# @voooooogel — 2024-05-24

♥3 ↻0 · https://x.com/voooooogel/status/1794150922345746606

@NickADobos @karan4d theoretically yes, assuming such a feature exists—the SAE extracts *every* feature in the model. e.g. here's all the features discovered by an SAE trained on gpt-2-sm (from Neuronpedia). theoretically you could clamp 773 and get gpt-2-sm to only talk about art, for example https://t.co/f6s78rtpAi

tags: author:voooooogel, kind:tweet, on:gpt-2, year:2024
cited on: _dossiers/gpt-2.md, gpt-2