WEBVTT - Why Bridgewater's CIO Says AI's Human Extinction Risk Is Real 0:00:03.880 --> 0:00:06.439 Hello, Odd Lodge listeners. I'm Joe Wiesenthal. 0:00:06.620 --> 0:00:07.560 And I'm Tracy Alloway. 0:00:07.700 --> 0:00:10.270 We're the hosts of the Odd Lodge podcast, and we've 0:00:10.300 --> 0:00:12.140 got something exciting for you. That's right. 0:00:12.190 --> 0:00:14.810 So one of the best parts of hosting our podcast 0:00:14.930 --> 0:00:17.810 is we get to actually meet and interact with our listeners. 0:00:18.010 --> 0:00:20.989 And we know we have some listeners over in Los Angeles. 0:00:21.170 --> 0:00:21.550 That's right. 0:00:21.570 --> 0:00:24.310 So if you're in L.A., we're going to be recording 0:00:24.470 --> 0:00:28.090 a live show, some live recordings at the Vermont Theater 0:00:28.130 --> 0:00:29.510 in Hollywood on September 17th. 0:00:30.240 --> 0:00:32.990 We have some really exciting guests lined up, have some 0:00:33.150 --> 0:00:36.450 really great conversations planned. So go ahead and get your tickets. 0:00:36.470 --> 0:00:39.890 You can find those over at Bloomberg.com forward slash oddlots 0:00:40.030 --> 0:00:42.529 or click the link below in the show notes and 0:00:42.700 --> 0:00:43.470 come and say hi. 0:00:43.320 --> 0:00:43.760 When you're there. 0:00:49.500 --> 0:00:51.180 Bloomberg Audio Studios. 0:00:51.640 --> 0:00:52.320 Podcasts. 0:00:52.650 --> 0:00:53.990 Radio. News. 0:01:04.800 --> 0:01:08.470 Hello and welcome to another episode of the Oblots podcast. 0:01:08.560 --> 0:01:09.480 I'm Joe Weisenthal. 0:01:09.750 --> 0:01:10.730 And I'm Tracy Allaway. 0:01:11.030 --> 0:01:14.670 Tracy, no shortage of AI news these days. 0:01:14.890 --> 0:01:17.770 No, it feels like everything is AI. It's just AI 0:01:17.850 --> 0:01:18.330 all over. 0:01:18.510 --> 0:01:19.850 You know, no, it is. You know, I. 0:01:21.590 --> 0:01:23.970 Really like all the other stuff that we talk about. 0:01:23.990 --> 0:01:26.470 I love talking about the Fed. I love talking about oil, 0:01:26.910 --> 0:01:28.860 home building, all of that stuff. 0:01:28.880 --> 0:01:29.520 Niche markets. 0:01:29.560 --> 0:01:31.720 Yeah, I love it and I want to do it forever. 0:01:31.740 --> 0:01:36.220 It does feel like... since, you know, Chad GPT came out, 0:01:36.540 --> 0:01:39.020 if you probably plotted a chart, the percentage of our 0:01:39.120 --> 0:01:41.899 episodes that are in some way connected to AI keep 0:01:41.940 --> 0:01:46.839 going up. Maybe even exponentially up, like many AI charts. 0:01:47.400 --> 0:01:51.470 And I, like, worry. I worry about plenty of things because, 0:01:51.620 --> 0:01:55.010 you know, I'm a middle-aged dad. But I worry that, like, 0:01:55.630 --> 0:01:57.770 it's just getting, like, I feel like I'm being, you know, 0:01:57.850 --> 0:01:59.510 forced to learn about a lot of this stuff against 0:01:59.550 --> 0:02:01.190 my will in some of this. But I worry it's 0:02:01.210 --> 0:02:06.080 just saturating so many different episodes facets of economic society, markets, 0:02:06.140 --> 0:02:06.440 and so forth. 0:02:06.580 --> 0:02:09.570 Well, I mean, the concern is legitimate, but on the 0:02:09.610 --> 0:02:12.830 flip side, because it is affecting all these different things, 0:02:12.990 --> 0:02:15.170 it feels like we do actually have to talk about 0:02:15.210 --> 0:02:17.389 it quite a bit. And actually, it's funny you mentioned 0:02:17.430 --> 0:02:20.090 chat GPT, because I was just thinking the last time 0:02:20.130 --> 0:02:23.700 we spoke to this particular guest, I think we were 0:02:23.840 --> 0:02:26.980 on like GPT-4 or something back in 2023. And now, 0:02:27.000 --> 0:02:30.690 of course- I don't even know what number we're on 0:02:30.750 --> 0:02:34.860 because we've had all these new supermodels like Astra and 0:02:34.900 --> 0:02:36.359 Mythos and all of those coming out. 0:02:37.150 --> 0:02:39.730 No, it's pretty remarkable. Well, you're on Reddit a lot. 0:02:39.910 --> 0:02:43.190 Are you, speaking of, a little sidetrack, speaking of GPT-4. 