In-Context Learning Hints at a 'GPT Moment' for Robotics: Demos as the New Interface
chris_j_paxton · x · 2026-09-04
Chris Paxton argues that for general-purpose robots to be useful and economical, they must learn new skills on the fly — and in-context learning is now making that viable.
Key points:
- The shift is from programming robots to teaching them: new skills come from a demonstration, not retraining.
- The prompt is a long-horizon demonstration mixing video and proprioception; language alone isn't precise enough to specify tasks.
- Companies like Generalist, Skild, Rhoda, and RobbyAnt are already showing in-context long-horizon video demos, and it's starting to really work.
Paxton suggests this could be robotics' "GPT moment": demonstrations become the interface for robotics.
Related event: In-Context Learning Could Be Robotics' GPT Moment, Researcher Says(2 posts)→
More from Embodied
- NHTSA opens probe into Tesla's steering-wheel-free Cybercab hours after Austin launch — LexiLove · 2026-09-04
- Shenzhen company integrates humanoid robots with industrial robot arms — LexiLove · 2026-09-04
- Passenger goes from scared to asleep in Tesla Robotaxi within two days, Musk amplifies — elonmusk · 2026-09-04
- Singapore's COVID-era Spot robot patrols: public safety or expensive security theatre? — xuanalogue · 2026-09-04
- Thousands of Reachy Minis Poised to Ship After Team Tours Seeed's Shenzhen Robot Lines — _Stocko_ · 2026-09-04
- NTU Singapore's Seed-Sized Magnetic Robot Switches Between 5 Surgical Functions in Under a Second — Olivier__OG · 2026-09-04