François Fleuret: we know training works, but not why — inductive bias, distillation and more remain opaque

francoisfleuret · x · 2026-09-09

François Fleuret pushed back on claims that we understand all theoretical and operational aspects of training models, listing open questions: inductive bias, overfitting, latent representations, transfer learning, distillation, RL, simplicity bias, and more. We know we have all the bricks, he argues, only because you can talk to the thing at the end — success by output, not by theoretical understanding.

Original post →

More from Research

Research channel →