1 00:00:10,240 --> 00:00:13,520 Speaker 1: Hello, and welcome to another episode of the Odd Lots Podcast. 2 00:00:13,640 --> 00:00:15,080 Speaker 1: I'm Tracy Alloway. 3 00:00:14,760 --> 00:00:15,960 Speaker 2: And I'm Joe Wisenthal. 4 00:00:16,360 --> 00:00:19,759 Speaker 1: Joe, I think it's fair to say there is a 5 00:00:19,760 --> 00:00:24,239 Speaker 1: lot of excitement about investing in AI. There is also 6 00:00:24,280 --> 00:00:27,160 Speaker 1: a lot of excitement about using AI to invest. 7 00:00:27,920 --> 00:00:31,400 Speaker 2: Yes, I mean it's I think there's like a new 8 00:00:31,440 --> 00:00:34,720 Speaker 2: like chat ETF I saw an ad for and there's like, oh, 9 00:00:34,720 --> 00:00:37,440 Speaker 2: we're getting No. I think I saw another like project. 10 00:00:37,440 --> 00:00:39,600 Speaker 2: It was like, we're gonna have chat GPT pick the 11 00:00:39,640 --> 00:00:42,639 Speaker 2: stocks for us, And you know, I get it. It's 12 00:00:42,720 --> 00:00:45,000 Speaker 2: kind of exciting and maybe there's some new way of 13 00:00:45,120 --> 00:00:48,839 Speaker 2: like these super advanced digital brains that can beat the market, 14 00:00:48,880 --> 00:00:51,400 Speaker 2: et cetera. But like, I don't totally get it. 15 00:00:51,880 --> 00:00:52,160 Speaker 3: Well. 16 00:00:52,280 --> 00:00:55,360 Speaker 1: I also feel like there's a tendency nowadays for people 17 00:00:55,400 --> 00:01:00,560 Speaker 1: to talk about artificial intelligence in a sort of abstract manner. 18 00:01:00,680 --> 00:01:04,600 Speaker 1: You hear people bring up AI almost as a synonym 19 00:01:04,800 --> 00:01:08,520 Speaker 1: for just software at this point. I think you pointed 20 00:01:08,560 --> 00:01:13,360 Speaker 1: out recently that the Kroger CEO mentioned AI like times 21 00:01:13,400 --> 00:01:16,520 Speaker 1: on the earnings call. So a supermarket chain, right. 22 00:01:16,480 --> 00:01:21,000 Speaker 2: Yeah, And you know it's like machine learning, tech, algebra, algorithms, 23 00:01:21,040 --> 00:01:24,440 Speaker 2: it's all existed for a long time quantitative investing, but 24 00:01:24,480 --> 00:01:27,839 Speaker 2: it feels like because of the excitement around a few 25 00:01:27,880 --> 00:01:31,480 Speaker 2: specific consumer phasing products that have been unveiled over the 26 00:01:31,600 --> 00:01:34,520 Speaker 2: last six months and the way they've captured people's attention, 27 00:01:35,000 --> 00:01:37,480 Speaker 2: people like you know, suddenly there's a lot of interest 28 00:01:37,480 --> 00:01:41,279 Speaker 2: in like, how are companies using this tech to do something? 29 00:01:41,480 --> 00:01:44,319 Speaker 1: Yeah, well, I'm glad you mentioned that because today we 30 00:01:44,400 --> 00:01:47,200 Speaker 1: really do have the perfect guest. This is someone we've 31 00:01:47,280 --> 00:01:52,480 Speaker 1: actually spoken to about AI before last year, in fact, 32 00:01:52,840 --> 00:01:56,480 Speaker 1: someone who is at a firm that has a lot 33 00:01:56,520 --> 00:02:01,720 Speaker 1: of experience using machine learning and of different types, and 34 00:02:01,760 --> 00:02:05,160 Speaker 1: we're going to get into the differences between all those technologies. 35 00:02:05,520 --> 00:02:07,680 Speaker 1: I'm very pleased to say we're going to be speaking 36 00:02:07,760 --> 00:02:11,280 Speaker 1: once again with Greg Jensen, the co chief investment officer 37 00:02:11,560 --> 00:02:14,560 Speaker 1: at Bridgewater Associates. So, Greg, thank you so much for 38 00:02:14,600 --> 00:02:16,359 Speaker 1: coming back on OD thoughts. 39 00:02:16,800 --> 00:02:18,799 Speaker 3: Yeah, it's great to be here. Exciting topic. 40 00:02:19,280 --> 00:02:23,240 Speaker 1: Yeah, So I actually revisited our conversation from last year, 41 00:02:23,520 --> 00:02:25,799 Speaker 1: I think it was in May of twenty twenty two, 42 00:02:25,880 --> 00:02:29,920 Speaker 1: and you said two things that stuck out in retrospect. 43 00:02:30,000 --> 00:02:32,480 Speaker 1: So number one, you said that markets had further to fall, 44 00:02:32,560 --> 00:02:36,880 Speaker 1: which turned out to be correct. And two you brought 45 00:02:36,960 --> 00:02:41,400 Speaker 1: up artificial intelligence as a major point of interest for Bridgewater, 46 00:02:41,480 --> 00:02:44,520 Speaker 1: And this was all before chat GPT really became a 47 00:02:44,560 --> 00:02:48,359 Speaker 1: thing and everyone started talking about AI at every single 48 00:02:48,440 --> 00:02:52,000 Speaker 1: conference and earnings call and so on. So I guess, 49 00:02:52,080 --> 00:02:54,400 Speaker 1: just to begin with, maybe you could lay the scene 50 00:02:54,639 --> 00:02:58,480 Speaker 1: and going back to Joe's point in the intro, we 51 00:02:58,520 --> 00:03:02,560 Speaker 1: are used to hearing these terms. So Bridgewater does machine 52 00:03:02,639 --> 00:03:07,920 Speaker 1: learning and systematic strategy strategies and quantitative trading strategies and 53 00:03:08,000 --> 00:03:11,840 Speaker 1: AI and things like that. What's the difference between all 54 00:03:11,880 --> 00:03:14,600 Speaker 1: of these things and how do they relate to each 55 00:03:14,600 --> 00:03:16,480 Speaker 1: other at a firm like Bridgewater. 56 00:03:18,919 --> 00:03:21,400 Speaker 3: Yeah, great question. So I think to answer that, let 57 00:03:21,440 --> 00:03:23,040 Speaker 3: me take a step back for a second and give 58 00:03:23,040 --> 00:03:24,640 Speaker 3: you a little bit of my background, because it all 59 00:03:25,280 --> 00:03:28,160 Speaker 3: kind of comes together in a way. You can connect 60 00:03:28,200 --> 00:03:32,240 Speaker 3: these different pieces. So you know, even as a kid 61 00:03:32,320 --> 00:03:36,000 Speaker 3: or whatever, I was certainly interested in kind of translating 62 00:03:36,440 --> 00:03:42,000 Speaker 3: and predicting things using some mix of my thinking and technology. 63 00:03:42,080 --> 00:03:45,200 Speaker 3: So I can think back to in the late eighties 64 00:03:45,560 --> 00:03:49,080 Speaker 3: using stratomatic baseball cards, know what they are, but programming 65 00:03:49,120 --> 00:03:51,160 Speaker 3: them into computers to try to calculate the way to 66 00:03:51,200 --> 00:03:55,960 Speaker 3: create the best baseball lineup and use that in fantasy 67 00:03:56,000 --> 00:04:00,880 Speaker 3: baseball type situations and similar things wither and whatever, and 68 00:04:00,920 --> 00:04:03,800 Speaker 3: try to learn how to kind of use technology to 69 00:04:03,800 --> 00:04:08,120 Speaker 3: combine with human intuition to get at what was different 70 00:04:08,120 --> 00:04:10,920 Speaker 3: ways to create edges. And then in college when I 71 00:04:10,920 --> 00:04:13,520 Speaker 3: heard about Bridgewater Bridge Order. It was a tiny place 72 00:04:13,520 --> 00:04:17,080 Speaker 3: at the time, but the basic idea that there was 73 00:04:17,120 --> 00:04:19,040 Speaker 3: a place where we were trying to understand the world, 74 00:04:19,440 --> 00:04:22,799 Speaker 3: trying to predict what was next, but doing that by 75 00:04:22,960 --> 00:04:27,239 Speaker 3: taking human intuition and translating that into algorithms to predict 76 00:04:27,240 --> 00:04:30,080 Speaker 3: what was next kind of mixed two things that I loved. 77 00:04:30,080 --> 00:04:31,960 Speaker 3: I love to try to understand the world, and I 78 00:04:32,000 --> 00:04:34,560 Speaker 3: love the idea of having the discipline to write down 79 00:04:34,600 --> 00:04:36,960 Speaker 3: what you believed and stress test what you believed and 80 00:04:37,120 --> 00:04:39,240 Speaker 3: utilize that. Right. So, if you go back, and this 81 00:04:39,279 --> 00:04:43,040 Speaker 3: is now in the nineties kind of where artificial intelligence 82 00:04:43,160 --> 00:04:46,000 Speaker 3: was at the time, most of the focus was still 83 00:04:46,040 --> 00:04:48,359 Speaker 3: on expert systems, was still on the notion that you 84 00:04:48,400 --> 00:04:53,800 Speaker 3: could take human intuition, you could translate that into algorithms, 85 00:04:53,839 --> 00:04:55,800 Speaker 3: and if you did enough of that, if you kept 86 00:04:55,920 --> 00:05:00,840 Speaker 3: kind of representing things in symbolic algorithm that you could 87 00:05:00,880 --> 00:05:04,440 Speaker 3: build enough human knowledge to get kind of a superpowered human. 88 00:05:05,160 --> 00:05:08,880 Speaker 3: And Bridgewater was a rare example of where that worked. 89 00:05:08,920 --> 00:05:10,960 Speaker 3: Where given the focus of trying to predict what was 90 00:05:11,040 --> 00:05:14,400 Speaker 3: next in markets, given the incredible investment that we made 91 00:05:14,920 --> 00:05:18,240 Speaker 3: into creating the technology to take human intuition and translate 92 00:05:18,320 --> 00:05:22,360 Speaker 3: that into algorithms and stress tests, that it's incredibly successful 93 00:05:22,920 --> 00:05:27,039 Speaker 3: expert system essentially that was built over the years, I'd 94 00:05:27,040 --> 00:05:32,040 Speaker 3: say probably the most profitable expert system out there. And 95 00:05:32,080 --> 00:05:34,960 Speaker 3: that's really what Bridgewater has been about, which is building 96 00:05:34,960 --> 00:05:37,800 Speaker 3: this great technology to help us take human intuition out 97 00:05:37,800 --> 00:05:40,920 Speaker 3: of the brain, get it into technology where it's both 98 00:05:40,960 --> 00:05:46,000 Speaker 3: then readable by let's say investment experts, but also runs 99 00:05:46,080 --> 00:05:49,280 Speaker 3: on a technology basis. And that's kind of where algorithms, 100 00:05:49,360 --> 00:05:52,880 Speaker 3: let's say, the mix of algorithms and human intuition it 101 00:05:52,960 --> 00:05:55,240 Speaker 3: was really important. You know, if you go through the 102 00:05:55,279 --> 00:05:58,280 Speaker 3: history of our competitors, they're littered by people that tried 103 00:05:58,279 --> 00:06:01,520 Speaker 3: to do something more statistical, meaning that they would take 104 00:06:01,560 --> 00:06:04,920 Speaker 3: the data, run regressions, and then after regressions, let's say 105 00:06:04,920 --> 00:06:07,880 Speaker 3: basic machine learning techniques to predict the future. And the 106 00:06:07,920 --> 00:06:11,480 Speaker 3: problem that always had is that there wasn't enough data like, 107 00:06:11,520 --> 00:06:15,960 Speaker 3: the truth is that market data isn't like the data 108 00:06:16,000 --> 00:06:18,039 Speaker 3: and the physical world in the sense that a you 109 00:06:18,080 --> 00:06:20,760 Speaker 3: only have one run through human history, you don't have 110 00:06:20,880 --> 00:06:23,840 Speaker 3: very many cycles, even cycles that debt cycles could take 111 00:06:23,839 --> 00:06:26,400 Speaker 3: seventy years to play out. Economic cycles tend to plan 112 00:06:26,520 --> 00:06:29,440 Speaker 3: around for seven years. There's just not enough data to 113 00:06:29,560 --> 00:06:35,120 Speaker 3: represent the world. And secondly that the game changes as 114 00:06:35,160 --> 00:06:38,839 Speaker 3: participants learned, So the existence of algorithms, as an example, 115 00:06:39,320 --> 00:06:42,400 Speaker 3: change the nature of markets such that the history that 116 00:06:42,640 --> 00:06:45,320 Speaker 3: preceded it was less and less relevant to the world 117 00:06:45,400 --> 00:06:49,279 Speaker 3: you're living in. So those are big problems with let's 118 00:06:49,279 --> 00:06:53,520 Speaker 3: say a more pure statistical technique to markets. So you 119 00:06:53,600 --> 00:06:57,120 Speaker 3: had to get to a world where statistical techniques or 120 00:06:57,120 --> 00:07:02,160 Speaker 3: machine learning could substitute for human intuition. And that's really 121 00:07:02,160 --> 00:07:06,480 Speaker 3: where kind of the exciting leaps are. Now that you're 122 00:07:06,520 --> 00:07:09,280 Speaker 3: getting closer. It's not totally there, but you're much closer 123 00:07:09,279 --> 00:07:13,400 Speaker 3: than you've ever been, where large language models actually allow 124 00:07:13,440 --> 00:07:16,520 Speaker 3: a path to something that at least mimics human intuition, 125 00:07:16,640 --> 00:07:21,880 Speaker 3: if not is human intuition, and that you can then 126 00:07:21,920 --> 00:07:25,280 Speaker 3: combine that with other techniques and suddenly you have a 127 00:07:25,400 --> 00:07:28,520 Speaker 3: much more powerful set of tools that can deal at 128 00:07:28,600 --> 00:07:30,680 Speaker 3: least in take a big leap forward on dealing with 129 00:07:30,720 --> 00:07:33,800 Speaker 3: the problem of very small data sets and the fact 130 00:07:33,800 --> 00:07:36,640 Speaker 3: that the world changes as people learn in a way 