Forget It
A joint study from the UK government and Eleuther AI shows that artificial intelligence may work better when it knows less. Researchers trained models with “deep ignorance,” stripping out sensitive instructions on bioweapons, hacking, and other risks while keeping general data intact. The result: systems still solved math, coding, and reasoning tasks effectively but failed at generating dangerous outputs. Safety came not from filters layered afterward but from reshaping what the model learned in the first place, teaching AI to forget strategically.