> EDIT: This last part isn't true. I think they are only looking at the intermed... | Hacker News

Hacker Newsnew | past | comments | ask | show | jobs | submit

		whimsicalism on May 9, 2023 \| parent \| context \| favorite \| on: Language models can explain neurons in language mo... > EDIT: This last part isn't true. I think they are only looking at the intermediate layer of the FFN which does have a privileged basis. it does?

sebzim4500 on May 9, 2023 [–]

Yeah, that's where they apply the activation function and that happens per neuron so you can't rotate everything and expect the same result.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact