"Supervised vs self-supervised" is a false dichotomy, argues UT RL lecturer

KyleMorgenstein · x · 2026-11-18

In his "intelligence is RL" thread, UT RL lecturer Kyle Morgenstein pushes back on the common dichotomy: "supervised vs self-supervised" is false. Even when the reference isn't an explicit label, it can be a latent property of the dataset — contrastive methods, for instance, still train against some reference.

This tweet is the opening move of an argument that continues into his claim that supervised/self-supervised learning are just shortcuts for bootstrapping priors, and that real intelligence ultimately comes down to reinforcement learning.

Related event: UT Lecturer Argues Intelligence Is RL, Other Paradigms Are Shortcuts(2 posts)→

Original post →

More from AGI Musings

AGI Musings channel →