MIT interpretability paper drama: willdepue calls the hype 'misleading' but backs weight space programming
aryaman2020 · x · 2026-11-07
A spat over the viral MIT interpretability paper: willdepue says it was presented in a foolish, misleading way but is glad it got people excited about weight space programming, which he calls "really cool." aryaman pushes back that the post itself is fine and Twitter hype caused misreadings, and accuses willdepue of being overconfident about model internals, citing his earlier incorrect claims about attention.
More from Fun
- Dev mocks apps that shipped 'garbage for a decade' until LLMs could code — ludwigABAP · 2026-11-07
- Infinite self-evolving Backrooms game built on Claude Opus is 'the coolest game experience' — noahsolomon · 2026-11-07
- Dug-up draft: alignment researchers around Eliezer on a bed of broken GPUs — akbirthko · 2026-11-07
- "Don't tell Sam Altman they sell RAM trucks": AI circle memes the memory crunch — djcows · 2026-11-07
- NYU professor: Claude now mimics my TikZ style from my own code, drafting figures instantly — alfcnz · 2026-11-07
- When AGI arrives, package delivery won't hold us hostage until 9pm — ZeroStateReflex · 2026-11-07