1 00:00:02,720 --> 00:00:18,919 Speaker 1: Bloomberg Audio Studios, Podcasts, Radio News. Hello and welcome to 2 00:00:18,960 --> 00:00:22,360 Speaker 1: another episode of The Odd Lots Podcast. I'm Joe Wisenthal, 3 00:00:22,520 --> 00:00:26,760 Speaker 1: and I'm Tracy Alloway. Tracy, I'm very interested in AI. 4 00:00:27,600 --> 00:00:30,320 Speaker 1: No really, really, Joe, I am. 5 00:00:30,360 --> 00:00:31,040 Speaker 2: That's a surprise. 6 00:00:31,280 --> 00:00:34,840 Speaker 1: I'm very interested in the investment context, specifically, I mean 7 00:00:35,720 --> 00:00:39,239 Speaker 1: the actual implementation of like how do investors use it? 8 00:00:39,280 --> 00:00:43,760 Speaker 1: Because I think obviously just sort of substantively incredibly important 9 00:00:43,840 --> 00:00:47,280 Speaker 1: question for reasons they need no explaining. But I also 10 00:00:47,360 --> 00:00:50,360 Speaker 1: think it like raises very interesting sort of like puzzles 11 00:00:50,440 --> 00:00:54,880 Speaker 1: about what the technology is used for. And I remember 12 00:00:55,040 --> 00:00:58,000 Speaker 1: like when chadgic BT came out and people were like 13 00:00:58,480 --> 00:01:00,840 Speaker 1: asking it like what stock should I or we did 14 00:01:00,840 --> 00:01:04,080 Speaker 1: that prediction market episode recently, and it's like one thing 15 00:01:04,120 --> 00:01:06,560 Speaker 1: you definitely can't get much value out of is saying 16 00:01:06,560 --> 00:01:09,920 Speaker 1: like which contract should I buy? Or what's inflation going 17 00:01:09,959 --> 00:01:12,680 Speaker 1: to be so I can trade this contract? So like, 18 00:01:13,319 --> 00:01:16,280 Speaker 1: But that doesn't mean that there aren't interesting ways. It 19 00:01:16,400 --> 00:01:19,560 Speaker 1: just seems like, you know, the sort of the most 20 00:01:19,680 --> 00:01:23,919 Speaker 1: crude version of quote using AI for investing is obviously 21 00:01:24,640 --> 00:01:25,240 Speaker 1: a dead end. 22 00:01:25,720 --> 00:01:27,600 Speaker 2: DA here's the question I have. 23 00:01:28,120 --> 00:01:32,680 Speaker 3: You know, we've been through technological revolutions and investing before, notably, 24 00:01:32,880 --> 00:01:35,680 Speaker 3: you know, we had robo advisors, that's right, remember that, 25 00:01:35,800 --> 00:01:40,240 Speaker 3: we had systematic investing wants, we had high frequency trading. 26 00:01:40,800 --> 00:01:43,520 Speaker 3: My big question is how much of the current use 27 00:01:43,560 --> 00:01:47,040 Speaker 3: of AI is basically an iteration or an improvement on 28 00:01:47,240 --> 00:01:50,080 Speaker 3: some of those kind of you know, machine learning dynamics, 29 00:01:50,160 --> 00:01:54,840 Speaker 3: let's say, versus something more substantial or more revolutionary. Is 30 00:01:54,880 --> 00:01:57,920 Speaker 3: it like a minor evolution or a moderate evolution or 31 00:01:57,960 --> 00:01:58,520 Speaker 3: something big. 32 00:01:58,640 --> 00:02:01,040 Speaker 1: Well, this came up a lit little bit in our 33 00:02:01,160 --> 00:02:05,600 Speaker 1: conversation with Ian from Hudson River Trading. You and I'm 34 00:02:05,600 --> 00:02:08,640 Speaker 1: glad you brought up the quant example because like, all right, 35 00:02:08,880 --> 00:02:11,840 Speaker 1: big data in some sense has been part of quant 36 00:02:12,080 --> 00:02:15,240 Speaker 1: since the very beginning. Like, how do you establish that 37 00:02:15,639 --> 00:02:19,400 Speaker 1: quote value stocks outperform expensive stocks. Well, you just need 38 00:02:19,400 --> 00:02:22,240 Speaker 1: like a lot of data and computers and running that math, 39 00:02:22,280 --> 00:02:27,040 Speaker 1: et cetera to establish that fact. But this idea of yes, 40 00:02:27,080 --> 00:02:31,680 Speaker 1: maybe cheap stocks outperform expensive ones, momentum stocks outperform stocks 41 00:02:31,680 --> 00:02:34,760 Speaker 1: with bad momentum things that seem to be true, but 42 00:02:34,840 --> 00:02:36,880 Speaker 1: we don't really know why. And there's a lot of 43 00:02:36,919 --> 00:02:39,880 Speaker 1: disagreement as to the source of these quant things, and 44 00:02:39,960 --> 00:02:42,440 Speaker 1: that really makes me think about AI because we can 45 00:02:42,480 --> 00:02:45,520 Speaker 1: get these outputs from AI models that are obviously remarkable. 46 00:02:45,680 --> 00:02:48,800 Speaker 1: You can recognize patterns, yes, but they they can't really 47 00:02:48,840 --> 00:02:51,640 Speaker 1: explain how they arrived at that pattern. And I'm very 48 00:02:51,680 --> 00:02:54,680 Speaker 1: interested in this sort of where this leads us with 49 00:02:54,840 --> 00:02:58,160 Speaker 1: investing and whether it's like, Okay, do we get these 50 00:02:58,520 --> 00:03:02,600 Speaker 1: ideas and strategies, et cetera that work, but we can't 51 00:03:02,639 --> 00:03:04,480 Speaker 1: really articulate them in plann English. 52 00:03:04,600 --> 00:03:07,320 Speaker 3: I'm glad you brought up data as well, because one 53 00:03:07,320 --> 00:03:11,320 Speaker 3: thing you hear at basically every finance conference nactas is 54 00:03:11,360 --> 00:03:14,760 Speaker 3: the importance of data when it comes to running llms 55 00:03:14,760 --> 00:03:17,280 Speaker 3: and making sure that your data is actually processed, it's clean, 56 00:03:17,360 --> 00:03:19,400 Speaker 3: that you have a lot of it, and that you 57 00:03:19,520 --> 00:03:22,760 Speaker 3: have hopefully proprietary data, and the people we're going to 58 00:03:22,760 --> 00:03:24,519 Speaker 3: talk to have a lot of data. 59 00:03:24,560 --> 00:03:28,639 Speaker 1: Actually totally. There was just one more thing. I don't 60 00:03:28,639 --> 00:03:31,079 Speaker 1: know if you caught it last week. Were recording this 61 00:03:31,280 --> 00:03:34,200 Speaker 1: July seventh, last week while you're on vacation. It was 62 00:03:34,200 --> 00:03:38,240 Speaker 1: a very interesting paper out for Bridgewater talking about their 63 00:03:38,360 --> 00:03:42,240 Speaker 1: use of proprietary data to fine tune an open source 64 00:03:42,360 --> 00:03:45,880 Speaker 1: version of Quinn and for a specific purpose of being 65 00:03:45,880 --> 00:03:49,840 Speaker 1: able to identify what is newsworthy financial information and not. 66 00:03:50,400 --> 00:03:53,400 Speaker 1: They found that the combination of open source models plus 67 00:03:53,440 --> 00:03:57,600 Speaker 1: proprietary data got them better results than the most frontier 68 00:03:57,920 --> 00:04:00,960 Speaker 1: US models. Yep, that is very interesting to me, and 69 00:04:01,040 --> 00:04:02,960 Speaker 1: like these are the types of things that I'm very 70 00:04:03,000 --> 00:04:03,560 Speaker 1: curious about. 71 00:04:03,560 --> 00:04:05,560 Speaker 3: Where does is the secret sauce or the alpha actually 72 00:04:05,560 --> 00:04:06,040 Speaker 3: come from? 73 00:04:06,200 --> 00:04:06,640 Speaker 4: Totally? 74 00:04:06,720 --> 00:04:09,280 Speaker 1: Well, yeah, especially when everyone is going to have access 75 00:04:09,320 --> 00:04:13,560 Speaker 1: to like really really intelligent models. Anyway, enough talk from us. 76 00:04:13,560 --> 00:04:15,880 Speaker 1: I'm really excited. We have the perfect guests to talk 77 00:04:15,920 --> 00:04:20,200 Speaker 1: about this actually understanding the implementation of AI within the 78 00:04:20,520 --> 00:04:23,800 Speaker 1: investment or asset management context. We're going to be speaking 79 00:04:23,839 --> 00:04:27,520 Speaker 1: with Gary Collier Man Group CTO, as well as Tushara Fernando, 80 00:04:28,080 --> 00:04:30,800 Speaker 1: the firm's head of data and AI. So, like I said, 81 00:04:30,920 --> 00:04:33,880 Speaker 1: really the perfect guests. Gary and Tushara. Thank you both 82 00:04:33,960 --> 00:04:39,000 Speaker 1: so much for coming on odlogs. Let's start with Gary, 83 00:04:39,320 --> 00:04:41,440 Speaker 1: or maybe you know, we'll go both of you one 84 00:04:41,480 --> 00:04:43,039 Speaker 1: after the other. Why don't you just give us a 85 00:04:43,120 --> 00:04:46,039 Speaker 1: very quick description of your roles at Man Group and 86 00:04:46,200 --> 00:04:47,200 Speaker 1: what specifically you. 87 00:04:47,120 --> 00:04:51,119 Speaker 5: Do, so when I can start with that, perhaps useful 88 00:04:51,160 --> 00:04:53,960 Speaker 5: to give a bit of context of the firm. I 89 00:04:54,080 --> 00:04:57,560 Speaker 5: describe Mangrove really as a full spectro back to asset manager, 90 00:04:57,640 --> 00:05:01,239 Speaker 5: So full spectrum being recover tips, we cover long only, 91 00:05:01,600 --> 00:05:04,599 Speaker 5: we cover public markets, we cover private markets, we cover 92 00:05:04,680 --> 00:05:08,680 Speaker 5: consystematic cover fundamental discretion investing. And we've got a solutions 93 00:05:08,720 --> 00:05:12,040 Speaker 5: business that can bring all that together into custom client mandates. 