AlphaProof Nexus turns AI math from persuasive text into checked proof
Formal proof as a machine cockpit: AlphaProof Nexus searches, Lean verifies.📷 AI-generated image / TECH&SPACE
- ★AlphaProof Nexus solved nine open Erdős problems, including two that were 56 years old.
- ★Unlike natural-language proof approaches, its results are checked through the Lean compiler.
- ★Inference cost is only a few hundred dollars per problem, but the overall success rate remains 2.5 percent.
Google DeepMind’s AlphaProof Nexus moves the AI-in-math debate away from “can a model sound convincing” and toward a stricter question: can it produce a proof that a compiler accepts. According to The Decoder, the system autonomously solved nine open Erdős problems, including two problems that had remained open for 56 years.
That matters because a mathematical proof is not a polished essay. In natural language, a model can skip an edge case, hide a weak lemma, or produce an argument that reads smoothly until someone checks it line by line. AlphaProof Nexus instead relies on Lean, a formal system where proofs are written in a way a computer can verify step by step. If a step fails, style cannot cover the gap.
The Erdős connection raises the bar. Paul Erdős left a deep imprint on combinatorics, number theory, and discrete mathematics, and many of his problems survive because they are easy to state but difficult to close. A useful overview of that mathematical legacy is available through the MacTutor biography of Paul Erdős, while the current technical context for machine-checked proofs is best tracked through the Lean prover community.
Google DeepMind’s system autonomously solved nine open math problems, but its 2.5 percent success rate shows how hard formal proof still is for AI.
The breakthrough sits in checkable steps, not persuasive text.📷 AI-generated image / TECH&SPACE
The striking part is not only the number of solved problems, but the cost of trying. The source says inference costs were only a few hundred dollars per problem. If that holds up under broader reproduction, formal mathematics gets a new kind of research instrument: not a replacement for mathematicians, but a system that can cheaply search proof spaces humans may not want or be able to exhaust manually.
The 2.5 percent overall success rate, however, cuts against any premature story about an autonomous mathematician. It means the system still failed on most attempts. In mathematics, that rate is not trivial when open problems are involved, but it is not general theorem-solving intelligence either. The real signal is narrower: when AI is paired with formal verification, the output is no longer just a plausible text artifact. It becomes something checkable.
That is where AlphaProof Nexus differs from approaches centered on natural language, including the route often associated with OpenAI-style reasoning models. A natural-language proof can be useful for intuition, but Lean introduces binary discipline. The proof passes, or it does not. For science, that distinction matters because it reduces room for persuasive hallucination and shifts the burden onto formally verified steps.
The remaining question is scale. Nine solved problems is impressive, but the mathematical value will depend on independent checking, the readability of the formal proofs, and whether human researchers can extract new methods from them rather than merely accept finished certificates. If AlphaProof Nexus is finding patterns the community can understand, this is not just a demonstration of AI compute. It is an early version of a different infrastructure for mathematical research.