131 00:07:36,760 --> 00:07:39,640 Speaker 3: that up until the big breakthroughs in large language models, 132 00:07:40,320 --> 00:07:44,000 Speaker 3: I think we're much further away. So that's a huge 133 00:07:44,160 --> 00:07:49,040 Speaker 3: change in the limits of ways that statistical machine learning 134 00:07:49,440 --> 00:07:53,480 Speaker 3: could affect something with small amounts of data, something where 135 00:07:53,520 --> 00:07:57,680 Speaker 3: the future varies from the past. All of those problems 136 00:07:57,680 --> 00:08:00,520 Speaker 3: we're closer to having. At least way is to take 137 00:08:00,560 --> 00:08:03,320 Speaker 3: on more and more of what humans have done at Bridgewater, 138 00:08:03,400 --> 00:08:06,320 Speaker 3: what humans generally do in investment management firms, And that's 139 00:08:06,520 --> 00:08:09,280 Speaker 3: that's a huge leap forward that's going on now. 140 00:08:09,640 --> 00:08:12,920 Speaker 2: I have one very short quick question. I realized just 141 00:08:13,040 --> 00:08:16,480 Speaker 2: know that not long after we talked to last year, 142 00:08:16,720 --> 00:08:19,360 Speaker 2: last spring, like a month later, you won your first 143 00:08:19,600 --> 00:08:22,920 Speaker 2: World Series of poker bracelets. So congratulations on that. I 144 00:08:22,920 --> 00:08:24,640 Speaker 2: at least say that because you mentioned poker, did you 145 00:08:24,680 --> 00:08:26,080 Speaker 2: play the World Series this year? 146 00:08:27,720 --> 00:08:30,240 Speaker 3: I'm heading out actually after this. 147 00:08:30,160 --> 00:08:34,880 Speaker 2: Because I know there are okay, congrats and good luck. 148 00:08:34,920 --> 00:08:36,680 Speaker 3: Yeah, And it kind of connects to this because I 149 00:08:36,720 --> 00:08:38,600 Speaker 3: never get to I don't get to play very much poker, 150 00:08:38,600 --> 00:08:44,800 Speaker 3: but I really studied what machines were learning about poker. 151 00:08:45,080 --> 00:08:46,760 Speaker 3: So much has been learned in the last five years, 152 00:08:46,760 --> 00:08:51,520 Speaker 3: ten years, and and one of the you know, basically 153 00:08:51,520 --> 00:08:55,480 Speaker 3: trying to translate that into intuitions that I could use, 154 00:08:55,600 --> 00:08:58,679 Speaker 3: you know that basically can't actually replicate Peter Place. Poker 155 00:08:58,679 --> 00:09:02,640 Speaker 3: are very complex way, but you can pull the concepts 156 00:09:02,679 --> 00:09:05,720 Speaker 3: out right. And this actually mirrors to what part of 157 00:09:05,760 --> 00:09:08,640 Speaker 3: what we're doing at Bridge Order, which is that as 158 00:09:08,679 --> 00:09:12,600 Speaker 3: you get to computer generated theories that if you can 159 00:09:12,679 --> 00:09:15,960 Speaker 3: pull the concepts out of these complex algorithms, you know, 160 00:09:16,040 --> 00:09:19,400 Speaker 3: you can make more of an assessment human assessment of 161 00:09:19,480 --> 00:09:21,720 Speaker 3: whether they make sense and what the problems might be. 162 00:09:22,200 --> 00:09:24,280 Speaker 3: And that's really a big deal. So there's actually a 163 00:09:24,320 --> 00:09:29,240 Speaker 3: link between what I'm doing in poker, imperfectly for sure, 164 00:09:29,360 --> 00:09:31,800 Speaker 3: and many of the concepts that we're trying to apply 165 00:09:32,920 --> 00:09:36,640 Speaker 3: at Bridge Order. And like you said, just we had 166 00:09:36,679 --> 00:09:40,640 Speaker 3: talked kind of before the lms had really hit the 167 00:09:40,640 --> 00:09:42,599 Speaker 3: public scene. But yeah, I mean, just to give you 168 00:09:42,600 --> 00:09:44,760 Speaker 3: a little bit of background for me the you know, 169 00:09:44,960 --> 00:09:48,440 Speaker 3: if you go back to twenty twelve, First off, we 170 00:09:48,559 --> 00:09:51,160 Speaker 3: brought Dave Ferriucci, who had run the Watson project at 171 00:09:51,160 --> 00:09:54,440 Speaker 3: IBM that had beat Jeopardy into Bridgewater, and that was 172 00:09:54,440 --> 00:09:56,920 Speaker 3: that was a time when I was trying to experiment with, okay, 173 00:09:56,920 --> 00:09:59,480 Speaker 3: what can we do with more machine learning techniques? And 174 00:09:59,600 --> 00:10:02,160 Speaker 3: Dave was trying to take what he had done to 175 00:10:02,160 --> 00:10:04,160 Speaker 3: win a Jeopardy but actually put in more of a 176 00:10:04,160 --> 00:10:07,920 Speaker 3: reasoning engine, because while what happened to on Jeopardy was impressive, 177 00:10:07,920 --> 00:10:09,880 Speaker 3: it was pure data. It had no idea why it 178 00:10:09,920 --> 00:10:13,600 Speaker 3: was doing what it was doing, and therefore really a 179 00:10:13,600 --> 00:10:16,000 Speaker 3: lot of the path with Watson or whatever was going 180 00:10:16,040 --> 00:10:21,000 Speaker 3: to be very hard to move forward with because because 181 00:10:21,360 --> 00:10:23,680 Speaker 3: at its end, it was just statistical and it didn't 182 00:10:23,720 --> 00:10:27,640 Speaker 3: really have any reasoning capability. So Dave came to Bridge 183 00:10:27,720 --> 00:10:29,599 Speaker 3: Order and later partnered with Bridge Order roll out of 184 00:10:29,640 --> 00:10:34,959 Speaker 3: company Elemental Cognition that's focused on using large language models, 185 00:10:34,960 --> 00:10:39,680 Speaker 3: et cetera, but overlaying a reasoning engine that essentially helps 186 00:10:39,679 --> 00:10:43,360 Speaker 3: with things like hallucinations that out that large language models 187 00:10:43,400 --> 00:10:46,040 Speaker 3: have and focus on how what is human reasoning and 188 00:10:46,040 --> 00:10:48,800 Speaker 3: how does it work and how does that limit views 189 00:10:48,840 --> 00:10:52,240 Speaker 3: that are unlikely to be true? So that's one thing, 190 00:10:52,240 --> 00:10:54,600 Speaker 3: And then in twenty sixteen or seventeen, I was introduced 191 00:10:54,640 --> 00:10:57,520 Speaker 3: to open ai and actually as they transition from a 192 00:10:57,720 --> 00:11:02,280 Speaker 3: charity to a company. I was one of the in 193 00:11:02,320 --> 00:11:04,439 Speaker 3: that first round, and it was like met a lot 194 00:11:04,440 --> 00:11:08,040 Speaker 3: of the people and looked hard at their vision to 195 00:11:08,720 --> 00:11:12,200 Speaker 3: using scale and technical scale to build general intelligence and 196 00:11:12,200 --> 00:11:16,280 Speaker 3: build reasoning. So I both was working with Dave Rucci 197 00:11:16,360 --> 00:11:18,440 Speaker 3: and sort of understood many of the people at open 198 00:11:18,480 --> 00:11:21,520 Speaker 3: ai at the time and moving forward with those things. 199 00:11:21,520 --> 00:11:23,719 Speaker 3: And then I was literally the first check for anthropic 200 00:11:23,840 --> 00:11:27,760 Speaker 3: and other large language model kind of people that had 201 00:11:27,800 --> 00:11:30,880 Speaker 3: been at open AI. And so I've been passionate about this, 202 00:11:31,040 --> 00:11:34,600 Speaker 3: realized trying to take different paths to how will we 203 00:11:34,720 --> 00:11:39,880 Speaker 3: build a reasoning engine to overlay on statistical things, and 204 00:11:40,080 --> 00:11:43,040 Speaker 3: a couple of different approaches that were being applied at 205 00:11:43,040 --> 00:11:46,120 Speaker 3: the time, and obviously different they panned out to a 206 00:11:46,120 --> 00:11:49,679 Speaker 3: different degree, but many things are coming together now to say, Okay, 207 00:11:49,760 --> 00:11:53,680 Speaker 3: you can actually in a way at a pace and 208 00:11:53,679 --> 00:11:57,200 Speaker 3: a speed humans can never do, you could replicate human reasoning. 209 00:11:57,840 --> 00:12:00,440 Speaker 3: And that's a huge deal. And if you could really 210 00:12:00,440 --> 00:12:02,680 Speaker 3: break through that, you could start to apply it in 211 00:12:02,720 --> 00:12:05,720 Speaker 3: so many ways in our industry, I believe, and obviously 212 00:12:05,760 --> 00:12:08,520 Speaker 3: way beyond our industry. 213 00:12:23,679 --> 00:12:27,600 Speaker 2: You talked about earlier generations trying to embed human knowledge. 214 00:12:27,600 --> 00:12:29,480 Speaker 2: And I'm wondering, you know if an analogy is like 215 00:12:29,880 --> 00:12:31,840 Speaker 2: I remember when Deep Blue came out and they had 216 00:12:31,880 --> 00:12:35,120 Speaker 2: all the grand masters sort of work with IBM to 217 00:12:35,240 --> 00:12:37,280 Speaker 2: like come up with this a great computer program that 218 00:12:37,400 --> 00:12:40,839 Speaker 2: was basically as good or eventually better than Gary Kasparov. 219 00:12:41,520 --> 00:12:44,679 Speaker 2: But then the next generation of a chess computers didn't 220 00:12:44,679 --> 00:12:46,960 Speaker 2: even have the grand masters playing it. It just learned the 221 00:12:47,000 --> 00:12:52,199 Speaker 2: game from ground up and crushed those crush those previous generation. 222 00:12:52,640 --> 00:12:54,920 Speaker 2: Is that sort of the what we're talking about here 223 00:12:55,000 --> 00:12:58,839 Speaker 2: with with the transition from earlier engines to the new 224 00:12:58,880 --> 00:13:01,480 Speaker 2: sort of LLM folks, which is like the sort of 225 00:13:01,559 --> 00:13:04,679 Speaker 2: reasoning come becomes comes out of the computer rather than 226 00:13:04,679 --> 00:13:07,199 Speaker 2: having to be taught directly by the experts. 227 00:13:08,440 --> 00:13:10,920 Speaker 3: Yeah, I think something like that is happening. Right. You 228 00:13:10,960 --> 00:13:13,920 Speaker 3: got that in chess because once you had the ability 229 00:13:14,640 --> 00:13:17,240 Speaker 3: you had enough data and enough compute, you were able 230 00:13:17,440 --> 00:13:21,679 Speaker 3: to do enough sampling that the pure that you got 231 00:13:21,720 --> 00:13:24,839 Speaker 3: to the point where the pure data process, with good 232 00:13:24,920 --> 00:13:27,160 Speaker 3: human intuition on how to build that data process, but 233 00:13:27,960 --> 00:13:31,880 Speaker 3: a data process, was able to beat that those rules 234 00:13:31,880 --> 00:13:36,000 Speaker 3: based things. Now, chess, unlike markets, is you know a 235 00:13:36,000 --> 00:13:39,080 Speaker 3: little bit more static in the sense that while while 236 00:13:39,080 --> 00:13:41,160 Speaker 3: there are adversaries, and the adversaries they'll try to learn 237 00:13:41,200 --> 00:13:43,400 Speaker 3: your weaknesses, it's more static in the rules of the 238 00:13:43,400 --> 00:13:45,880 Speaker 3: game are steady and those types of things, so that 239 00:13:45,880 --> 00:13:48,800 Speaker 3: that sampling could work right. Although it was interesting, I 240 00:13:48,840 --> 00:13:50,760 Speaker 3: love the like because it is an analogy to some 241 00:13:50,840 --> 00:13:52,920 Speaker 3: of the problems that pop up and will pop up 242 00:13:52,920 --> 00:13:55,520 Speaker 3: if you take Alpha go right on. The Go game 243 00:13:56,360 --> 00:14:00,920 Speaker 3: Go got you also after Chess, obviously, but the Google 244 00:14:01,040 --> 00:14:04,400 Speaker 3: was able to create this game that was beating the 245 00:14:04,440 --> 00:14:07,920 Speaker 3: pros and radically beating the pros, killing everybody and getting 246 00:14:07,920 --> 00:14:11,880 Speaker 3: better and better and better, although you know, I don't 247 00:14:11,880 --> 00:14:13,160 Speaker 3: know how up to the day you are. But then 248 00:14:13,200 --> 00:14:15,520 Speaker 3: there was this loophole in it where that's that another 249 00:14:15,559 --> 00:14:20,080 Speaker 3: person who was a mediocre Go player, but a computer 250 00:14:20,120 --> 00:14:23,120 Speaker 3: scientiists who thought there might be a hole in this 251 00:14:23,240 --> 00:14:27,280 Speaker 3: super AI used a little program to find the hole. 252 00:14:28,000 --> 00:14:30,400 Speaker 3: And what it illustrated was the a I had no 253 00:14:30,440 --> 00:14:33,040 Speaker 3: idea how to play the game, because what a six 254 00:14:33,080 --> 00:14:35,720 Speaker 3: year old wouldn't The mistake the AI was prone to 255 00:14:35,880 --> 00:14:37,560 Speaker 3: was a mistake of six year old playing Go would 256 00:14:37,600 --> 00:14:39,920 Speaker 3: never make where if you made a large enough in circling, 257 00:14:40,240 --> 00:14:43,720 Speaker 3: if you now go works, but if you encircle the 258 00:14:43,800 --> 00:14:49,520 Speaker 3: other guy's pieces, right, you eliminate them all. And something 259 00:14:49,560 --> 00:14:51,240 Speaker 3: that would never work in a human game is you 260 00:14:51,360 --> 00:14:54,320 Speaker 3: make a really big circle. And because it never came 261 00:14:54,400 --> 00:14:57,960 Speaker 3: up in human games, and