On September 8, OpenAI announced that a swarm of roughly 10,000 coordinated AI agents had produced a proof — checked in Lean, a system that verifies proofs mechanically — of a long-open Millennium Prize problem about the equations that describe how fluids move. Later the same day, the mathematician Terence Tao posted a four-part thread that reframes the story: the real question isn’t whether AI can solve hard problems, but what mass “solution extraction” does to the field that grows good problems in the first place.

His central claim: good open problems are being mined in a non-renewable fashion.

  • The set of possible math questions is infinite, but almost all are worthless — working out the 10^10^10th digit of pi teaches you nothing. A problem is only good if solving it promises new insights or connections.
  • Experienced mathematicians navigate by a “difficulty landscape”: a felt sense of which questions are easy, which are hard but reachable, and which are walls. That map is what makes problem-finding a craft.
  • Every advance flattens part of the landscape. AI flattens it without revealing where the new frontier is — and labs that refuse to publish negative results or the path to their solutions make the map unreadable from the outside.
  • The scarce resource has shifted from solutions to the problems themselves: even a rumor that someone is working on a question can now trigger a massive AI effort to flatten it first.
  • His proposed norm: treat raw solutions without analysis as near-worthless for designated problems — demand answers that also extract insight about why they work and what they imply nearby, like a food drive that refuses contributions that are technically edible but useless.

Tao is close to the best-positioned person alive to make this argument — problem-finding is arguably his signature skill, and he is describing the depletion of his own raw material. The essay is worth reading as the sharpest statement yet of an idea that applies well beyond math: when answers become cheap, the bottleneck moves to asking good questions and actually understanding what the answers mean.

It also raises a harder, more political point. If even the rumor of a research direction invites an automated land-grab, the rational move for human researchers is to stop sharing promising ideas — reversing centuries of open-science norms. The same incentive structure is playing out in code, writing, and every other field where frontier models now operate, and Tao’s warning is that the casualty is not individual problems but the commons that produces them.