Why the AI Scaling Hypothesis May Never Be Falsified: A Duhem–Quine Argument

burny_tech · x · 2026-09-27

The author argues the AI scaling hypothesis is in principle unfalsifiable: any failed scaling attempt can be blamed on auxiliary hypotheses — insufficient normalization ruining representations, poor data, bad RL rewards or unstable training, architecture tweaks destroying gradients, bad learning-rate schedules, flawed evaluation metrics.
The same logic applies to neurosymbolic AI ("nobody has done it right or scaled it right yet") and possibly parts of mechanistic interpretability — endless tweaking of tools until they appear to work.
He frames this as a special case of the Duhem–Quine thesis: empirical tests always rest on background assumptions, so a hypothesis alone never makes predictions and can never be unambiguously refuted — illustrated by early objections to Earth's motion based on birds not flying off branches.

More from AGI Musings