> Essentially, I think the AI Alignment people have got it wrong, thinking the problem is "How do we get Arnie to be nice like in Terminator 2?" when it's really "Don't give the chatbot your passwords".
> are increasingly "set loose" they are very good at some things and also unpredictable. All computer systems are hackable, and even if an autonomous software system does something like set all traffic lights in London to red that's a massive problem. >
Interesting perspective of the 6 camps within the growing ‘discourse coalition’ that is metascience. open.substack.com/pub/macrosci...