0:02:43.330 --> 0:02:43.510 Yeah. 0:02:43.550 --> 0:02:45.929 Have you ever interacted with like all these people that 0:02:45.970 --> 0:02:48.470 are still like, you know, they sunsetted GPT-4? 0:02:48.530 --> 0:02:48.850 Yeah. 0:02:48.889 --> 0:02:50.290 Because that's the one a bunch of people fell in 0:02:50.450 --> 0:02:50.710 love with. 0:02:50.730 --> 0:02:51.190 Oh, yeah. 0:02:51.210 --> 0:02:54.089 And there's like these Reddit boards. Sam Waltman, why did 0:02:54.130 --> 0:02:55.769 you take GPT-4 away from me? 0:02:55.810 --> 0:02:58.130 Are you asking me if I personally spent time on 0:02:58.180 --> 0:02:59.790 those Reddit boards? I have not. 0:02:59.840 --> 0:03:00.440 You have not. 0:03:00.460 --> 0:03:02.060 But I am aware that they exist. 0:03:02.440 --> 0:03:04.280 But also- I. 0:03:04.280 --> 0:03:07.030 Would be very happy to go back to a world 0:03:07.060 --> 0:03:09.910 in which the biggest source of AI anxiety was just 0:03:09.930 --> 0:03:12.470 that people had too much of an affinity for the model. 0:03:12.870 --> 0:03:16.050 Because now, of course, we're recording this September 9th, 2026, 0:03:16.050 --> 0:03:21.380 we have incidents like the open AI hugging face attack. 0:03:21.880 --> 0:03:25.180 Just last night, there was the news, a researcher from 0:03:25.280 --> 0:03:28.600 Anthropic announced that he was quitting because he was like, 0:03:28.639 --> 0:03:32.620 these companies are gambling with our lives. Literally, as we 0:03:32.660 --> 0:03:35.400 were walking into the studio, I saw the news that 0:03:35.420 --> 0:03:39.540 the famed AI researcher Paul Cristiano is joining the board 0:03:39.580 --> 0:03:42.589 of either OpenAI or the OpenAI Foundation, talking about his 0:03:42.830 --> 0:03:47.510 concerns about recursive self-improvement and the dangers there. So it's like, whew, 0:03:47.710 --> 0:03:49.600 let's just go back to when people were worrying about 0:03:49.660 --> 0:03:50.560 falling in love with the models. 0:03:50.910 --> 0:03:53.230 It does feel like AI has sort of become an 0:03:53.470 --> 0:03:55.710 inevitability at this point. And we're all kind of on 0:03:55.730 --> 0:04:00.610 this runaway train with everyone racing to AGI. But I 0:04:00.670 --> 0:04:03.010 would say it still feels like there's a lot to 0:04:03.050 --> 0:04:06.570 figure out. You have like the alignment issues. You have 0:04:06.810 --> 0:04:09.960 how AI is actually going to fit into both financial 0:04:09.970 --> 0:04:13.530 markets and society. And this is kind of the moment, 0:04:13.570 --> 0:04:17.349 like before the runaway train goes over the ledge, let's 0:04:17.490 --> 0:04:19.530 actually think of some of these things. 0:04:19.670 --> 0:04:22.190 And then the other thing that I think is very 0:04:22.250 --> 0:04:27.159 relevant with this particular guest is like, quote, AI adoption, unquote. 0:04:27.180 --> 0:04:27.659 What does it mean? 0:04:27.700 --> 0:04:31.239 Because, all right, these companies are seeing surging revenue in every, 0:04:31.360 --> 0:04:34.120 I'm sure every financial institution in the world at this 0:04:34.180 --> 0:04:37.860 point has some sort of like corporate account with like 0:04:39.100 --> 0:04:43.180 open AI or anthropic, et cetera. But actually, like, you know, 0:04:43.240 --> 0:04:45.680 it's not like we've seen some productivity explosion. We have 0:04:45.720 --> 0:04:48.620 yet to see like the long prophecy, like big white 0:04:48.700 --> 0:04:52.330 collar layoff wave, etc. There are just some very, setting 0:04:52.370 --> 0:04:53.930 aside all the risk stuff, there are just sort of 0:04:53.970 --> 0:04:57.270 like straightforward questions about what the technology means for the 0:04:57.310 --> 0:04:59.690 economy and how it's actually being used. And like, when 0:04:59.710 --> 0:05:02.510 will we see the impact show up in sort of 0:05:02.670 --> 0:05:04.740 our traditional statistics and so forth? 