Interpretability researcher: even a 432-parameter model resists full understanding
JacobHHilton · x · 2026-10-17
Jacob Hilton argues against optimistic takes on fully understanding LLMs. After analyzing a 432-parameter model, he found interpretable circuit structure is "absolutely everywhere" hard to decode in LLMs — we just haven't discovered enough interpretable structure to think otherwise.
He expects critics to produce cleaner circuits from big models on toy tasks like two-digit addition, but says that won't be encouraging, since big models can do far more. Hope from image classifiers is misplaced: their better trend came from larger, more diverse datasets (ImageNet vs MNIST), whereas language data is already rich — scaling just draws more out of it.
More from AGI Musings
- AI gives you a whole team of researchers, and some swear by it — Pascallisch · 2026-10-17
- Musk: merging with machines is the 'only escape route' from AI elimination — Puzzleheaded-King584 · 2026-10-17
- Reddit user argues "rogue AI" is a psyop giving owners plausible deniability — Zappy_Oh · 2026-10-17
- Malware author slams 'AI worms' hype as weaponized misinformation — dyn___ · 2026-10-17
- Harvard Magazine spotlights 'First Encounters with AI,' writers' essays on generative text — begusgasper · 2026-10-17
- Viral meme mocks EA doomers: why not stage a false-flag AI hack to prove the risk? — NathanpmYoung · 2026-10-17