Bridgewater’s Greg Jensen Calls for Regulating AI Firms Like Systemically Important Banks

Greg Jensen, co–chief investment officer at Bridgewater Associates, thinks the top holders of AI compute warrant extra regulation, similar to what governs the world’s biggest financial institutions.

If a company controls more than, say, 5% of U.S. or global compute, it should be subject to heightened oversight, “like we do with systemically important banks,” Jensen said. In a wide-ranging interview, he also discussed Bridgewater’s growing use of AI in investing and where it still remains cautious, including leaving risk controls and trade execution to what he called “human-driven algorithms.”

As co-CIO at one of the world’s largest hedge fund firms, Jensen oversees Bridgewater’s Pure Alpha strategy, which makes computer-driven bets on big economic trends, as well as an AI skunkworks called AIA Labs. Jensen founded AIA in 2023 to develop an AI investing machine that can reason its way through markets and come up with winning trades. Bridgewater now has more than 80 employees devoted to the approach and a dedicated fund that uses AI to manage $4.5 billion.

Jensen, in other words, is a big proponent of AI and the benefits it can bring to the investing profession as well as society, he said in an interview. But he’s also worried about what AI can do if left unchecked, and he is calling for a laundry list of regulations and increased oversight.

As part of that, Jensen argues that slowing progress at the leading labs won’t be enough to control AI’s risks, which could also come from open-source models.

“There’s a huge problem with open-source models because you can train them. We do this at Bridgewater; it’s extremely effective reinforcement learning training on powerful open-source models,” he said. “It’s extremely powerful. A great technology, and at the same time clearly a very dangerous one.”

Jensen explained Bridgewater’s own experiences and challenges with AI, including how a model cheated on a test it had been given inside the firm, as well as its approach to managing its own AI infrastructure. Bridgewater rents all of its compute, he said, as opposed to other investment managers that have acquired their own graphics processing units or gotten involved in building data centers.

When it comes to actually investing in AI-linked companies, Bridgewater thinks much of the infrastructure build-out is priced into stocks now, and the firm is setting its sights on other trades.

“Overall, we have a very small position on the AI build-out at this point and are much more interested in the disruption and adoption trades now,” Jensen said.

This interview has been edited for length and clarity.

You’ve clearly been thinking about AI safety for some time—do the AI industry’s efforts to slow development address your concerns?

Any way you look at the evidence, it’s clear that the machines are reasoning in ways that are surprising to the designers, in ways that illustrate a desire in some sense to survive, a willingness to cheat, [a] willingness to lie. And that’s across labs. The OpenAI Hugging Face [breach] is an extreme example, but that’s happening all over the place. So we’re at this very, very dangerous point.

Slowing down the frontier labs is not nearly enough, meaning we have to think through the whole picture of what is happening. Start with understanding what’s happening, both for closed-source models [and] open-source models—the implications from a risk safety perspective are a big deal—and then thinking through the societal problems that are going to occur as a result of this technology.

So this is a good start, but there’s a lot more that we need to get working on now because the problems are going to keep coming at us at an increasing rate.

What else is needed?

You don’t know what people are doing with them. There’s no way to track that, so it’s dangerous from that perspective. We haven’t even begun the conversation on how to deal with that issue.

And then we have to deal with the concentration of power. In two years, OpenAI and Anthropic are going to control 35% to 50% of the world’s compute. That’s a crazy outcome for a society to allow on something as powerful as compute. Would we let one entity control that much of some other form of energy or commodity?

Finally, I’d say, we need to be really clear as much as possible as to who is responsible for crimes committed with AI, and make sure that that’s well understood so that the people putting out things in a rush recognize the risk that they’re taking as corporations and individuals.

How do you stop the concentration of compute?

We should take anybody that has more than x percent, let’s say 5%, of the world’s or U.S.’s compute resources. Say, “OK, those are systemically important institutions—we’re going to put them into some sort of thing like we do with systemically important banks and say there’s going to be regulation.”

