Great post on using measuring logit distribution for eval awareness. Surprisingly effective, even for unverbalised eval awareness! ✍️

https://www.lesswrong.com/posts/PK7ZvFZxrgpYtrpF4/logits-as-a-new-monitor-for-evaluation-awareness-1

saw an interesting post for those also trying to transition into technical AI safety :)

https://www.lesswrong.com/posts/xiTBpBDwubnr4MLRe/trees-are-mostly-made-of-air-and-a-generalizable-lesson-for

Greenblatt post trying out open weight NLAs showing the limitations for reconstructing activations

https://www.lesswrong.com/posts/QQQAcKuWK6k98FivY/can-activation-verbalizers-surface-an-internal-chain-of-1