Can RL-trained LLMs design DSLs from scratch? keenan revisits the Bitter Lesson for 3D
eatnow240008 · x · 2026-09-19
A substantive thread questioning the Bitter Lesson's applicability to 3D learning: the colloquial "scale always beats problem-specific modeling" (not actually Sutton's claim, the author argues) ignores Kolmogorov complexity, and impressive LLM/agent-generated 3D content often relies on problem-specific structure.
Asked whether LLMs could soon use RL to extend or even design DSLs from scratch, keenanisalive says that's very likely but hard to define — e.g. given images and a renderer, the task of inventing a language for faithful 3D reconstruction is ill-posed, the renderer already encodes priors, and a language nobody writes in doesn't help: the LLM still has to model next-token probabilities in the target language.
Related event: 'Bitter Lesson' challenged in 3D generation debate(4 posts)→
More from AGI Musings
- Andreessen: Agents Are Entering Enterprises as Chaos, Just Like PCs and Spreadsheets Did — Dan_Jeffries1 · 2026-11-04
- 1000+ mathematicians sign pledge to shun AI companies and models — Shesawthat · 2026-11-04
- Von Neumann meme: "You don't have to be responsible for the world that you're in" — akbirthko · 2026-11-04
- Joscha Bach on the dream chocolate that stopped being real — Machine Learning Street Talk · 2026-11-04
- François Chollet: modeling brain wiring as an engineering problem constrained by genome size and time — CatAstro_Piyush · 2026-11-04
- Most US Couples Now Meet in a Data Centre, Doubling Office Matches — menhguin · 2026-11-04