94 00:05:12,600 --> 00:05:15,200 Speaker 5: And what that means in terms of CTO role I 95 00:05:15,320 --> 00:05:17,720 Speaker 5: described as as like full spectrum too, because all of 96 00:05:17,760 --> 00:05:21,120 Speaker 5: the above, we're very opinionated in what every part of 97 00:05:21,160 --> 00:05:23,760 Speaker 5: the text that looks like, right down from choice of servers, 98 00:05:23,800 --> 00:05:26,960 Speaker 5: networking devices, all the way through to the end user 99 00:05:27,360 --> 00:05:30,680 Speaker 5: facing pieces of software, and simply left to right. We 100 00:05:30,800 --> 00:05:33,960 Speaker 5: build an awful lot of our own technology right from 101 00:05:34,080 --> 00:05:37,680 Speaker 5: the custom data feeds all the way through research frameworks, 102 00:05:37,800 --> 00:05:41,960 Speaker 5: trading systems, and the middle and back office operating platform 103 00:05:42,040 --> 00:05:44,640 Speaker 5: So very much force spectrum. 104 00:05:44,200 --> 00:05:45,520 Speaker 1: Role to sure. 105 00:05:45,880 --> 00:05:46,080 Speaker 6: Yeah. 106 00:05:46,160 --> 00:05:49,080 Speaker 7: So on my side, I look after data and AI. 107 00:05:49,480 --> 00:05:53,960 Speaker 7: So that's everything from market data to alternative data, to 108 00:05:54,640 --> 00:05:56,960 Speaker 7: all of the context and knowledge that we have that 109 00:05:57,040 --> 00:06:00,800 Speaker 7: we plug into our AI models. Then from an perspective, 110 00:06:00,839 --> 00:06:04,000 Speaker 7: it's about how do we give the great capabilities that 111 00:06:04,040 --> 00:06:07,400 Speaker 7: we now have to our quants, to our researchers and 112 00:06:07,440 --> 00:06:09,960 Speaker 7: to our fundamental portfolio managers. 113 00:06:10,240 --> 00:06:12,200 Speaker 3: Yeah, Tshara, I wanted to ask you about this because 114 00:06:12,200 --> 00:06:14,359 Speaker 3: I think your title actually used to be head of 115 00:06:14,440 --> 00:06:16,640 Speaker 3: Data and Machine Learning and now it's Head of Data 116 00:06:16,800 --> 00:06:19,400 Speaker 3: and AI, Like, what exactly how do you think about 117 00:06:19,400 --> 00:06:23,560 Speaker 3: the difference between those two terms machine learning versus AI. 118 00:06:24,760 --> 00:06:27,040 Speaker 7: That's a really good question, and I think it linked 119 00:06:27,120 --> 00:06:30,920 Speaker 7: to one of the earlier points that you had around 120 00:06:31,120 --> 00:06:34,279 Speaker 7: whether it was that we were using traditional machine learning 121 00:06:34,320 --> 00:06:38,039 Speaker 7: techniques or whether we were using generative AI. And I 122 00:06:38,040 --> 00:06:44,359 Speaker 7: think historically machine learning in a quantum was really about 123 00:06:44,760 --> 00:06:48,000 Speaker 7: using these more traditional machine learning techniques that were looking 124 00:06:48,040 --> 00:06:51,320 Speaker 7: at things like linear aggressions, they were looking at neural 125 00:06:51,400 --> 00:06:56,240 Speaker 7: nets to try to predict some future, whereas. 126 00:06:55,880 --> 00:06:58,920 Speaker 4: Now with the onset of generative AI, is more than that. 127 00:06:59,000 --> 00:07:02,680 Speaker 7: It's really about enablement of people to create things as 128 00:07:02,680 --> 00:07:06,640 Speaker 7: well as using these traditional techniques. So that's why we've 129 00:07:06,720 --> 00:07:10,280 Speaker 7: changed the title of the role, because it encompasses both 130 00:07:10,320 --> 00:07:13,600 Speaker 7: the generative AI aspects as well as the more traditional 131 00:07:13,640 --> 00:07:14,760 Speaker 7: machine learning methods. 132 00:07:14,920 --> 00:07:15,800 Speaker 2: Okay, that makes sense. 133 00:07:15,840 --> 00:07:18,760 Speaker 3: So if I was sitting within a man group office 134 00:07:18,760 --> 00:07:22,160 Speaker 3: today and I take Gary laid it out perfectly well, 135 00:07:22,200 --> 00:07:23,840 Speaker 3: you guys have a lot of different roles, So you 136 00:07:23,880 --> 00:07:27,200 Speaker 3: have discretionary sort of traditional portfolio managers, and then you 137 00:07:27,280 --> 00:07:30,280 Speaker 3: have more systematic quants and all of that. But if 138 00:07:30,320 --> 00:07:35,400 Speaker 3: I'm shadowing one of your traders or pms, and I'm 139 00:07:35,400 --> 00:07:37,920 Speaker 3: watching what they're doing on a screen today versus what 140 00:07:37,960 --> 00:07:40,119 Speaker 3: they were doing on a screen in let's say twenty 141 00:07:40,200 --> 00:07:43,760 Speaker 3: twenty two or something like that, what exactly has changed 142 00:07:43,920 --> 00:07:45,840 Speaker 3: with the use of generative AI? 143 00:07:46,000 --> 00:07:47,160 Speaker 2: What are you doing differently? 144 00:07:48,280 --> 00:07:50,960 Speaker 5: I think if you were to walk across the floor, 145 00:07:51,920 --> 00:07:56,840 Speaker 5: you would say different windows, different tools in use now, 146 00:07:56,960 --> 00:08:00,640 Speaker 5: regardless of role, whether it be trading, quant research, or 147 00:08:00,760 --> 00:08:04,680 Speaker 5: discretionary investing. And so chatting to a couple of discretionary 148 00:08:04,720 --> 00:08:08,160 Speaker 5: analysts last week and they're commenting to me on how 149 00:08:08,520 --> 00:08:12,920 Speaker 5: everywhere you look across the discretionary floor, everybody has got 150 00:08:12,920 --> 00:08:16,000 Speaker 5: an AI focused like window on their screen, and so 151 00:08:16,560 --> 00:08:21,760 Speaker 5: it's affecting, improving, like augmenting pretty much every role we 152 00:08:21,880 --> 00:08:25,080 Speaker 5: have in the firm, of course, in different ways depending 153 00:08:25,120 --> 00:08:27,680 Speaker 5: on what the role is, but there's no role that's 154 00:08:27,760 --> 00:08:29,800 Speaker 5: unaffected by AI. 155 00:08:30,120 --> 00:08:32,960 Speaker 1: What are they doing again it Rather than talk about 156 00:08:34,480 --> 00:08:39,079 Speaker 1: let's say I'm like a sort of traditional long only 157 00:08:39,920 --> 00:08:42,640 Speaker 1: stock selector, like the sort of classic thing that in 158 00:08:42,679 --> 00:08:45,240 Speaker 1: our minds we often think of it as an asset manager, 159 00:08:45,240 --> 00:08:48,600 Speaker 1: and it's like, Okay, I have now access to some 160 00:08:48,679 --> 00:08:51,880 Speaker 1: of the world's most advanced models. What am I doing 161 00:08:51,920 --> 00:08:52,240 Speaker 1: with them? 162 00:08:52,960 --> 00:08:56,720 Speaker 7: So if you think about traditionally how PM like that 163 00:08:56,760 --> 00:09:00,800 Speaker 7: would work. They want to look across broad set of 164 00:09:00,880 --> 00:09:04,040 Speaker 7: names and they want to get access to as much 165 00:09:04,240 --> 00:09:08,200 Speaker 7: data as possible for those names. So they want to 166 00:09:08,240 --> 00:09:11,000 Speaker 7: look at earnings reports, they want to look at broker research, 167 00:09:11,080 --> 00:09:13,920 Speaker 7: they want to look at alternative data. They want to 168 00:09:13,960 --> 00:09:16,760 Speaker 7: look at podcasts. But there's only so much time in 169 00:09:16,800 --> 00:09:18,800 Speaker 7: the day. There's a few hours in the day, and 170 00:09:18,960 --> 00:09:21,640 Speaker 7: there's fifty names. How are they going to cover them all? 171 00:09:22,200 --> 00:09:26,600 Speaker 7: How do they get to the really important insight? And 172 00:09:27,600 --> 00:09:30,320 Speaker 7: what AI has allowed us to do is allowed us 173 00:09:30,360 --> 00:09:33,720 Speaker 7: to access all of those different types of data, lots 174 00:09:33,760 --> 00:09:41,640 Speaker 7: of different modalities, podcasts, alternative data, things like broker research reports, 175 00:09:42,040 --> 00:09:46,320 Speaker 7: and synthesize them into what's actually meaningful so that a 176 00:09:46,400 --> 00:09:50,640 Speaker 7: PM can asynchronously get that inside. They could go away 177 00:09:50,640 --> 00:09:53,319 Speaker 7: for a coffee, they can come in overnight and an 178 00:09:53,360 --> 00:09:57,880 Speaker 7: agent that has actually distilled a change that's happened on 179 00:09:57,920 --> 00:10:03,319 Speaker 7: the internet and given them the insider's meaningful to their portfolio. 180 00:10:02,760 --> 00:10:04,079 Speaker 4: To their investment pieces. 