Mechinterp drama: critics doubt new Claude interpretability claim, citing same-suffix neuron mapping
inductionheads · x · 2026-10-16
tenobrus says Nick's claimed mechinterp breakthrough is looking bad, suspecting Claude may have been showing him what he wanted to see. inductionheads pushes back: the paper itself says words with the same ending map to the same neurons, so Colorado and Tornado clustering together is unsurprising — a technical spat over an interpretability result.
More from Fun
- Tesla's Cybertruck kickflip video: jet controls right the truck midair like a cat — astralmatrix · 2026-10-16
- As coding agents run longer autonomously, one dev's tip: don't parallel-session, just breathe — teropa · 2026-10-16
- 1998 Shockwave code art classic 'The Square Root of -1' ported to run today — floguo · 2026-10-16
- Flying DeLorean renders go viral: the flying-car dream in Back to the Future style — sergeantsref · 2026-10-16
- MiniMax-generated horror series goes viral: the toy moves every time you blink — fa6637 · 2026-10-16
- Professor shares son's college essay written entirely without LLM help — and it was great — arjunrajlab · 2026-10-16