Thanks, Chris. Auto-regressive prediction has failed for video in the past but not obviously for language. @ylecun has emphasised this drift, but I wonder how much (1) redundancy and (2) discreteness can provide self-correction.

→ View original post on X — @nandodf