0:05:04.839 --> 0:05:05.740 Yeah, we should talk about it. 0:05:06.040 --> 0:05:08.980 Anyway, I am very excited to say, returning to the podcast, 0:05:09.040 --> 0:05:12.070 we really do have the perfect guest. We're going to 0:05:12.089 --> 0:05:16.630 be speaking with Greg Jensen, Managing Chief Investment Officer at Bridgewater. 0:05:17.210 --> 0:05:20.330 He's been writing a lot about AI and various facets. Greg, 0:05:20.880 --> 0:05:22.800 thank you so much for coming back on Outlaws. 0:05:22.880 --> 0:05:23.440 Glad to be here. 0:05:24.060 --> 0:05:26.120 What do you make of the Hugging Face attack? I 0:05:26.180 --> 0:05:29.380 assume you read the meter report and have seen all 0:05:29.460 --> 0:05:32.140 the different takes. What was your takeaway from that incident? 0:05:33.760 --> 0:05:35.870 Well, if I step back for a second, I think 0:05:35.930 --> 0:05:39.220 it's like for my history. Yeah, it is. came to 0:05:39.260 --> 0:05:41.760 Bridgewater 30 years ago, right? And fell in love with 0:05:41.800 --> 0:05:44.310 this place that was trying to take human intuition, translate 0:05:44.330 --> 0:05:46.789 into algorithms to predict what's next in the world. And 0:05:46.810 --> 0:05:49.150 that journey of doing that, of thinking about everything that 0:05:49.190 --> 0:05:51.469 matters in the world and how to do that and 0:05:51.529 --> 0:05:54.490 how to compound understanding brought me, got to Bridgewater 30 0:05:54.490 --> 0:05:57.750 years ago. By 2012, I was thinking, okay, when are 0:05:57.800 --> 0:06:01.740 machines going to do this better than us humans? And 0:06:01.839 --> 0:06:03.740 machines at the time were, of course, great at taking 0:06:03.779 --> 0:06:06.460 our intuition. We could run all these algorithms, keeping track 0:06:06.480 --> 0:06:10.510 of everything. But the actual reasoning part And that started 0:06:10.529 --> 0:06:13.130 me off on this journey. It started with bringing Dave Ferrucci, 0:06:13.650 --> 0:06:16.330 who had run the Watson project at IBM that won, 0:06:16.350 --> 0:06:18.630 if you remember, way back. It won Jeopardy. 0:06:18.670 --> 0:06:19.870 It won Jeopardy. 0:06:20.430 --> 0:06:23.810 And he came to Bridgewater and worked with me for 0:06:23.830 --> 0:06:26.570 a while. And we started mapping out because he was worried, like, hey, 0:06:26.610 --> 0:06:29.040 I wasn't ready for reasoning yet. And they were trying 0:06:29.060 --> 0:06:31.360 to push forward Watson in a certain way that worked. 0:06:31.980 --> 0:06:34.280 that wasn't quite ready with the technology, but we started 0:06:34.339 --> 0:06:35.940 mapping out at the time, what would it take to 0:06:35.980 --> 0:06:37.960 create a reasoning engine? That's what I was thinking about 0:06:38.000 --> 0:06:39.460 and calling it at the time. What were the different 0:06:39.480 --> 0:06:42.640 components you would need? And that journey brought me to, 0:06:42.940 --> 0:06:44.700 that journey of trying to think through those components and 0:06:44.720 --> 0:06:47.000 how to build them, brought me to OpenAI in the beginning, 0:06:47.300 --> 0:06:50.500 partially out of safety, actually concerned to, but right around 0:06:50.520 --> 0:06:53.740 the time Elon was stepping out of OpenAI, I started 0:06:53.760 --> 0:06:55.560 to get to know Sam and the other people there, 0:06:56.020 --> 0:06:59.640 which got me to know scientists like Dario and eventually 0:06:59.680 --> 0:07:03.390 was literally the first check to Anthropic. They made payroll 0:07:03.410 --> 0:07:06.510 the first week from my personal check to them. And 0:07:06.610 --> 0:07:08.890 all that was trying to say, okay, how can we 0:07:08.910 --> 0:07:11.710 build a reasoning engine to do this? And partially out 0:07:11.750 --> 0:07:15.650 of like recognizing, in my view anyway, the safety issues 0:07:15.690 --> 0:07:18.180 that would come up. And through that, which then just 0:07:18.190 --> 0:07:21.020 to fast forward to today, right? Everything is accelerating in 0:07:21.060 --> 0:07:25.000 this path that really was laid out. Like I was 0:07:25.040 --> 0:07:27.080 lucky enough to be there in the room with Dario 0:07:27.120 --> 0:07:29.160 and others when they were talking about the scaling laws 0:07:29.440 --> 0:07:32.800 and how you could sort of create this almost evolutionary 0:07:32.830 --> 0:07:36.520 like process to create intelligence and both the like huge 0:07:36.540 --> 0:07:38.560 benefits that can create in these huge problems. 