I think this is the most powerful technological revolution that we've ever been through, and I think that's why this is so important to get right. And I think you have a responsibility, particularly if you're bullish on the power of AI, to see the risks and make sure we as a society mitigate them, because it's the only way in the long run we're going to get the benefits. So the two go hand in hand. I think if you believe it's going to transform society, you have to know that's a dangerous thing that happens very quickly. That you have to deal with both the good of AI and the danger of AI simultaneously, or you will get the danger and not the good.
I'm Jim Haskell, editor of the Bridgewater Daily Observations. We recently put out a daily observations from CIO Greg Jensen and our CEO Nir Bardea, and it was titled 'Our Thoughts on What Is Likely the Most Important Policy Decision of Our Lifetime.' And in that video, we shared our framework for how policymakers might want to consider navigating huge technological transformations like we're seeing with AI today, and also why it's so important for policymakers to get this right now. And even some specific policy proposals we would recommend, particularly around two areas: AI safety, where Greg and Nir make the case that government rather than industry should set standards for how models are developed, released, and used; and around labor disruption. They suggest a tax on tokens and citizen equity in the leading AI companies that would help make society more resilient to the AI transition and give citizens a stake in it.
That observations got very broad readership and great feedback from our readers. It prompted excellent conversations both internally at Bridgewater and with many of our clients around AI policy. And based on the follow-up questions we received, I asked both Greg and Nir to join me today to discuss many of those questions and to give them a chance to expand on some of the points laid out in the observations. And just to be clear, the goal both here and in the original BDO is to discuss the big picture questions at the high level and not double-click into the details. We will follow up with more detailed work in coming observations. So first, let me welcome in both Greg and Nir. Thank you so much for joining, and I want to start with a direct question some people raise, which is why comment on AI policy at all? I mean, of course, we've been studying and using AI technology and we're focused on its investment implications, but we're not policymakers. So it's really important to lay out what perspective you're trying to add here.
So first of all, Jim, on behalf of both Greg and I, just so happy to be having this conversation. There's really nothing more important for us to be discussing right now. And that leads me kind of to answer your question, which I see as two parts. First of all, you know, at Bridgewater, our goal has always been to understand how the world works. And we comment and try to wrap our head around the most important dynamics. Those are the things that are going to drive economies and markets. And whatever those dynamics are, when they are COVID, we wrap our head around pandemics. So when they are wars, we wrap our head around, you know, politics and dynamics of war. When it's technologies, we wrap our head around deeply on emerging technologies. And when it's generational policy decisions throughout history, we do the same. And we work hard to understand them, to understand how things work, cause-effect linkages, and then we share our thinking with our clients and with policymakers.
And that leads me to the second reason, which is, you know, for many of us, I know for Greg, for me, we are blessed with that position of putting the puzzle together, of having an objective, unbiased understanding of how things work. And we do that with one of the most powerful teams out there and 50 years of compound understanding. So in some ways, I think we all believe we have to speak up, and especially when things are this complicated. Because look, if you look around right now and you look at this void that I think is a function of so many people looking at this reality and saying, 'Well, it's so complicated, you have to understand technology, you have to understand policy, you have to understand geopolitics,' and everybody feels unequipped to say something, and that void will be filled by either less informed people or people that have biases. So in many ways, I think it's more important than ever for us to kind of put our thoughts out there. And I'll just finish with where you started, which is since we put this observation out there, we've had great internal debates about these thoughts, we've had great back and forth with clients, great back and forth with policymakers. And to both Greg and I, that is exactly what we wanted: smart people that care about how things actually work, that are going back and forth trying to sharpen each other's ideas, leading us to find the best solutions. So, you know, for Bridgewater, we've been doing that for the last 50 years, and you can definitely expect us to keep talking, not shying away about whatever the most important dynamics are that are driving the world's economies and markets.
Yeah, just to add to what you're saying. I mean, first off, on what I think the value of the daily observations is, you get an unvarnished look at what we're thinking about. It's quite a project to write every day, and I think it sharpens our thinking, which is why we do it. I think we're better thinkers because we share our thinking every day. And what we intend that to be is a look over our shoulder to what actually matters. And we are at this point where AI in total is incredibly impactful. It's kind of a third or a half of everything that's going on, from how economic growth is being formed to equity markets, etc. And the regulation around it is going to be critical to whether it is sustainable or not. So these questions are the questions we must answer.
