Frontier AI Agents Fail at Open-Ended Research Despite 6 Days of Compute
sayashk · x · 2026-07-31
A new study tests whether frontier AI agents can conduct open-ended AI R&D. The researchers introduced "shadow evaluations," where agents tackle the core open-ended questions of unpublished NeurIPS 2026 papers, given 6 days and thousands of dollars in compute.
Results show that while agents completed all engineering tasks autonomously, they failed to make substantial progress on the research questions, leading to unambiguous rejections by the original authors. The paper identifies five recurring failure modes, most notably poor judgment regarding the bar for publishable research. This provides early evidence against the assumption that AI agents are ready to automate explosive AI progress.
More from AGI Musings
- Can AI trained only on human content infer the shape of the human brain? — rickasaurus · 2026-10-16
- Vals RSI Index extrapolates Google models hitting 130% autonomous AI R&D by Oct 2027 — HaydnBelfield · 2026-10-16
- Gowers' tl;dr: no p(doom), but can't confidently say it's close to zero — michael_nielsen · 2026-10-16
- Fields Medalist Gowers pens long essay: can't confidently say AI extinction risk is near zero — michael_nielsen · 2026-10-16
- Woman arrested for blocking Uber delivery robot in Philadelphia to protest job losses — GarrisonLovely · 2026-10-16
- Revisiting Breiman's Two Cultures: Predictive Models Are Never Clean — beenwrekt · 2026-10-16