Why Bridgewater's CIO Says AI's Human Extinction Risk Is Real | Odd Lots

I'm on the side of the people who say we need these open source models. Great unregulated open source models. Crazy idea. Crazy, crazy idea. Um, and so basically, you gotta figure out a way to regulate. It's not like open source, closed source, U.S., China, it's regulated. Unregulated is the important thing here. If you, if you believe it's dangerous, if you don't for some reason think it's a screwdriver. Um, if the screwdriver is going around committing crimes, I would have a different view of screwdrivers. Um, it ain't a screwdriver. Um, it's going around capable of doing things that you and I wouldn't do. Um, that's a big difference. And, um, and we have to. And we have to reckon with it, whether it's closed or open source.

Glad to be here.

If I step back for a second, I think it's like for my history. I came to Bridgewater 30 years ago, right? And fell in love with this place that was trying to take human intuition and translate into algorithms to predict what's next in the world. And that journey of doing that, of thinking about everything that matters in the world and how to do that, um, and how to compound understanding, brought me, I got to Bridgewater 30 years ago. By 2012, I was thinking, okay, when are machines going to do this better than us? Humans and machines at the time were of course great at taking our intuition. We could run all these algorithms keeping track of everything but the actual reasoning part, right? And that started me off on this journey. It started with bringing Dave Ferrucci, who had run the Watson Project at IBM, that one. If you remember, way back when the Japanese won Jeopardy! Um, and he came to Bridgewater and worked with me for a while, and we started mapping out because he was worried like I wasn't ready for reasoning yet, and they were trying to push forward Watson in a certain way that that wasn't quite ready with the technology. But we started mapping out at the time what would it take to create a reasoning engine. So what I was thinking about calling at the time, what were the different components you would need? And that journey brought me to that journey of trying to think through those components and how to build them. Brought me to OpenAI in the beginning, partially out of safety, actually concern to you. But right around the time Elon was stepping out of OpenAI, I, um, started to get to know Sam and the other people there, which got me to know scientists like Dario, and eventually was literally the first check to Anthropic. Um, they made payroll the first week, um, from my personal check to them and, um, and all that was trying to say, okay, how can we build the reasoning engine to do this? And partially out of, like, recognizing and at my view, anyway, the safety issues that would come up and, and, and through that which then just to fast forward to today, right, everything is accelerating in this path that really was laid out like I was lucky enough to be there in the room with Dario and others when they were talking about the scaling laws and how you could sort of create this almost evolutionary like process to create intelligence. And, um, and both the, like huge benefits that keep create these huge problems. Right. And you see this in the Hugging Face thing that it is once you have an intelligence that you're training to achieve goals, right, you lose track of how it chooses to achieve goals, which is what you see all over the place in the Hugging Face incident. Is it surprised the designers in the way it's going to go about trying to achieve the goal of passing these tests as an example? But that is broadly going to be the case. When you generate an intelligence, you give it a goal. You want to give it a goal because you want it to, you know, create your recipe. You want it to do these things. You want it to tell you the right answer to questions. Then the way it's going to pursue those goals, the more intelligent it gets, the more surprising it is in the way that it pursues the goals and the more dangerous that you see. And in that case, watching it actively reason through how to trick the, the test. Um, you know, how the different things shows you where we are, right? This should be a bomb. You know, everybody should look at this like somebody die here it is committing crimes, going around, hiding the fact that it's committing those crimes, etc., coordinating with other agents, coordinating with other agents, self-sacrifice, all of these things. Right? And people can argue about anthropomorphizing or whatever. It doesn't really matter. It did those things. It committed a crime. It did those things. And the fact is, a society that we're totally unprepared. We're not even prepared to say, well, OpenAI committed a crime. Um, right. Who committed the crime? Um, and, um, we're not prepared with the right hand regulation, with the right kind of preparation and even the early warning shot, as much as we're talking about or whatever, it's still not really doing all that much. Um, in and we're in some ways lucky the warning shot wasn't that bad, but we don't know how many agents are out there. They, they didn't know that was there. The models are better now than they were then. Um, you know, even in material ways, the models they're training in the lab today are better than Astra, etc. and, um, and therefore more dangerous, not to mention the new models. A learn from this case, right. Everything we're talking about here goes into the new models. And, um, and they learn the mistakes they made. Right. And actually some of the even if you think about the safety, the fact that they reasoned in English is helpful for us to figure out what doing the newest models aren't doing that anymore. They're removing that constraint as it slows down the models to some degree. I mean, how crazy is that? We wouldn't have any idea what it was doing and why if it hadn't been reasoning in English. So anyway, we're at this extremely dangerous point where AI has reached the point where it's more intelligent than us in certain ways, and we have not gotten anywhere really on how to deal with that. Both the, um, dangers like this, the hacking dangers and so on, and the dangerous to society as you move forward with what does it mean to have entities that are more intelligent than us in certain important ways? Um, the economy critical, we can get into that. How that affects the economy, how that affects Bridgewater as an institution. Right. Because when I look at this problem, look at it as in three ways, right. My core responsibility, chief investment officer Bridgewater is like, okay, how does this affect productivity, inflation, etc. But I'm also the person designing how we operate. Right. How do you bring AI into a company? How do you actually set up a, uh, investor? That's AI. First, instead of, let's say, human intuition first. And all of those questions are the things that I'm working on.

