Google DeepMind's AlphaProof Nexus solved dozens of open math problems

Google DeepMind's AlphaProof Nexus has solved dozens of open mathematical problems — including nine Erdős problems, two of which had gone unsolved for 56 years — according to a peer-reviewed study published in the journal Science.

The peer-review detail is the story. DeepMind submitted its machine-generated proofs to formal scientific review before publication, a step its main competitor skipped. The publication lands at a moment when OpenAI is facing a boycott from parts of the mathematics community after releasing AI-generated proofs on GitHub without any peer review.

That contrast — one lab publishing through Science, another posting unreviewed proofs to a code repository — marks the current fault line in how AI-generated mathematics enters the scientific record.

What did AlphaProof Nexus actually solve?

According to the study in Science, the system solved dozens of open problems in mathematics. The headline results are the nine Erdős problems — questions associated with the prolific twentieth-century mathematician Paul Erdős, whose conjectures have long served as benchmark challenges across combinatorics and number theory.

Two of those problems had remained open for 56 years before AlphaProof Nexus closed them out.

Erdős problems occupy a special status in mathematics. They are typically easy to state and notoriously hard to prove, and mathematicians have measured progress in the field for decades by which of them fall. A single new solution would normally draw attention. Nine at once, produced by a machine and validated by human referees, is the kind of result that forces a reassessment of where automated theorem-proving stands.

Why does peer review matter here?

Because the alternative is already playing out. OpenAI is currently facing a boycott after dumping AI-generated proofs onto GitHub without submitting them to peer review.

The boycott reflects a genuine methodological dispute, not professional pique. A proof's value lies entirely in its verification. A claim posted to GitHub, however plausible, has not been checked by independent experts in any formal sense. When the author is a neural system whose reasoning process cannot be directly inspected, the verification question becomes sharper, not weaker.

DeepMind took the opposite path. As The Decoder reports, the company "had its results reviewed before publishing." The study in Science means referees examined the proofs and judged them sound enough for one of the world's most selective scientific journals.

For the research community, that distinction sets the precedent. If AI labs want credit for mathematical discovery, the AlphaProof Nexus publication suggests the route runs through the existing institutions of science rather than around them.

What does this mean for AI in research?

The stakes extend past mathematics. Automated proof systems sit at the intersection of two of AI's most consequential research frontiers: reasoning and verifiability.

Mathematics is the one domain where an AI system's output can, in principle, be checked with certainty. A proof either holds or it does not. That makes it the cleanest available test of whether large reasoning models can produce genuinely new knowledge rather than plausible-sounding text — a question that hangs over every claim about AI scientific discovery.

Solving nine Erdős problems, two open since the late 1960s, is among the strongest signals yet that these systems can cross from recombining known results into extending the frontier itself. The Science publication converts that signal from an announcement into established findings.

The competitive context sharpens the picture. Two leading labs pursued AI-generated mathematics and arrived at two different publication ethics. One is being boycotted. The other is now in the peer-reviewed record. Researchers deciding whose tools to trust — and institutions deciding whose output to count — have a concrete comparison to point to.

How did the community react to OpenAI's GitHub release?

With a boycott, per the report. The mathematics community's response to OpenAI's unreviewed proof dump signals that volume is not a substitute for verification.

The episode also clarifies what mathematicians are actually being asked to accept. Reviewing machine-generated proofs is labor. Asking referees to do that work after the fact, on material already publicized, inverts the scientific process. DeepMind's sequencing — solve, review, then publish — preserves the referee's role as gatekeeper rather than notary.

What comes next?

The AlphaProof Nexus study establishes that machine proofs can clear the highest bar of formal scientific validation, on problems that resisted human mathematicians for more than half a century. The open question is whether peer-reviewed publication becomes the standard all labs must meet — or whether the market splits between results that carry institutional validation and results that do not. Right now, only one of the two leading efforts has that validation.