181 00:10:04,960 --> 00:10:09,320 Speaker 7: So just I suppose an example might be recently we 182 00:10:09,760 --> 00:10:13,600 Speaker 7: had a PM that was covering the AI trade and 183 00:10:13,640 --> 00:10:17,360 Speaker 7: really they're one of the important things is about where's 184 00:10:17,360 --> 00:10:20,200 Speaker 7: the bottomneck? Where's the bottomneck here? 185 00:10:19,520 --> 00:10:22,160 Speaker 4: Yea in is it? 186 00:10:22,280 --> 00:10:22,480 Speaker 6: Yeah? 187 00:10:23,040 --> 00:10:25,320 Speaker 4: And it's hard to know exactly where it is. 188 00:10:25,360 --> 00:10:29,480 Speaker 7: And there was a podcast cast from one of the 189 00:10:29,520 --> 00:10:33,400 Speaker 7: heads of engineering from a large hyper scaler and he 190 00:10:33,679 --> 00:10:39,439 Speaker 7: was saying that it was increasingly important to have more 191 00:10:39,440 --> 00:10:43,040 Speaker 7: and more GPS to train models, but data centers were 192 00:10:43,200 --> 00:10:47,320 Speaker 7: becoming more scarce, and it was becoming increasingly difficult to 193 00:10:48,800 --> 00:10:50,719 Speaker 7: find data centers that were actually. 194 00:10:50,480 --> 00:10:51,800 Speaker 4: Big enough to train these models. 195 00:10:51,840 --> 00:10:54,520 Speaker 7: So what you could see there was a couple of 196 00:10:54,559 --> 00:10:57,080 Speaker 7: things like one that there's GPU scarce see and the 197 00:10:57,120 --> 00:10:59,640 Speaker 7: other thing is that it's likely that we're going to 198 00:10:59,720 --> 00:11:03,079 Speaker 7: need better networking between these large data sensors in the future. 199 00:11:03,200 --> 00:11:06,040 Speaker 7: And that was something that came on a podcast from 200 00:11:06,080 --> 00:11:09,440 Speaker 7: a head of engineering. This isn't someone who PM would 201 00:11:09,520 --> 00:11:14,080 Speaker 7: usually interact with. They don't go to the investicles, that're 202 00:11:14,559 --> 00:11:18,120 Speaker 7: not somebody they often have access to. But through AI, 203 00:11:18,400 --> 00:11:21,800 Speaker 7: the PM was able to have an AI agent transcribe 204 00:11:21,840 --> 00:11:25,240 Speaker 7: that podcast that synthesized that data so that they could 205 00:11:25,280 --> 00:11:27,920 Speaker 7: get better insight into that investment idea. 206 00:11:28,840 --> 00:11:31,559 Speaker 3: Yeah, I feel like podcasts are an important source of 207 00:11:31,600 --> 00:11:32,160 Speaker 3: Alpha Joe. 208 00:11:32,240 --> 00:11:33,880 Speaker 2: Everyone shall be listening to podcasts, you. 209 00:11:33,920 --> 00:11:37,280 Speaker 1: Know, the non joke version of the which is one 210 00:11:37,320 --> 00:11:40,439 Speaker 1: of the things I people say, we had Alex Namas 211 00:11:40,559 --> 00:11:42,200 Speaker 1: on the podcast, and it's like, what's going to be 212 00:11:42,240 --> 00:11:43,480 Speaker 1: scarceyftter agi. 213 00:11:43,600 --> 00:11:43,840 Speaker 6: Yeah. 214 00:11:43,880 --> 00:11:46,520 Speaker 1: People who are able to get good guests on their podcast. 215 00:11:46,559 --> 00:11:49,160 Speaker 1: People who are able to get those people, you know. 216 00:11:49,400 --> 00:11:52,199 Speaker 2: Okay, so we're clearly talking our own book. Yes, okay. 217 00:11:52,200 --> 00:11:54,840 Speaker 3: Setting that aside for a second, it sounds like what 218 00:11:55,160 --> 00:11:58,040 Speaker 3: most of your doing, and when you describe that process 219 00:11:58,120 --> 00:12:02,880 Speaker 3: is augmenting research, you're allowing your investors, your managers to 220 00:12:02,920 --> 00:12:05,839 Speaker 3: be more efficient in their research process. There's a lot 221 00:12:05,840 --> 00:12:09,440 Speaker 3: of talk nowadays about agentic workflows and the actual idea 222 00:12:09,480 --> 00:12:12,520 Speaker 3: that instead of having humans drive every step of a 223 00:12:12,559 --> 00:12:16,280 Speaker 3: particular trade or project, you have a system that can 224 00:12:16,360 --> 00:12:18,920 Speaker 3: you know, do the research. It can generate ideas, it 225 00:12:18,960 --> 00:12:21,480 Speaker 3: can test the hypotheses, and then it can even execute 226 00:12:21,520 --> 00:12:25,320 Speaker 3: on them. Is that something that you're sort of working 227 00:12:25,480 --> 00:12:29,440 Speaker 3: towards or is it still too far away for you 228 00:12:29,480 --> 00:12:31,520 Speaker 3: guys to even be contemplating. 229 00:12:32,440 --> 00:12:33,960 Speaker 6: No, that's not too far away at all. 230 00:12:34,000 --> 00:12:36,559 Speaker 5: In fact, that's something we've been working on for quite 231 00:12:36,559 --> 00:12:39,480 Speaker 5: some time now. And if we shift focus to the 232 00:12:39,559 --> 00:12:41,320 Speaker 5: quant the systematic parts. 233 00:12:40,960 --> 00:12:44,280 Speaker 6: Of the business, what we've been doing now. 234 00:12:44,280 --> 00:12:46,640 Speaker 5: Look, if you take a step back and think, well, 235 00:12:47,720 --> 00:12:52,120 Speaker 5: technology plus quant techniques, that gives you the ability to 236 00:12:52,200 --> 00:12:55,840 Speaker 5: build systematic strategies. So what do we get if we 237 00:12:55,920 --> 00:12:59,800 Speaker 5: add AI in that mix. Well, we have the ability 238 00:12:59,840 --> 00:13:03,679 Speaker 5: to think about systematizing the way that we build systematic 239 00:13:03,720 --> 00:13:08,280 Speaker 5: strategies to begin with, and in effect giving a big force. 240 00:13:08,120 --> 00:13:10,400 Speaker 4: Multiplier, big leverage multiplier to our. 241 00:13:10,679 --> 00:13:12,120 Speaker 6: Quants, our researchers. 242 00:13:12,600 --> 00:13:16,360 Speaker 5: So what we've been doing there for well over a year, 243 00:13:16,480 --> 00:13:19,960 Speaker 5: actually a year and a half is building a system 244 00:13:20,120 --> 00:13:24,199 Speaker 5: that can actually take all of those different parts of 245 00:13:24,240 --> 00:13:28,120 Speaker 5: the quant research process. So the idea formation part looking 246 00:13:28,200 --> 00:13:33,520 Speaker 5: at academic papers, looking at carefully labeled data sets, reasoning 247 00:13:34,040 --> 00:13:36,959 Speaker 5: about the content of those papers, the content of those 248 00:13:37,080 --> 00:13:41,080 Speaker 5: data sets, are the economic hypotheses in there that could 249 00:13:41,120 --> 00:13:43,680 Speaker 5: be real or at least could be worth testing out, 250 00:13:44,240 --> 00:13:48,839 Speaker 5: and then having other agents build the code to encapsulate 251 00:13:48,920 --> 00:13:52,600 Speaker 5: those ideas, get the right market data, run the right 252 00:13:52,640 --> 00:13:57,160 Speaker 5: back tests, etc. And then further agents that evaluate the 253 00:13:57,200 --> 00:13:58,959 Speaker 5: output of the prior agent. 254 00:13:59,080 --> 00:13:59,760 Speaker 6: So this is. 255 00:14:00,200 --> 00:14:03,720 Speaker 5: Really one of these earlier ideas that we thought would 256 00:14:03,720 --> 00:14:06,000 Speaker 5: give potentially great bang per book, and we're working on 257 00:14:06,120 --> 00:14:09,000 Speaker 5: some time and just to make it real, to demonstrate 258 00:14:09,080 --> 00:14:11,400 Speaker 5: this is not just something that's happening in the lab 259 00:14:11,440 --> 00:14:15,000 Speaker 5: but doesn't have any real or practical consequences. There's been 260 00:14:15,080 --> 00:14:18,040 Speaker 5: a bunch I think fifteen twenty models at the last 261 00:14:18,160 --> 00:14:21,440 Speaker 5: count that had gone all the way through. They started 262 00:14:21,480 --> 00:14:25,920 Speaker 5: off as models that were ideated by AI, went all 263 00:14:25,960 --> 00:14:31,520 Speaker 5: the way through the signal construction, the validation process, were 264 00:14:31,600 --> 00:14:37,320 Speaker 5: reviewed and validated by a human investment committee and deemed 265 00:14:37,440 --> 00:14:40,840 Speaker 5: fit and proper to trade our client's assets with. 266 00:14:41,080 --> 00:14:43,040 Speaker 6: So this is very real. It's not to make believe 267 00:14:43,040 --> 00:14:43,240 Speaker 6: at this. 268 00:14:43,280 --> 00:15:02,760 Speaker 1: Point, should you ma mention sort of hoovering up information 269 00:15:02,880 --> 00:15:06,680 Speaker 1: that might appear on a podcast where someone identifies a 270 00:15:06,760 --> 00:15:10,440 Speaker 1: particular bottleneck in their business. We are on a podcast. 271 00:15:10,840 --> 00:15:13,400 Speaker 1: What is your bottleneck? What would you like to have 272 00:15:13,680 --> 00:15:17,280 Speaker 1: more of right now in order to improve your process? 