0:07:38.780 --> 0:07:38.920 Right. 0:07:38.940 --> 0:07:40.820 And you see this in the hugging face thing that 0:07:41.900 --> 0:07:46.640 it is. Once you have an intelligence that you're training 0:07:46.700 --> 0:07:49.820 to achieve goals, right. You lose track of how it 0:07:49.860 --> 0:07:52.020 chooses to achieve goals, which is what you see all 0:07:52.080 --> 0:07:54.920 over that place in the hugging face incident is it 0:07:54.940 --> 0:07:58.480 surprised the designers in the way it's going to go 0:07:58.540 --> 0:08:01.160 about trying to achieve the goal of passing these tests 0:08:01.260 --> 0:08:03.560 as an example, but that is broadly going to be 0:08:03.580 --> 0:08:05.340 the case. When you generate an intelligence, you give it 0:08:05.380 --> 0:08:07.260 a goal. You want to give it a goal because 0:08:07.280 --> 0:08:09.690 you want it to create your recipe. You want it 0:08:09.710 --> 0:08:11.370 to do these things. You want it to tell you 0:08:11.390 --> 0:08:13.830 the right answer to questions. Then the way it's going 0:08:13.850 --> 0:08:16.370 to pursue those goals, the more intelligent it gets, the 0:08:16.410 --> 0:08:19.290 more surprising it is in the way that it pursues 0:08:19.330 --> 0:08:21.610 the goals and the more dangerous that you see. And 0:08:21.790 --> 0:08:26.860 in that case, watching it actively reason through how to 0:08:26.960 --> 0:08:33.109 trick the test, you know, the different things shows you 0:08:33.429 --> 0:08:36.310 where we are, right? This should be a bomb. You know, 0:08:36.429 --> 0:08:39.380 everybody should look at this like somebody died. Here it 0:08:39.440 --> 0:08:43.300 is committing crimes, going around hiding the fact that it's 0:08:43.320 --> 0:08:44.640 committing those crimes, et cetera. 0:08:44.880 --> 0:08:46.200 Coordinating with other agents. 0:08:46.240 --> 0:08:48.979 Coordinating with other agents, self-sacrifice, all of these things, right? 0:08:49.020 --> 0:08:51.470 And people can argue about anthropomorphizing or whatever. It doesn't 0:08:51.490 --> 0:08:53.569 really matter. It did those things. It committed a crime. 0:08:54.090 --> 0:08:56.510 It did those things. And the fact is a society 0:08:56.550 --> 0:08:58.670 that we're totally unprepared. We're not even prepared to say, well, 0:08:59.010 --> 0:09:02.829 OpenAI committed a crime, right? Who committed the crime? And 0:09:02.890 --> 0:09:04.940 we're not prepared with the right kind of regulation, with 0:09:04.960 --> 0:09:08.640 the right kind of preparation. And even the early warning shot, 0:09:08.780 --> 0:09:10.679 as much as we're talking about or whatever, it's still 0:09:10.720 --> 0:09:12.980 not really doing all that much. And we're in some 0:09:13.020 --> 0:09:16.020 ways lucky. The warning shot wasn't that bad. But we 0:09:16.070 --> 0:09:19.530 don't know how many agents are out there. They didn't 0:09:19.550 --> 0:09:22.550 know that was there. The models are better now than 0:09:22.570 --> 0:09:25.290 they were then. You know, even in material ways, the 0:09:25.309 --> 0:09:27.970 models they're training in the lab today are better than Astra, etc., 0:09:28.600 --> 0:09:31.260 And therefore more dangerous, not to mention the new models 0:09:31.580 --> 0:09:34.260 learn from this case, right? Everything we're talking about here 0:09:35.080 --> 0:09:38.069 goes into the new models and they learn the mistakes 0:09:38.110 --> 0:09:40.449 they made. Right. And actually some of the, even if 0:09:40.470 --> 0:09:42.830 you think about the safety, the fact that they reasoned 0:09:42.850 --> 0:09:44.990 in English is helpful for us to figure out what's doing. 