"There is no science of transformers" — researcher slams irreproducible AI research

gerardsans · x · 2026-11-19

In a pointed exchange with Alex G. Dimakis, gerardsans argues there is currently "no science of transformers": AI research is done poorly, with no real agents — just different inference runs. With hidden variables everywhere, systems can't be studied or diagnosed; no null hypothesis means no conclusions are possible.

He adds that running inference and reporting results is a wasteful way to describe a setup, since no reproducible configuration can be isolated — "This is a survey. Not science."

Original post →

More from Research

Research channel →