Terence Tao: AI is harvesting open math problems unsustainably, and 'Math 2.0' must reward exposition and community
- Terence Tao's four-post thread argues that AI prompters now solve open problems autonomously and then walk away, with no interest in the field once the initial target is solved.
- The fallout he lists: far fewer seminars, workshops, and collaborations than traditional breakthroughs generate, few newcomers joining the community, and promising open directions withheld from the public for fear of being scooped.
- Tao says the process cannot be reversed: a solved problem cannot become unsolved again, and even knowing a solution exists contaminates human and AI attempts to find alternate routes that reveal more insight, so open problems are being harvested at large scale in an unsustainable way.
- His proposed fix, Math 2.0, decenters raw problem solving and elevates exposition, community building, and opening new directions of study, and he says the community must re-evaluate its criteria for education, publication, and career advancement.
- Tao holds that AI can contribute positively to exposition and community building as well, but that takes more imagination than pointing a favorite agent at a set of open problems.
Hacker News opinions
How would you even formalize what makes a problem interesting, so a machine can generate new ones from the existing corpus? That's the part nobody has an answer for.
For me that path just means our job as humans is understanding intelligence and everything else gets automated. Hard pill to swallow.
Or we stop letting pure mathematicians decide what's interesting and reward the practical applications instead.
Simple test: measure how long a problem takes to solve. If it takes more than 5 minutes, we don't understand that area well and it's worth investing in.
Let me know when UCLA stops hiring whoever has the most top-tier journal papers. More likely a PI hires one PhD student instead of two and spends the other $50K on AI.
Frontier labs will just hand out huge EDU discounts if you let them train on your data. And top math programs haven't hired on journal pubs for a while anyway, it's reputation from one or two big results spreading through arXiv and talks.
He's only right if model intelligence stalls. If it keeps improving, the prompter doesn't need to understand anything, the model can walk a mathematician through the proof step by step. Why would it explain new proofs worse than old ones?
If the mathematician doesn't understand anything, are they even a mathematician?
Tao is great at explaining things, but he could spend months explaining one of his proofs to me and I still wouldn't get it. Superhuman explanation doesn't guarantee human understanding.
Related: the AHM statement on OpenAI's October 6 release of mathematical documents.
We should stop listening to the early nay-sayers and just wait a bit. The stochastic parrot crowd went quiet. Tao is the wisdom provider here, not a nay-sayer.
Even after a problem is solved there's always value in simpler proofs and corollaries that build intuition. Sounds like Tao says that becomes the main job now, and I assume models will beat us at it.
There's value but it isn't rewarded anywhere near the original contribution. Same as peer review.
Will they though? Code from these models isn't getting more elegant, the architectures get more baroque. You can RL for correctness, optimizing for taste is much harder.
This is what people in tech have been saying for a year. Swap math for any field and it still reads true.
Not that simple. Three buckets: fields that say the human experience is the whole point, like marathon runners and poets; fields that are split, and math is right there; and fields that chase results and extract as much as they can before it collapses.
Balanced take. Dumping proofs on the community and expecting others to verify and refine them isn't advancing math, it's hitting a benchmark. But it's a maniacal race with too much money in it.
Short sighted. A few years ago models couldn't do this at all, and there's no evidence refining results stays out of reach. Just have another chatbot verify, like we do with code.