Stanford's After42: experts rate 45 of 135 AI-proposed research questions worth pursuing
aigclink · x · 2026-10-29
After42: Teaching AI to ask questions, not just solve them
Stanford's After42 project asks whether frontier AI models can propose research questions worth scientists' time. With AI already solving problems like Navier-Stokes and Erdős problems — and Tao warning that good open problems are being "mined non-renewably" — question selection becomes the new bottleneck.
Method: discover questions by attempting to solve them
- No brute brainstorming: starting from a broad goal set by researchers, an agentic workflow reads literature, batch-generates candidate questions, and iteratively attempts to solve them, refining questions through the process.
- Covers math, statistics, theoretical CS, and physics.
Results
- Of 135 final questions, 29 experts judged 45 interesting and worth pursuing, with wide subfield variation: math 33/70, black-hole ringdown 3/8, online algorithms 4/12, statistics 5/35, plasma physics 0/10.
- AI judge shortlists align with expert consensus but not individual experts' tastes, showing how much research taste varies.
- The expert-assessed questions are released as an open resource for the community.
More from AGI Musings
- Mathematicians face an AI existential crisis — art history has the answers — jonippolito · 2026-10-29
- Dev says agents wrote 90% of his 5,000 commits this year — McDonaghMatthew · 2026-10-29
- The singularity's week: 80-year-old cells reset, 25-year-old math disproved by AI labs — rand_longevity · 2026-10-29
- Father praises 11-year-old son's drawing talent, worries about his future in the AI era — CtrlAltDwayne · 2026-10-29
- Gary Marcus echoes PIMCO's AI bubble warning: 'writing on the wall' — GaryMarcus · 2026-10-29
- Manufacturing AI will spawn the next generational companies, says 5-layer startup map — McDonaghMatthew · 2026-10-29