Technology

We should probably tease apart AGI and what I prefer to call “large-scale computing” (LLMs, SD, or any statistical ML approach

AGI has a pretty good chance of killing us all, or creating massive problems pretty much on its own. See: instrumental convergence.

Large-scale computing has the potential to cause all sorts of problems too. Just not the same kinds of problems as AGI.

I don’t think he sees LSC as an x-risk. Except maybe in the sense that a malicious actor who wants to provoke nuclear war could do so a bit more efficiently by using LSC, but it’s not like an LSC service is pulling a “War Games”.

What he’s proposing is:

• Since AGI is an extinction risk…
• and the companies pursuing it are pushing LSC along the way…
• and some of the problems caused by LSC will continue to be problems with AGI…
• and we have zero international groundwork for this so far…
• then we should probably start getting serious about regulating LSC now before AGI progress skyrockets as quickly as LSC did

And why not? LSC already poses big epistemic/economic/political/cultural problems on its own, even if nobody had any ambitions toward AGI.

Or C) Actually understand that alignment is a very hard problem we probably won't be able to solve in time.