New framework argues dense shallow generalization may suffice for economically meaningful AGI
gleech · x · 2026-11-18
In a drive-by statement for a new issue (retweeted by gleech), Peligrietzer lays out his view of AI progress:
- Dense, shallow generalization can "explain away" many apparent cases of strong OOD generalization, yet has no obvious practical limits on its own.
- Combined with increasing automation of RL-ifying domains, it may plausibly deliver AIs performant in the economist's sense of AGI — matching top humans at most economically measurable tasks.
- Three major open axes within the LLM transformers+RL paradigm: (1) can RL pipelines achieve extreme density where ground-truth signal is scarce (simulation, powerful critics)? (2) do transfer/meta-learning effects reduce sample complexity as the catalog of mastered domains grows? (3) do real-world distribution shifts limit one-off dense shallow generalization long-term?
He plans to write more on this soon.
More from AGI Musings
- Field experiment: Google AI Overviews cut publisher clicks 39.8% with no UX gain — paulnovosad · 2026-11-18
- Academia Has the Strongest Anti-AI Norms, Yet Grad Students Use AI Anyway — CatAstro_Piyush · 2026-11-18
- "College and the 9-5 path is over" — AI era anxiety goes mainstream — rand_longevity · 2026-11-18
- Models Will Be Commoditized: Burkov Backs Efficient Open-Weight Models, Doubts Overpriced IPOs — marlene_zw · 2026-11-18
- Commentary on Christiano/Dwarkesh: training-pipeline automation is desirable and safety-helpful — sebkrier · 2026-11-18
- "If AI brings intelligence and opportunity to billions, dying industries deserve no sympathy" — VraserX · 2026-11-18