273 00:15:17,640 --> 00:15:20,480 Speaker 1: Is it compute? Is it data? Or if you could 274 00:15:20,480 --> 00:15:22,520 Speaker 1: snap your fingers and have more of something, what would 275 00:15:22,520 --> 00:15:24,240 Speaker 1: it be, Well. 276 00:15:24,280 --> 00:15:26,880 Speaker 5: You can always use more compute and you can always 277 00:15:26,960 --> 00:15:29,480 Speaker 5: use more data. But I think what the bigger problem 278 00:15:29,960 --> 00:15:34,560 Speaker 5: is the world of opportunity that's supported by the AI 279 00:15:34,680 --> 00:15:37,600 Speaker 5: capabilities that we're seeing and of course being delivered all 280 00:15:37,600 --> 00:15:42,600 Speaker 5: the time. That envelope is expanding so quickly that keeping 281 00:15:42,680 --> 00:15:45,000 Speaker 5: up with it from a human perspective can be. 282 00:15:45,200 --> 00:15:45,880 Speaker 6: Like quite hard. 283 00:15:45,880 --> 00:15:49,320 Speaker 5: There's no sources of good ideas. Filtering ideas is important, 284 00:15:49,760 --> 00:15:53,360 Speaker 5: but of course, whilst we're used to making lots of 285 00:15:53,480 --> 00:16:01,040 Speaker 5: change mang Group, the level of organizations or change that 286 00:16:01,120 --> 00:16:03,240 Speaker 5: we need to get involved in to put all of 287 00:16:03,320 --> 00:16:07,000 Speaker 5: this stuff into effect, I think that's the bottleneck making 288 00:16:07,040 --> 00:16:09,520 Speaker 5: sure because we're a regulated business, we've got produciary duty. 289 00:16:09,920 --> 00:16:11,680 Speaker 5: We need to make sure that the things that we're doing, 290 00:16:11,720 --> 00:16:14,880 Speaker 5: the things were deploying, are done with the minimum amounts 291 00:16:14,920 --> 00:16:17,960 Speaker 5: of risk. And there's a lot of work in this 292 00:16:18,280 --> 00:16:22,000 Speaker 5: field that really I think the industry doesn't have firm 293 00:16:22,000 --> 00:16:27,360 Speaker 5: answers to yet. I mean evaluations testing the results of 294 00:16:27,400 --> 00:16:31,440 Speaker 5: AI processes are good and proper. That's a rapidly expanding field. 295 00:16:31,920 --> 00:16:35,680 Speaker 5: Thinking about how we run agents across the business in 296 00:16:35,760 --> 00:16:38,920 Speaker 5: a safe and controlled fashion. You know, we all, I'm 297 00:16:38,960 --> 00:16:41,240 Speaker 5: sure have seen and smiled at the well the AI 298 00:16:41,320 --> 00:16:44,120 Speaker 5: deleted my inbox or the AI deleted all my photos 299 00:16:44,360 --> 00:16:47,000 Speaker 5: and I can't get them back. We can't afford to 300 00:16:47,040 --> 00:16:49,960 Speaker 5: have that type of thing happen at an enterprise level. 301 00:16:50,080 --> 00:16:53,920 Speaker 5: So it's making sure that we can move quickly but safely. 302 00:16:54,000 --> 00:16:56,160 Speaker 5: I think is the bottleneck if you like. 303 00:16:56,400 --> 00:16:59,239 Speaker 1: Yeah, that makes sense. Let's get into a couple of specifics, 304 00:16:59,560 --> 00:17:01,160 Speaker 1: and I want to get back to the sort of 305 00:17:01,240 --> 00:17:04,280 Speaker 1: how you move forward safely. But both of you have 306 00:17:04,480 --> 00:17:08,359 Speaker 1: now mentioned filtering, and that actually is precisely what I 307 00:17:08,480 --> 00:17:11,520 Speaker 1: mentioned in the intro. There was a paper from Bridgewater 308 00:17:11,840 --> 00:17:15,080 Speaker 1: about fine tuning a version of Quen on their own data, 309 00:17:15,359 --> 00:17:19,720 Speaker 1: precisely for better filtering, so that the pms could just be, 310 00:17:19,920 --> 00:17:22,159 Speaker 1: you know, have more efficient use of their time to 311 00:17:22,200 --> 00:17:24,879 Speaker 1: see what signal. How do you build that? Is this 312 00:17:25,680 --> 00:17:28,520 Speaker 1: an off the shelf thing? What is the tech the model, 313 00:17:28,600 --> 00:17:32,520 Speaker 1: et cetera that you have to like this ingestion pipeline, like, 314 00:17:32,800 --> 00:17:33,720 Speaker 1: what is it consist of? 315 00:17:34,680 --> 00:17:35,760 Speaker 6: Yeah, I can say that one. 316 00:17:35,880 --> 00:17:38,600 Speaker 7: So I think for us it really depends on the 317 00:17:38,680 --> 00:17:42,800 Speaker 7: type of data. There's broadly three types of data that 318 00:17:42,920 --> 00:17:45,520 Speaker 7: we look at the first one is market data that's 319 00:17:45,600 --> 00:17:50,240 Speaker 7: often very structured tick data, things like order books. We 320 00:17:50,880 --> 00:17:55,800 Speaker 7: ingest every tick from most exchanges, almost a terraby of 321 00:17:55,920 --> 00:18:01,000 Speaker 7: data just from ticks per day. And then we have 322 00:18:02,080 --> 00:18:08,480 Speaker 7: alternative and unstructured data that is much more messy, it's 323 00:18:08,600 --> 00:18:13,119 Speaker 7: much more malformed, and there it's really about how we 324 00:18:13,240 --> 00:18:16,160 Speaker 7: structure the data, how we tag it, how we connect 325 00:18:16,200 --> 00:18:16,960 Speaker 7: it with each other. 326 00:18:17,280 --> 00:18:19,480 Speaker 4: What's the knowledge layer on top of it. 327 00:18:19,680 --> 00:18:23,800 Speaker 7: How do you think about the connectivity between tickers and 328 00:18:23,920 --> 00:18:29,120 Speaker 7: sectors and companies. How does AI actually interrogate that data 329 00:18:29,280 --> 00:18:34,640 Speaker 7: in a way that is uniform. What's the common language 330 00:18:34,680 --> 00:18:38,240 Speaker 7: between all of those data sets? And then the last 331 00:18:38,280 --> 00:18:42,919 Speaker 7: piece is really something that's quite new. It's our institutional knowledge, 332 00:18:42,960 --> 00:18:46,320 Speaker 7: it's our context, and this is becoming a new layer 333 00:18:46,400 --> 00:18:50,399 Speaker 7: in our data architecture. How do we tell AI about 334 00:18:50,720 --> 00:18:53,879 Speaker 7: our processes, how do we make it speak mang group, 335 00:18:53,960 --> 00:18:56,200 Speaker 7: how do we tell it the right way to run 336 00:18:56,280 --> 00:18:56,880 Speaker 7: a back test? 337 00:18:57,359 --> 00:19:00,800 Speaker 4: So those are really the three areas that we focus on. 338 00:19:02,000 --> 00:19:06,080 Speaker 1: Sorry, just pushing on this point further, Like all of 339 00:19:06,160 --> 00:19:09,400 Speaker 1: that makes a ton of sense to me, and especially 340 00:19:09,520 --> 00:19:11,680 Speaker 1: you know the second two or like big problems and 341 00:19:12,080 --> 00:19:15,680 Speaker 1: we know the generative AI in particular solves a lot 342 00:19:15,840 --> 00:19:18,800 Speaker 1: of the unstructured data problem. You can actually get a 343 00:19:18,840 --> 00:19:22,240 Speaker 1: lot of signal from you know, yeah, a bunch of 344 00:19:22,320 --> 00:19:25,000 Speaker 1: texts dump dumped in a file. But what is what 345 00:19:25,119 --> 00:19:27,040 Speaker 1: are you building? Like how does it work? And like 346 00:19:27,400 --> 00:19:31,480 Speaker 1: do you have to can off the shelf frontier models? 347 00:19:31,560 --> 00:19:33,920 Speaker 1: Are they the best for the job? Do you do 348 00:19:34,119 --> 00:19:37,320 Speaker 1: your own in house fine tuning of open source models? 349 00:19:37,359 --> 00:19:40,520 Speaker 1: Like right today? What is the best tech stack for 350 00:19:40,680 --> 00:19:41,560 Speaker 1: that part of the process. 351 00:19:42,400 --> 00:19:45,080 Speaker 7: So we have historically looked at fine tuning for a 352 00:19:45,200 --> 00:19:47,960 Speaker 7: couple of cases, but where we're seeing the most bang 353 00:19:48,080 --> 00:19:51,359 Speaker 7: for our book at the moment is proper tagging and 354 00:19:51,800 --> 00:19:56,119 Speaker 7: structuring of the data, so pre processing. What we found is, though, 355 00:19:56,119 --> 00:19:58,560 Speaker 7: if you take a data set like credit card data, 356 00:19:58,640 --> 00:20:02,200 Speaker 7: for example, AI can look at that, It can see 357 00:20:02,240 --> 00:20:04,160 Speaker 7: the columns, it can see the rows, but it doesn't 358 00:20:04,200 --> 00:20:07,560 Speaker 7: really understand the nuances of it. It doesn't really understand 359 00:20:08,040 --> 00:20:10,639 Speaker 7: what it is. So what we're having to do is 360 00:20:10,800 --> 00:20:15,480 Speaker 7: invest in trying to add extra color, extra metadata to 361 00:20:15,840 --> 00:20:20,399 Speaker 7: that information. We want to have descriptors to tell it 362 00:20:20,560 --> 00:20:25,239 Speaker 7: the nuances. We will say things like when you look 363 00:20:25,280 --> 00:20:29,320 Speaker 7: at this data set, each row really means that a 364 00:20:29,400 --> 00:20:31,960 Speaker 7: person has gone into a shop and bought something, and 365 00:20:32,080 --> 00:20:34,440 Speaker 7: you tell AI that in plain English, you give it 366 00:20:34,560 --> 00:20:35,439 Speaker 7: those descriptors. 367 00:20:35,520 --> 00:20:37,919 Speaker 4: Yeah, and you do that for all of your data sets. 368 00:20:38,200 --> 00:20:41,680 Speaker 7: And then the second piece is how do you connect 369 00:20:42,000 --> 00:20:47,080 Speaker 7: lots of data sets together. They use very different language, 370 00:20:47,160 --> 00:20:51,000 Speaker 7: they use very different semantics. So investing in a shared 371 00:20:51,080 --> 00:20:55,080 Speaker 7: language is said semantic layer are shared way of doing 372 00:20:55,240 --> 00:20:57,760 Speaker 7: things is something that we've had to do so that 373 00:20:57,880 --> 00:21:00,560 Speaker 7: we tag all of the columns with this unified language 374 00:21:01,040 --> 00:21:05,840 Speaker 7: so the AI can connect different data sets together, because 375 00:21:05,880 --> 00:21:10,120 Speaker 7: that's really an important thing in idea generation. That knows 376 00:21:10,280 --> 00:21:13,399 Speaker 7: that a field in one data set is linked to 377 00:21:13,760 --> 00:21:16,240 Speaker 7: a data set to a field in another data set, 378 00:21:16,680 --> 00:21:22,240 Speaker 7: it can quickly navigate between tickers and sectors for macro. 