J Rosser

My Neel Nanda MATS 10.0 Ap­pli­ca­tion: Study­ing Fea­ture Split­ting in SAEs via Train­ing Data Attribution

Data fil­ter­ing works a lot worse than you would ex­pect

In­fu­sion: Re­v­erse en­g­ineer­ing In­fluence Func­tions to craft train­ing doc­u­ments...

• J Rosser 8 Feb 2026 13:37 UTC

  1 point

  1

  in reply to: joseph_c’s comment on: Speedrun­ning a Mech In­terp Re­search Setup (Re­mote GPU, Torch, Trans­formerLens, Cuda, SSH, VS Code)

  +1 to this!
  
  Maybe a neat solution might be to instead use a deploy key (gives read/​write access just to one repo) instead of adding a full ssh key?
  
  https://​​docs.github.com/​​en/​​authentication/​​connecting-to-github-with-ssh/​​managing-deploy-keys

Speedrun­ning a Mech In­terp Re­search Setup (Re­mote GPU, Torch, Trans­formerLens, Cuda, SSH, VS Code)

Trans­mit­ting Misal­ign­ment with Sublimi­nal Learn­ing via Paraphrasing

The Case for White Box Control