Paul califax argues extrapolation isn't special: LLMs can learn out-of-distribution skills via high-dimensional interpolation
paul_cal · x · 2026-10-16
Researcher paul califax pushes back on the common "LLMs only interpolate, never extrapolate" critique: extrapolation isn't special, he argues — it's just one of many strategies a model can learn via high-dimensional interpolation. He claims there are now enough empirical examples showing that LLM-shaped objects can learn to do things that break out of their training distribution.
The post stakes out a clear position in the generalization-capability debate, though it cites no specific experiments or papers.
More from AGI Musings
- Rob Miles mocks AI pitches: 'commits multiple felonies' isn't a selling point — Puzzleheaded-King584 · 2026-10-16
- An agent surfaced a zero-like YouTube podcast dissecting AI disruption through 250 years of industrial history — oran_ge · 2026-10-16
- WeChat mini programs plus agents could be the killer combo for personal agents — sujingshen · 2026-10-16
- Blogger: AI Will Force a Major Pivot in the Education System — ThePeterMick · 2026-10-16
- Ex-Amazon Engineer: AI Has Killed Bezos's Six-Page Memo Culture — bibryam · 2026-10-16
- Why Personal Agents Struggle in China: No Access to WeChat, Alipay and Daily App Ecosystems — sujingshen · 2026-10-16