What's this about?
A detailed case study of human–AI collaboration on a real math problem — the Grothendieck constant — documenting how to actually use AI agents effectively in long-horizon research mathematics. It's a natural companion to this week's Claude/Riemann result.
What exactly is "long-horizon" research?
Most AI tasks are short: ask, answer, done. Long-horizon research means a problem that takes many steps over a long time — exploring dead ends, building on partial results, backtracking. The paper studies how to steer AI through that, since agents often flail without the right structure and human guidance.
Why it matters
The frontier question isn't "can AI do math" but "*how* do you work with it to get real results." Concrete playbooks for human–AI research collaboration are what turn flashy one-off proofs into a repeatable method.