379 00:21:22,160 --> 00:21:25,440 Speaker 3: Insight, what do you think is most important having the 380 00:21:25,560 --> 00:21:30,560 Speaker 3: latest frontier model or a beautiful set of structured, tagged 381 00:21:30,840 --> 00:21:31,680 Speaker 3: and labeled data. 382 00:21:32,480 --> 00:21:34,440 Speaker 7: Yeah, I'd say it depends on the task that you're 383 00:21:34,440 --> 00:21:36,640 Speaker 7: trying to do. If you're looking at a coding task, 384 00:21:37,240 --> 00:21:41,440 Speaker 7: you really want the latest frontier models. But for a 385 00:21:41,760 --> 00:21:47,080 Speaker 7: quan research task, I think that looking at the underlying 386 00:21:47,200 --> 00:21:52,600 Speaker 7: data is what fuels alpha research. Trying to just use 387 00:21:52,640 --> 00:21:54,320 Speaker 7: a frontier model will get you nowhere. 388 00:21:55,520 --> 00:21:57,359 Speaker 1: So this is something that comes up in a lot 389 00:21:57,440 --> 00:22:00,879 Speaker 1: of our AI deployment questions, and maybe both of you 390 00:22:01,119 --> 00:22:03,800 Speaker 1: could take this. We have all these different teams, and 391 00:22:03,880 --> 00:22:09,399 Speaker 1: I assume everybody wants really, including people who don't know 392 00:22:09,520 --> 00:22:12,760 Speaker 1: much about AI, intuitively think, oh, I want the strongest 393 00:22:12,880 --> 00:22:15,200 Speaker 1: version of the model. I want to opus four point a, 394 00:22:15,359 --> 00:22:20,400 Speaker 1: and I want fable whatever. And I assume, okay, yes, 395 00:22:20,520 --> 00:22:25,920 Speaker 1: for like deep computational tasks or solve engineering problems, coding problems, Yes, 396 00:22:26,000 --> 00:22:27,720 Speaker 1: probably those are the best, but there are probably a 397 00:22:27,800 --> 00:22:30,879 Speaker 1: lot of people who do not need anything like that 398 00:22:31,040 --> 00:22:33,680 Speaker 1: at all. How do you think about the question of 399 00:22:34,160 --> 00:22:40,040 Speaker 1: internal provision of token consumption and not wasting money by 400 00:22:40,119 --> 00:22:43,520 Speaker 1: having people use the most advanced models, but also giving 401 00:22:43,640 --> 00:22:47,280 Speaker 1: people enough green space to explore and figure out what 402 00:22:47,480 --> 00:22:50,520 Speaker 1: is the maximal potential value that they can get from AI. 403 00:22:51,520 --> 00:22:55,440 Speaker 5: Yeah, the economics questions and interesting one. It's an important one. 404 00:22:56,000 --> 00:22:59,520 Speaker 5: What we've done there is well, towards the end of 405 00:22:59,600 --> 00:23:03,199 Speaker 5: last year, when we knew growth was likely to accelerate rapidly, 406 00:23:03,320 --> 00:23:06,440 Speaker 5: we modeled what we thought the company would start to 407 00:23:06,640 --> 00:23:10,720 Speaker 5: look like in terms of different classifications of users with 408 00:23:10,880 --> 00:23:16,879 Speaker 5: different use cases, came up with certain budgets and then 409 00:23:16,920 --> 00:23:19,080 Speaker 5: what we've done this year is federated that. 410 00:23:19,240 --> 00:23:22,480 Speaker 4: Out to all of the business units within the firm. 411 00:23:22,840 --> 00:23:27,200 Speaker 5: I mean strong believer in pushing decision making down to 412 00:23:27,520 --> 00:23:30,399 Speaker 5: the also the lowest possible level that it makes sense 413 00:23:30,480 --> 00:23:33,560 Speaker 5: to do so it allows people to be agile and 414 00:23:34,000 --> 00:23:36,840 Speaker 5: use the economics at work for them and their departments. 415 00:23:36,880 --> 00:23:39,639 Speaker 5: And of course budgets are pungible if departments wants to 416 00:23:40,440 --> 00:23:43,000 Speaker 5: move money from some other spend and say right, I 417 00:23:43,080 --> 00:23:45,840 Speaker 5: think we should buy more tokens than they are free 418 00:23:45,960 --> 00:23:49,320 Speaker 5: to do that. And the platform that we've built, the 419 00:23:49,359 --> 00:23:54,359 Speaker 5: air platform, has a very rich, not one hundred percent complete, 420 00:23:54,359 --> 00:23:57,080 Speaker 5: but a very very rich set of models that can 421 00:23:57,160 --> 00:24:00,119 Speaker 5: be chosen, including all of the main frontier models and 422 00:24:00,200 --> 00:24:02,679 Speaker 5: a number of the different open source open white models. 423 00:24:03,600 --> 00:24:07,240 Speaker 1: Do you build a model that routes queries to the 424 00:24:07,359 --> 00:24:10,280 Speaker 1: optimal sort of like cost efficient model. I know there's 425 00:24:10,320 --> 00:24:13,360 Speaker 1: a lot of interest in this. The sort of classifiers 426 00:24:13,400 --> 00:24:16,000 Speaker 1: and the big AI labs have them themselves, but the 427 00:24:16,119 --> 00:24:20,480 Speaker 1: classifiers that route queries, is that something that you built 428 00:24:20,640 --> 00:24:22,320 Speaker 1: that you have in house or how do you solve 429 00:24:22,359 --> 00:24:22,800 Speaker 1: their problem. 430 00:24:23,800 --> 00:24:25,399 Speaker 4: We could do that we've chosen not to. 431 00:24:25,760 --> 00:24:28,639 Speaker 7: I think that the reason is that we want people 432 00:24:28,720 --> 00:24:32,520 Speaker 7: to understand the dynamics of how best to use AI 433 00:24:32,680 --> 00:24:36,520 Speaker 7: and which models to use. Okay, So what we've leaned 434 00:24:36,560 --> 00:24:40,040 Speaker 7: on instead is education. So we have a very rich 435 00:24:40,200 --> 00:24:42,920 Speaker 7: data set of how people are using AI, so you 436 00:24:43,040 --> 00:24:45,440 Speaker 7: have great insight, and some of the things that we 437 00:24:45,600 --> 00:24:48,480 Speaker 7: saw were just quite basic mistakes. 438 00:24:48,840 --> 00:24:53,320 Speaker 4: So often people would be doing multiple tasks in the 439 00:24:53,400 --> 00:24:54,760 Speaker 4: same context window. 440 00:24:54,880 --> 00:24:57,639 Speaker 7: They'd be trying to figure out the best way to 441 00:24:58,280 --> 00:25:00,760 Speaker 7: write an investment thesis, and then they were trying to 442 00:25:00,760 --> 00:25:02,439 Speaker 7: figure out where to go to lunch, and then they 443 00:25:02,480 --> 00:25:03,400 Speaker 7: were trying to ask. 444 00:25:03,400 --> 00:25:05,080 Speaker 2: Fable what the weather will be tomorrow. 445 00:25:05,320 --> 00:25:08,280 Speaker 4: Yeah, exactly, And that is. 446 00:25:09,080 --> 00:25:12,879 Speaker 7: I suppose, like a well known problem if you're in 447 00:25:13,040 --> 00:25:16,440 Speaker 7: the weeds of AI. But we have seventeen eighteen hundred 448 00:25:16,440 --> 00:25:21,000 Speaker 7: people at my group, and the technical understanding of how 449 00:25:21,119 --> 00:25:23,280 Speaker 7: things work at a fundamental. 450 00:25:22,840 --> 00:25:26,640 Speaker 4: Level is varied. So we've been really focusing on education. 451 00:25:26,920 --> 00:25:30,560 Speaker 7: We're very transparent about the budgets that people have and 452 00:25:30,640 --> 00:25:34,399 Speaker 7: how they're being spent, and we talk to them about 453 00:25:34,640 --> 00:25:38,119 Speaker 7: the different classes of models, and it's through that that 454 00:25:38,240 --> 00:25:43,399 Speaker 7: we have seen better results and we've actually seen people 455 00:25:43,880 --> 00:25:49,920 Speaker 7: finding quite creative ways to reduce token spend and contributing 456 00:25:50,040 --> 00:25:51,720 Speaker 7: that back to the platform. 457 00:25:51,320 --> 00:25:52,080 Speaker 4: As a result of it. 458 00:25:52,400 --> 00:25:53,320 Speaker 2: Wait, say more about that. 459 00:25:54,840 --> 00:25:56,359 Speaker 4: So, for example, if. 460 00:25:56,280 --> 00:26:01,879 Speaker 7: You are using a coding agent and you want to 461 00:26:01,920 --> 00:26:04,920 Speaker 7: do some basic get commands to interact with the version 462 00:26:04,960 --> 00:26:10,639 Speaker 7: control system, often the coding agent will send the command, 463 00:26:10,720 --> 00:26:13,480 Speaker 7: it will process the entire result that comes back from 464 00:26:13,520 --> 00:26:17,800 Speaker 7: that command as tokens. Instead, there's some really simple tools 465 00:26:17,880 --> 00:26:21,800 Speaker 7: and simple basic steps that you could use the instead 466 00:26:21,840 --> 00:26:25,320 Speaker 7: of calling an LLLM to do inference and tool calls, 467 00:26:25,800 --> 00:26:28,320 Speaker 7: it can just intercept that tool call and do it 468 00:26:28,440 --> 00:26:30,040 Speaker 7: outside of the agentic loop. 469 00:26:30,440 --> 00:26:32,800 Speaker 3: If I was looking at a chart of your overall 470 00:26:32,920 --> 00:26:36,359 Speaker 3: token consumption, what would I mean? I assume it's upward 471 00:26:36,680 --> 00:26:40,800 Speaker 3: sloping despite some of these efficiency efforts, but like how 472 00:26:40,920 --> 00:26:42,480 Speaker 3: steep is the slope at the moment? 