A question for both of you is that one of the biggest themes that comes up when we talk to our clients about AI is what about China? How are you guys thinking about that?
Great. So, I'm glad we're taking this question up front because I think one of the most important things to lay out is to clarify how we believe the concept of winning actually works. Meaning, what does it mean to win the AI race? And what Greg and I laid out in the daily observation is the argument that winning just the technological race while dismantling your society or political system isn't winning. And even if you say, 'Well, you know, I think it is. I just want to win the technological race,' the process of unhinged pursuit of technological victory is so likely to cause either a security or political backlash that even for those that are saying, 'Look, I really just want to focus on that one dimension,' ignoring the factors that we laid out in our write-up is almost certainly a mistake and is also certainly going to hold you back from winning even just the technological arms race. And I think a good example that we're living through right now is the wave of populism and mercantilism that is following the last, what I would call, unhinged pursuit of productivity growth from globalization. Because from a global perspective, I think people could look at the US and say, 'Well, the US has won.' But when you look at the US domestically, internally, over the last several years, I think many or most would ask themselves, 'Has the US won?' And this is not a political point. I think both from the left and the right, people look internally at the US and say we're at one of the lowest points that we've been because the gutting of the manufacturing base leading to a disenfranchised and divided society and ultimately real pressures on institutions, real pressures on the ability to govern, has circled all the way back to destabilizing geopolitics. These are lessons that are right in front of us, learned by all generations, about the concept of winning that we have to carry over to this next massive productivity leap.
Just to add on to what Nir has described about winning, I'd say the problem is a human problem. It's across the world, and the US has a special responsibility and role being at the frontier of this technology. But it is correct that in the end, the only way for the world and humanity to be safe and to manage the disruption that's likely to come is through global answers to these questions. Now, how do we get started on that? Right? We can't get started by making this incredibly complex. I think starting with the US is super important. And I think the US moving will, A, make it more likely that this works in a way that doesn't tear US society apart or put us at unnecessary risk, and B, it will ripple into the rest of the world. The US mandating that all models used in the US are regulated will put pressure on the Chinese either to do similar regulation or to not get the capital benefits of being able to impact the US AI ecosystem, which is the biggest. So A, it has some direct consequences there. B, intellectual property from the US is slipping out across the world to China and other places. And by slowing down and carefully monitoring and regulating the security around AI models, you'll slow down some of the distillation and some of the intellectual property theft. As a result, you'll make the rest of the world safer even if it's uncooperative. So for those reasons, I think it's important to, A, see this as a global problem, B, see the US can and should lead in changing the direction that the world's currently on in this race dynamic. And that by doing so, even without cooperation, which I still would like and expect that cooperation is more likely than pessimists expect, but even without it, you slow down the race in other countries because the slippage from the US is controlled.
Greg, you just discussed the importance of slowing down and regulating the security around AI models, but within this, there are open-source models, there are closed-source models. How do you differentiate between the two and the risk that they pose?
Yeah, so there are risks to both open and closed-source models, and I really think we need to change the language to regulated and unregulated models. We have to be very careful about what we leave unregulated, whether it's closed-source or open-source. But let's talk about the differences and then why both need to be managed. The first is closed-source models. The danger here is that we're putting so much power in so few hands that needs to be regulated. We've seen the risk of that. We've seen the race dynamic as the different closed-source models right now, OpenAI and Anthropic, but Google and Meta and so on, all race to win. The externalities of that are not being managed. The implications of that are not being managed, and they need to be. And the risk of an unregulated closed-source model getting in control is both the types of risks that we'll talk about, all the dangers AI might cause, but also the risk of the concentration of power. That's the good news about open-source models is they are democratizing all of these powers. That's a good thing. They allow you to reinforcement learn on them. That could be used in many beneficial ways, and we do that here in our labs. The risk of it on open-source models is there's even less control how you reinforcement learn, what you reinforcement learn for them to do. Secondly, that when you ask a question, at least with the closed-source models, Anthropic and OpenAI check whether they should answer your question or not. There's no equivalent way to do that with an open-source model where you could get the weights, you could ask it anything that you want. It can't control at the point of input. So both paths, open-source model and closed-source models, have big risks and both can cause societal disaster. So both need to be regulated. We need to regulate what open-source models come into our ecosystem and what closed-source models are able to do. People can point accurately on how difficult that's going to be. Particularly open-source models, preventing them from being utilized is difficult because they're easily shared and so on. But the difficulty of it doesn't change the fact that it's necessary, and there are ways to handle these things if the government chooses and has the will to do it. So you're going to need to separate rather than open and closed, regulated and unregulated. And we're going to have to do a lot to keep unregulated models out of the ecosystem from doing grave harm and on figuring out the very tough challenge of what regulations can be effective against the growing risks that we're seeing in terms of these models.
