Nate Thomas

• Nate Thomas 26 Oct 2023 14:50 UTC

  LW: 2 AF: 1

  0

  AF

  in reply to: Neel Nanda’s comment on: Ap­ply to the Con­stel­la­tion Visit­ing Re­searcher Pro­gram and As­tra Fel­low­ship, in Berkeley this Winter

  Thanks, Neel! It should be fixed now.

Ap­ply to the Con­stel­la­tion Visit­ing Re­searcher Pro­gram and As­tra Fel­low­ship, in Berkeley this Winter

Causal scrub­bing: re­sults on in­duc­tion heads

Causal scrub­bing: re­sults on a paren bal­ance checker

Causal scrub­bing: Appendix

Causal Scrub­bing: a method for rigor­ously test­ing in­ter­pretabil­ity hy­pothe­ses [Red­wood Re­search]

Ap­ply to the Red­wood Re­search Mechanis­tic In­ter­pretabil­ity Ex­per­i­ment (REMIX), a re­search pro­gram in Berkeley

• Nate Thomas 17 Sep 2022 15:04 UTC

  LW: 18 AF: 9

  12

  AF

  in reply to: Quintin Pope’s comment on: Take­aways from our ro­bust in­jury clas­sifier pro­ject [Red­wood Re­search]

  Note that it’s unsurprising that a different model categorizes this correctly because the failure was generated from an attack on the particular model we were working with. The relevant question is “given a model, how easy is it to find a failure by attacking that model using our rewriting tools?”

High-stakes al­ign­ment via ad­ver­sar­ial train­ing [Red­wood Re­search re­port]

We’re Red­wood Re­search, we do ap­plied al­ign­ment re­search, AMA