473 00:26:43,680 --> 00:26:48,399 Speaker 7: So since January, I think token consumption has gone up 474 00:26:49,080 --> 00:26:50,159 Speaker 7: eighty six times. 475 00:26:50,440 --> 00:26:50,680 Speaker 2: Wow. 476 00:26:51,040 --> 00:26:56,119 Speaker 7: So it's really really quite quite incredible. We were not 477 00:26:56,600 --> 00:27:01,159 Speaker 7: expecting usage to be the way there has been, and 478 00:27:01,480 --> 00:27:04,080 Speaker 7: it's been across the board where it's not just been 479 00:27:04,160 --> 00:27:07,480 Speaker 7: in the tech and tech adjacent departments. We have seen 480 00:27:08,480 --> 00:27:12,119 Speaker 7: people in finance, people in operations, people in the people 481 00:27:12,240 --> 00:27:16,600 Speaker 7: team using agentic coding workflows, and as a technologist, that's 482 00:27:16,640 --> 00:27:20,159 Speaker 7: just super exciting to give this new capability, this new 483 00:27:20,280 --> 00:27:23,240 Speaker 7: power to people who haven't been able to use it before. 484 00:27:23,400 --> 00:27:26,520 Speaker 1: Well, this actually gets the question that I've been wondering about, 485 00:27:26,720 --> 00:27:30,159 Speaker 1: and it definitely feels like December and January we're just 486 00:27:30,240 --> 00:27:34,720 Speaker 1: like a very pivotal period. And from your perspective, that 487 00:27:35,240 --> 00:27:39,199 Speaker 1: sharp inflection point up, how much was it driven by 488 00:27:39,280 --> 00:27:42,439 Speaker 1: the capability of the models themselves, whether going from an 489 00:27:42,480 --> 00:27:45,160 Speaker 1: opus four point seven or four point five to four 490 00:27:45,160 --> 00:27:50,200 Speaker 1: point six and beyond, or this sort of discovery of 491 00:27:51,080 --> 00:27:54,040 Speaker 1: these very high quality harnesses like a cloud code or 492 00:27:54,240 --> 00:27:57,000 Speaker 1: cowork or whatever. I think that's what it's called. That 493 00:27:57,760 --> 00:28:01,399 Speaker 1: really allows someone to do things with AI that are 494 00:28:01,440 --> 00:28:06,359 Speaker 1: beyond asking questions and actually manipulate real work. The model 495 00:28:06,520 --> 00:28:09,520 Speaker 1: or the harness, which would you describe as the key 496 00:28:09,720 --> 00:28:11,600 Speaker 1: driver of that huge acceleration. 497 00:28:12,640 --> 00:28:13,720 Speaker 6: I think the two are couple. 498 00:28:14,280 --> 00:28:17,719 Speaker 7: I think one of the interesting benchmarks to look at 499 00:28:17,840 --> 00:28:21,000 Speaker 7: for this is a benchmark called meter and what that 500 00:28:21,160 --> 00:28:25,800 Speaker 7: tries to do is it looks at tasks that humans 501 00:28:25,840 --> 00:28:29,520 Speaker 7: would do for different time periods from a couple of 502 00:28:29,600 --> 00:28:33,080 Speaker 7: minutes to many hours. And what we're seeing is that 503 00:28:33,359 --> 00:28:36,399 Speaker 7: every seven months or so, the amount of time that 504 00:28:37,200 --> 00:28:40,600 Speaker 7: an ergentic workflow can go away and do a task 505 00:28:40,800 --> 00:28:43,560 Speaker 7: is doubling. So now you can ask an agent to 506 00:28:43,640 --> 00:28:46,600 Speaker 7: do a task that would take a human sixteen hours, 507 00:28:47,160 --> 00:28:51,200 Speaker 7: and that changes the way that you think about teams. 508 00:28:51,280 --> 00:28:53,800 Speaker 7: It changes the way that you think about interacting with 509 00:28:53,920 --> 00:28:54,640 Speaker 7: these agents. 510 00:28:54,800 --> 00:28:57,400 Speaker 4: You go from a place where you're. 511 00:28:58,800 --> 00:28:59,440 Speaker 6: In the loop. 512 00:28:59,560 --> 00:29:04,040 Speaker 7: You're sking an agent to solve a problem like writing 513 00:29:04,080 --> 00:29:07,880 Speaker 7: a unit test, to a problem like building an entire 514 00:29:07,960 --> 00:29:11,840 Speaker 7: feature of an application, or an entire application in itself. 515 00:29:12,600 --> 00:29:18,720 Speaker 7: So I think it's the scalability that has allowed larger 516 00:29:18,800 --> 00:29:23,240 Speaker 7: tasks to be completed. There's resulted in larger token usage. 517 00:29:39,000 --> 00:29:42,200 Speaker 3: I want to go back to I guess the oversight question. 518 00:29:42,520 --> 00:29:46,840 Speaker 3: And we all know that finance is highly regulated environment 519 00:29:47,720 --> 00:29:51,560 Speaker 3: and you've touched on this earlier, but you're still approving 520 00:29:51,840 --> 00:29:55,040 Speaker 3: a lot of these model outputs through humans, So there's 521 00:29:55,080 --> 00:29:58,600 Speaker 3: some oversight there. And I would assume that when you're approving, 522 00:29:59,320 --> 00:30:02,120 Speaker 3: if a human is improving a new model or an output, 523 00:30:02,280 --> 00:30:05,160 Speaker 3: that they have to understand what's actually coming out of it, right, 524 00:30:05,240 --> 00:30:08,240 Speaker 3: there has to be some explainability there. If I'm a 525 00:30:08,360 --> 00:30:11,880 Speaker 3: PM or I don't know, a quant sitting in front 526 00:30:12,000 --> 00:30:16,840 Speaker 3: of a risk management committee or a regulator, what does 527 00:30:16,960 --> 00:30:21,840 Speaker 3: explainability actually look like? And how am I translating the 528 00:30:21,960 --> 00:30:26,280 Speaker 3: model outputs into something that is understandable and also I 529 00:30:26,320 --> 00:30:27,240 Speaker 3: guess defensible. 530 00:30:28,320 --> 00:30:29,040 Speaker 6: Yeah, you're right. 531 00:30:29,200 --> 00:30:33,040 Speaker 5: Explainability is super importance to us. And to be clear, 532 00:30:33,080 --> 00:30:36,040 Speaker 5: the sort of business that we're in is not the 533 00:30:36,240 --> 00:30:40,480 Speaker 5: high frequency trading business where people are constructing huge neural 534 00:30:40,600 --> 00:30:44,640 Speaker 5: nets and looking through like multi dimensional spaces and the 535 00:30:44,720 --> 00:30:48,080 Speaker 5: output not being at all in serviceable. We're not in 536 00:30:48,440 --> 00:30:53,640 Speaker 5: that space trading and holding period horizons specifically days to weeks, 537 00:30:53,760 --> 00:31:00,600 Speaker 5: two months, so all of our trading decisions are ultimately explainable. 538 00:31:01,240 --> 00:31:04,040 Speaker 5: We never want to be in a position where we 539 00:31:04,080 --> 00:31:05,360 Speaker 5: don't know why that trade happened. 540 00:31:05,400 --> 00:31:06,640 Speaker 6: The AI did it. 541 00:31:07,320 --> 00:31:09,360 Speaker 5: And going back to some of the things that we 542 00:31:09,560 --> 00:31:13,160 Speaker 5: talked about earlier, take the example of the AI coming 543 00:31:13,440 --> 00:31:17,200 Speaker 5: up with brand new trading hypotheses based on what it's 544 00:31:17,280 --> 00:31:19,800 Speaker 5: seen in terms of content of a data set, what 545 00:31:19,920 --> 00:31:22,640 Speaker 5: it's seen in the content of an academic paper, and 546 00:31:22,800 --> 00:31:24,720 Speaker 5: the model will go as far the system, the agent 547 00:31:24,840 --> 00:31:30,640 Speaker 5: goes far as naming and writing the investment hypothesis. 548 00:31:31,120 --> 00:31:33,440 Speaker 4: It's one of the first things it does before it 549 00:31:33,640 --> 00:31:35,320 Speaker 4: moves on to writing codes. 550 00:31:35,320 --> 00:31:39,960 Speaker 5: It's giving us a real English description of what it thinks. 551 00:31:40,600 --> 00:31:45,480 Speaker 5: The rationale for the trading signal is I'm curious. 552 00:31:45,720 --> 00:31:48,560 Speaker 1: You know, obviously we haven't even gotten to the question 553 00:31:48,640 --> 00:31:52,160 Speaker 1: of like what is the future of labor and the 554 00:31:52,280 --> 00:31:54,240 Speaker 1: humans in the loop and how many humans in the 555 00:31:54,280 --> 00:31:56,840 Speaker 1: loop will we need in the future. But one thing 556 00:31:56,880 --> 00:31:59,840 Speaker 1: I'm curious about is a way to ask this question differently. 