Study: AI agents lack the taste for open-ended research, slowing recursive self-improvement
ArtificialOther · x · 2026-09-14
A new study led by Peter Kirgis and Sayash Kapoor at Princeton, covered by MIT Technology Review, finds AI agents can handle the engineering tasks of AI research — writing code, generating synthetic data, optimizing chips — but fall short at open-ended investigation requiring judgment and taste. Their outputs don't reach the caliber of papers accepted at top ML conferences, suggesting hyped timelines for recursive self-improvement may be running ahead of the evidence.
More from AGI Musings
- LeCun mocks Dario's 2019 GPT-2 danger claim: "everyone should make fun of them now" — robleclerc · 2026-09-14
- Skeptic doubts labs would truly pause without lobbying hard against open-source models — wightmanr · 2026-09-14
- zetalyrae: anti-human AI ideas stem from ideologies over self-interest — zetalyrae · 2026-09-14
- beffjezos: Beliefs about the future are hyperstitious, hope beats fear on AGI — beffjezos · 2026-09-14
- zetalyrae pushes back on anti-human ideas in AI discourse: valuing humans over AI is fine — zetalyrae · 2026-09-14
- Dario and Wenfeng: the only two 'serious' AI leaders, and their mirrored motives — nabeelqu · 2026-09-14