Comment by Stuart J. Russell

I think at some point it will become politically obvious that there is only one way forward: passing a law that says you may not build AI systems exceeding a certain level of capability until you prove they are safe.

Frontier AI companies should grant embedded evaluators access equivalent to that of their own highly privileged employees for the purposes of their evaluations, and with exceptions to protect sensitive data belonging to the company’s customers and ot...

Below middling levels of capability, probably a net benefit. As they become more capable, open-weight models become much more dangerous than closed-weight models. For instance, imagine if five billion people had free access to an open-weight model mo...