Breaking the Supervised Learning Curse: Models Must Learn 'In the Wild'

tensorqt · x · 2026-08-09

Responding to John Schulman's point about models struggling with controversial topics from first principles, the commentary argues this is the inherent curse of supervised (including self-supervised) training. While highly effective, this paradigm is information-bound by its pre-text tasks. To break this barrier, models must learn 'in the wild', accepting all the implied risks that come with it.

Original post →

More from AGI Musings

AGI Musings channel →