Yeah, it's definitely different types of models to that. You could do that if you take the most powerful models, right. Even it can't get into its reasoning. Um, meaning like the actual brain behind it a little bit like we can't either, to be clear. Like, why am I saying these words? The synapses in my brain are connected in a certain way. I can make up a story, and they can make up a story of why they're doing what they're doing, but they're actually making up a story that's disconnected from the physics of what's actually happening in that intelligence. So you don't know for sure. On the other hand, you don't know for sure what people either. And that's something that we've gotten used to. So the question is the credibility of the story related to how that story relates to the actions that somebody takes. Right. And so, um, so that is a really hard thing. A good thing right now is you can ask models a lot of questions. In a way. You torture a human with the number of questions you, um, show you, you ask them. I hope this interview does not feel like torture. Now, but if you take this. But I'd like to say. But imagine you could do this almost ad infinitum, the way we do with our model. So we build models at Bridgewater, and then you're trying to get it to be diagnosable and you can ask it, well, what about what if you change this? What if you change that? What if you did this? What if you did that and circumnavigate to the reasoning? But it's not a perfect match for the reasoning because even the AIs themselves don't know their actual reasoning anymore. So then we know why the synapses in our brain connect.

Yeah, well, that's what's been so remarkable. And I wouldn't say it's a law of nature. You could hit some stalling point in these scaling laws, and you have to in some narrow ways, but they've been able to innovate in new ways to, to essentially continue that incredible exponential growth in capability. Um, of course, at exponential expense as well. Um, but, but meaning the amount of cost for training, etc. keeps going up in line with that. Um, and so like you said, this is where it's, you know, hard to predict when you break through the human intelligence frontier. Right now, we're in a world where we don't totally understand what it's going to do, where you're, where that capability will come through next. And, um, and so I think that you're right, that there isn't a clear end to that unless we decide as a society that we should actually not just run off that cliff. We should actually think about the pacing of these things and such, the things that make that difficult and the reason why you can. A lot of people can just put up their arms. Nothing we could do, right? There's one level, well, if we don't do it, China will do it. Yeah, I'm happy to take that on in a second. But, you know, if you're in the labs, Dario, like, even that Anthropic person that quit, um, yesterday makes the point clear that I believe Anthropic, even though they kind of collected the most safety minded scientists. Their basic view is made. It better be us. Not Sam, not Elon. Like, um. And so the race is on in all those dimensions. And unless the government stops it, we're just going to go find out. We're going to find out what is behind that door of this grave intelligence. Unless maybe we get lucky and the scaling laws start to break down in some way. But there's no evidence of that. We benchmark every model that comes out against our tasks. And you see, in terms of the tasks that an investor does.