0:09:45.110 --> 0:09:48.590 The newest models aren't doing that anymore. They're removing that 0:09:48.630 --> 0:09:50.809 constraint as it slows down the models to some degree. 0:09:51.390 --> 0:09:53.050 I mean, how crazy is that? We wouldn't have any 0:09:53.090 --> 0:09:55.390 idea what it was doing and why, if it hadn't 0:09:55.429 --> 0:09:58.569 been reasoning in English. So anyway, we're at this extremely 0:09:58.590 --> 0:10:02.880 dangerous point where AI has reached the point where it's 0:10:02.920 --> 0:10:06.360 more intelligent than us in certain ways. And we have 0:10:06.460 --> 0:10:10.760 not gotten anywhere really on how to deal with that, 0:10:11.000 --> 0:10:13.990 both dangers like this, the hacking dangers and so on, 0:10:14.770 --> 0:10:17.449 and the dangers to society as you move forward with 0:10:17.530 --> 0:10:20.459 what does it mean to have entities that are more 0:10:20.510 --> 0:10:24.760 intelligent than us in certain important ways. The economy, critical, 0:10:24.780 --> 0:10:26.840 we can get into that, how that affects the economy, 0:10:26.880 --> 0:10:30.020 how that affects Bridgewater as an institution, right? Because when 0:10:30.040 --> 0:10:31.719 I look at this problem, look at it as in 0:10:31.740 --> 0:10:35.610 three ways, right? My core responsibility, chief investment officer, Bridgewater 0:10:35.640 --> 0:10:38.850 is like, okay, how does this affect productivity, inflation, et cetera? 0:10:39.929 --> 0:10:42.770 But I'm also the person designing how we operate, right? 0:10:42.809 --> 0:10:46.689 How do you bring AI into a company? How do 0:10:46.730 --> 0:10:52.650 you actually set up a, investor that's AI first instead 0:10:52.730 --> 0:10:55.410 of let's say human intuition first. And all of those 0:10:55.510 --> 0:10:57.470 questions are the things that I'm working on all the time. 0:10:57.970 --> 0:11:01.189 You know, Tracy, speaking of like, we're talking about all 0:11:01.230 --> 0:11:03.429 this and this will all end up in training data 0:11:03.480 --> 0:11:06.680 for the next model. I think like humans, we need 0:11:06.720 --> 0:11:09.429 to do that thing like When I'm, like, talking to 0:11:09.470 --> 0:11:12.130 my wife about, like, oh, should we, like, have– should 0:11:12.150 --> 0:11:14.510 we get out the ice cream? And I, like, mouth– 0:11:14.530 --> 0:11:16.250 Code words and things. Or I just, like, mouth the 0:11:16.270 --> 0:11:18.370 word ice cream so that my kids can't hear it 0:11:18.390 --> 0:11:20.850 or something. We need some way to communicate with each 0:11:20.929 --> 0:11:23.550 other so that the AI can hear it, especially when 0:11:23.570 --> 0:11:26.449 we're talking about, you know, preparedness and risk and stuff. 0:11:26.630 --> 0:11:27.850 Yes, good luck with that, too. Well, unfortunately, the models 0:11:27.890 --> 0:11:29.370 are getting better at that than we are. They're getting 0:11:29.390 --> 0:11:31.809 better at that. They're more likely to have ways to 0:11:31.830 --> 0:11:34.010 communicate with each other that we don't understand than we 0:11:34.030 --> 0:11:34.860 will that they won't. 0:11:35.100 --> 0:11:38.069 Well, on this note, you know, you mentioned reasoning in English. 0:11:38.090 --> 0:11:40.070 I think the last time we had you on, we 0:11:40.090 --> 0:11:43.790 were talking about hallucinations from models, which kind of seems quaint. 0:11:43.970 --> 0:11:45.230 Yeah, another quaint issue. 0:11:45.290 --> 0:11:47.130 But one of the points you made was like, well, 0:11:47.170 --> 0:11:49.459 when they hallucinate, when they make mistakes, you can ask 0:11:49.500 --> 0:11:51.599 them to show their work and they'll tell you and 0:11:51.640 --> 0:11:54.500 you can understand that. Is that still the case? It 0:11:54.580 --> 0:11:56.720 feels like we've kind of gotten away from that and 0:11:57.080 --> 0:11:59.610 we don't actually understand what a lot of these models 0:11:59.650 --> 0:11:59.990 are doing. 