Sentdex: a raw multimodal LLM drives a robot with zero training, VLA era may be over
Sentdex · x · 2026-10-09
Sentdex demos a robot running on a raw multimodal LLM (DSV4.1F) with just a prompt — no training, no sim2real, unseen cameras, motors and form factor. Currently 10x slower than shown but he argues it's the slowest this intelligence level will ever be. His take: ACT/VLA-style task training is a sideshow that general-purpose models will eat, though RL still rules low-level skills like gait, and decision models like Clef and Jev remain useless for robotics.
Related event: Sentdex demos zero-training robot control via multimodal LLM prompts(2 posts)→
More from AGI Musings
- Ethan Mollick Revisits His 2005 Hacking History Paper to Draw a Parallel With AI — emollick · 2026-10-09
- Thought Experiment: Would Serial-Processing Aliens Change Our Consciousness Judgments? — eschwitz · 2026-10-09
- EA Critic's Challenge: Show a Scarce-Resource Allocation That Avoids Monstrous Conclusions — livgorton · 2026-10-09
- Hadfield-Menell: taking contested values seriously is a natural brake for alignment — dhadfieldmenell · 2026-10-09
- Foresight Institute Turns 40: Drexler and Peterson's Early Bet on Transformative Tech — allisondman · 2026-10-09
- EA Debate: Any Framework for Allocating Scarce Resources Yields Monstrous Cases — livgorton · 2026-10-09