AI now autonomously resolves open math problems, mathematician argues profession must adapt

The claim

An essay published on the blog proofsandprompts.com, currently on the front page of Hacker News, makes a stark claim about the pace of AI progress in mathematics: three years ago AI systems could not reliably add two numbers, a year ago internal models at OpenAI and DeepMind achieved the equivalent of gold-medal scores on the International Mathematical Olympiad, and now such systems are autonomously resolving major open questions. The author, a mathematician, writes that it is hard to imagine the trend continuing for another year but expects it will, and that the profession faces a radical rethinking.

A capability claim and an institutional argument

The capability trajectory is the hook, but the essay's core argument is institutional. The author takes as a premise that AI systems robustly superhuman at most or all aspects of mathematics will arrive soon, while noting that the proposed changes only require accepting a weaker premise: that producing mathematical text is becoming increasingly disconnected from mathematical understanding.

The essay points out that the community has long operationalized its goals through proving theorems, and that this metric was always hollow in isolation. A program could enumerate consequences of the ZFC axioms, or conjecture every proposition in alphabetical order, with no understanding whatsoever. What mathematicians actually value, the author argues, is resolving problems that resisted substantial effort, building theories, asking good questions and generating human understanding — none of which raw proof production captures.

Text no longer certifies expertise

The sharpest practical observation concerns credentialing. It is now possible, the author writes, to produce a PhD thesis one has not even read, which means mathematical text no longer reliably signals anything about the person who produced it. Academic institutions have historically used that single signal to certify two different things: mathematical progress and mathematical expertise. The essay argues those functions must now be separated, and that interesting results should be welcomed regardless of provenance.

What the author proposes

The essay proposes reconceptualizing the mathematics PhD: the goal becomes becoming a world expert on an interesting, deep topic and being able to convey that interest and understanding to others. A thesis might still play a part, but the degree would be awarded primarily on the basis of a rigorous defense in which the student explains the topic.

On institutions more broadly, the author urges protecting values rather than the current shape of the system. Preserving journals, peer review, arXiv moderation or gatekeeping roles is, in the essay's view, absurd once anyone with a laptop and a few hundred dollars can generate what would have been an Annals paper the year before. The author also warns against chasing the moving edge of model capabilities — rewarding only the skills AI still lacks, such as theory-building, asking questions and exposition — because the academy adapts far more slowly than models improve.

Finally, the essay argues any plan must not depend on AI capabilities disappearing. Whatever one thinks of AI labs, the author writes, the underlying issue is the technology itself, not the behavior of the labs.

A positive vision, not an ending

The essay follows a talk the author recently gave titled The End of Mathematics, which laid out a gloomy scenario in which superhuman AI exists but institutional failure causes human understanding of mathematics to stall. The author considers that future plausible if academic mathematics does not adapt, but avoidable. The counter-vision is an enormous expansion of mathematics with more need for human mathematicians than ever, organized around the things the author wants to preserve: learning seminars, serendipitous conversations that spark ideas, students knocking on professors' doors, and large communities collectively working through their confusion.

Why it matters

If the essay's account is accurate, autonomously resolving open research questions is a categorically different milestone from competition-level performance, because open problems are precisely where human effort had stalled. The consequences are immediate and practical: doctoral credentialing, hiring, peer review and mathematical education all rest on the assumption that written mathematics certifies understanding, and that assumption is exactly what the essay says has broken. The piece also reads as a preview for other fields whose credentials and workflows are built on producing documents. One caveat: this is a single author's account, and the IMO results it cites concern internal, unreleased models at OpenAI and DeepMind rather than publicly available systems.