Frontier LLMs solved robot in-context learning as a side effect, says researcher
YouJiacheng · x · 2026-11-22
YouJiacheng reshares a robot in-context learning demo: in August it was hailed as the first time ICL worked for robots, but three months later frontier LLMs solved it as a side effect of general intelligence, without any robotics training. He argues instruction understanding, planning, ICL, memory and long-horizon skills should all live in LLMs rather than dedicated VLA/world models, calling it the fastest he's seen a research direction age. Also references OpenAI shutting down its robotics team as general models proved sufficient.
More from AGI Musings
- Tao expected the Riemann Hypothesis to fall by 2050 — it took AI only 35 days — SydSteyerhart · 2026-11-22
- Owner of first-gen Tesla GPU argues local inference is increasingly just cope — generativist · 2026-11-22
- After AI cracked Riemann and Fermat in Lean, Reddit asks: Turing Award for Lean creator de Moura? — coconautico · 2026-11-22
- 'Anyone who thinks there is moat in source code are delusional' — krishnan · 2026-11-22
- tszzl: AI lab executives' most aggressive progress predictions were too conservative — tszzl · 2026-11-22
- Nietzschean roast of Dario: "sick, priestly, duty-driven" will to power — teortaxesTex · 2026-11-22