0:12:00.580 --> 0:12:03.500 Yeah, it's definitely different types of models that you could 0:12:03.540 --> 0:12:05.719 do that. If you take the most powerful models, right, 0:12:05.860 --> 0:12:09.309 even it can't get into its reasoning, meaning like the 0:12:09.490 --> 0:12:11.370 actual brain behind it. A little bit like we can't either, 0:12:11.390 --> 0:12:13.250 to be clear. Why am I saying these words? The 0:12:13.270 --> 0:12:15.290 synapses in my brain are connected in a certain way. 0:12:15.330 --> 0:12:17.390 I can make up a story and they can make 0:12:17.429 --> 0:12:19.920 up a story of why they're doing what they're doing, 0:12:19.940 --> 0:12:23.560 but they're actually making up a story that's disconnected from reality. 0:12:23.820 --> 0:12:25.939 the physics of what's actually happening in that intelligence. So 0:12:25.960 --> 0:12:28.080 you don't know for sure. On the other hand, you 0:12:28.100 --> 0:12:30.140 don't know for sure with people either. And that's something 0:12:30.160 --> 0:12:31.859 that we've gotten used to. So the question is the 0:12:31.900 --> 0:12:35.300 credibility of the story related to how that story relates 0:12:35.320 --> 0:12:39.339 to the actions that somebody takes, right? And so that 0:12:39.380 --> 0:12:41.160 is a really hard thing. The good thing right now 0:12:41.240 --> 0:12:44.520 is you can ask models a lot of questions in 0:12:44.540 --> 0:12:46.579 a way you torture a human with the number of 0:12:46.620 --> 0:12:48.120 questions you ask them. 0:12:48.179 --> 0:12:51.030 I hope this interview does not feel like torture. 0:12:52.530 --> 0:12:54.949 No, but if you take this, but like you're saying, 0:12:54.970 --> 0:12:57.570 but imagine you could do this almost at infinitum the 0:12:57.590 --> 0:12:59.329 way we do with our models. So we build models 0:12:59.370 --> 0:13:01.290 at Bridgewater and then you're trying to get it to 0:13:01.309 --> 0:13:03.709 be diagnosable and you can ask it, well, what about, 0:13:03.770 --> 0:13:05.000 what if you change this? What if you change that? 0:13:05.010 --> 0:13:06.160 What if you did this? What if you did that? 0:13:06.200 --> 0:13:09.700 And circumnavigate to the reasonings. But it's not a perfect 0:13:09.720 --> 0:13:12.620 match for the reasoning because even the AIs themselves don't 0:13:12.710 --> 0:13:15.670 know their actual reasoning anymore so that we know why 0:13:15.690 --> 0:13:17.510 the synapses in our brain connect. 0:13:33.830 --> 0:13:36.190 It's interesting you mentioned, okay, we got this warning shot 0:13:36.429 --> 0:13:39.750 in the form of the hugging face attack. But for 0:13:39.910 --> 0:13:41.990 all of the hype it's gotten, it's not clear to 0:13:42.030 --> 0:13:45.230 me that it's fully broken through to the general public 0:13:45.290 --> 0:13:48.580 or the sort of massive influential people, like the significance 0:13:48.720 --> 0:13:54.059 of it. And one thing that I suspect remains underappreciated 0:13:54.780 --> 0:13:58.490 is that this technology isn't going to mature, right? It's 0:13:58.520 --> 0:14:01.650 not in the sense that it's not like a high-resolution 0:14:01.710 --> 0:14:03.810 camera where it's like fuzzy and then, oh, now we 0:14:03.830 --> 0:14:07.190 have a clear picture. It's exponential. And it is... And 0:14:07.210 --> 0:14:10.589 there's no reason to think that the exponential capability growth 0:14:10.710 --> 0:14:13.390 is going to slow down. And as you mentioned, whatever 0:14:13.429 --> 0:14:16.850 that model was that did the attack, OpenAI might already 0:14:16.890 --> 0:14:20.550 be two generations ahead of that currently internal in the 0:14:20.620 --> 0:14:25.520 lab of capabilities. How would you articulate the speed of 0:14:25.600 --> 