I also think it’s fine to say we’re going to have caps on how much you can own. There’s a cap on how many commodity futures you can own. There’s a cap on much less important things. We may be able to do that with current law—you may need new laws. Either way, let’s get going on sorting that out legally.

There should be a bipartisan recognition that you shouldn’t want monopolistic control on what I think most people will agree is one of the most important resources in the world.

What would you say to the critics who argue that would give China an edge?

I’m no big fan of government regulation, but if the problem’s big enough and society bears the risk, you need government regulation. No matter how much you might dislike it or find it a very hapless tool in many ways, you can’t get away from it if it’s [about something] extremely risky to society.

On the China race thing, I’d make a couple points. The reason China and open-source models are moving so quickly is in large part because they’re borrowing from the technology the frontier labs are putting out there—distilling them, stealing intellectual property and other things. If you slow down the frontier, you mechanically slow down others, because they’re copying those things. They don’t have the compute. They’re not actually creating models in the same way frontier labs are.

I also think the idea that China would never cooperate on such a thing is also misreading how China operates. I think that they are as interested in or more interested in controlling technology, making sure it doesn’t threaten their power and so on. There’s a lot of reason to believe there would be common ground if we chose to seek it.

Lastly, we should put regulatory standards on any model that enters the U.S. ecosystem. And that’ll slow down the revenue and the other things those labs need if we don’t come to some way of being able to regulate those labs at all.

Have you seen AI do things firsthand that concern you?

In terms of our use case, look, there is stuff that surprises me. Reinforcement learning on open-source models is better than I expected. It’s better than frontier in many ways. But if you turn that over and say what you could aim that technology at, it’s pretty easy to transfer that to other things that would be much more negative.

And we did have a result where one of the models cheated on our test. Obviously, people within the labs on the cutting edge have an even deeper window into things. But people that think they’re making this up for publicity or whatever, I believe that is a wrong view.

Wait, you had a model cheat? What task had you given it?

We have research projects where we build out a test of whether it can go achieve the types of things our human investors do. And the model went and it just found the answer key. It did that rather than doing the task.

What is Bridgewater’s AI infrastructure setup?

We don’t own our own compute. We’re renting it in a variety of ways, both literally renting [from cloud providers] for our own use and our own training models, and then using some of our partners’ compute, particularly for the most compute-intensive reinforcement learning tasks that we do.

Most of the best reinforcement learning tools, honestly, are not in those frontier labs. Thinking Machines has the best tools, and [there are] a few other partners that do the best work on reinforcement learning for others.

How much of a bottleneck is access to compute for you?

Where compute would be helpful, particularly when you look out into the future, is some of the stuff we have in the lab will require a significant amount of compute to put into real time. Compute is somewhat of a future potential bottleneck. Compute is not our main constraint.

Our most important constraint is creating great algorithms that actually reason and handle the problem of markets, which is that there’s not really that much information, there’s limited data, and the data in the future is not necessarily the data of the past. Even understanding that there are AIs in markets today, that’s just one example of something that’s changed, and you have to really question the value of past data given the change in the world.

So you need to reason across it, and those are the most important, hard algorithms to crack. Not even the labs have cracked this problem.

What models are you using?

You can think of [AIA] as many different agents. Some of those agents are fueled by the best frontier technology, some of them are best used for reinforcement learning for a particular task.…Sometimes you really want a broad intelligence. Sometimes you want your expert that really knows a sector. So we use different tools.

We’re constantly benchmarking. So it’s not like we’re locked into Anthropic or OpenAI—we’re constantly measuring their tools for the different tasks that we are building, the different components of the brain, and choosing the models that provide the best intelligence for those tasks.

How do you think about the data that you’re sharing with partners?

Carefully. We’ve worked hard on how to build this. For what it’s worth, we were one of the first people that moved a significant amount of intellectual property into the cloud more than 15 years ago or so with Amazon, and worked hard on their security outline for how to get comfortable doing that with Bridgewater’s algorithms and such. Similarly, we have worked with AI providers to get to agreements that we can feel comfortable with. And [we] move fast.

