Vince Sitzmann: interaction-learned small models may hit the Hz needed for dexterous robot control

vincesitzmann · x · 2026-09-08

Responding to latency concerns, Vince Sitzmann argues LLMs and video models will become much faster — the tech is barely a decade old. Large models may not reach the hundreds of Hz needed for dexterous dynamic control, but smaller models actively learned via real-world interaction just might. He also speculates OpenAI may have added lots of robot and computer data to training, so some behavior may not be emergent.

Original post →

More from Models

Models channel →