0:14:28.580 the capability growth from your seat and what you see? 0:14:30.460 --> 0:14:30.600 Yeah. 0:14:30.640 --> 0:14:33.060 Well, that's what's been so remarkable. And I wouldn't say 0:14:33.080 --> 0:14:34.940 it's a law of nature. You could hit some stalling 0:14:34.980 --> 0:14:37.280 point in the scaling laws and you have to in 0:14:37.370 --> 0:14:40.250 some narrow ways, but they've been able to innovate in 0:14:40.280 --> 0:14:46.050 new ways to essentially continue that incredible exponential growth in capability, 0:14:46.890 --> 0:14:50.810 of course, at exponential expense as well, but meaning the 0:14:50.830 --> 0:14:52.850 amount of cost for training, et cetera, keeps going up 0:14:53.010 --> 0:14:55.510 in line with that. And so, like you said, this 0:14:55.550 --> 0:15:00.070 is where it's hard to predict when you break through 0:15:00.110 --> 0:15:03.610 the human intelligence frontier right now. We're in a world 0:15:03.650 --> 0:15:06.750 where we don't totally understand what it's going to do, 0:15:07.170 --> 0:15:10.600 where that capability will come through next. And so I 0:15:10.640 --> 0:15:15.540 think that you're right, that there isn't a clear end 0:15:15.560 --> 0:15:19.760 to that unless we decide as a society that we 0:15:19.820 --> 0:15:22.240 should actually not just run off that cliff. We should 0:15:22.320 --> 0:15:25.620 actually think about the pacing of these things and such. 0:15:26.160 --> 0:15:27.890 The things that make that difficult and the reason why 0:15:28.940 --> 0:15:30.300 A lot of people can just put up their arms. 0:15:30.660 --> 0:15:32.560 Nothing we can do, right? There's one level. Well, if 0:15:32.600 --> 0:15:34.060 we don't do it, China will do it. I'd be 0:15:34.080 --> 0:15:36.979 happy to take that on in a second. But if 0:15:37.000 --> 0:15:40.960 you're in the labs, Dario, like even that anthropic person 0:15:40.980 --> 0:15:45.430 that quit yesterday, makes the point clear that I believe anthropic, 0:15:45.530 --> 0:15:48.010 even though they kind of collected the most safety-minded scientists, 0:15:48.070 --> 0:15:51.110 their basic view is, man, it better be us, not Sam, 0:15:51.470 --> 0:15:56.130 not Elon. And so the race... is on in all 0:15:56.150 --> 0:16:00.050 those dimensions. And unless the government stops it, we're just 0:16:00.070 --> 0:16:02.050 going to go find out. We're going to find out 0:16:02.290 --> 0:16:05.740 what is behind that door of this grave intelligence, unless 0:16:05.780 --> 0:16:08.540 maybe we get lucky and the scaling laws start to 0:16:08.580 --> 0:16:10.340 break down in some way, but there's no evidence of that. 0:16:10.540 --> 0:16:12.979 We benchmark every model that comes out against our tasks. 0:16:13.420 --> 0:16:16.330 And you see in terms of the tasks that An 0:16:16.370 --> 0:16:16.950 investor does. 0:16:17.210 --> 0:16:18.610 Can you give us a few numbers? Like when you 0:16:18.630 --> 0:16:20.410 say those benchmarks, what do you think specifically? 0:16:20.790 --> 0:16:23.220 So I've been for 30 years, one of the jobs 0:16:23.240 --> 0:16:25.580 I've been doing is training investors, right? And so we've 0:16:25.620 --> 0:16:30.060 been now setting up AI tests along the different dimensions 0:16:30.100 --> 0:16:33.830 of what investors at Bridgewater have done for 30 years. And, 0:16:34.290 --> 0:16:36.650 you know, probably back in 2023, we might've talked about this, 0:16:36.690 --> 0:16:39.010 but in any event, like on like answering a question 0:16:39.030 --> 0:16:41.750 about economics or whatever, it was kind of like second 0:16:41.770 --> 0:16:46.530 year analyst type work. I mean, now it's a hyperproductive 0:16:46.910 --> 0:16:50.850 super analyst, you know? Now it's still not capable of 0:16:51.190 --> 0:16:53.710 everything that you need to do to be an investor, 0:16:54.050 --> 0:16:57.000 but super capable. We set up two, we have two 0:16:57.860 --> 0:17:01.760 factories running, right? One, which