X · @teortaxesTex
· X / Twitter
RT Goodfire: Re This is worth being careful about! Naive setups can degrade interp tools’ validity, but these probes are run on a *frozen copy* of th…
RT GoodfireRe This is worth being careful about!Naive setups can degrade interp tools’ validity, but these probes are run on a *frozen copy* of the model during training. Empirically, they don’t degrade, even on the trained model.You can read our detailed thoughts on this here: https://www.goodfire.ai/blog/intentional-