Yoav Goldberg: Editing model weights is mechanical once we can read them

yoavgo · x · 2026-11-03

Responding to a question about whether editing weights is more than changing numbers, AI researcher Yoav Goldberg argues the difficulty lies in reading, not writing. Weights aren't "just an array of floats" — they're structured into matrices and vectors with known computations, and once humans can truly interpret them, editing becomes mechanical enough to be a course assignment (he notes such questions have already been used in exams).

Related event: Yoav Goldberg: Reading Weights Is More Impressive Than Writing Them From Scratch(3 posts)→

Original post →

More from AGI Musings

AGI Musings channel →