because when they perturbed human 262 00:14:58,000 --> 00:15:02,520 Speaker 3: games and started playing computer against computer, they basically started 263 00:15:02,720 --> 00:15:06,160 Speaker 3: with a seed of human games, they never perturbed it 264 00:15:06,280 --> 00:15:09,040 Speaker 3: enough to try this out, to try a massive circle, 265 00:15:09,680 --> 00:15:11,840 Speaker 3: and a human would never let the massive circle have 266 00:15:11,920 --> 00:15:15,560 Speaker 3: it. It's so easy to defend against. But actually the best 267 00:15:15,880 --> 00:15:19,040 Speaker 3: Go algorithm in the world allowed it to happen, right, 268 00:15:19,080 --> 00:15:23,000 Speaker 3: And now a mediocre Go player with a little bit 269 00:15:23,000 --> 00:15:26,400 Speaker 3: of AI found a way to beat this incredible Go 270 00:15:26,480 --> 00:15:30,480 Speaker 3: game again because the Go algorithm at that time had 271 00:15:30,560 --> 00:15:32,560 Speaker 3: this tremendous amount of data, but the things that weren't 272 00:15:32,600 --> 00:15:34,840 Speaker 3: in this data wasn't aware of and it wasn't in 273 00:15:34,880 --> 00:15:38,400 Speaker 3: any deep sense understanding the principles of the game. So 274 00:15:38,560 --> 00:15:40,680 Speaker 3: that's the type of you know, data problem you can 275 00:15:40,720 --> 00:15:42,720 Speaker 3: have even with a massive amount of data played, you know, 276 00:15:43,160 --> 00:15:45,600 Speaker 3: millions and millions of games, but to play every possible 277 00:15:45,640 --> 00:15:48,440 Speaker 3: Go board, you'd have to there's more possible Go boards 278 00:15:48,480 --> 00:15:50,400 Speaker 3: than there are atoms in the universe. So it was 279 00:15:50,440 --> 00:15:53,640 Speaker 3: never going to calculate every possibility and it never got 280 00:15:53,680 --> 00:15:57,920 Speaker 3: to reasoning, right, and that therefore that was a weakness, right. 281 00:15:57,960 --> 00:16:01,600 Speaker 3: And on the other hand, you mix that blend that 282 00:16:01,680 --> 00:16:04,000 Speaker 3: even with a basic reason error, that a language model 283 00:16:04,000 --> 00:16:06,040 Speaker 3: could come up with understanding the rules of GO and 284 00:16:06,040 --> 00:16:08,800 Speaker 3: being able to talk about it. There's an element of 285 00:16:09,080 --> 00:16:12,480 Speaker 3: knowing those things that humans already know that's possible with 286 00:16:12,520 --> 00:16:16,120 Speaker 3: a blend of let's say a statistical technique like alpha 287 00:16:16,160 --> 00:16:21,600 Speaker 3: GO was using and a reasoner to prevent these types 288 00:16:21,600 --> 00:16:22,200 Speaker 3: of mistakes. 289 00:16:22,960 --> 00:16:24,720 Speaker 1: I like that story because it makes me think I 290 00:16:24,760 --> 00:16:29,800 Speaker 1: have a chance against the super smart supercomputer. Okay, that's 291 00:16:30,000 --> 00:16:33,000 Speaker 1: kind of comforting, But I definitely want to ask you 292 00:16:33,040 --> 00:16:38,160 Speaker 1: more about weaknesses in AI and large language models, but 293 00:16:38,440 --> 00:16:41,400 Speaker 1: maybe before we do, you know, just sort of setting 294 00:16:41,400 --> 00:16:44,720 Speaker 1: the groundwork once again. But when we see headlines like 295 00:16:45,120 --> 00:16:50,880 Speaker 1: Bridgewater restructures will put more focus on AI, what does 296 00:16:50,920 --> 00:16:54,000 Speaker 1: that mean exactly? What does it mean for a firm, 297 00:16:54,040 --> 00:16:59,000 Speaker 1: an investment firm like Bridgewater to build up resources in AI? 298 00:16:59,280 --> 00:17:02,800 Speaker 1: And then secondly, could you walk us through a concrete 299 00:17:02,960 --> 00:17:08,080 Speaker 1: example of how AI would be deployed in a particular 300 00:17:08,440 --> 00:17:11,720 Speaker 1: trading strategy. I feel like the more concrete we can 301 00:17:11,760 --> 00:17:14,000 Speaker 1: get with this, the more helpful it'll be. 302 00:17:15,000 --> 00:17:19,040 Speaker 3: Yeah. Great. So I think as we restructured, one of 303 00:17:19,040 --> 00:17:21,879 Speaker 3: the things that as we've made the transition at Bridgewater, 304 00:17:22,280 --> 00:17:25,680 Speaker 3: you know, from Ray having the key ownership to ownership 305 00:17:25,680 --> 00:17:29,560 Speaker 3: at a board level and that transition, we have done 306 00:17:29,560 --> 00:17:31,639 Speaker 3: something we hadn't done in the past, which is essentially 307 00:17:32,200 --> 00:17:34,879 Speaker 3: retain earnings in a very significant way, which allows us 308 00:17:34,880 --> 00:17:38,119 Speaker 3: to invest in things that you know, are aren't going 309 00:17:38,160 --> 00:17:40,720 Speaker 3: to be part profitable right right away, but are the 310 00:17:40,720 --> 00:17:44,960 Speaker 3: big long term bats that we're making, and certainly recognizing 311 00:17:45,000 --> 00:17:47,280 Speaker 3: that there's a way to reinvent a lot of what 312 00:17:47,359 --> 00:17:54,359 Speaker 3: we do using AI machine learning techniques to improve what 313 00:17:54,359 --> 00:17:57,920 Speaker 3: we're doing to understand the world, accelerate that, and specifically 314 00:17:58,240 --> 00:18:01,119 Speaker 3: what we've done on the aimlside is we've set up 315 00:18:01,119 --> 00:18:04,440 Speaker 3: this venture. Essentially they're seventeen of us with me leading it. 316 00:18:04,800 --> 00:18:06,960 Speaker 3: You know, I'm still very much involved in core bridge Order, 317 00:18:06,960 --> 00:18:10,680 Speaker 3: but the sixteen others are one hundred percent dedicated to 318 00:18:12,280 --> 00:18:15,359 Speaker 3: kind of reinventing Bridge Order in a way with machine learning. 319 00:18:15,359 --> 00:18:17,760 Speaker 3: We're going to have a fund specifically run by machine 320 00:18:17,800 --> 00:18:19,919 Speaker 3: learning techniques which will take me into tracy what kind 321 00:18:19,920 --> 00:18:22,520 Speaker 3: of strategies you can do, you know, that's what we're 322 00:18:22,520 --> 00:18:26,760 Speaker 3: working on right now in that lab and pressing the 323 00:18:27,080 --> 00:18:29,760 Speaker 3: edges of what AI is capable of now a like 324 00:18:29,840 --> 00:18:32,639 Speaker 3: machine learning is capable of now right now. There are 325 00:18:32,680 --> 00:18:36,640 Speaker 3: big problems right A. You take large language models and 326 00:18:37,200 --> 00:18:39,800 Speaker 3: they have two types of problems. One thing is the 327 00:18:39,840 --> 00:18:43,480 Speaker 3: basic problem is there they are trained on the structure 328 00:18:43,480 --> 00:18:46,399 Speaker 3: of language, so they usually return something that looks like 329 00:18:47,160 --> 00:18:50,880 Speaker 3: good structure of language. They don't always return accurate answers, 330 00:18:51,080 --> 00:18:53,520 Speaker 3: so that's a problem. It hallucinates, It makes things up 331 00:18:53,720 --> 00:18:56,160 Speaker 3: because it's more focused on the structure of what word 332 00:18:56,480 --> 00:18:59,440 Speaker 3: or what concept would come next, then whether it's accurate 333 00:18:59,560 --> 00:19:00,520 Speaker 3: in what comes. 334 00:19:00,760 --> 00:19:04,199 Speaker 1: Can I just say when I hear AI hallucinations, it 335 00:19:04,320 --> 00:19:07,840 Speaker 1: becomes so science fiction for me. It's very like robot 336 00:19:07,920 --> 00:19:11,600 Speaker 1: stream of electric cheap kind of it's just so surreal. 337 00:19:13,320 --> 00:19:15,520 Speaker 3: Yeah, well, I mean in this case, you can imagine 338 00:19:15,600 --> 00:19:17,720 Speaker 3: what's happening, right because it's just what it's what it's 339 00:19:17,800 --> 00:19:22,560 Speaker 3: trained on. Right. So if you're just if basically the 340 00:19:22,600 --> 00:19:25,639 Speaker 3: basic concept is give it any stream of words and 341 00:19:25,720 --> 00:19:28,560 Speaker 3: it'll predict based on having read everything that's ever been read. 342 00:19:28,880 --> 00:19:32,440 Speaker 3: What comes next, right, and that if it's a little 343 00:19:32,480 --> 00:19:37,320 Speaker 3: bit wrong in what comes next, it can misfire and 344 00:19:37,359 --> 00:19:40,120 Speaker 3: give you something that sounds like something that could come next, 345 00:19:40,119 --> 00:19:42,480 Speaker 3: but actually wrong, you know. And it's just what it's 346 00:19:42,480 --> 00:19:44,520 Speaker 3: trained on, right, It's trained to predict the next word. 347 00:19:44,760 --> 00:19:48,360 Speaker 3: Slight errors in that create those types of issues. Now, 348 00:19:48,359 --> 00:19:53,000 Speaker 3: the algorithm is pretty remarkable, particularly like we as I said, 349 00:19:53,000 --> 00:19:57,040 Speaker 3: I've been tracking in AI as an investor for a 350 00:19:57,040 --> 00:19:59,840 Speaker 3: long time and looking at their technology for a long 351 00:19:59,880 --> 00:20:05,880 Speaker 3: time time. And you know, up until there's GPT one, two, three, 352 00:20:05,920 --> 00:20:08,640 Speaker 3: and many versions of between, and then at GPT three 353 00:20:08,720 --> 00:20:10,760 Speaker 3: it started to have some use. GPT one and two 354 00:20:10,760 --> 00:20:14,280 Speaker 3: were you know, barely coherent, GPT three was you know, 355 00:20:14,400 --> 00:20:17,240 Speaker 3: somewhat usable for certain tasks. Three and a half, which 356 00:20:17,280 --> 00:20:20,359 Speaker 3: is what CHAT GPT is, you know, got to a 357 00:20:20,359 --> 00:20:23,520 Speaker 3: certain level, like on Bridgewater's internal tests, you suddenly got 358 00:20:23,520 --> 00:20:27,000 Speaker 3: to the point where it was able to answer our 359 00:20:27,160 --> 00:20:29,960 Speaker 3: investment associate tests at the level of a first year 360 00:20:30,400 --> 00:20:33,240 Speaker 3: IA right around with chat GPT three point five and 361 00:20:34,359 --> 00:20:39,200 Speaker 3: anthropics most recent quad and then GPT four was able 362 00:20:39,280 --> 00:20:42,199 Speaker 3: to do significantly better. And these are you know, at 363 00:20:42,280 --> 00:20:45,320 Speaker 3: least what we thought were conceptual tests significantly better than 364 00:20:45,359 --> 00:20:47,800 Speaker 3: our average you know, first year investment associate that went 365 00:20:47,840 --> 00:20:52,200 Speaker 3: through training. And similarly, it's able to take the LSAD 366 00:20:52,240 --> 00:20:55,320 Speaker 3: and do well, et cetera. So it can be basically 367 00:20:55,680 --> 00:20:58,480 Speaker 3: pretty smart. It is pretty smart in a wide variety 368 00:20:58,520 --> 00:21:00,880 Speaker 3: of things with errors, but pretty smart on a wide 369 00:21:00,960 --> 00:21:05,240 Speaker 3: variety of whether it's BMCAT or the LSAD or Bridgewaters 370 00:21:05,359 --> 00:21:07,960 Speaker 3: internal tests or whatever, a whole wide variety of things. 371 00:21:08,359 --> 00:21:10,920 Speaker 3: This is a big deal that it can achieve all 372 00:21:10,960 --> 00:21:15,200 Speaker 3: of those kind of academic things, and yet it's still 373 00:21:15,200 --> 00:21:17,480 Speaker 3: eightieth percentile kind of thing on a lot of those things, 374 00:21:17,520 --> 00:21:20,440 Speaker 3: which is remarkable to be eightieth percentile on many many 375 00:21:20,440 --> 00:21:23,640 Speaker 3: different things. But at the same time, it's eightieth percentile 376 00:21:23,680 --> 00:21:26,560 Speaker 3: for a reason. There are flaws, meaning it's not one 377 00:21:26,640 --> 00:21:30,640 Speaker 3: hundred percentile, and so that leads to like you need 378 00:21:30,680 --> 00:21:34,960 Speaker 3: to find a way to work through those flaws, right, 379 00:21:35,000 --> 00:21:38,280 Speaker 3: and that's really where you know. So if somebody's going 380 00:21:38,359 --> 00:21:41,000 Speaker 3: to use large language models to pick stocks, I think 381 00:21:41,080 --> 00:21:44,760 Speaker 3: that's hopeless. That is a hopeless path. But if you 382 00:21:44,880 --> 00:21:49,760 Speaker 3: use large language models to create some theories which it 383 00:21:49,800 --> 00:21:53,919 Speaker 3: can theorize about things, and you use other techniques to 384 00:21:54,160 --> 00:21:57,880 Speaker 3: judge those theories and you iterate between them to create 385 00:21:57,920 --> 00:22:01,160 Speaker 3: a sort of an artificial reasoner. Where language models are 386 00:22:01,160 --> 00:22:04,800 Speaker 3: good at certainly generating theories any theories that already exist 387 00:22:04,880 --> 00:22:09,120 Speaker 3: in human knowledge, and putting those things connect together, they're 388 00:22:09,160 --> 00:22:12,320 Speaker 3: bad at determining whether they're true. But there are other 389 00:22:12,359 --> 00:22:16,560 Speaker 3: ways to pair it with statistical models and other types 390 00:22:16,560 --> 