Meg

Karma: 579

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training

evhub, Carson Denison, Meg, Monte M, David Duvenaud, Nicholas Schiefer and Ethan Perez

12 Jan 2024 19:51 UTC

288 points

94 comments3 min readLW link

(arxiv.org)

Steering Llama-2 with contrastive activation additions

Nina Rimsky, Wuschel Schulz, NickGabs, Meg, evhub and TurnTrout

2 Jan 2024 0:47 UTC

119 points

29 comments8 min readLW link

(arxiv.org)

Towards Understanding Sycophancy in Language Models

Ethan Perez, mrinank_sharma, Meg and Tomek Korbak

24 Oct 2023 0:30 UTC

66 points

0 comments2 min readLW link

(arxiv.org)

Paper: LLMs trained on “A is B” fail to learn “B is A”

lberglund, Owain_Evans, Meg, Maximilian Kaufmann, Mikita Balesni, Asa Cooper Stickland and Tomek Korbak

23 Sep 2023 19:55 UTC

120 points

73 comments4 min readLW link

(arxiv.org)

Paper: On measuring situational awareness in LLMs

Owain_Evans, Daniel Kokotajlo, Mikita Balesni, Tomek Korbak, lberglund, Asa Cooper Stickland, Meg and Maximilian Kaufmann

4 Sep 2023 12:54 UTC

106 points

16 comments5 min readLW link

(arxiv.org)