AlphaProof Nexus Solves Nine Erdős Problems, Passes Peer Review

What happened

Google DeepMind's AlphaProof Nexus has solved dozens of previously open mathematical problems, including nine long-standing Erdős problems, two of which had resisted proof for 56 years, according to a peer-reviewed study published in the journal Science. The publication is notable not just for the mathematical results but for the process behind them: DeepMind submitted its findings for independent peer review before releasing them.

The announcement lands at a pointed moment. OpenAI is currently facing criticism and a boycott from parts of the mathematics community after it published a batch of AI-generated proofs directly to GitHub without subjecting them to peer review. DeepMind's approach — verification first, publication second — stands in contrast to that episode and has become part of the story itself, as reported by The Decoder.

Why it matters

The substance of the result — an AI system contributing genuine progress on problems that had stumped mathematicians for decades — is a meaningful marker for how far machine reasoning has advanced on open-ended, creative problem-solving rather than narrow pattern-matching. That matters for any organisation weighing where AI can realistically take on complex analytical work previously thought to require specialist human expertise.

But the more durable lesson for leaders in AI and digital transformation may be procedural rather than mathematical. As AI systems increasingly generate claims, proofs, code or decisions that others must trust, the credibility of the verification process behind those outputs becomes as important as the outputs themselves. DeepMind's decision to route its results through peer review — while a rival faced backlash for skipping that step — is a live case study in how trust is built, or eroded, when AI-generated work enters domains governed by established standards of rigor.

By the numbers

• Dozens of open mathematical problems solved by AlphaProof Nexus, per the Science study.
• Nine Erdős problems were among those resolved.
• 56 years was the length of time two of those Erdős problems had remained unsolved.

The Renascence take

Most coverage of this story will fixate on the headline achievement — AI cracking decades-old maths problems. The more instructive detail, for anyone designing AI-enabled services or experiences, is the choice DeepMind made about how to earn belief in its claims.

Capability is no longer the scarce resource in AI — credibility is. The gap between DeepMind's peer-reviewed release and OpenAI's unreviewed GitHub drop is really a gap in how each organisation thinks about the person on the receiving end of an AI claim: do they have to simply take it on faith, or has the hard work of earning trust already been done for them? That is a service-design question as much as a scientific one. Any operator deploying AI into customer- or employee-facing decisions should treat verification, transparency and independent scrutiny not as a bureaucratic afterthought but as the actual product — because in a market this noisy, the organisations that bother to prove their claims are the ones whose claims people will eventually choose to believe.