00:22:19,359 Speaker 3: of AI to combine those together. And that's really what 391 00:22:19,400 --> 00:22:22,360 Speaker 3: we're focused on, which is combining large language models that 392 00:22:22,800 --> 00:22:26,400 Speaker 3: are bad at precision with statistical models that are good 393 00:22:26,440 --> 00:22:29,560 Speaker 3: at being precise about the past but terrible about the future, 394 00:22:30,119 --> 00:22:33,760 Speaker 3: and combining those together you start to build an ecosystem 395 00:22:34,200 --> 00:22:38,760 Speaker 3: that can achieve I believe can achieve the types of 396 00:22:38,840 --> 00:22:42,800 Speaker 3: things that bridge order analysts combined with our stress testing 397 00:22:42,840 --> 00:22:46,320 Speaker 3: process and compounding understanding process at Bridgeworker can do, but 398 00:22:46,400 --> 00:22:48,520 Speaker 3: it can do it at so much more scale, because 399 00:22:48,520 --> 00:22:51,040 Speaker 3: all of a sudden, if you have an eightieth percentile 400 00:22:51,119 --> 00:22:55,119 Speaker 3: investment associate, technologically you have millions of them at once, 401 00:22:55,600 --> 00:23:00,760 Speaker 3: and if you have the ability to control their hallucinations 402 00:23:00,760 --> 00:23:05,280 Speaker 3: in their errors by having a rigorous statistical backdrop, you 403 00:23:05,280 --> 00:23:08,719 Speaker 3: could do a tremendous amount at a rapid rate. And 404 00:23:08,720 --> 00:23:10,879 Speaker 3: that's that's really what we're doing in our lab and 405 00:23:11,040 --> 00:23:13,760 Speaker 3: proving out that that process can work. I see. 406 00:23:13,840 --> 00:23:18,199 Speaker 1: So, so is the idea that AI could possibly generate 407 00:23:18,800 --> 00:23:24,879 Speaker 1: theses or ideas that can then be rigorously, you know, 408 00:23:24,960 --> 00:23:29,479 Speaker 1: statistically fact checked by either the humans or you know, 409 00:23:29,600 --> 00:23:32,439 Speaker 1: existing algorithms and data sets. Is that the idea? 410 00:23:33,520 --> 00:23:35,760 Speaker 3: Yeah? And then yes, and but the idea goes further, 411 00:23:35,840 --> 00:23:38,280 Speaker 3: But yes, that's the start. Language models could do that. 412 00:23:38,440 --> 00:23:42,520 Speaker 3: Statistical AI can then take theories and generate whether like 413 00:23:42,880 --> 00:23:44,520 Speaker 3: those have at least been true in the past, and 414 00:23:44,600 --> 00:23:47,200 Speaker 3: what the flaws with them are and refine them, offer 415 00:23:47,240 --> 00:23:50,760 Speaker 3: suggestions on how to do them differently, which then you 416 00:23:50,760 --> 00:23:54,119 Speaker 3: could dialogue with. So then the other strength of language 417 00:23:54,119 --> 00:23:57,800 Speaker 3: model has that that humans are weaker at is now 418 00:23:57,920 --> 00:24:02,120 Speaker 3: take a complex statistical model and talk about what it's doing, 419 00:24:03,520 --> 00:24:06,680 Speaker 3: and there's ways to train language models to do that. 420 00:24:06,680 --> 00:24:10,879 Speaker 3: That then allow sort of a judgment to say, okay, 421 00:24:10,960 --> 00:24:13,920 Speaker 3: now let's think about what's happening here and reason over 422 00:24:13,960 --> 00:24:17,280 Speaker 3: what's happening. So you use the way we've modeled this 423 00:24:17,359 --> 00:24:20,399 Speaker 3: kind of out as language models can come up with 424 00:24:20,480 --> 00:24:23,399 Speaker 3: potential theories. Now there's a limit to that. It's not 425 00:24:23,440 --> 00:24:26,000 Speaker 3: the most creative thing in the world, although it's met 426 00:24:26,119 --> 00:24:31,680 Speaker 3: theory at scale for sure. And then there's and again 427 00:24:31,720 --> 00:24:33,840 Speaker 3: that's language models with good you know, you got to 428 00:24:33,840 --> 00:24:35,760 Speaker 3: tune your language models in a certain way so it's 429 00:24:35,800 --> 00:24:37,840 Speaker 3: not straight out of the box. But then you can 430 00:24:37,920 --> 00:24:41,439 Speaker 3: use statistical things to control that. Then you can use 431 00:24:41,480 --> 00:24:43,840 Speaker 3: language models again to take what's coming out of that 432 00:24:43,880 --> 00:24:46,480 Speaker 3: statistical engine and talk about it with a human or 433 00:24:46,560 --> 00:24:50,119 Speaker 3: other machine learning agents, and we kind of report back 434 00:24:50,680 --> 00:24:53,639 Speaker 3: on what you're finding and what that is and the 435 00:24:53,680 --> 00:24:55,760 Speaker 3: types of theories that are out there that might run 436 00:24:55,760 --> 00:24:59,040 Speaker 3: contrary to what you believe, which can lead to more 437 00:24:59,080 --> 00:25:03,240 Speaker 3: tests and and other thing. So that's the loop that 438 00:25:04,240 --> 00:25:06,760 Speaker 3: you know that I'm very excited about. And as I said, 439 00:25:06,960 --> 00:25:10,760 Speaker 3: up until the thing that statistical AI was limited because 440 00:25:10,760 --> 00:25:14,640 Speaker 3: it was focused on the data of markets, where language 441 00:25:14,680 --> 00:25:16,480 Speaker 3: models the good thing is it has a much better 442 00:25:16,520 --> 00:25:19,359 Speaker 3: sense of something that a statistco model wouldn't really have. 443 00:25:19,480 --> 00:25:23,400 Speaker 3: Statistical model markets doesn't get the concept of greed. Language 444 00:25:23,400 --> 00:25:26,320 Speaker 3: models pretty much understand the concept of greed. They've read 445 00:25:26,320 --> 00:25:29,520 Speaker 3: everything that's ever been written about greed and fear and whatever. 446 00:25:29,760 --> 00:25:32,480 Speaker 3: So now it can start to think about statistical results 447 00:25:32,480 --> 00:25:35,760 Speaker 3: in the context of the human condition that generates those results. 448 00:25:36,520 --> 00:25:39,400 Speaker 3: Big deal and really a radical difference. 449 00:25:39,760 --> 00:25:41,879 Speaker 2: Let me ask you one very simple question, and it 450 00:25:41,960 --> 00:25:45,000 Speaker 2: might be one that speaks to an anxiety of listeners. 451 00:25:45,480 --> 00:25:50,359 Speaker 2: If already GPT can perform at maybe the type of 452 00:25:50,480 --> 00:25:54,560 Speaker 2: level that high quality first year or second year associator 453 00:25:54,640 --> 00:25:58,960 Speaker 2: Analystic Bridgewater can do, does it mean fewer highers in 454 00:25:59,000 --> 00:26:02,359 Speaker 2: the future humans being hired at Bridgewater or does it 455 00:26:02,440 --> 00:26:06,280 Speaker 2: mean the same number or more humans doing even more? 456 00:26:06,359 --> 00:26:08,040 Speaker 2: Like do use is it a replacement? Like what does 457 00:26:08,080 --> 00:26:11,119 Speaker 2: it mean for like the type of person that would 458 00:26:11,119 --> 00:26:15,040 Speaker 2: have been the ten years ago? First your employee at Bridgewater. 459 00:26:16,119 --> 00:26:19,040 Speaker 3: What I think people should expect at bridgeworder but and 460 00:26:19,160 --> 00:26:21,840 Speaker 3: just generally at bridgeworker in a hurry is things are 461 00:26:21,920 --> 00:26:26,320 Speaker 3: changing quick that it really requires people to be capable 462 00:26:27,240 --> 00:26:32,120 Speaker 3: of playing whatever role is necessary in order to do that. Right, 463 00:26:32,200 --> 00:26:34,800 Speaker 3: Like if you go back at the clock at Bridgewater 464 00:26:34,840 --> 00:26:38,000 Speaker 3: when I started or just before that, right, we were 465 00:26:38,200 --> 00:26:40,800 Speaker 3: you know, we were using egg time, Like we had 466 00:26:40,880 --> 00:26:42,480 Speaker 3: rules on how to trade, but we were using egg 467 00:26:42,520 --> 00:26:45,280 Speaker 3: timers and humans to like do these things. And over 468 00:26:45,359 --> 00:26:47,080 Speaker 3: time computers could do more and more of that. We 469 00:26:47,160 --> 00:26:48,880 Speaker 3: kind of got to this point where it was i'd say, 470 00:26:48,960 --> 00:26:53,439 Speaker 3: kind of humans settled into the role of intuition and 471 00:26:53,520 --> 00:26:56,960 Speaker 3: idea generation, and we use computers for memory and for 472 00:26:58,200 --> 00:27:02,159 Speaker 3: constantly running those rules accurate, et cetera. That was a 473 00:27:02,200 --> 00:27:05,840 Speaker 3: transition half like something it got to fifty to fifty 474 00:27:05,920 --> 00:27:09,640 Speaker 3: technology and people. And now this is another leap, right, 475 00:27:09,680 --> 00:27:12,600 Speaker 3: And it's definitely true that it's going to change the 476 00:27:12,720 --> 00:27:16,639 Speaker 3: roles that investment associates play now exactly how and you 477 00:27:16,720 --> 00:27:20,359 Speaker 3: still need the for as far foreseeable future. You're going 478 00:27:20,400 --> 00:27:25,120 Speaker 3: to want people around that out that working on those things. 479 00:27:25,200 --> 00:27:29,240 Speaker 3: There's edges that these techniques I'm describing certainly won't do 480 00:27:29,400 --> 00:27:32,159 Speaker 3: well for an extented period of time, and there's how 481 00:27:32,160 --> 00:27:37,520 Speaker 3: to build the ecosystem of these machine learning agents, et cetera. 482 00:27:38,080 --> 00:27:40,600 Speaker 3: And so what I've found is certainly the people in 483 00:27:40,640 --> 00:27:43,080 Speaker 3: the lab. You want people who are curious about these 484 00:27:43,080 --> 00:27:46,600 Speaker 3: new technologies, you want with to utilize them, and that's 485 00:27:46,880 --> 00:27:49,400 Speaker 3: that's going to be really part of the future of work. 486 00:27:49,440 --> 00:27:50,840 Speaker 3: I think. I think it's going to be very hard 487 00:27:50,880 --> 00:27:55,000 Speaker 3: in any knowledge industry to not utilize these And we're 488 00:27:55,000 --> 00:27:59,480 Speaker 3: seeing this huge breakthrough encoding, right that is so democratizing 489 00:27:59,520 --> 00:28:02,760 Speaker 3: in a sense that you don't you really need to 490 00:28:02,800 --> 00:28:05,080 Speaker 3: know what you want to code more than you need 491 00:28:05,119 --> 00:28:08,000 Speaker 3: to know coding, you know, And that's a big breakthrough. 492 00:28:08,000 --> 00:28:09,879 Speaker 3: So a bunch of people that weren't as well trained 493 00:28:09,920 --> 00:28:13,439 Speaker 3: or as capable in C plus plus or in Python 494 00:28:13,520 --> 00:28:16,560 Speaker 3: or whatever can suddenly get what they want so much faster. 495 00:28:16,720 --> 00:28:19,240 Speaker 3: So all of a sudden, the skill sets are changing, 496 00:28:19,240 --> 00:28:21,560 Speaker 3: and they're changing in ways that I think are as 497 00:28:21,600 --> 00:28:25,320 Speaker 3: surprise to many because it's actually a lot of the 498 00:28:25,480 --> 00:28:29,760 Speaker 3: knowledge work, a lot of the things where you content 499 00:28:29,800 --> 00:28:33,280 Speaker 3: creating and whatever that that I think people thought would 500 00:28:33,320 --> 00:28:37,719 Speaker 3: be later in computer replacement that are happening faster. So 501 00:28:37,760 --> 00:28:39,720 Speaker 3: the main thing is, i'd say, right now, there's so 502 00:28:39,800 --> 00:28:43,200 Speaker 3: much in flux that having flexible the more you need 503 00:28:43,240 --> 00:28:47,280 Speaker 3: flexible generalists who can have an eye towards this and 504 00:28:47,360 --> 00:28:49,720 Speaker 3: eye towards the goal and be able to utilize whatever 505 00:28:49,760 --> 00:28:52,680 Speaker 3: tools are necessary to get there. That's really where I think, 506 00:28:52,800 --> 00:28:55,280 Speaker 3: you know, you're seeing a fair amount of change quickly. 507 00:28:56,280 --> 00:29:00,880 Speaker 1: So you mentioned earlier that just the existence of machine 508 00:29:00,920 --> 00:29:04,920 Speaker 1: learning can impact both the current environment and the future. 509 00:29:05,120 --> 00:29:08,520 Speaker 1: So I think you said the future data points aren't 510 00:29:08,520 --> 00:29:11,240 Speaker 1: going to look like the past data points simply because 511 00:29:11,280 --> 00:29:16,920 Speaker 1: machine learning exists. Does that sort of reflexivity between machine 512 00:29:16,960 --> 00:29:21,680 Speaker 1: learning slash AI and markets become more of an issue 513 00:29:22,040 --> 00:29:26,080 Speaker 1: as AI and machine learning becomes more and more popular 514 00:29:26,160 --> 00:29:27,160 Speaker 1: and more entrenched. 