I attended Max's recent lecture at Harvard, where they were recruiting for their AI safety fellowship; a friend of mine captured a good video, 54 minutes:
I think it's a good approach, we should use AI to write proof carrying code. I think we'll want AI to do things beyond what we can prove is safe, but maybe we shouldn't.
I think some commenters here are being too dismissive, Max came across to me as a provacatuer, making proposals that are not sure things, but whose rebuttals will advance the field.