James Sullivan

What Sen­tences Cause Align­ment Fak­ing?

Are we al­ign­ing the model or just its mask?

• James Sullivan 21 Mar 2026 17:00 UTC

  1 point

  0

  in reply to: jdp’s comment on: Ter­rified Com­ments on Cor­rigi­bil­ity in Claude’s Constitution

  And the trick is that there is no trick. The way “general intelligence” works is that you are a narrow intelligence with limited out-of-distribution generalization, and this is obscured from you by the fact that while you are asleep, your brain is rearranging itself to try to meet whatever challenges it thinks you’re going to face the next day.

  Would we really say that a human is a “narrow intelligence” when trying any new task until they sleep on it? I think the only thing that would meet the definition of “general intelligence” that this implies is something that generalize to all situations, no matter how foreign. By that definition, I’m not sure if general intelligence is possible.

Play­ing Dumb: De­tect­ing Sand­bag­ging in Fron­tier LLMs via Con­sis­tency Checks

Jailbreak­ing Claude 4 and Other Fron­tier Lan­guage Models

How do AI agents work to­gether when they can’t trust each other?

• James Sullivan 14 Apr 2025 16:35 UTC

  2 points

  0

  on: MATS Spring 2024 Ex­ten­sion Retrospective

  Of the people that wanted to go to a frontier lab, how many had a mentor that worked at a frontier lab? I assume that would make finding a role easier.

Devel­op­men­tal Stages in Multi-Prob­lem Grokking