515 00:29:28,400 --> 00:29:30,120 Speaker 3: Yeah, I think it's a big deal, right, And I 516 00:29:30,120 --> 00:29:34,120 Speaker 3: think it's both something that's going to cause act and 517 00:29:34,200 --> 00:29:36,440 Speaker 3: something I'm super excited about. Obviously, I'm excited about the 518 00:29:36,440 --> 00:29:38,600 Speaker 3: power of this that I think there's ways to utilize 519 00:29:38,600 --> 00:29:41,800 Speaker 3: it really well. And it'll also there will be a 520 00:29:41,840 --> 00:29:44,080 Speaker 3: lot of mistakes. Like you're saying, there will be funds 521 00:29:44,680 --> 00:29:49,160 Speaker 3: that will use GPD to pick stocks and not really 522 00:29:49,560 --> 00:29:52,160 Speaker 3: deeply understanding what's happening and why or why what the 523 00:29:52,240 --> 00:29:55,320 Speaker 3: weaknesses that might be there are already plenty of times 524 00:29:55,360 --> 00:29:59,280 Speaker 3: where statistical pure statistical because there's not enough data. You're 525 00:29:59,280 --> 00:30:03,320 Speaker 3: not building with those fundamental issues in mind. You know, 526 00:30:03,560 --> 00:30:05,800 Speaker 3: not that it was directly markets, but in the housing market, 527 00:30:06,120 --> 00:30:09,960 Speaker 3: what Zilo did is a great example. Right. Zillo goes 528 00:30:10,000 --> 00:30:12,680 Speaker 3: out and uses an AI technique that wasn't fit for 529 00:30:12,760 --> 00:30:14,400 Speaker 3: purpose for when it's worth but they use an AI 530 00:30:14,560 --> 00:30:17,840 Speaker 3: technique to predict housing prices and then go into the 531 00:30:17,840 --> 00:30:20,360 Speaker 3: market to start buying houses that they think are undervalued, right, 532 00:30:20,400 --> 00:30:23,120 Speaker 3: And they have a couple problems. One is, while they 533 00:30:23,120 --> 00:30:26,160 Speaker 3: had a ton of housing data, it was over a 534 00:30:26,200 --> 00:30:29,160 Speaker 3: relatively short period of time. So even though they had 535 00:30:29,360 --> 00:30:31,400 Speaker 3: tons what looked like tons of data points because they 536 00:30:31,400 --> 00:30:33,760 Speaker 3: have the price of every house and everywhere or whatever, 537 00:30:34,440 --> 00:30:37,719 Speaker 3: there's still a macro cycle that affects everything that was 538 00:30:38,120 --> 00:30:41,440 Speaker 3: underestimated in what they did. And secondly, they underestimated what 539 00:30:41,440 --> 00:30:43,960 Speaker 3: it would be like in theory versus in practice, whether 540 00:30:44,120 --> 00:30:46,760 Speaker 3: it's actually an adversarial market. Every time they won an auction, 541 00:30:47,320 --> 00:30:50,240 Speaker 3: there was something about that particular lot that the other 542 00:30:50,360 --> 00:30:53,400 Speaker 3: people bidding on that lot knew that they didn't, and 543 00:30:53,480 --> 00:30:57,080 Speaker 3: so it ended up obviously being a huge problem for Zillo, 544 00:30:57,160 --> 00:30:59,600 Speaker 3: and they kind of had a big impact on the 545 00:30:59,640 --> 00:31:03,640 Speaker 3: real estate market and then a big failure. And that's 546 00:31:03,640 --> 00:31:04,840 Speaker 3: the kind of thing you're going to see over and 547 00:31:04,840 --> 00:31:08,800 Speaker 3: over again. If because the basic problem that the data 548 00:31:08,880 --> 00:31:11,680 Speaker 3: that you're looking at isn't necessarily the data you'll face 549 00:31:11,720 --> 00:31:14,960 Speaker 3: in real world. You're not facing the adversarial problem when 550 00:31:15,000 --> 00:31:17,800 Speaker 3: you're looking at that data the way they were. You're 551 00:31:17,840 --> 00:31:21,560 Speaker 3: not a statistical technique that's very good at seasonality and 552 00:31:21,600 --> 00:31:25,400 Speaker 3: trend following might not be very good at understanding macro 553 00:31:25,480 --> 00:31:29,560 Speaker 3: cycles and so on. So that was another case where 554 00:31:30,000 --> 00:31:31,520 Speaker 3: Zillow is a case and I think we'll see it 555 00:31:31,560 --> 00:31:34,560 Speaker 3: over and over again where the recognition that it's not 556 00:31:34,680 --> 00:31:37,000 Speaker 3: as simple as taking machine learning out of the pack 557 00:31:37,280 --> 00:31:40,160 Speaker 3: and applying it to this problem. Even when there's a 558 00:31:40,160 --> 00:31:42,560 Speaker 3: ton of data, right some of the places where there 559 00:31:42,600 --> 00:31:44,480 Speaker 3: is a lot more machine learning going on, very short 560 00:31:44,560 --> 00:31:48,440 Speaker 3: term trading arguably is better for machine learning because there's 561 00:31:48,440 --> 00:31:50,920 Speaker 3: a lot of data and you can learn faster over 562 00:31:50,960 --> 00:31:53,880 Speaker 3: that data, and there's some merit to that. And in 563 00:31:54,000 --> 00:31:57,000 Speaker 3: terms of tangible places this is now years ago, But 564 00:31:57,040 --> 00:31:59,760 Speaker 3: where we started applying some of these techniques were in 565 00:31:59,800 --> 00:32:02,760 Speaker 3: things like monitoring our transaction costs and looking for patterns 566 00:32:02,760 --> 00:32:05,480 Speaker 3: and shorter term data because there's a lot more data. 567 00:32:05,640 --> 00:32:09,240 Speaker 3: But on the other hand, the data often it's like 568 00:32:09,320 --> 00:32:11,840 Speaker 3: having the data of your heart rate for your whole life. 569 00:32:11,880 --> 00:32:14,520 Speaker 3: You could feel like, wow, this is a yeah, I've 570 00:32:14,560 --> 00:32:18,800 Speaker 3: got every heartbeat for you know, you know, forty nine years. 571 00:32:18,800 --> 00:32:20,440 Speaker 3: That seems like a lot of data, but its not. 572 00:32:20,600 --> 00:32:24,360 Speaker 3: It's totally irrelevant when you've artitacked. So that even when 573 00:32:24,400 --> 00:32:27,160 Speaker 3: there's lots of data can be misleading. And that those 574 00:32:27,200 --> 00:32:29,480 Speaker 3: are those are the types of issues that will lead 575 00:32:29,560 --> 00:32:33,600 Speaker 3: to these techniques having huge problems, which means it's not 576 00:32:33,960 --> 00:32:35,719 Speaker 3: out of the box AI is going to solve all 577 00:32:35,720 --> 00:32:38,680 Speaker 3: these problems. You really, and this comes back to you 578 00:32:38,760 --> 00:32:41,360 Speaker 3: have to understand the tools, what they're good at, what 579 00:32:41,360 --> 00:32:43,640 Speaker 3: they're bad at, and put them together in a way 580 00:32:44,080 --> 00:32:46,280 Speaker 3: that use what they're good at and protects them from 581 00:32:46,280 --> 00:32:48,680 Speaker 3: what they're bad at. Now, nothing, no process work coming 582 00:32:48,760 --> 00:32:50,960 Speaker 3: up with will do that perfectly. But the more and 583 00:32:51,000 --> 00:32:52,720 Speaker 3: more you could do that, I think, the more and 584 00:32:52,760 --> 00:32:56,400 Speaker 3: more you could become, let's say, better than humans at that, 585 00:32:56,440 --> 00:32:59,560 Speaker 3: because humans have many of those fallibilities or versions of 586 00:32:59,600 --> 00:33:04,000 Speaker 3: those abilities that these processes will have. And that's like 587 00:33:04,080 --> 00:33:06,719 Speaker 3: that'll be the question of how far we can how 588 00:33:06,760 --> 00:33:10,080 Speaker 3: far we could take that and how how much human 589 00:33:10,160 --> 00:33:13,080 Speaker 3: judgment is better than those things, which is stuff you 590 00:33:13,080 --> 00:33:15,600 Speaker 3: know we'll be experimenting with as we as we go along. 591 00:33:32,080 --> 00:33:35,360 Speaker 2: So you know, one thing that you know, your founder 592 00:33:35,880 --> 00:33:38,560 Speaker 2: Ray Dalio years ago, like sort of he wrote down 593 00:33:38,560 --> 00:33:40,680 Speaker 2: a set of rules. You've talked about this before. He 594 00:33:40,720 --> 00:33:43,320 Speaker 2: wrote down a set of rules about how he understood 595 00:33:43,360 --> 00:33:45,640 Speaker 2: the sort of the machine of the markets to work, 596 00:33:45,800 --> 00:33:48,560 Speaker 2: and one of the issues with AI, and I think 597 00:33:48,600 --> 00:33:51,400 Speaker 2: you're sort of been hit getting at this is that 598 00:33:51,480 --> 00:33:55,760 Speaker 2: like AI legibility and the understanding of like, okay, you 599 00:33:55,800 --> 00:33:58,920 Speaker 2: put in that you pose a query to a large 600 00:33:58,960 --> 00:34:02,240 Speaker 2: language model, it creates some output you don't really know 601 00:34:02,520 --> 00:34:05,440 Speaker 2: what it did to get there, and so that's you know, 602 00:34:05,480 --> 00:34:07,640 Speaker 2: that's sort of different than dealing with a human analyst. 603 00:34:07,680 --> 00:34:09,040 Speaker 2: Do you get say, well, what did you think about that? 604 00:34:09,120 --> 00:34:11,200 Speaker 2: Did you think about that? Can you talk a little 605 00:34:11,239 --> 00:34:13,920 Speaker 2: bit more about like the sort of I don't know 606 00:34:13,920 --> 00:34:15,880 Speaker 2: if that's a weakness or how do you sort of 607 00:34:15,920 --> 00:34:18,759 Speaker 2: get around the fact that, like it's still difficult to 608 00:34:19,640 --> 00:34:22,880 Speaker 2: query an AI model and say like how did you 609 00:34:22,960 --> 00:34:24,640 Speaker 2: arrive at X or Y conclusion? 610 00:34:25,160 --> 00:34:28,000 Speaker 3: Yeah, and I think that's really important and that we 611 00:34:28,200 --> 00:34:30,719 Speaker 3: but also something that's more and more breakable because even 612 00:34:30,719 --> 00:34:32,160 Speaker 3: with humans. One of the things like one of the 613 00:34:32,200 --> 00:34:34,200 Speaker 3: places where I think there are a lot of areas 614 00:34:34,239 --> 00:34:36,880 Speaker 3: where Bridgewater has a strength, right, Bridgewater has a strength 615 00:34:36,920 --> 00:34:39,520 Speaker 3: And we never went from a statistical model, So we 616 00:34:39,560 --> 00:34:41,759 Speaker 3: built data based on what we needed for reasoning, and 617 00:34:41,800 --> 00:34:44,880 Speaker 3: as a result, we have a better, longer, cleaner database 618 00:34:45,520 --> 00:34:48,800 Speaker 3: than I think anybody has. We've been thinking through this 619 00:34:48,920 --> 00:34:50,880 Speaker 3: problem that you're referring, which is how do you actually 620 00:34:50,960 --> 00:34:53,359 Speaker 3: get out what somebody means? You'd be surprised how hard 621 00:34:53,400 --> 00:34:56,640 Speaker 3: it is to truly get from a human. Humans don't 622 00:34:56,640 --> 00:35:00,239 Speaker 3: actually know why their synapses do what they do. They 623 00:35:00,280 --> 00:35:02,759 Speaker 3: actually like when you ask somebody to describe something, you 624 00:35:02,800 --> 00:35:06,160 Speaker 3: get some partial version of what they're thinking. If you 625 00:35:06,520 --> 00:35:09,120 Speaker 3: took like an intuitive trader and you start peeling back 626 00:35:09,160 --> 00:35:11,600 Speaker 3: all the reasons, that's very hard. We've been doing that 627 00:35:11,640 --> 00:35:14,000 Speaker 3: for a long time and have an expertise in doing it, 628 00:35:14,040 --> 00:35:15,920 Speaker 3: and I would say that humans don't even know what 629 00:35:15,960 --> 00:35:20,600 Speaker 3: they're doing often, but there are ways to you know, 630 00:35:20,760 --> 00:35:23,359 Speaker 3: like you're saying, query and force questions and what about 631 00:35:23,360 --> 00:35:25,239 Speaker 3: this and what about that? That will help pull out 632 00:35:25,280 --> 00:35:28,120 Speaker 3: human intuition. And what you find with machine learning algorithms 633 00:35:28,160 --> 00:35:29,960 Speaker 3: if you get good at this and this is you know, 634 00:35:30,000 --> 00:35:31,480 Speaker 3: going back to two thousand and sixty. Two thousand and 635 00:35:31,480 --> 00:35:34,040 Speaker 3: seventy has been critical to my work is there's a 636 00:35:34,080 --> 00:35:39,479 Speaker 3: way that you can query machine learning algorithms like query 637 00:35:40,040 --> 00:35:42,799 Speaker 3: like it's different, but the concepts the same as how 638 00:35:42,800 --> 00:35:45,240 Speaker 3: you query humans to get at why they really believe 639 00:35:45,280 --> 00:35:48,359 Speaker 3: what they believe. And as I was saying I think 640 00:35:48,360 --> 00:35:53,480 Speaker 3: there's actually elements of large language models interpreting what statistical 641 00:35:54,040 --> 00:35:58,080 Speaker 3: AI is doing that allows that process to accelerate. And 642 00:35:58,120 --> 00:36:00,160 Speaker 3: I think it's very critical you really want to know 643 00:36:00,239 --> 00:36:02,000 Speaker 3: because that's the way you find the flaws. If you 644 00:36:02,000 --> 00:36:04,400 Speaker 3: go back to my go example and you say you 645 00:36:04,440 --> 00:36:06,759 Speaker 3: can think about if you can querry a model and 646 00:36:06,760 --> 00:36:08,399 Speaker 3: think about what it's done and what it hasn't done, 647 00:36:08,760 --> 00:36:10,839 Speaker 3: then you can figure out what data is missing, right, 648 00:36:10,880 --> 00:36:14,640 Speaker 3: and you need to set up adversarial techniques