Human reasoning is not based on auto-regressive discrete symbol (token) prediction.
It is based on the manipulation of mental models in continuous representations spaces.
It is based on *searching* for a set of manipulations of this model to arrive at a particular result.

→ View original post on X — @ylecun,