UT lecturer: intelligence is RL — supervised and self-supervised are just shortcuts
KyleMorgenstein · x · 2026-11-18
Kyle Morgenstein, an RL lecturer at UT, posted a strong take: "when we talk about intelligence, we're talking about RL. all other learning paradigms are shortcuts and hacks."
In the follow-up thread he argues:
- "Supervised vs self-supervised" is a false dichotomy — you always train against some reference, even if it's a latent property of the dataset (e.g. contrastive methods)
- Supervised/self-supervised learning is fine for bootstrapping priors, but doing something interesting with them eventually means RL
- "This isn't some big secret" — it's slide one of his UT RL lecture notes: supervised learning ⇔ function approximation, RL ⇔ function discovery. If you can write down a (proxy) objective, you can optimize it; intelligence is about function discovery
Attached screenshots show his lecture slides formalizing the comparison and the Agent-Environment loop.
Related event: UT Lecturer Argues Intelligence Is RL, Other Paradigms Are Shortcuts(2 posts)→
More from AGI Musings
- Professor: non-native students write better with Grammarly than with ChatGPT or Claude — arvindsatya1 · 2026-11-18
- signulll: attention converts to money easily, but money no longer buys clout — signulll · 2026-11-18
- "Supervised vs self-supervised" is a false dichotomy, argues UT RL lecturer — KyleMorgenstein · 2026-11-18
- 'Selling fire insurance via arson': snarky take on AI labs' business model — ivan_bezdomny · 2026-11-18
- Viral Take: Half the Population Fried by Short Videos, the Other Half Outsourcing Thinking to LLMs — yungcontent · 2026-11-18
- AI Safety and Governance Work Often Pays Barely Above Subsistence, Insiders Say — JacquesThibs · 2026-11-18