Researchers from the University of Oxford, the UK AI Safety Institute, and EleutherAI demonstrate a way to get AI models to avoid learning unsafe knowledge, as an alternative to trying to block them from outputting it post training.

→ View original post on X — @willknight,