557 00:32:00,160 --> 00:32:03,520 Speaker 1: Has AI allowed you to look at markets that you 558 00:32:03,560 --> 00:32:07,760 Speaker 1: wouldn't have had the bandwidth before. So, for example, I'm at, 559 00:32:07,880 --> 00:32:11,160 Speaker 1: you know, let's take the Ethiopian stock exchange or something 560 00:32:11,280 --> 00:32:14,320 Speaker 1: like that. There's a certain amount of human labor that 561 00:32:14,400 --> 00:32:18,320 Speaker 1: would be required to gain any familiarity with it whatsoever, 562 00:32:18,640 --> 00:32:21,760 Speaker 1: setting aside everything else, And no matter how much potential 563 00:32:21,840 --> 00:32:25,280 Speaker 1: profit there is in say frontier market, there is a 564 00:32:25,360 --> 00:32:28,280 Speaker 1: minimum amount that's going to cost and that might take 565 00:32:28,360 --> 00:32:31,920 Speaker 1: some investment opportunities off the board because the potential profit 566 00:32:32,400 --> 00:32:35,000 Speaker 1: isn't big enough to justify the spend to getting up 567 00:32:35,040 --> 00:32:38,600 Speaker 1: to speed. This strikes me as something that AI could 568 00:32:38,640 --> 00:32:42,160 Speaker 1: potentially help with or solve for of creating a new 569 00:32:42,280 --> 00:32:46,960 Speaker 1: opportunity set by reducing some of the upfront human labor costs. 570 00:32:47,040 --> 00:32:49,560 Speaker 1: Is that something that you think about or have seen 571 00:32:49,680 --> 00:32:54,040 Speaker 1: specifically so far in terms of Man group, I. 572 00:32:54,080 --> 00:32:58,760 Speaker 5: Think it's likely happening incrementally at the margins if we 573 00:32:58,880 --> 00:33:02,800 Speaker 5: start to sum up all of the different micro augmentations 574 00:33:03,000 --> 00:33:04,400 Speaker 5: that we see across the firm. 575 00:33:04,560 --> 00:33:08,560 Speaker 6: For example, and here is a related one, someone had. 576 00:33:08,480 --> 00:33:15,000 Speaker 5: Built a relatively simple aisystem to take data from complex instruments, PDFs, 577 00:33:15,280 --> 00:33:21,240 Speaker 5: specifications and automatically populate our reference data store with that. 578 00:33:21,720 --> 00:33:25,360 Speaker 5: So yes, I think it's it's absolutely happening, but it's 579 00:33:25,360 --> 00:33:27,720 Speaker 5: just some of a lot of different parts across the firm. 580 00:33:28,400 --> 00:33:31,200 Speaker 7: I think what was seeing as well in the systematic 581 00:33:31,320 --> 00:33:34,960 Speaker 7: space is that you need a few prerequisites to build 582 00:33:35,000 --> 00:33:39,360 Speaker 7: a systematic strategy. You need to have some connectivity to 583 00:33:39,440 --> 00:33:42,440 Speaker 7: trade the instrument, and you need to be able to 584 00:33:42,600 --> 00:33:44,960 Speaker 7: understand what the price of the market is. Those are 585 00:33:45,040 --> 00:33:49,720 Speaker 7: two real fundamental things, and for developed. 586 00:33:49,400 --> 00:33:51,480 Speaker 4: Markets that's really easy you go and look at. 587 00:33:51,400 --> 00:33:54,760 Speaker 7: The order book, but for less developed markets, things like 588 00:33:55,320 --> 00:34:02,200 Speaker 7: cryptosecuritized credit, they're often harder to can too. There may 589 00:34:02,280 --> 00:34:08,800 Speaker 7: be traded more verbally, or the contracts are complicated and 590 00:34:08,880 --> 00:34:12,719 Speaker 7: it's difficult to understand the price where there's these unstructured 591 00:34:12,840 --> 00:34:17,399 Speaker 7: data nuances in the derivation of that price. We've seen 592 00:34:17,560 --> 00:34:22,680 Speaker 7: that AI allows us to think about accessing that market 593 00:34:22,840 --> 00:34:25,720 Speaker 7: in a systematic way earlier than we could before. 594 00:34:26,520 --> 00:34:29,239 Speaker 3: Just going back to the labor market side of things, 595 00:34:29,320 --> 00:34:31,560 Speaker 3: I guess you know if I'm a PM at Man Group, 596 00:34:31,600 --> 00:34:34,359 Speaker 3: and more and more of my job is using AI 597 00:34:34,560 --> 00:34:40,120 Speaker 3: for research or even basically supervising agents that I maybe 598 00:34:40,239 --> 00:34:44,120 Speaker 3: help develop. What does that mean for what you're looking 599 00:34:44,239 --> 00:34:48,359 Speaker 3: for in terms of talent? Are you looking for engineers 600 00:34:48,560 --> 00:34:51,680 Speaker 3: who can tweak these models? Are you looking for more 601 00:34:51,800 --> 00:34:56,279 Speaker 3: traditional investors who I guess have stronger intuition about how 602 00:34:56,360 --> 00:34:59,880 Speaker 3: these things might play out, or some sort of combination 603 00:35:00,360 --> 00:35:01,279 Speaker 3: of characteristics. 604 00:35:01,360 --> 00:35:02,279 Speaker 2: What do you look for now? 605 00:35:03,200 --> 00:35:06,359 Speaker 5: I think very fair to saying this is something I've 606 00:35:06,400 --> 00:35:10,200 Speaker 5: been making a strong caseful that everyone we hire now 607 00:35:10,640 --> 00:35:15,880 Speaker 5: into the firm should up the bar with respect to AI, 608 00:35:16,160 --> 00:35:20,239 Speaker 5: regardless of the role that they're coming in to do. 609 00:35:20,719 --> 00:35:23,440 Speaker 5: And I think that applies as much in the operation 610 00:35:23,560 --> 00:35:26,080 Speaker 5: space as it does in the front office space. 611 00:35:27,080 --> 00:35:27,680 Speaker 1: What does that mean? 612 00:35:27,760 --> 00:35:27,800 Speaker 6: So? 613 00:35:27,880 --> 00:35:31,840 Speaker 1: What like someone like Okay, I'm capable of upping the 614 00:35:31,960 --> 00:35:35,319 Speaker 1: bar with respect to air like, what does that say 615 00:35:35,400 --> 00:35:38,560 Speaker 1: more about Okay? In the recruiting process, what does that 616 00:35:38,680 --> 00:35:39,279 Speaker 1: person look like? 617 00:35:40,680 --> 00:35:46,759 Speaker 5: That means being as familiar with the technology as you 618 00:35:46,880 --> 00:35:51,160 Speaker 5: could reasonably expect a person to be, given the wealth 619 00:35:51,200 --> 00:35:53,959 Speaker 5: of information that's out there. When I'm when I'm hiring people, 620 00:35:54,120 --> 00:35:55,879 Speaker 5: you know what sort of people don't want. I want 621 00:35:55,960 --> 00:35:58,880 Speaker 5: bright people, and I want people who are motivated, They 622 00:35:58,960 --> 00:36:02,759 Speaker 5: get things done and passionate about the subject matter their 623 00:36:02,800 --> 00:36:06,120 Speaker 5: field of expertise. And I think it's very hard to 624 00:36:07,080 --> 00:36:10,879 Speaker 5: fulfill all of those criteria chicking nowadays and say, well, 625 00:36:11,360 --> 00:36:13,279 Speaker 5: you know, AI, I don't know much about it. I 626 00:36:13,360 --> 00:36:15,799 Speaker 5: don't really use it as part of my job. 627 00:36:16,000 --> 00:36:16,160 Speaker 6: Yeah. 628 00:36:16,200 --> 00:36:18,680 Speaker 7: I think the other piece that you want is somebody 629 00:36:18,680 --> 00:36:21,200 Speaker 7: who's a bit more of a long term thinker, somebody 630 00:36:21,239 --> 00:36:24,960 Speaker 7: who really wants to automate a process from end to end. 631 00:36:25,239 --> 00:36:28,520 Speaker 7: They are happy not to be in the weeds in 632 00:36:28,600 --> 00:36:33,320 Speaker 7: the loop, And when you think about technologists, that's actually 633 00:36:33,400 --> 00:36:35,600 Speaker 7: quite difficult. People love being in the weeds, they love 634 00:36:35,680 --> 00:36:40,759 Speaker 7: debugging issues, getting into the nitty gritty, but actually, fundamentally, 635 00:36:40,960 --> 00:36:42,920 Speaker 7: you want to level up. 636 00:36:43,040 --> 00:36:47,640 Speaker 4: You want to be a kind of a conductor of 637 00:36:47,719 --> 00:36:48,360 Speaker 4: these agents. 638 00:36:48,480 --> 00:36:50,160 Speaker 7: You want to be in charge of the end to 639 00:36:50,360 --> 00:36:55,120 Speaker 7: end process rather than necessarily being in the weeds, almost 640 00:36:55,200 --> 00:36:58,480 Speaker 7: like a manager who has a lot of technical expertise, 641 00:36:59,040 --> 00:37:04,840 Speaker 7: so who is thinking about things in broader strategic terms 642 00:37:05,160 --> 00:37:07,080 Speaker 7: is much more value than they used to be. 643 00:37:07,680 --> 00:37:09,719 Speaker 3: So this leads to the other thing I wanted to ask, 644 00:37:09,800 --> 00:37:14,080 Speaker 3: which is you could see the arrival of AI tools 645 00:37:14,360 --> 00:37:17,719 Speaker 3: generating like two different outcomes here, where you have some 646 00:37:17,920 --> 00:37:20,800 Speaker 3: people who are just really really good at using AI 647 00:37:21,440 --> 00:37:24,759 Speaker 3: and they become superstars and orchestrators of a bunch of 648 00:37:24,840 --> 00:37:28,200 Speaker 3: different agents, as you put it, or you could have 649 00:37:28,320 --> 00:37:32,640 Speaker 3: this sort of democratizing effect where maybe you're a junior 650 00:37:32,680 --> 00:37:35,040 Speaker 3: employee with not as much experience, but now you can 651 00:37:35,120 --> 00:37:37,520 Speaker 3: automate a bunch of tasks, you can learn from AI, 652 00:37:37,719 --> 00:37:39,600 Speaker 3: you can use it for research and things like that. 653 00:37:40,480 --> 00:37:42,239 Speaker 3: What are you seeing more of at the moment, the 654 00:37:42,360 --> 00:37:48,200 Speaker 3: superstar dynamic or the democratization of skill sets throughout man group. 