SAEs don't just mean interpretable text models—they mean, if you can train a model on *any generating process*, you can blow it up into interpretable features, then use those features to query and…

Posters: voooooogel. voooooogel

4 likes recorded by the source corpus.

Topics: AI & LLMs

View Post: https://x.com/voooooogel/status/1853360436483412162