factual

Language Models May Generalize Beyond Training Data

Despite worries that language models cannot extrapolate outside their training set the way humans extrapolate via common sense, examples like a model suggesting sawing a table in half to fit it through a doorway show the model learned real-world facts and can reason in non-obvious, 'outside the box' ways, suggesting it may generalize better than skeptics expect.

factualpending

Speaker

Tai-Danae Bradley

Evidence Quote

“it's showing you that it learned something about the Earth... maybe there is a lot in there in these kind of billions of parameters”

Source

174 | Tai-Danae Bradley on Algebra, Topology, Language, and Entropy— Sean Carroll's Mindscape
Created: 6/13/2026, 12:11:28 AM

Your notes

Loading notes...