Frank Calegari on Terry Tao's blog: OpenAI's 700-paper math release is a watershed, and mathematicians should explore rather than mourn
- OpenAI released over 700 research papers on October 6, which guest author Frank Calegari calls a watershed in the history of mathematics; the papers carry strong ideas but are badly written, break community attribution norms, and some are outright wrong.
- Hugo Duminil-Copin's reaction, quoted in the post: not a single major open problem he mentioned publicly across his career, in talks, lectures, articles, or grant proposals, was left untouched by the announcement.
- Calegari's historical precedent is the classification of finite simple groups: Gorenstein announced the program in 1972, declared the classification done in 1983 (the quasithin case took Aschbacher and Smith two more decades), and the completion cooled interest in finite group theory even though questions like Conway's why the Monster group exists remain.
- Calegari argues mathematicians should spend effort on what they control, hiring, publishing, and training students, and should reward work that opens new questions instead of pronouncing a subject done.
- The essay is a guest post on Terence Tao's blog dated November 1, 2026, and was converted from a different file format using AI.
Hacker News 의견들
Solving one problem just creates new ones, that's not the point. The point is AI will eventually advance on any problem better than any human can.
We don't actually know that yet. A problem that needs a whole new paradigm might be exactly where AI stalls, and nobody has shown an AI can invent one.
Better in what sense? Those 700 solutions aren't accepted until humans sign off on them. It's like AlphaStar beating pros at SC2, the pros just kept playing.
How does the AI even know which problems are worth working on? Traditionally you figure that out by grinding through the previous ones.
The real value is the machinery, the human-comprehensible framework a proof builds and the further questions it suggests. A proof no human understands doesn't do that.
Any AI produce a new interesting conjecture yet? That's the part I want to see.
I'll take a working mathematician's read on his own field over a commenter calling it pants on head stupid.
I didn't find it well argued either, the classification story almost cuts the other way. But you misread him: with a theorem prover you use Langlands to generate new conjectures instead of speculating for 50 years.
It's the same ten people on Tao's blog citing each other. The guest author even runs a UChicago seminar about AI in math.
It's a seminar, not a cult. Math research is shifting from a craft to a technology, CS went through this decades ago.
Technology extended the craft, sure, but building applications and communicating results is still the job.
The industrialization analogy is imperfect because there was no profit motive inside math pushing this. The push is entirely exogenous, from companies that see proofs as a means to something else.
People aren't dreaming big enough. It's a car engine and everyone's mourning the horses instead of noticing it also lets us fly.
The fast road robs you of the slow one, the random detours and unrelated things you notice along the way. I'd trade a little progress for that.