Every safety taxonomy today is a list of scenarios that already went wrong. Coverage lags the incident that created the category.
Getting ahead of safety incidents takes a stated view on how a model should behave, not a long list of failures. Many labs today have a list and not a view.
Cyber is the current example. Code and cyber safety matter now, two years after RL made agents good at offensive coding.
显示更多
alright who’s creating an alignment data factory DMs open