in order 649 00:36:14,680 --> 00:36:18,600 Speaker 3: to keep querying an algorithm for what it's doing. And again, 650 00:36:18,680 --> 00:36:21,279 Speaker 3: I think that's still an area of research, but a 651 00:36:21,320 --> 00:36:25,560 Speaker 3: process that's moving along quickly to basically get to the 652 00:36:25,600 --> 00:36:29,960 Speaker 3: point where the standard is even though a machine learning 653 00:36:29,960 --> 00:36:32,000 Speaker 3: technique might be doing something very different than a human 654 00:36:32,120 --> 00:36:36,640 Speaker 3: is that it can still explain itself, and it might 655 00:36:36,680 --> 00:36:39,800 Speaker 3: not perfectly explain itself, just like humans don't perfectly explain themselves, 656 00:36:40,120 --> 00:36:42,680 Speaker 3: but to a very high degree of confidence across a 657 00:36:42,719 --> 00:36:45,040 Speaker 3: wide range of outcomes that you have a sense of 658 00:36:45,080 --> 00:36:48,840 Speaker 3: what's going on is possible. And that's the you know, 659 00:36:48,960 --> 00:36:50,839 Speaker 3: that's part of the design of what we're putting in, 660 00:36:50,840 --> 00:36:53,120 Speaker 3: which is, well, how do you query it, how do 661 00:36:53,160 --> 00:36:56,319 Speaker 3: you give it more information, remove information, etc. See how 662 00:36:56,320 --> 00:36:59,840 Speaker 3: it changes its mind to determine roughly what's going on. 663 00:37:00,640 --> 00:37:03,160 Speaker 1: You know, you mentioned the data sets there, and I 664 00:37:03,160 --> 00:37:05,600 Speaker 1: guess it's a cliche nowadays to say, well, a model 665 00:37:05,640 --> 00:37:07,560 Speaker 1: is only as good as the data that it's trained on. 666 00:37:08,000 --> 00:37:11,239 Speaker 1: But it's a cliche because it's true. Do you use 667 00:37:11,280 --> 00:37:14,960 Speaker 1: your own internal data for the large language models or 668 00:37:15,000 --> 00:37:17,880 Speaker 1: where are you actually pulling a data from? And then secondly, like, 669 00:37:18,000 --> 00:37:21,359 Speaker 1: what type of data have you found so far? Is 670 00:37:21,640 --> 00:37:24,759 Speaker 1: most useful for these types of projects? 671 00:37:25,320 --> 00:37:28,799 Speaker 3: Well, I think the things that are most interesting to 672 00:37:29,040 --> 00:37:31,839 Speaker 3: us A we're trying to learn things that we don't 673 00:37:31,880 --> 00:37:34,040 Speaker 3: already know. So we're being careful about what kind of 674 00:37:34,040 --> 00:37:38,160 Speaker 3: Bridgewader knowledge we put in here, because it's not that 675 00:37:38,200 --> 00:37:41,200 Speaker 3: helpful if we reinvent Bridgewater somewhat helpful, but it's about 676 00:37:41,200 --> 00:37:44,360 Speaker 3: it as helpful as let's say, reinventing everything that we 677 00:37:44,360 --> 00:37:46,480 Speaker 3: don't know about that other people have thought about, etc. 678 00:37:47,280 --> 00:37:49,879 Speaker 3: And so point one in the lab right now, at least, 679 00:37:49,880 --> 00:37:53,400 Speaker 3: we're focused on not making this through Bridgewader centric on purpose, 680 00:37:53,440 --> 00:37:55,759 Speaker 3: because it's in that way learn things that we don't 681 00:37:55,760 --> 00:37:59,200 Speaker 3: already know and if you just fed a bridgeworder information, 682 00:37:59,400 --> 00:38:01,360 Speaker 3: which we may well do, that could be a productivity 683 00:38:01,440 --> 00:38:06,480 Speaker 3: enhancing thing, but you'll quickly produce something very similar to Bridgewater. 684 00:38:06,520 --> 00:38:09,360 Speaker 3: Where what's been amazing so far as we're producing good 685 00:38:09,400 --> 00:38:13,920 Speaker 3: results by Bridgewater standards, but different, very very different conclusions 686 00:38:13,920 --> 00:38:16,759 Speaker 3: and different thoughts than what we have internally. So I 687 00:38:16,760 --> 00:38:20,479 Speaker 3: think that's zero point one choice now on raw data 688 00:38:20,480 --> 00:38:23,279 Speaker 3: and cleaning data and how you put together data. Now 689 00:38:23,360 --> 00:38:26,160 Speaker 3: we are benefiting from Bridgewater scale on that. That's been 690 00:38:26,200 --> 00:38:28,600 Speaker 3: a big that's a big deal that over the years, 691 00:38:28,840 --> 00:38:32,040 Speaker 3: again precisely because we took human intuition and said, what 692 00:38:32,120 --> 00:38:34,480 Speaker 3: data do we need to replicate that intuition. We have 693 00:38:34,560 --> 00:38:37,040 Speaker 3: a unique database where if everybody else is pulling from 694 00:38:37,120 --> 00:38:40,120 Speaker 3: data stream Bloomberg, et cetera, we put together the data 695 00:38:40,160 --> 00:38:43,399 Speaker 3: we needed to feed our intuitions. Oftentimes that data didn't exist. 696 00:38:43,480 --> 00:38:45,000 Speaker 3: We had to figure out the way to create it. 697 00:38:46,239 --> 00:38:49,480 Speaker 3: And also we're big believers that you need to stress 698 00:38:49,520 --> 00:38:52,239 Speaker 3: us across a very long period of time, so we 699 00:38:52,280 --> 00:38:55,000 Speaker 3: have much longer data histories now. Those things are certainly 700 00:38:55,080 --> 00:38:59,080 Speaker 3: valuable in a context of small data, any quantity of data, 701 00:38:59,400 --> 00:39:05,080 Speaker 3: any like the understanding the data being able to therefore 702 00:39:05,480 --> 00:39:10,520 Speaker 3: for a given theory find appropriate unoptimized data. Those are 703 00:39:10,520 --> 00:39:13,640 Speaker 3: big deals and that that we are using and and 704 00:39:13,800 --> 00:39:17,759 Speaker 3: you know that does allow us to move forward more 705 00:39:17,960 --> 00:39:20,640 Speaker 3: and on the land large language models. You know, there's 706 00:39:20,640 --> 00:39:21,960 Speaker 3: still a lot of work to be done, but you 707 00:39:22,040 --> 00:39:27,000 Speaker 3: certainly can train through reinforcement learning to you know, to 708 00:39:27,080 --> 00:39:29,600 Speaker 3: make sure that they're not making mistakes that you know about. 709 00:39:30,680 --> 00:39:33,839 Speaker 3: And so there's ways to to do that. Now we've 710 00:39:33,880 --> 00:39:36,160 Speaker 3: been trying to avoid that for the reasons I was 711 00:39:36,160 --> 00:39:38,520 Speaker 3: describing before, avoid doing too much of that of ejecting 712 00:39:38,560 --> 00:39:41,480 Speaker 3: our own knowledge and use external sources to do that. 713 00:39:42,000 --> 00:39:45,000 Speaker 3: But that's still part of uh, you know, part of 714 00:39:45,040 --> 00:39:47,880 Speaker 3: the tool set that will be available that yes, you 715 00:39:47,920 --> 00:39:51,240 Speaker 3: could train it more directly on things you already believe 716 00:39:51,280 --> 00:39:53,239 Speaker 3: to be true if you want to do that, and 717 00:39:53,320 --> 00:39:57,360 Speaker 3: that certainly will lead to answers that replicate your thinking 718 00:39:57,440 --> 00:39:58,360 Speaker 3: more quickly. 719 00:39:59,239 --> 00:40:02,080 Speaker 1: So just on this point, one thing I wanted to 720 00:40:02,600 --> 00:40:06,040 Speaker 1: get your opinion on is how good is AI at 721 00:40:06,280 --> 00:40:11,640 Speaker 1: predicting big turning points or structural breaks in market regimes? 722 00:40:11,719 --> 00:40:14,000 Speaker 1: Because I don't know about you, Joe, but one of 723 00:40:14,040 --> 00:40:16,239 Speaker 1: the first things I did with chat GPT was I 724 00:40:16,280 --> 00:40:19,240 Speaker 1: asked it to write, you know, a financial news article 725 00:40:19,480 --> 00:40:23,040 Speaker 1: about inflation, just to see whether whether our jobs were 726 00:40:23,800 --> 00:40:27,000 Speaker 1: in danger, and you could tell that it was trained 727 00:40:27,239 --> 00:40:30,520 Speaker 1: on not quite current data. It was talking about how 728 00:40:30,560 --> 00:40:33,200 Speaker 1: inflation has been stubbornly low for many years and the 729 00:40:33,239 --> 00:40:35,440 Speaker 1: FED is trying to get it to the two percent target. 730 00:40:35,920 --> 00:40:39,840 Speaker 1: But how good is AI at predicting those regime changes? 731 00:40:39,880 --> 00:40:43,399 Speaker 1: Because if you're running, you know, a macro fund, I 732 00:40:43,400 --> 00:40:46,640 Speaker 1: imagine that's one of the important things that you need 733 00:40:46,680 --> 00:40:49,200 Speaker 1: to do, is try to figure out when something is 734 00:40:49,280 --> 00:40:50,920 Speaker 1: fundamentally changing in the market. 735 00:40:51,760 --> 00:40:54,960 Speaker 3: Yeah, and I'd say terrible if you use it in 736 00:40:54,960 --> 00:40:57,120 Speaker 3: the sense that you're using it right like that. It's 737 00:40:57,160 --> 00:40:58,920 Speaker 3: a little bit like saying, well, how good are people 738 00:40:58,960 --> 00:41:01,600 Speaker 3: at that? Well, people are pretty darn bad at that, right. 739 00:41:01,680 --> 00:41:04,359 Speaker 3: That doesn't mean that there isn't a way where some 740 00:41:04,440 --> 00:41:07,840 Speaker 3: people who could do such a thing right, So AI 741 00:41:08,760 --> 00:41:10,799 Speaker 3: like it. It's hard to just think about AI as 742 00:41:10,800 --> 00:41:12,560 Speaker 3: a thing or think of like, Okay, well, if I'm 743 00:41:12,560 --> 00:41:14,400 Speaker 3: just gonna use chat Gypt for that, You're exactly right. 744 00:41:14,520 --> 00:41:16,759 Speaker 3: Chatgypt as it comes out of the box is only 745 00:41:16,800 --> 00:41:19,759 Speaker 3: trained over to a certain history and it doesn't care 746 00:41:20,160 --> 00:41:23,080 Speaker 3: like unless you know how to make it care. It 747 00:41:23,080 --> 00:41:25,120 Speaker 3: doesn't care that it's you know, it's just to answer 748 00:41:25,160 --> 00:41:27,279 Speaker 3: your question about inflation. Based on everything it's ever read 749 00:41:27,280 --> 00:41:30,040 Speaker 3: about inflation, time isn't even that important unless you make 750 00:41:30,080 --> 00:41:33,080 Speaker 3: time be very important to it and predicting, and so 751 00:41:33,960 --> 00:41:36,160 Speaker 3: you have to know how to use the tools to 752 00:41:36,400 --> 00:41:40,319 Speaker 3: generate the type of outcome that you're describing. So do 753 00:41:40,440 --> 00:41:43,359 Speaker 3: I think like AI out of the box will do that. No, 754 00:41:43,400 --> 00:41:46,680 Speaker 3: absolutely not, It'll be awful at that. Are there ways 755 00:41:46,719 --> 00:41:50,359 Speaker 3: to take what's embedded in AI to come up with 756 00:41:50,400 --> 00:41:54,280 Speaker 3: a way to do that? I was embedded in language models, 757 00:41:54,480 --> 00:41:57,759 Speaker 3: and if you combine that with statistical tools, yeah, there's 758 00:41:57,760 --> 00:41:59,480 Speaker 3: a path there. But it's not going to be as 759 00:41:59,480 --> 00:42:03,000 Speaker 3: simple as open up JATGPD and ask it that question. 760 00:42:03,080 --> 00:42:06,160 Speaker 3: It's it's a there's more involved. But if you basically 761 00:42:06,160 --> 00:42:08,960 Speaker 3: it is helpful to have an analyst that's read everything 762 00:42:08,960 --> 00:42:12,360 Speaker 3: that was ever produced, even if they stopped reading in 763 00:42:12,400 --> 00:42:15,000 Speaker 3: twenty twenty two in twenty twenty one, I should say 764 00:42:15,200 --> 00:42:17,400 Speaker 3: it's there's a way to use that, but you have 765 00:42:17,480 --> 00:42:20,759 Speaker 3: to use it correctly and not misuse it in order 766 00:42:20,760 --> 00:42:22,000 Speaker 3: to try to generate that answer. 767 00:42:22,200 --> 00:42:24,840 Speaker 2: All right, So I can't just ask a large language 768 00:42:24,880 --> 00:42:29,040 Speaker 2: model when will inflation get back to the Fed's target. 769 00:42:29,920 --> 00:42:32,919 Speaker 2: But I'm speaking I'm not speaking to a large language model. 770 00:42:32,920 --> 00:42:37,440 Speaker 2: I'm speaking to Cio Bridgewater. And you know, I do 771 00:42:37,560 --> 00:42:39,080 Speaker 2: I am curious, you know, I do want to talk 772 00:42:39,120 --> 00:42:40,880 Speaker 2: a little. We do want to talk a little macro 773 00:42:41,000 --> 00:42:43,120 Speaker 2: and I, you know, before we sort of like, I'm 774 00:42:43,120 --> 00:42:45,560 Speaker 2: not going to directly ask you when inflation will be 775 00:42:45,600 --> 00:42:48,360 Speaker 2: back to the Fed's target. But what strikes me about 776 00:42:48,360 --> 00:42:51,040 Speaker 2: the last year and since the last time we talked, 777 00:42:51,080 --> 00:42:54,160 Speaker 2: that's really blowing my mind is that rate hikes have 778 00:42:54,160 --> 00:42:59,000 Speaker 2: been a lot faster than people expected. Inflation is hotter 779 00:42:59,080 --> 00:43:02,319 Speaker 2: than people expect, did the unemployment rate is lower than 780 00:43:02,400 --> 00:43:06,239 Speaker 2: people expected. What is it that people misunderstood a year 781 00:43:06,280 --> 00:43:10,000 Speaker 2: ago about the economic machine? Such that the FED has 782 00:43:10,040 --> 00:43:13,640 Speaker 2: hyped rates much faster than people expected, and yet it's 783 00:43:13,719 --> 00:43:17,200 Speaker 2: been surprisingly ineffective at cooling things down. And to this 784 00:43:17,320 --> 00:43:20,719 Speaker 2: day there seems to be a surprising amount of economic 785 00:43:20,760 --> 00:43:23,520 Speaker 2: momentum with FED funds at like five and a half percent. 