OpenAI Publishes 722 Math Papers Written by AI

On October 6, OpenAI published 722 math papers prepared by AI. They include works with proofs related to the quasi-Riemann hypothesis. For comparison: an entire mathematical subfield produces hundreds of publications a year through the work of thousands of people, while here that volume arrived in one package. The same day, Chinese media, including Jiqizhixin, began talking about RSI, recursive self-improvement: according to their reports, one team of Chinese origin has already completed a scalable closed loop. The available excerpt gives no details about the team itself, so below we separate what is confirmed from what remains a claim.

What is known about the release

The facts are modest but weighty. Date: October 6. Volume: 722 papers. Subject: mathematics, including proofs related to the quasi-Riemann hypothesis. This is not a solution to the Riemann hypothesis itself, and the source does not claim it is. The two results cannot be compared: this is a related, weaker formulation.

The loudest voice came from Christian Szegedy, a former Google scientist and xAI co-founder. In his words, the summits mathematicians spent their lives climbing can be taken by the machine "arriving by plane." He immediately added a caveat: mathematics is not coming to an end; the way research is done is changing. Szegedy has long worked on autoformalization, that is, translating mathematics into a machine-verifiable form, so his assessment comes from a practitioner, not an outside observer.

Where RSI comes in

RSI refers to a loop in which a model improves its own outputs or its own training, and the next iteration starts from a higher point. Mathematics is an almost ideal proving ground for such a loop: a proof has a formal criterion of correctness, and the answer can be checked without a human. If verification is automatic, the generate-and-select loop closes and scales with compute.

This, judging by the headline, is what the Chinese team in question calls "completed." The text does not disclose which models were used, how many GPUs, or what share of false proofs the verification filters out. Without these figures, the claim remains a signpost, not a measured result.

What we know and what we do not:

• Publication date: October 6, 2026.
• Release volume: 722 papers.
• Subject: mathematics, including proofs around the quasi-Riemann hypothesis.
• Independent review of all 722 papers: no data.
• Share of papers verified formally (in Lean or similar tools): not stated.
• Metrics of the Chinese team (models, compute, number of loop iterations): not disclosed.

What to make of this news

First: do not mistake volume for quality. 722 papers prove nothing by themselves until reviewers and formal verifiers work through a sample. A sensible indicator here is the share of papers whose proofs passed machine verification, not the number of titles on a list.

Second: watch who does the verifying. If strong results arrive with formal verification, that signals the "generate, verify, select" loop works in practice. If not, what we have is closer to hypotheses dressed up as papers.

Third: for researchers and agent developers, the takeaway is practical. Domains with cheap automatic verification (formal mathematics, parts of programming, verifiable code) will be the first where RSI-like schemes deliver real gains. Where the criterion of correctness is fuzzy, as in biology, economics and law, such a loop will run into a human bottleneck.

And finally. Szegedy's words about a change in the "way research is done" deserve to be read literally: the mathematician of the future will increasingly formulate problems and evaluate results rather than derive every step by hand. What happens next depends on how openly OpenAI and the Chinese teams show the data behind their claims.