Defense Agents Will Beat Rogue Swarms | Rohit Krishnan

Rohit Krishnan (0:00)

We are in a fascinating time, because I would not have suspected running MTS to be a harder problem than solving a Millennium Prize that has stood unsolved for like nine decades. But it turns out it is.

Turns out, number goes up, you can do some things really well, way better than anybody ever could, but many things you just can't do at all. So to me, the fact that we are able to solve Millennium problems is freaking amazing. It is amazing that we are able to do that. It is one of the highest achievements of humanity.

SPEAKER_2 (0:33)

All right, we're back. We are live with Rohit Krishnan, who is, how do we begin to describe Rohit? He was an investor for some time. He is building a stealth startup. He's a senior fellow at Wharton. He's the author of the excellent Strange Loop Canon on Substack, and former guest of the Theo Jaffee podcast. So Rohit, welcome to MTS.

Rohit Krishnan (0:53)

That is the highest honor one can achieve. So I was very happy to get there.

SPEAKER_3 (0:56)

I haven't been a guest yet.

SPEAKER_2 (0:58)

You're a guest every day. Highest honor.

Yeah, of course.

Rohit Krishnan (1:02)

Anyway. Welcome. Thank you very much. Pleasure to be here.

SPEAKER_2 (1:05)

So you are one of the very best doom skeptics. Most doom, AI doom skeptics, anti-doomers have, I would say pretty bad epistemics, and their arguments are quite poor, but not you. Strange Loop Canon has been very good in this domain. In the last day, to say nothing of the last couple months, we've seen a remarkable preference cascade of people coming out in support of believing that AI could pose a risk of doom.

And so what would you say to those people?

Rohit Krishnan (1:39)

I mean, we are in a fascinating time, because we have AI that can solve Navier-Stokes, not Navier-Stokes as I was just told, but would still have a really hard time running MTS.

I would suspect that running MTS is a, you know, I would not have suspected running MTS to be a harder problem than solving a Millennium Prize that has stood unsolved for like nine decades, but it turns out it is.

So one question we should ask ourselves is like, is our conception of what is intelligence and what it can provide us actually accurate? Because if you roll back the clock, every argument about intelligence kind of treated it like a scalar. It's a number, number goes up and you can do many things. Turns out number goes up, you can do some things really well, way better than anybody ever could, but many things you just can't do at all. And you still can't do, right? I mean, we hate reading AI slop for a reason. So to me, like the fact that we are able to make San think and solve Millennium problems is freaking amazing. It is amazing that we are able to do that. One of the highest achievements of humanity. I don't particularly see why we should tarry it with the fact that another story about how it's going to go kill everybody. I don't think those two stories are necessarily related, even if they're held by the same people.

SPEAKER_2 (3:05)

Yeah, so I mean, I think the counterargument to this is like, yes, it is true that AI can't automate white collar work yet. AI could not run MTS yet. However, as AI progresses further, there will be basically a threshold after which it will be able to do that.

Rohit Krishnan (3:23)

Yeah.

SPEAKER_2 (3:23)

Just like there's a threshold before, it could automate Millennium Prize problems, and similarly, after, now it can just do them. Like, there are rumors that opening AI has also solved the Hodge conjecture. I have no private information here, but I've heard this as well, or one of the labs, maybe.

Rohit Krishnan (3:40)

I mean, one way to think about it, again, like, I remember GPT-2, which people were a little scared of, you know, because of misinformation, and this could take off through RSI, right? I mean, Dario famously was extremely worried about this.

We have gone many, many orders of magnitude now. Like, the latest models, you know, Astra is way bigger, the next one, Bell, is supposed to be another order of magnitude bigger. Like, these are giant models that have, like, the crazy thing here is that the scaling hypothesis has held over, I don't even know how many orders of magnitude at this point, 10, 14, 15, something of that nature. So, it is bad epistemics to constantly look at it and go like, the next one is the hard one, when you've been wrong every time before. So at some point, you need to kind of update.