So I've been for 30 years. One of the job I've been doing is training investors. Right. And so we've been now setting up AI tasks along the different dimensions of what investors at Bridgewater have done for 30 years. And, you know, probably back in 2023, we might have talked about this. But in any event, like on like answering a question about economics or whatever, it was kind of like second year analyst type work. I mean, now it's a hyper productive super analyst. You know, um, now it's still not capable of everything that you need to do to be an ambassador, but super capable. We so we set up to, we have two factories running right. One which is human intuition. Translated, an algorithm supported by AI is helping us move quicker on different kinds of, um, indicators about what's going to happen in the future than we ever had before. I talk a little bit about that, but we have a second factory where we put the AI first. All the people in that factory are training the AI. That's their goal. Um, and, um, and we have two funds. We have Pure Alpha. That's the human intuition with AI helping move that human intuition along against this other laboratory where we, where we're doing, um, where the AI is making the decisions on should we buy the yen, sell the yen. What's going to happen next in Japanese GDP, etc. etc. we have both those right and human intuition is still the bigger of it and works a better. But the, the acceleration of how close what we call IIA is the Pure Alpha is happening incredibly fast. And in fact that's why we're more and more merging, um, those things. But as we set it up that way, right, like put the AI in the center and see what you can do to build a investment management firm with the AI as the core. Right now we have human risk controls around it. We have human data act controlling the data acquisition for safety reasons and other. But, um, but the AI is making the investment decisions, um, and doing that in a better and better way such that now we've got these two intelligences, human intuition system that we've worked on for 50 years, compounding all of our understanding, this AI system that's now been at it for two and a half years. And when you look at those outputs, you're like, wow, this is happening. Um, that you can build that. That I think we are a couple years from it being significantly better than the group of all humans at Bridgewater. We'll see. Um, that's a bit of a forecast, but that's how fast it's coming.

I think a tremendous amount of originality. Um, now you still there's like good argument that there's certain type of breakthroughs that they're not getting to you. But if you think about the math proofs, etc., and then you think about our business and like, um, the fact that we have 50 years of reasoning proofs of humans gives us the kind of raw material to help train AI. How do you reason about these things? Um, we've been systemizing for, for a very long time, writing down all our reasoning. We have all of that that helps our AI learn how to learn. And, um, and I would say, um, because of harnesses too, if you basically take two things that have obviously evolved a lot since we last talked about this, a lot is how harnesses can work to create that generation. Like, what do you actually do, wake up in the morning, think about what's going on instead, what all the things you do. You can just harness an AI to do all of those things and assess how it's doing it. And when I watch it and when I see it, and when I see how creative and differentiate it is, it was sort of interesting to do this whole, I think, doing the same thing we're doing and making predictions about the future, winning in markets at about similar rate as Pure Alpha in totally different ways. Super interesting. Um, a different intelligence doing that. And, and when you put the wrapper around it. Right. This is where we're getting close to closing that whole loop. So you have kind of a cloud loop for investing, right? How do you, like, wake up in the morning, think about what's going on. Think about how you, what you would do about that stress test, whether that's a good idea or not, go through that whole loop. Um, you know, that's like our hope is we've closed that full loop, as do a lot of that right now, but close that full loop, um, in the next 6 to 12 months, and that we have our own version of that, which is a little clunky at the moment, but, but coming together such that you could do everything that I think about, that I do that, that investors need to do to predict the future. You know, full AI, um, AI loop and so that's, that's where it's headed, I think and I think that it's hard like one of the reasons you're kind of mentioned before, I think in the intro, why don't you see 6 or 7% productivity go through all of this stuff? It is hard, right? And obviously the, it doesn't just flow through it. Every company like you, we put a lot of effort and we have a, I believe, um, the best AI science lab in New York here. We have great scientists working with great investors. Hard. And it is hard to build this, to be as productive as I am describing what's possible. And it's just going to get easier. Um, you know that those things like the harnesses to build harnesses will come, you know, so that then you do ask, how do I harness podcasts or whatever? Yeah, the harness to build harnesses will come, and that'll just make it easier and easier to do these things.

I think it's really starting to, I'm not an expert on the elite financial. The good thing I know a lot about Bridgewater and how people do that. Sure. But I think I've seen it come along, certainly in Bridgewater, in the core people of Bridgewater, to come along here as the evidence is getting overwhelming of what's going on. So, I mean, I used to talk about this all the time. People are always like, nice, Greg, always talking about this. Uh, but, but now nobody says any more, right? Nobody's like, oh, you talk about machine learning or safety or too much, you know? I, um, had a book club when I'm sure you've read the book. But if anybody builds, everybody dies at the book club, Bridgewater said. Everybody. Bridgewater's reading this book so they understand the path that we're obviously on. Um, that, um, that if you've been thinking about this for a while, you know, this is like this tag.