786 00:43:25,160 --> 00:43:27,360 Speaker 3: Yeah, it's a great question. I have a bunch of 787 00:43:27,560 --> 00:43:29,480 Speaker 3: thoughts on it. You know, certainly I can't speak for 788 00:43:29,480 --> 00:43:31,160 Speaker 3: all people, but I can speak for myself. I've been 789 00:43:31,160 --> 00:43:33,520 Speaker 3: wrong about a bunch of a bunch of those things. 790 00:43:32,920 --> 00:43:36,359 Speaker 3: So just to talk about what I certainly and let's 791 00:43:36,400 --> 00:43:39,520 Speaker 3: say we at Bridgewater didn't now like you're saying. I 792 00:43:39,600 --> 00:43:43,440 Speaker 3: thought the degree of and certainly are everything that we 793 00:43:43,480 --> 00:43:45,719 Speaker 3: had understood in our statistic models or whatever that we 794 00:43:45,800 --> 00:43:48,400 Speaker 3: knew that we could easily be wrong, but that the 795 00:43:48,480 --> 00:43:53,080 Speaker 3: degree of tightening was fast and high relative to history, 796 00:43:53,120 --> 00:43:55,920 Speaker 3: and that any tightening like this in the past had 797 00:43:56,040 --> 00:44:00,640 Speaker 3: led to significant downturns. Although the lead life is somewhat 798 00:44:00,680 --> 00:44:02,919 Speaker 3: variable and its still possible that's right. But I think 799 00:44:03,200 --> 00:44:07,640 Speaker 3: a lot of things that happened different than I expected. 800 00:44:07,800 --> 00:44:11,200 Speaker 3: Was a Usually, when let's say, as they were last year, 801 00:44:11,200 --> 00:44:15,359 Speaker 3: stocks were falling and short rates were rising, that formula 802 00:44:15,640 --> 00:44:18,720 Speaker 3: in history always led to the personal savings rate rising, 803 00:44:18,800 --> 00:44:24,200 Speaker 3: people seeing higher interest rates available to them, asset prices falling, 804 00:44:24,280 --> 00:44:27,439 Speaker 3: housing slowing down, etc. Usually people save more money, which 805 00:44:27,480 --> 00:44:29,480 Speaker 3: meant there was less revenue for companies, which meant there 806 00:44:29,480 --> 00:44:32,759 Speaker 3: were layoffs, which meant savings rates rose more when the 807 00:44:32,760 --> 00:44:36,399 Speaker 3: employment market weakened, and you know, our recession was caused 808 00:44:36,440 --> 00:44:39,920 Speaker 3: through that mechanism. And what's happened in this period is 809 00:44:39,960 --> 00:44:42,920 Speaker 3: that I think now I could be wrong, that normal 810 00:44:43,560 --> 00:44:46,239 Speaker 3: let's say impact of the higher interest rate and wealth 811 00:44:46,280 --> 00:44:50,560 Speaker 3: effect impact was offset by the fact that wealth had 812 00:44:50,600 --> 00:44:53,399 Speaker 3: been changed so radically in the twenty twenty twenty one 813 00:44:53,480 --> 00:44:56,840 Speaker 3: period by fiscal policy, and that we have fiscal policy 814 00:44:58,160 --> 00:45:01,320 Speaker 3: as extreme as the war, and the ripple of the 815 00:45:01,440 --> 00:45:05,640 Speaker 3: length to which that disrupted let's say, those other relationships 816 00:45:05,960 --> 00:45:08,000 Speaker 3: was interesting. The degree of it was interesting. I think 817 00:45:08,000 --> 00:45:10,920 Speaker 3: there's ways we should have, you know, looking back now, 818 00:45:11,480 --> 00:45:13,759 Speaker 3: I think there are reasons that we should have I 819 00:45:13,800 --> 00:45:17,480 Speaker 3: should have known that, and some people were pointing to that, 820 00:45:18,280 --> 00:45:22,480 Speaker 3: but that created much less of a reaction in household balance, 821 00:45:22,560 --> 00:45:25,120 Speaker 3: in household savings rates as you normally did. You came 822 00:45:25,160 --> 00:45:27,239 Speaker 3: out of the recession with better balance sheets than ever. 823 00:45:27,920 --> 00:45:30,640 Speaker 3: People were willing to dissave. So even as rates climbed 824 00:45:30,719 --> 00:45:35,160 Speaker 3: and actually debt growth collapsed as it normally would. But 825 00:45:35,239 --> 00:45:38,800 Speaker 3: what simultaneously collapsed outside of debt is let's say, increase, 826 00:45:38,920 --> 00:45:42,080 Speaker 3: was the willingness to spend down the cash that that 827 00:45:42,200 --> 00:45:46,360 Speaker 3: households had built up. And that cash doesn't just disappear 828 00:45:46,440 --> 00:45:48,680 Speaker 3: when one person spends it, it goes on to others. 829 00:45:48,760 --> 00:45:51,520 Speaker 3: Balance sheets whether it's corporate balance sheets, other household balance sheets, 830 00:45:51,760 --> 00:45:53,799 Speaker 3: and so that what's been happening, it appears, is that 831 00:45:53,880 --> 00:45:57,360 Speaker 3: money's been spinning around in a way that made the 832 00:45:57,440 --> 00:46:00,279 Speaker 3: rate hike have much less impact than I believe would 833 00:46:00,320 --> 00:46:03,280 Speaker 3: have had pre COVID, if you had anything like that, 834 00:46:03,280 --> 00:46:06,240 Speaker 3: that rate hike. On top of that, within the US economy, 835 00:46:06,280 --> 00:46:09,840 Speaker 3: in particular, corporates had extended their duration, So the impact 836 00:46:09,920 --> 00:46:13,040 Speaker 3: is taking longer on the effect on corporates, although I 837 00:46:13,080 --> 00:46:15,880 Speaker 3: think it's happening, but it is taking longer, and so 838 00:46:16,520 --> 00:46:19,000 Speaker 3: there are a few other things. And then obviously the 839 00:46:19,080 --> 00:46:23,600 Speaker 3: benefit of when nominal what did happen is rate rise 840 00:46:23,719 --> 00:46:27,000 Speaker 3: is created a decline in nominal demand, but that's mostly 841 00:46:27,000 --> 00:46:29,839 Speaker 3: shown up in inflation. So nominal demands fallen pretty much 842 00:46:29,840 --> 00:46:32,960 Speaker 3: as much as I've expected. It's been more inflation falling 843 00:46:33,520 --> 00:46:37,320 Speaker 3: than real growth falling, which again I think there's reasons 844 00:46:37,360 --> 00:46:40,480 Speaker 3: that that that's the case. But before there was this 845 00:46:40,600 --> 00:46:44,960 Speaker 3: massive demand shock from the what the Fed, what the 846 00:46:45,440 --> 00:46:48,720 Speaker 3: central banks and the Treasury had done to get everybody's 847 00:46:48,760 --> 00:46:51,799 Speaker 3: balance sheets up, and supply was struggling to keep up 848 00:46:51,840 --> 00:46:56,080 Speaker 3: with this massive demand shock, and now demand's falling. But 849 00:46:56,160 --> 00:46:58,600 Speaker 3: supply is still catching up to that old level, so 850 00:46:58,719 --> 00:47:01,479 Speaker 3: in on net, real growth has come out stronger. Now 851 00:47:02,280 --> 00:47:04,440 Speaker 3: I could see all that in the rear view mirror 852 00:47:05,120 --> 00:47:07,840 Speaker 3: by anything predict that that would be the way it 853 00:47:07,840 --> 00:47:11,319 Speaker 3: would play out. But I think that's why you've had 854 00:47:11,360 --> 00:47:16,080 Speaker 3: this stubborn strengthen the economy and that, you know, and 855 00:47:16,120 --> 00:47:19,640 Speaker 3: that's created a certain amount of stability. Now equities have 856 00:47:19,719 --> 00:47:22,600 Speaker 3: rallied significantly since then, there's like some of the negative 857 00:47:22,640 --> 00:47:25,680 Speaker 3: wealth effects have eased. At the same time, though a 858 00:47:25,760 --> 00:47:28,280 Speaker 3: lot of that excess cash that was on balance sheets 859 00:47:28,280 --> 00:47:31,280 Speaker 3: have been distributed, so there's a mix of pressures here 860 00:47:31,360 --> 00:47:34,200 Speaker 3: that looking forward, you know, we do think inflation is 861 00:47:34,239 --> 00:47:37,240 Speaker 3: still coming down a bit, although on net we've entered 862 00:47:37,280 --> 00:47:40,359 Speaker 3: what we think is a more inflationary environment, such that 863 00:47:40,760 --> 00:47:42,799 Speaker 3: two percent inflation probably more likely to be more of 864 00:47:42,800 --> 00:47:46,960 Speaker 3: a bottom than a cap. And we do think fiscal 865 00:47:47,000 --> 00:47:49,880 Speaker 3: policy as the way to deal with the recessions is 866 00:47:49,920 --> 00:47:53,120 Speaker 3: probably the politically the more likely outcome. Then let's say 867 00:47:53,160 --> 00:47:55,680 Speaker 3: moving back in the next recession to more que and 868 00:47:55,719 --> 00:47:59,040 Speaker 3: fiscal policy is a lot more inflationary and effective in 869 00:47:59,080 --> 00:48:01,839 Speaker 3: a sense of stimulating growth quickly as we've seen. So 870 00:48:01,880 --> 00:48:04,880 Speaker 3: I think you're going to see a world where we 871 00:48:04,960 --> 00:48:08,359 Speaker 3: are still adjusting to a higher inflation, world that's de globalizing. 872 00:48:08,840 --> 00:48:11,600 Speaker 3: Although everything we're talking about on the productivity front, maybe 873 00:48:11,640 --> 00:48:16,200 Speaker 3: machine learning changes that we'll see, but largely X a 874 00:48:16,520 --> 00:48:21,320 Speaker 3: major productivity miracle. I think deglobalization, the move towards fiscal 875 00:48:21,360 --> 00:48:27,319 Speaker 3: policy has changed the long term inflation path in a 876 00:48:27,360 --> 00:48:31,680 Speaker 3: way that markets haven't fully adjusted to. Because markets right 877 00:48:31,719 --> 00:48:35,480 Speaker 3: now believe the fat is totally credible that inflation is 878 00:48:35,480 --> 00:48:38,759 Speaker 3: going to return to target basically with very little problems. 879 00:48:39,280 --> 00:48:42,600 Speaker 3: When we measure the pressures, we don't think, so we 880 00:48:42,680 --> 00:48:45,240 Speaker 3: think it's going to be much more challenging to get 881 00:48:45,360 --> 00:48:48,680 Speaker 3: inflation where markets expected. The impact on earnings is going 882 00:48:48,719 --> 00:48:51,239 Speaker 3: to be a lot more negative than the markets are 883 00:48:51,280 --> 00:48:54,560 Speaker 3: currently expecting, and it's going to take longer and be harder. 884 00:48:54,640 --> 00:48:58,680 Speaker 3: So big differences between what we're seeing and expecting and 885 00:48:58,719 --> 00:48:59,959 Speaker 3: what the markets are currently priced. 886 00:49:00,719 --> 00:49:03,279 Speaker 1: So I think last year you were talking about the 887 00:49:03,320 --> 00:49:06,919 Speaker 1: possibility of a recession in twenty twenty three. Is that 888 00:49:07,520 --> 00:49:10,880 Speaker 1: off the table now? So you're still positioned. It sounds 889 00:49:10,920 --> 00:49:15,200 Speaker 1: like for a level of higher inflation, but it sounds 890 00:49:15,239 --> 00:49:18,480 Speaker 1: like maybe you're a bit more optimistic on the growth front. 891 00:49:20,040 --> 00:49:22,239 Speaker 3: Yeah, we've been wrong on growth, So I'd say, look, 892 00:49:22,280 --> 00:49:24,319 Speaker 3: we think it's going to be a struggle. We're in 893 00:49:24,360 --> 00:49:27,400 Speaker 3: a state of disequilibrium in the sense that relative to 894 00:49:27,440 --> 00:49:29,120 Speaker 3: a given level of growth, we think the level of 895 00:49:29,120 --> 00:49:31,080 Speaker 3: inflation to the bad target that they're going to have 896 00:49:31,080 --> 00:49:34,440 Speaker 3: a difficulty achieving growth and inflation at the levels they 897 00:49:34,480 --> 00:49:36,280 Speaker 3: want and are going to have to give on something 898 00:49:36,960 --> 00:49:40,080 Speaker 3: in the short run. I think that's leading to you know, 899 00:49:40,400 --> 00:49:44,719 Speaker 3: higher rates. The expectation that the massive easing's coming is unlikely. 900 00:49:45,239 --> 00:49:47,120 Speaker 3: The Fed's going to continue to have to be tighter 901 00:49:47,280 --> 00:49:50,840 Speaker 3: longer than the markets expected. So that's bad for you know, 902 00:49:50,920 --> 00:49:54,440 Speaker 3: let's say bonds and long dated short rates. It's also 903 00:49:54,520 --> 00:49:57,279 Speaker 3: probably bad for equities. And at the same time, we 904 00:49:57,280 --> 00:50:02,920 Speaker 3: think growth will be struggling. It's nominal growth slowing. Penomenal 905 00:50:02,960 --> 00:50:05,160 Speaker 3: growth is going to continue to slow, and as nominal 906 00:50:05,200 --> 00:50:09,920 Speaker 3: growth slows, while you're more in stick your inflation, things 907 00:50:09,960 --> 00:50:12,839 Speaker 3: like wage growth and some of the service areas more 908 00:50:12,880 --> 00:50:15,720 Speaker 3: sticky inflation, you get more of a challenge. It's nominal 909 00:50:15,760 --> 00:50:17,760 Speaker 3: growth falls for it to just flow through to inflation. 