Prediction Is Not Perception: Why Video Models Fall Short for Embodied AI

PierceLilholt · x · 2026-07-30

Pierce Lilholt emphasizes that 'prediction' and 'perception' are fundamentally distinct concepts in AI applications, a distinction that will prove more critical than currently anticipated.

In the referenced discussion, @taiyasaki adds that relying on a generic video model as the core perception system for embodied AI is insufficient. While current approaches might seem viable, they require additional, specialized modules and mechanisms ('bells and whistles') to function properly. However, he notes that over time, these advanced perception capabilities will eventually be taken for granted.

Original post →

More from Embodied

Embodied channel →