danwil

Karma: 71

Most Current Model Organisms Leak: Perplexity Differencing Often Reveals Finetuning Objectives

Luca Baroni, Mo Abu Baker and danwil

1 Jul 2026 10:07 UTC

26 points

0 comments7 min readLW link

Widening AI Safety’s talent pipeline by meeting people where they are

Ruben Castaing, yanni kyriacos, Nelson Gardner-Challis and danwil

25 Sep 2025 20:50 UTC

34 points

3 comments8 min readLW link

Tokenized SAEs: Infusing per-token biases.

tdooms and danwil

4 Aug 2024 9:17 UTC

20 points

20 comments15 min readLW link