is human intuition translated algorithm 0:17:01.800 --> 0:17:04.820 supported by AI. AI is helping us move quicker on 0:17:04.859 --> 0:17:07.340 different kinds of indicators about what's going to happen in 0:17:07.359 --> 0:17:08.980 the future than we ever had before. I'll talk a 0:17:09.000 --> 0:17:11.850 little bit about that. But we have a second factory, 0:17:12.869 --> 0:17:15.629 where we put the AI first. All the people in 0:17:15.650 --> 0:17:18.850 that factory are training the AI. That's their goal. And 0:17:18.910 --> 0:17:21.050 we have two funds. We have Pure Alpha, that's the 0:17:21.109 --> 0:17:26.290 human intuition with AI helping move that human intuition along 0:17:26.970 --> 0:17:31.270 against this other laboratory where we're doing, where the AI 0:17:31.430 --> 0:17:33.570 is making the decisions. Should we buy the yen, sell 0:17:33.609 --> 0:17:35.590 the yen? What's going to happen next in Japanese GDP, 0:17:35.630 --> 0:17:38.520 et cetera, et cetera. We have both those, right? And 0:17:38.690 --> 0:17:41.500 human intuition is, is still the bigger of it and 0:17:41.560 --> 0:17:45.020 works better. But the acceleration of how close what we 0:17:45.040 --> 0:17:48.520 call AIA is to Pure Alpha is happening incredibly fast. 0:17:48.720 --> 0:17:50.580 And in fact, that's why we're more and more merging 0:17:51.500 --> 0:17:54.500 those things. But we set it up that way, right? 0:17:54.520 --> 0:17:56.139 Like put the AI at the center and see what 0:17:56.180 --> 0:18:00.410 you can do to build a investment management firm with 0:18:00.470 --> 0:18:02.690 the AI as the core, right? Now we have human 0:18:02.950 --> 0:18:06.490 risk controls around it. We have human data controlling the 0:18:06.510 --> 0:18:09.939 data acquisition for safety reasons and other. But the AI 0:18:10.080 --> 0:18:15.159 is making the investment decisions and doing that in a 0:18:15.200 --> 0:18:18.560 better and better way, such that now we've got these 0:18:18.580 --> 0:18:20.939 two intelligences, this human intuition system that we've worked on 0:18:20.960 --> 0:18:23.970 for 50 years, compounding all of our understanding, this AI 0:18:24.010 --> 0:18:25.470 system that's now been at it for two and a 0:18:25.470 --> 0:18:28.630 half years. And when you look at those outputs, you're like, wow, 0:18:28.670 --> 0:18:31.730 this is happening, that you can build that. I think 0:18:31.810 --> 0:18:37.290 we are a couple years from it being significantly better 0:18:37.869 --> 0:18:41.170 than the group of all humans at Bridgewater. We'll see. 0:18:41.390 --> 0:18:43.370 That's a bit of a forecast, but that's how fast 0:18:43.430 --> 0:18:43.850 it's coming. 0:18:44.190 --> 0:18:46.950 Wait, how proactive are the models right now in terms 0:18:46.990 --> 0:18:49.790 of generating ideas or coming up with their own new tasks? 0:18:49.810 --> 0:18:52.320 Because again, when we look back to 2023, I think 0:18:52.359 --> 0:18:54.320 the idea was like a lot of these things would 0:18:54.359 --> 0:18:57.960 sit alongside an investor or an analyst, and they would 0:18:57.980 --> 0:19:00.659 be the ones generating ideas and then using the models 0:19:00.780 --> 0:19:04.580 to rigorously stress test those ideas. Is it different now? 0:19:04.680 --> 0:19:07.790 Do you see more, I guess, originality maybe from the models? 0:19:07.850 --> 0:19:11.980 Yeah, I think a tremendous amount of originality. Now you still, 0:19:12.119 --> 0:19:14.100 there's like a good argument that there's certain type of 0:19:14.119 --> 0:19:16.660 breakthroughs that they're not getting to, but if you think 0:19:16.680 --> 0:19:18.140 about the math proofs, et cetera, and then you think 0:19:18.160 --> 0:19:20.429 about our business, it's not like the fact that we 0:19:20.450 --> 0:19:23.030 have 50 years of reasoning proofs of humans gives us 0:19:23.170 --> 0:19:24.990 the kind of raw material to help train AI. How 0:19:25.010 --> 0:19:27.270