what's really surprising to me about the new anthropic paper is they have all these fancy interpretability tools, yet still just sample multiple times and say "this shit happens 12% of the time"…

Posters: voooooogel. voooooogel

4 likes recorded by the source corpus.