Comment by Miles Kodama

The mainstream take is that an AI lab should assess risk before deploying a new model externally. But I think risk assessment should come way earlier in the model lifecycle. Compare the situation with dangerous bio. We wouldn't find it acceptable for a virology lab to synthesize a superplague & test it on mice in the lab without doing any risk assessment. It's no defense if the lab says they were going to do risk assessment before releasing the superplague. Better to assess risk before you deploy a new model internally on sensitive tasks, such as developing new frontier models. Even better to assess risk before you initiate a training run that could yield a frontier model. And best of all to assess risk before you run an experiment that could yield dangerous algorithmic insights.