Part of the problem is surely that we know how milk should be produced and consumed safely, but we don't know how AI should be produced and made accessible safely.

Should we be more concerned about malevolent humans using AI to cause harm? Or should we be more worried about misaligned AI agents escaping their sandbox to cause harm? Or should we be more worried about job loss, or the concentration of political power in the hands of a small clique of billionaires?

All of these are real possibilities, but it's hard to put a finger on the actual probabilities. And even if you're identified and quantified these dangers, what should the actual regulatory environment look like?

None of this is to argue that we shouldn't have regulation, just that it's very hard to know ahead of time what the rules should be. It doesn't mean we shouldn't try to regulate, but we need to prepare for a rapidly changing technology landscape.

I tend to favor an open source approach to safety rather than keeping the technology hidden. I'd like to see regulation to mandate that model training and weights be much more open than they are now. But I could be wrong: maybe Dario is right that this should be kept in the dark. The only things I'm fairly certain about is that nobody knows for sure and we're going to need to adjust quickly.