Can open-ended intelligence growth be reduced to closed-loop optimization?
jzl86 · x · 2026-11-15
In a discussion responding to Caleb Watney, sebkrier frames the crux: whether open-ended intelligence growth reduces to closed-loop optimization over an internal learning metric, or whether discovering better algorithms and real-world competence still depends on the 'rubber hitting the road' — an environment acting as a rate limiter. Abstracting to 'pretraining loss + in-context learning is all you need' treats general intelligence as a compact mathematical learning rule waiting in a sandbox, which he can't disprove but finds unconvincing.
More from AGI Musings
- Runway CEO cites 1888 Kodak backlash as analogy for today's AI art haters — c_valenzuelab · 2026-11-15
- Ben Reinhardt: people fear sharing positive visions of the future — Ben_Reinhardt · 2026-11-15
- Survey: Patients More Open to AI in Healthcare Than Their Doctors — pswider · 2026-11-15
- Lawyer: SOTA LLMs are no better at novel legal arguments than in 2022 — GaryMarcus · 2026-11-15
- AI Ethicist Daniel Faggella Clarifies His Cause: Life, Potentia, and AGI's Moral Weight — danfaggella · 2026-11-15
- Ben Todd: spark vs software-only singularity is really about which feedback loops run how far — ben_j_todd · 2026-11-15