910 00:50:18,600 --> 00:50:21,480 Speaker 3: So my views, you end up with growth disappointing a bit, 911 00:50:22,040 --> 00:50:25,200 Speaker 3: and inflation disappointing on the high side a bit ending 912 00:50:25,280 --> 00:50:28,719 Speaker 3: up you know, probably bad for bonds and probably you 913 00:50:28,719 --> 00:50:33,000 Speaker 3: know a little bit bad for equities, and generally weak, 914 00:50:33,760 --> 00:50:38,120 Speaker 3: weak growth, and if that weak growth starts to translate 915 00:50:38,200 --> 00:50:40,600 Speaker 3: into rising savings rate, you could easily end up into 916 00:50:40,680 --> 00:50:42,520 Speaker 3: it into a recession, and one that's going to be 917 00:50:42,600 --> 00:50:45,080 Speaker 3: difficult to deal with, you know. But yeah, I'd say 918 00:50:45,080 --> 00:50:48,600 Speaker 3: we've teamed. I've tamed, and we've tamed a bridgewater some degree. 919 00:50:48,640 --> 00:50:51,600 Speaker 3: Our view on growth, while still negative, not as extreme 920 00:50:51,680 --> 00:50:55,520 Speaker 3: as it appeared, and and it's a more gradual process 921 00:50:55,560 --> 00:50:59,239 Speaker 3: that's unfolding. And then on the inflation front, while we've 922 00:50:59,239 --> 00:51:02,239 Speaker 3: had a week I did a quick decline inflation as 923 00:51:02,320 --> 00:51:04,799 Speaker 3: novel GDP foul. We do think we're in the range 924 00:51:04,840 --> 00:51:07,080 Speaker 3: where you're in the much more stubborn part of inflation. 925 00:51:07,080 --> 00:51:10,840 Speaker 3: It's be harder to continue to get those inflation falls 926 00:51:10,880 --> 00:51:11,640 Speaker 3: going forward. 927 00:51:11,800 --> 00:51:13,960 Speaker 2: So just to be clear, though, you do think there 928 00:51:14,040 --> 00:51:17,120 Speaker 2: is a gap between either what the market sees in 929 00:51:17,200 --> 00:51:18,960 Speaker 2: terms of how much more work the FED is going 930 00:51:19,000 --> 00:51:21,600 Speaker 2: to have to do or what the FED thinks how 931 00:51:21,680 --> 00:51:23,480 Speaker 2: much more work the Fed is going to have to do, 932 00:51:24,080 --> 00:51:27,200 Speaker 2: and what basically you think the FED is going to 933 00:51:27,280 --> 00:51:29,800 Speaker 2: have to do if it actually is serious about getting 934 00:51:29,840 --> 00:51:32,719 Speaker 2: inflation back to something resembling its target. 935 00:51:32,960 --> 00:51:34,479 Speaker 3: Yeah, I think so. I mean, I'd say the FED 936 00:51:34,560 --> 00:51:36,600 Speaker 3: seems a little bit more realistic than the markets do 937 00:51:36,680 --> 00:51:38,680 Speaker 3: on what it's going to take. But right that, we 938 00:51:39,080 --> 00:51:40,759 Speaker 3: think that's right that when you look at what the 939 00:51:40,760 --> 00:51:44,520 Speaker 3: markets are saying, that it's super optimistic, it could come true. 940 00:51:44,760 --> 00:51:47,279 Speaker 3: You do need essentially to get an equity rally from here, 941 00:51:47,360 --> 00:51:50,799 Speaker 3: you have to have lower rates fairly quickly into a 942 00:51:50,840 --> 00:51:53,960 Speaker 3: world where earnings are pretty good. That's kind of the 943 00:51:53,960 --> 00:51:56,920 Speaker 3: discounted line. To get above that, you need even more 944 00:51:56,960 --> 00:51:59,520 Speaker 3: than that. And I think that line is super optimistic 945 00:51:59,600 --> 00:52:02,160 Speaker 3: relative to what we're you know, what we measure and 946 00:52:02,200 --> 00:52:05,880 Speaker 3: again are I'm using the words, but I'm describing the 947 00:52:06,000 --> 00:52:09,719 Speaker 3: process that's based on studying, you know, hundreds of years 948 00:52:09,760 --> 00:52:12,120 Speaker 3: of economic history and how these linkages work and building 949 00:52:12,160 --> 00:52:15,080 Speaker 3: all of that into a systematic process. But just spitting 950 00:52:15,080 --> 00:52:17,880 Speaker 3: out kind of the output of that is that it 951 00:52:17,920 --> 00:52:21,480 Speaker 3: doesn't appear that you'll that the FED will be able 952 00:52:21,520 --> 00:52:25,479 Speaker 3: to achieve that, and that we're in this disequilibrium where 953 00:52:25,480 --> 00:52:28,000 Speaker 3: you still have more inflation relative to growth, and you 954 00:52:28,040 --> 00:52:31,000 Speaker 3: don't have an easy way to close that gap. So 955 00:52:32,160 --> 00:52:35,680 Speaker 3: we'll see we've been wrong about that in terms of 956 00:52:35,680 --> 00:52:38,279 Speaker 3: at least what the market outcomes have been for the 957 00:52:38,360 --> 00:52:41,919 Speaker 3: last six months or so, after having been incredibly right 958 00:52:42,000 --> 00:52:44,319 Speaker 3: for an extended period of time. And that's part of it. 959 00:52:44,400 --> 00:52:46,280 Speaker 3: We get a lot of things wrong, and that's normal. 960 00:52:47,080 --> 00:52:49,080 Speaker 3: But I think when you break down why we got 961 00:52:49,120 --> 00:52:51,680 Speaker 3: it wrong and the ways in which that you know, 962 00:52:51,719 --> 00:52:53,600 Speaker 3: we've learned from that, and the ways in which our 963 00:52:53,640 --> 00:52:58,360 Speaker 3: processes have taken in new information, still leads to this 964 00:52:58,360 --> 00:53:03,640 Speaker 3: this view that that the markets are overly optimistic about 965 00:53:03,680 --> 00:53:04,759 Speaker 3: how easy that's going to be. 966 00:53:05,320 --> 00:53:08,120 Speaker 1: All right, Well, Greg, we appreciate you coming on and 967 00:53:08,520 --> 00:53:12,080 Speaker 1: outlining your thought process both around the markets and AI 968 00:53:12,280 --> 00:53:14,919 Speaker 1: and how you're actually deploying this new technology. So really 969 00:53:15,000 --> 00:53:17,200 Speaker 1: appreciate it. Thanks for coming back on the show. 970 00:53:17,800 --> 00:53:19,000 Speaker 3: My pleasure, good to talk to you. 971 00:53:19,040 --> 00:53:22,960 Speaker 2: Good luck in Vegas. Yeah, bring home another bracelet. 972 00:53:23,840 --> 00:53:24,359 Speaker 3: We'll try. 973 00:53:25,400 --> 00:53:26,399 Speaker 2: Thanks Greg. That was great. 974 00:53:39,360 --> 00:53:42,520 Speaker 1: So Joe, I feel like I have a slightly better 975 00:53:43,040 --> 00:53:47,600 Speaker 1: conception of exactly how this kind of technology can be 976 00:53:47,680 --> 00:53:50,680 Speaker 1: used for investing. So the idea of maybe you have 977 00:53:51,360 --> 00:53:56,040 Speaker 1: the AI models come up with the cs or ideas 978 00:53:56,200 --> 00:53:59,880 Speaker 1: that could then be rigorously fact checked because all they 979 00:54:00,080 --> 00:54:04,080 Speaker 1: the eyes are hallucinating and things like that. That makes 980 00:54:04,080 --> 00:54:04,600 Speaker 1: some sense. 981 00:54:05,200 --> 00:54:07,320 Speaker 2: Yes, absolutely, And I think you know you asked the 982 00:54:07,400 --> 00:54:10,080 Speaker 2: question it's like, can AI do our jobs? And I 983 00:54:10,120 --> 00:54:13,239 Speaker 2: don't think the answer is yes. And I think it's 984 00:54:13,239 --> 00:54:16,560 Speaker 2: like can AI replace the stock picker? It doesn't sound 985 00:54:16,640 --> 00:54:19,280 Speaker 2: like the AI is yes, But like, can the AI 986 00:54:20,200 --> 00:54:24,440 Speaker 2: augment augment the way someone's thinking, test come up with 987 00:54:24,560 --> 00:54:28,399 Speaker 2: theories that then can be rapidly tested. Have that sort 988 00:54:28,400 --> 00:54:30,680 Speaker 2: of go back and forth and sort of do some 989 00:54:30,760 --> 00:54:33,560 Speaker 2: of the work that you currently sort of like junior 990 00:54:33,600 --> 00:54:36,920 Speaker 2: analysts do in terms of like theory testing ideas and 991 00:54:36,960 --> 00:54:39,239 Speaker 2: stuff like that. You could see how it could be 992 00:54:39,640 --> 00:54:42,160 Speaker 2: a force multiplier at at a large fund. 993 00:54:42,440 --> 00:54:46,160 Speaker 1: Yeah, but I mean to that sort of turning point 994 00:54:46,360 --> 00:54:50,000 Speaker 1: question that also seems to be maybe the big weakness 995 00:54:50,120 --> 00:54:53,560 Speaker 1: here is that if you have an algorithm or a 996 00:54:53,600 --> 00:54:56,600 Speaker 1: model that's been trained on years and years of prior data, 997 00:54:56,760 --> 00:55:01,760 Speaker 1: so rates going lower and lower, in inflation staying below 998 00:55:01,920 --> 00:55:07,200 Speaker 1: two percent seems very difficult to project what might change. 999 00:55:07,000 --> 00:55:10,080 Speaker 2: Which, to Greg's point, humans aren't very good at that either. 1000 00:55:10,080 --> 00:55:11,600 Speaker 2: But you would hope, like, right, like, that's what we 1001 00:55:11,640 --> 00:55:13,960 Speaker 2: won't want to just be able to ask ch GPT 1002 00:55:15,440 --> 00:55:17,440 Speaker 2: or whatever. You know, I'm using that as like a 1003 00:55:17,480 --> 00:55:18,200 Speaker 2: stand in for. 1004 00:55:18,400 --> 00:55:21,600 Speaker 1: This, Yeah, or maybe maybe you ask Ai, like what 1005 00:55:21,719 --> 00:55:24,760 Speaker 1: would you need to see in order to start taking 1006 00:55:24,760 --> 00:55:26,799 Speaker 1: the prospect of regime change seriously? 1007 00:55:27,920 --> 00:55:29,719 Speaker 2: Yeah, I like, I mean you talk about this idea 1008 00:55:29,719 --> 00:55:32,760 Speaker 2: of like the sort of like adversarial way of thinking 1009 00:55:32,800 --> 00:55:34,600 Speaker 2: about it, which I think is really important. And you 1010 00:55:34,640 --> 00:55:37,239 Speaker 2: pointed out the sort of like disaster of the how 1011 00:55:37,560 --> 00:55:43,000 Speaker 2: the home eye buyers and then they got adversely selected 1012 00:55:43,080 --> 00:55:45,640 Speaker 2: because it's like, well, if Zillo is in the market, 1013 00:55:45,680 --> 00:55:48,240 Speaker 2: we know they're going to overpay, and so everyone suddenly 1014 00:55:48,280 --> 00:55:50,799 Speaker 2: dumps all the homes on Zillo and it was not 1015 00:55:51,040 --> 00:55:54,440 Speaker 2: anticipating its own role in the market. In response to 1016 00:55:54,480 --> 00:55:56,719 Speaker 2: your question, which I think is like a really interesting 1017 00:55:56,960 --> 00:55:57,839 Speaker 2: dimension to all. 1018 00:55:57,760 --> 00:56:01,120 Speaker 1: Of this, Yeah, that's sort of reflexivity between the models 1019 00:56:01,160 --> 00:56:03,520 Speaker 1: and the markets. I think we're probably going to be 1020 00:56:03,520 --> 00:56:06,560 Speaker 1: hearing a lot more about in the future. On that note, 1021 00:56:06,640 --> 00:56:07,440 Speaker 1: shall we leave it there? 1022 00:56:07,480 --> 00:56:08,560 Speaker 2: Let's leave it there, all right? 1023 00:56:08,640 --> 00:56:11,719 Speaker 1: This has been another episode of the Odd Thoughts podcast. 1024 00:56:11,800 --> 00:56:14,160 Speaker 1: I'm Tracy Alloway. You can follow me on Twitter at 1025 00:56:14,239 --> 00:56:15,440 Speaker 1: Tracy Alloway. 1026 00:56:15,160 --> 00:56:17,720 Speaker 2: And I'm Joe Wisenthal. You can follow me on Twitter 1027 00:56:17,800 --> 00:56:21,680 Speaker 2: at the Stalwart. Follow our producers on Twitter Carmen Rodriguez 1028 00:56:21,719 --> 00:56:25,360 Speaker 2: at Carmen Arman and Dashel Bennett at Dashbot. Follow all 1029 00:56:25,440 --> 00:56:29,040 Speaker 2: of the Bloomberg podcasts under the handle at podcasts. And 1030 00:56:29,080 --> 00:56:31,760 Speaker 2: for more odd Lots content, go to Bloomberg dot com 1031 00:56:31,920 --> 00:56:34,840 Speaker 2: slash odd Lots, where we have transcripts, a blog, and 1032 00:56:34,880 --> 00:56:36,920 Speaker 2: a newsletter, and for even more. If you want to 1033 00:56:37,000 --> 00:56:39,640 Speaker 2: chat with fellow listeners about all these topics, there's even 1034 00:56:39,680 --> 00:56:42,279 Speaker 2: an AI channel in there. Check out our discord twenty 1035 00:56:42,320 --> 00:56:44,840 Speaker 2: four to seven people talk about all these things Discord 1036 00:56:44,920 --> 00:56:47,040 Speaker 2: dot gg slash odd lots. 1037 00:56:47,239 --> 00:56:51,560 Speaker 1: And if you enjoy odd Lots. If you like these conversations, 1038 00:56:51,640 --> 00:56:54,960 Speaker 1: please leave us a review, a positive review please on 1039 00:56:55,080 --> 00:56:58,799 Speaker 1: your favorite podcast platform. We'd really appreciate it. Thanks for 1040 00:56:58,840 --> 00:57:05,120 Speaker 1: listening in 1041 00:57:17,800 --> 00:57:17,840 Speaker 3: In