Adam Karvonen @a_karvonen · 21 Aug 2026
We test three activation-based tools that succeeded in prior auditing games: activation oracles, natural language autoencoders, and SAEs. None beat just reading the transcript. The result is robust to hyperparameters, prompting, and a Fable goal loop attempting to improve it. https://t.co/4VfUpONRCr
5 001Views
33Likes
2Reposts
2Replies
4Quotes
6Bookmarks
Is that a lot?
4.09×vs this author's median1 223 views is typical
73Percentile for this authorof 11 recent posts
0.80×vs under 10K median6 244 views is typical
115.63%Reachviews ÷ followers
0.82%Engagement rateof viewers reacted