Anthropic's Joshua Saxe: deep learning's core science questions are being abandoned

joshua_saxe · x · 2026-09-04

Joshua Saxe (Anthropic) argues that as scaling became an industrial operation and labs closed up, prominent researchers stopped publicly investigating fundamental questions: Is the lottery ticket hypothesis true? Why does test-time scaling work without explicit optimization? What happens in the double descent phase shift? Why does a little SFT generalize so well? He also calls many lab mech interp papers non-reproducible with a 'recruitment brand feel,' and urges academia to fill this gap.

Original post →

More from AGI Musings

AGI Musings channel →