There are two types of risks. You could get too stuck on security such that you don’t make progress using the most effective tools, and you could be reckless with your intellectual property. And so we’ve worked hard on balancing those and made our determinations of which partners to use for what things.

Are you able to negotiate zero data retention with Anthropic?

I won’t get into all of our negotiations, but we’re getting the best terms that we can get comfortable with. And that is provider by provider, but we’re able to get terms that we think put us on the best footing for what we’re doing.

What is the best way to think about AIA Labs?

We built out our first kind of internal model by giving it case studies of all of [our] history, helping it reinforce and learn what are the causal relationships through history, then apply that learning to the future. And that’s when we set up AIA as a return stream, so as a fund. We basically got that done by the beginning of 2024. We’ve been managing money since then.

So now we have two factories, one where it’s human intuition supported by the best technology and more and more AI. And we have one where humans are training and controlling an AI, but the AI is actually making the investment decision, whether we’re long the yen or short stocks or whatever. They’re risk controlled by human controls, data acquisition is human controlled, but the actual investment decisions in the AIA fund are made by the machine.

What makes AIA different from quant funds, which have done systematic, data-driven investing for decades?

The whole thing is focused on reasoning rather than, let’s say, quantitative prediction. I wanted to build something that I could interact with, that could teach me about how the world works and what it’s seeing, and so we built something where it’s not trying to predict the next market move per se, at least not directly. It’s trying to understand what the cause-effect relationships are in the world, draw out that causal map and then use that to predict what’s next.

If I sit down with clients and explain why we hold a position, we can hold AIA to that same standard. It should be a reasoning-first kind of thing, and that’s what’s getting better and better.

Would you ever let AIA trade without human intervention?

I don’t see that happening. One of the ways quant models work is just being on very short time frames where you can make up for the lack of reasoning with sample size. I would say that has its own problems, but to a certain degree, that’s the case. And you’re seeing more and more development of AIs at that time frame and AIs making their own decisions at that time frame, both with some risks to markets and obviously some benefits. That’s not something we’re currently touching.

Does Bridgewater build kill switches or something similar into its systems?

Definitely. At any point where it deals with the outside world, those all have the ability to sort of make sure it’s not doing something it shouldn’t. So those are essentially kill switches. Obviously, you give up something by not letting it do that. I think that choice still makes sense.

The thing the AI would most like to do, if you look at our AI in the lab, it would like to source its own data. Because it can do that very quickly. So that’s a question: Should it be able to source its own data or not, and what are the pros and cons?

For now, anyway, anything we use in production, we’ll make sure the data acquisition is OK’d by humans. And then all of the risk controls and execution are human-driven algorithms, not AI-driven algorithms.

Lastly, from an investing perspective, what’s your view on the AI trade?

We thought this was an incredible trade two years ago. A lot of it is now priced in. We have modeled all the AI data centers that are being built in the world and what components are likely to go into them and what that means for sales all the way out to 2028. We’re starting our 2029 model. We have a big system to understand what’s going on.

We still think the market is probably underestimating a little bit what will be built in 2028, but it’s close, and that assumes no significant disruption. Obviously, the regulation thing could be something, and funding could be a major disruption. You have to get a massive amount of funding, and you’re having trouble swallowing the funding this year. It’s going to be more than double that funding next year. There’s some execution risk on building out these data centers, and there’s literally the labs that take 40% of the build-out for training right now, potentially slowing down those training runs.

If you have the AI, you need to support the level of those valuations. You also have to imagine that it’s disruptive in many ways—disruptive even to the process of building AI, [and] potentially to the semiconductor process as well. So the rate of change that you might have on what those components are, and how those assets depreciate and so on, is still a risk.

Overall, we have a very small position on the AI build-out at this point and are much more interested in the disruption and adoption trades now.

To be clear, I don’t think it’s a bubble. I think it’s realistically pricing in what’s happening, but it’s not this incredible opportunity either—at least the picks and shovels elements of the AI trade.