Gurkenglas comments on Self-fulfilling misalignment data might be poisoning our AI models

Gurkenglas 4 Mar 2025 20:45 UTC
4 points
0
Have you seen https://www.lesswrong.com/posts/ifechgnJRtJdduFGC/emergent-misalignment-narrow-finetuning-can-produce-broadly ? :)
- aggliu 11 Mar 2025 16:38 UTC
  1 point
  0
  Parent
  Yep, that’s why I mentioned evil numbers specifically.