure volume; 50 million lines is a size, not a quality, and no independent review of any of it exists yet. The tool you can use today is /teamwork and /boost in Antigravity and Deep Review; the formal proof pattern behind the math is "upcoming." And the offer attached, six months free at the highest usage tier for 10,000 academic researchers, is by application, with no criteria or window published. We read the Google Research post and two secondary summaries. Here is what held up.


What Google Reported

ResultGoogle's figureWho
Weil conjecturesFormalized "fully autonomously"; about 1.8M lines of Lean; 28,800 subagent invocationsVendor claim
Hilbert's tenth problem over general rings of integersMore than 919,000 lines formalizing a recent proofVendor claim
Birkhoff conjecture (1927); 3D Bochner-RieszFirst formalizations; a gap in a recent informal proof found and fixedVendor claim
Courtade-Kumar conjecture (2013)Proved in full generality; 50M lines of Lean 4; Google, CUHK, USC, CMUVendor claim
Human involvement"All of this work is done in collaboration with human experts"Verified quote (via TSN, matching the post's framing)
Available today/teamwork and /boost in Antigravity and Deep Review; formal proof pattern "upcoming"Verified (vendor post)
Free tier10,000 academic researchers; six months at the highest usage tier; applications open; no criteria or windowVerified (vendor post)
InfrastructureMemory savings above 1 PiB (estimate); 24% less TPU preemption on some hardwareVendor claim
Plan gatingHydrogen for everyone; Carbon and Argon for Enterprise and Ultra or paidSecondary (Mirrokni via ORIENS); not in the post
Deep ReviewFree up to twice a weekSecondary (ORIENS)

Formalization is not discovery. Translating a known proof into Lean so a machine can check it is valuable and hard. It is a different claim from proving something new, and the post's "firsts" are mostly the first kind.

Humans on every result. By Google's own sentence. The agents did the lines; the experts did the mathematics, or at least stood next to it.

Fifty million lines is a size. It tells you how much Lean the agents produced, not whether a referee would accept the paper. None of this has been independently reviewed yet, and the arXiv identifiers were not in the post we read.

The product and the result are different things. /teamwork ships today. The formal proof pattern that produced the math does not. A reader who buys Antigravity for the Weil conjectures gets the workflow, not the pipeline.


Verified vs Unconfirmed: The Scorecard

ClaimVerdict
Post published Nov 13, 2026Verified (Google Research)
All mathematical results and line countsVendor claims; no independent review
Human experts on every resultVerified quote
/teamwork and /boost available; formal proof pattern upcomingVerified (vendor post)
10,000 free academic seats, six months, by applicationVerified (vendor post); criteria and window not published
Infrastructure savingsVendor claims
Plan gating by model; Deep Review twice weeklySecondary; held
arXiv identifiers for the Courtade-Kumar and Bochner-Riesz workNot pulled

What To Do

Use /teamwork for what it is, a multi-agent workflow in Antigravity that you can try today, and apply for the academic tier if you qualify, knowing the criteria are unpublished.

Skip repeating the line counts as achievements, and skip "AI proved" for the four formalization results; the post itself calls them formalizations and puts humans in the loop on all of them.

The frontier is elsewhere: on a board with error bars, Gemini 4 Carbon leads MathArena by a margin inside them, and Antigravity 2 is the product this workflow lives in. Our model comparison has the current standings side by side.


Sources


*Not sure whether a multi-agent workflow or a single strong model fits your research? Take the free 2-minute quiz and get matched. Then read about the model doing the formalizing: Gemini 4 Argon.*

Which model should you be using?

Three minutes, twelve questions, one defensible answer.

Take the quiz →