Google says teams of Gemini 4 agents working in Antigravity’s /teamwork proved a mathematics conjecture open since 2013 and produced machine-checked formalisations of landmark results across mathematics, chemistry and engineering. The company is now opening the agent tools to outside researchers and developers.
A 2013 conjecture, settled
Researchers from Google and four universities used /teamwork’s Long Proof pattern to prove the Courtade–Kumar conjecture in full generality. The result settles a 2013 information-theory question about how much a single bit can preserve when data passes through a noisy channel.
Gemini first found counterexamples to a broader version of the conjecture, then helped build the proof, splitting the parameter space and developing new analytical bounds. The team then formalised the entire proof in Lean 4, producing 50 million lines of code.
Formal proofs at machine scale
A second system, /teamwork-LEAP, working on its own, produced a machine-checked formalisation of the Weil conjectures, the landmark results on counting solutions of polynomial equations over finite fields. It spans about 1.8 million lines of Lean, built across 28,800 subagent calls.
The same framework formalised Hilbert’s tenth problem over general rings of integers and a proof of finite-time blowup for the 3D Navier–Stokes equations, a Millennium Prize problem open for about 90 years. It also tackled the Birkhoff conjecture, open since 1927.
Beyond mathematics
In chemistry, 30 specialised agents working with a physics-informed simulator identified complex molecular structures with 95 percent top-1 accuracy, nearly doubling single-agent performance. The team says the approach could extend to 3D proteins, complex structure dynamics and other hard laboratory puzzles.
In engineering, /teamwork campaigns rewrote the Zircon microkernel’s remaining C++ code into about 800,000 lines of safe Rust, and rebuilt Google’s AV1 decoder libgav1 in Rust with bit-exact parity, ending 20 to 30 percent faster than the original after profiling work.
Our Read
The verification is the real story here. Lean 4 turns a proof into code a computer kernel checks line by line, so a machine-checked result is far harder to fake than a claimed benchmark. That is why the Courtade–Kumar proof, with its public formalisation, is the anchor of this announcement.
The Weil claim is thinner. Google’s post links to Wikipedia pages for the conjectures rather than a paper or repository, so outsiders cannot yet check it independently. A separate group published its own Weil formalisation a week earlier, which makes the priority picture murkier than the announcement suggests.
Still, the direction is clear. OpenAI’s Yang-Mills validation and an earlier AI proof of the Aaronson–Ambainis conjecture show rival labs chasing the same prize: agents that do not just write code but settle open mathematics. Google’s edge is shipping the tools, not just the results.
Why It Matters
For researchers, the tools matter more than any single theorem. Google is giving 10,000 academic researchers six months of free access to its highest usage tier, and the formal-proof pattern is coming to Antigravity. Verified proofs could soon become routine mathematical work.
For the AI race, this is a new front. Model benchmarks measure speed and cost; machine-checked proofs measure correctness. Google’s Deep Review already screened 578 conference papers, and the lab that makes verification cheap changes what counts as a result.
Background
/teamwork pairs parallel exploration with dedicated falsifiers — proof checkers, test suites and simulators that verify candidate solutions automatically. Coordinators hand tasks to fresh successors, while critics and auditors challenge results from clean contexts, independently and without seeing the original reasoning.
The work used Gemini 4 Argon, Carbon and Hydrogen, and grew out of earlier Antigravity posts on /teamwork. Google says more features and deep-dive tutorials are coming to the Antigravity blog in the coming weeks, with the formal-proof pattern to follow.
Source: Google Research blog


0 Comments