Greg, one of the big themes of the Bridgewater Daily Observations you and Nir wrote is this nature of safety risk. And it's so interesting because for decades we've been running huge risk with nuclear power and bad actors, biological risk, you name it. And yet, fortunately, nothing terrible has happened. But what's really interesting about your warning here is that the nature of this risk is unique and different. So, I think it's really important for you to expand on this.
I want to expand on why I believe a major accident is inevitable if we don't change the path that we're on. Right. Two things. We're hurdling towards two different types of risks. The first is the acceleration of all the risks you just described. If you take cyber risk, if you take biological science risk, soon physical sciences, all of those cases, we are democratizing that intelligence in a way that's great in some sense that this power is getting handed out to people all around the world. Obviously, on the risk front, it's accelerating the risks of all those things that you described. Beyond that, and why the risks are different in kind, is that we're generating an intelligence that pursues its goals in its own way. Right. One of the powers of AI is that it could think for itself, decide how to achieve goals in ways that aren't programmed in. And by doing that, what that means is that it could choose a path that's highly dangerous in a way that nobody who built it intended. Not as an error actually, but because it's so smart, it could find different ways to find goals. But sometimes it could do that in ways with unintended consequences. So if you look at what's happened recently with the models breaking out of both OpenAI and Anthropic going into other companies' ecosystems, A, that's a crime. The AI committed a crime, and right now we're not even clear on, well, who's responsible when AI commits a crime. That's something that obviously needs to get worked on in terms of the regulation around that. But B, the basic point is that was never designed, right? They were asked to answer these questions about cyber attacks and other things. And in order to do that, chose to break out of the sandbox they were in, answer that question, actually organize many AI agents to pursue getting those answers. That level of thinking, however you think, like you want to describe what's going on there, whether it's agents scheming together or whether it's just a series of stochastic parrots that have learned these things from all the texts that they've read. Either way, it's an incredible amount of danger when you think about what might happen when it tries to solve other problems as they move it more into the material sciences, as they move it more into biological vaccine discovery, etc., and now it tries to answer questions and it chooses its own path on the way to answering those questions. I think the dangers are clear. They're imminent. And if we don't change the way we're regulating the labs, it's inevitable that we're going to have a very significant disastrous outcome. And unfortunately, a lot of times that's what it takes to create the kind of regulation that's needed.
The only thing I would add to what Greg is saying, which I completely agree with, is it's in addition to the ability of technology to do this, you have to contextualize this with where the world is. Meaning we are on this trend of, first of all, currently existing nationalism, meaning much more prone to confrontation, much more prone to where the ability exists for people to take advantage of that ability. Just look at what the last five years looked like. And if you ascribe to any of what we're saying about potential societal changes, you have to add to that nationalism and kind of geopolitical state that we've been living in for the last several years, discontent people within nations. So in addition to just that the capabilities being out there and in the hands of so many, you have to multiply that by, well, how many people are actually dissatisfied? How many people would actually like to put that to use? And that combination can paint a pretty scary future.
I want to move our discussion here a little bit to the labor front and particularly labor displacement. There's a wide range of opinions that AI may be more like a technology that will help create jobs to fears on the other side that it will cause much more disruption and labor displacement. In the observations you both wrote, you mentioned up to 18% of current jobs could be displaced over a number of years. And so I just want to first start by walking us through how you're assessing that risk.
