In-context learning hints at a "GPT moment" for robotics
chris_j_paxton · x · 2026-09-04
- Chris Paxton argues robots won't take off like ChatGPT did if they require post-training on every user; instead they need in-context learning — teaching new skills on the fly.
- The paradigm: prompt with a demonstration mixing video and proprioception; language alone doesn't specify the task well enough. Companies like Generalist, Skild, Rhoda, and RobbyAnt are already showing in-context long-horizon video demos, and it's starting to really work.
- Deepak Pathak amplified the point, tying it to S1 and the case for this new paradigm.
More from AGI Musings
- At 300,000 human births a day, robots matching birth-rate output could start replacing us — JosephJacks_ · 2026-09-04
- Scoble interview: Tesla autonomy, humanoid robots, SpaceX Starbase, Europe falling behind — Scobleizer · 2026-09-04
- Worklife VC: learning AI feels like Twitch, not textbooks — and it pays — briannekimmel · 2026-09-04
- David Krueger: Dismantle the AI Compute Supply Chain for an Indefinite International Pause — DavidSKrueger · 2026-09-04
- "Coordination overhang": AI agents as technology to close the group collaboration gap — anselm · 2026-09-04
- As Models Improve, Your Baseline Quietly Rises With Them — pvncher · 2026-09-04