4 ms·

Agent ability is far outpacing alignment though

• lukeschlather 25d ago
  That doesn't really seem true. The HF hack happened with a model that had all the alignment safeguards disabled intentionally. I think there's a good case for that kind of research, but also, OpenAI could just not do that if everyone thinks it's too dangerous.