AI ALIGNMENT FORUM
AF

Connor Kissane
Ω202700
Message
Dialogue
Subscribe

Posts

Sorted by New

Wikitag Contributions

Comments

Sorted by
Newest
No Comments Found
No wikitag contributions to display.
39White Box Control at UK AISI - Update on Sandbagging Investigations
10d
5
34SAEs are highly dataset dependent: a case study on the refusal direction
8mo
0
23Open Source Replication of Anthropic’s Crosscoder paper for model-diffing
9mo
3
35Base LLMs refuse too
10mo
10
28SAEs (usually) Transfer Between Base and Chat Models
1y
0
18Attention Output SAEs Improve Circuit Analysis
1y
0
34We Inspected Every Head In GPT-2 Small using SAEs So You Don’t Have To
1y
0
28Attention SAEs Scale to GPT-2 Small
1y
0
35Sparse Autoencoders Work on Attention Layer Outputs
2y
3