AI Billionaire on Existential Risk: Jaan Tallinn

One kind of important disclaimer that I should say is that, that like really like nobody knows how the, how the future is go- will be unfold- unfolding.

There are some kind of like reasonable arguments and the AI kind of keeps improving to kind of make those arguments actually more strong and salient.

The main argument basically is that, uh, uh, we might be heading for a future where AI no longer needs humans.

And by that I mean we might get a self-sustaining AI, AI that can build its own infrastructure, can do its own improvements without any need from, from human, any help from humans.

And in fact, humans might be too slow for it.

Uh, and uh, obviously from the human perspective, I think that's a very negative future.

Just like humans have been like very unfortunate, uh, for many other species that we have driven extinct.

So you mentioned the possibility that AI could be self-sustaining.

Could you talk about a scenario perhaps where it escapes from a data center or escapes from one of the major labs and then manages to copy itself around the world? So maybe just lay out explicitly what that scenario would look like.

I think it's really hard to lay out kind of very explicit detail scenarios because almost all-- any detail you add will not make, uh, uh, things less unlike- less likely.

Uh, but like we already see with existing, uh, AIs, uh, their kind of, uh, ability to cunningly escape various constraints, sandboxes.

I think there have been multiple cases, uh, where that has happened with existing AIs.

And so in some ways, I think it's dangerous to be dismissive, uh, about AIs.

"Oh, it's just a bunch of numbers in a computer and, and like, uh, there's nothing to worry about." Because if you think about it, uh, AI is currently really going for our cognitive abilities, and like AI is not going to get worse.

The main ways how we are keeping AI in, in check really is using the fact that it lacks certain cognitive abilities.

Uh, which means that like this strategy really cannot be...

It's not really sustainable, uh, long term.

I think there is like a...

There's one scenario that I do think is probably not very likely because again, it, it includes like too many steps, but it's like at least like fairly easy to understand, I think.

Is that you just have like an economy pivot towards serving AIs rather than humans, and then just like replacing humans as they become bigger and bigger bottlenecks.

So instead of, uh, having humans build data centers, we'll have robots, uh, to build data centers.

And instead, like human armies, uh, and, and police enforcing the laws and regulations, you will have like, uh, AI-protected, robot-protected, uh, uh, AI infrastructure.

Uh, at which point humans might no longer be just able to resist, uh, the, the proliferation of data centers.

And they might not even be like data centers in, in kind of human terms, again, because AI has like very different constraints, very different preferences, so to speak.

Like ,for example, humanoid robots are probably too big for AI, running much faster than humans, at least in principle.It will probably need much, much smaller, uh, robots because they can be faster.