655 00:37:48,560 --> 00:37:52,760 Speaker 5: I think we're saying both genuinely both, I think given 656 00:37:52,840 --> 00:37:55,680 Speaker 5: the size of the firm, though with saying more of 657 00:37:55,840 --> 00:37:58,040 Speaker 5: the latter, I mean, as I said at the start 658 00:37:58,120 --> 00:38:00,920 Speaker 5: of the chat almost or everybody using it to day 659 00:38:00,960 --> 00:38:07,120 Speaker 5: to day, but examples of huge, genuine right at the 660 00:38:07,719 --> 00:38:11,040 Speaker 5: cutting edge of thinking. I think they're naturally more rare, 661 00:38:11,120 --> 00:38:13,520 Speaker 5: but there's a fair view of them, even saying that. 662 00:38:14,480 --> 00:38:16,960 Speaker 4: They often cut across multiple teams as well. 663 00:38:17,000 --> 00:38:20,560 Speaker 7: And that is hard because you need to move from 664 00:38:20,600 --> 00:38:24,960 Speaker 7: a space where you spend most of your time executing 665 00:38:25,560 --> 00:38:28,400 Speaker 7: to a space where you spend most of your time planning. 666 00:38:29,000 --> 00:38:31,680 Speaker 7: The time to execute is just going down and down. 667 00:38:32,600 --> 00:38:34,799 Speaker 7: It's becoming cheaper and cheaper to do that. You can 668 00:38:34,920 --> 00:38:37,800 Speaker 7: build code, you can build features very quickly, so the 669 00:38:37,880 --> 00:38:40,839 Speaker 7: focus really needs to be on what should we build, 670 00:38:40,920 --> 00:38:45,600 Speaker 7: how does it connect together, and what is the process 671 00:38:45,760 --> 00:38:50,480 Speaker 7: that we want to develop across multiple teams. And it's 672 00:38:50,600 --> 00:38:53,080 Speaker 7: very difficult sometimes to take people out of the seat 673 00:38:53,120 --> 00:38:56,680 Speaker 7: and get them to collaborate and plan a workflow rather 674 00:38:56,800 --> 00:38:59,480 Speaker 7: than just going and trying to build a proof of 675 00:38:59,560 --> 00:39:02,040 Speaker 7: concept and execute on the idea. 676 00:39:03,320 --> 00:39:05,920 Speaker 1: Let's talk more about the eighty six x increase in 677 00:39:06,040 --> 00:39:09,560 Speaker 1: token spend. If we were having this conversation back in 678 00:39:10,040 --> 00:39:14,520 Speaker 1: February or January, a lot of the chat would have 679 00:39:14,640 --> 00:39:17,640 Speaker 1: been very much about like, what does this mean for 680 00:39:17,800 --> 00:39:20,920 Speaker 1: legacy software? Companies because that's when the big software company 681 00:39:21,280 --> 00:39:24,840 Speaker 1: software is all off was you know, really was quite intense. 682 00:39:25,239 --> 00:39:29,200 Speaker 1: But I feel like this conversation in July is becoming like, no, 683 00:39:29,360 --> 00:39:32,320 Speaker 1: these companies are not going after the world of software, 684 00:39:32,600 --> 00:39:34,919 Speaker 1: They're going after the world of labor. And you see 685 00:39:34,920 --> 00:39:37,800 Speaker 1: a lot of these conversations like people talk about the 686 00:39:38,000 --> 00:39:43,040 Speaker 1: ratio of token spend to employee salaries, et cetera, and 687 00:39:43,160 --> 00:39:45,760 Speaker 1: that that is the tam that it's like all human labor, 688 00:39:45,800 --> 00:39:47,759 Speaker 1: and maybe we're not going to get there for a while. 689 00:39:47,880 --> 00:39:48,520 Speaker 4: I kind of hope not. 690 00:39:49,000 --> 00:39:53,040 Speaker 1: But is token spend part of your tech budget or 691 00:39:53,239 --> 00:39:55,800 Speaker 1: is it like something that is a true line item 692 00:39:55,880 --> 00:39:58,399 Speaker 1: that's distinct, that's more on par with labor. And when 693 00:39:58,440 --> 00:40:00,600 Speaker 1: you think about Man Group in twenty two, twenty seven, 694 00:40:01,080 --> 00:40:04,400 Speaker 1: twenty twenty eight, do you talk about expected ratios of 695 00:40:04,640 --> 00:40:05,880 Speaker 1: salaries to token spend. 696 00:40:08,080 --> 00:40:12,359 Speaker 5: We genuinely haven't started having that conversation yet. I mean, 697 00:40:12,400 --> 00:40:15,640 Speaker 5: I would get at some point it will come, And 698 00:40:15,800 --> 00:40:18,480 Speaker 5: the company is set up in such a way that 699 00:40:18,800 --> 00:40:21,880 Speaker 5: we want to direct resources to where they produce the 700 00:40:22,000 --> 00:40:24,640 Speaker 5: best economic outcome for us. 701 00:40:25,200 --> 00:40:27,239 Speaker 6: So I'm sure that time will come. 702 00:40:28,400 --> 00:40:33,560 Speaker 7: One of the interesting things in the token budgeting process. 703 00:40:33,560 --> 00:40:36,320 Speaker 7: I suppose is that what we're increasingly seeing is that 704 00:40:37,600 --> 00:40:40,840 Speaker 7: the spend is actually not by people, is by agents, 705 00:40:41,000 --> 00:40:45,960 Speaker 7: and the agents relate to workflows, and who owns the workflows? 706 00:40:46,080 --> 00:40:48,320 Speaker 4: Is it this department? Is it that department? 707 00:40:48,800 --> 00:40:50,800 Speaker 7: And that's a new problem for us, one that we 708 00:40:50,920 --> 00:40:54,080 Speaker 7: haven't solved yet, but it's a great problem to have. 709 00:40:55,360 --> 00:40:59,520 Speaker 3: I brought up earlier, those sort of machine learning parallel 710 00:40:59,640 --> 00:41:01,560 Speaker 3: I guess, and I know you guys don't do a 711 00:41:01,600 --> 00:41:04,640 Speaker 3: lot of high frequency trading, but we've certainly seen a 712 00:41:04,719 --> 00:41:07,960 Speaker 3: dynamic in HFT where everyone's competing and it's sort of 713 00:41:08,160 --> 00:41:10,200 Speaker 3: a race to the bottom where I. 714 00:41:10,200 --> 00:41:12,240 Speaker 2: Can't even remember where we where we're at. 715 00:41:12,160 --> 00:41:13,640 Speaker 1: In terms of like micrometers. 716 00:41:13,800 --> 00:41:17,279 Speaker 3: Yeah, in terms of second increments or time increments, but 717 00:41:17,719 --> 00:41:20,879 Speaker 3: it's the sort of race to the bottom dynamic. Would 718 00:41:20,960 --> 00:41:25,840 Speaker 3: you expect AI driven alpha to get sort of competed 719 00:41:26,120 --> 00:41:30,640 Speaker 3: or arbitraged away relatively quickly as everyone seems to be 720 00:41:30,760 --> 00:41:34,839 Speaker 3: hopping on the same bandwagon, or are there certain advantages 721 00:41:35,560 --> 00:41:38,680 Speaker 3: you know, we talked about data for instance, scale perhaps 722 00:41:38,840 --> 00:41:41,120 Speaker 3: that you would expect to stick around for some time. 723 00:41:42,239 --> 00:41:45,120 Speaker 7: Yeah, I'd say it goes back to your question around 724 00:41:45,560 --> 00:41:48,239 Speaker 7: where is the alpha? And what is true is that 725 00:41:48,480 --> 00:41:52,879 Speaker 7: it is easier for people to onboard data sets, analyze them, 726 00:41:53,040 --> 00:41:55,759 Speaker 7: and build features, But that doesn't mean that they can 727 00:41:55,880 --> 00:42:00,279 Speaker 7: trade them. We've been doing this for decades. So what 728 00:42:00,440 --> 00:42:03,440 Speaker 7: we have are capabilities to actually access these markets. We 729 00:42:03,560 --> 00:42:07,920 Speaker 7: have the relationships with the brokers, We have access to 730 00:42:08,040 --> 00:42:12,000 Speaker 7: data that isn't just available off the shelf. So it's 731 00:42:12,040 --> 00:42:15,040 Speaker 7: putting all of those things together, putting those expertise together, 732 00:42:15,640 --> 00:42:18,160 Speaker 7: the access to markets, all of the data that we have, 733 00:42:18,520 --> 00:42:22,160 Speaker 7: the rich market data, and then giving AI access to 734 00:42:22,520 --> 00:42:27,840 Speaker 7: the capabilities such as back testing, compute. It's this whole network, 735 00:42:27,960 --> 00:42:31,839 Speaker 7: this whole ecosystem that together I think drives alpha. There 736 00:42:31,920 --> 00:42:34,960 Speaker 7: isn't one code, repository and man group that I can 737 00:42:35,040 --> 00:42:37,040 Speaker 7: point out and say that's where the alpha is. It's 738 00:42:37,239 --> 00:42:42,440 Speaker 7: really this network of different systems that interact with each other. 739 00:42:42,920 --> 00:42:47,919 Speaker 7: So I think the definitely some features of data sets 740 00:42:47,960 --> 00:42:52,279 Speaker 7: will become table stakes. They go from being alpha to 741 00:42:52,520 --> 00:42:56,840 Speaker 7: being a risk factor. But because everybody has them and 742 00:42:57,120 --> 00:43:02,640 Speaker 7: it moves the market, but it isn't the only way 743 00:43:02,719 --> 00:43:06,360 Speaker 7: that we make money. We are able to connect different 744 00:43:06,840 --> 00:43:09,400 Speaker 7: data sets, different ways of doings together and then actually 745 00:43:09,480 --> 00:43:11,040 Speaker 7: trade on those signals. 746 00:43:11,520 --> 00:43:13,359 Speaker 1: Garry, I want to go back to something you said 747 00:43:13,400 --> 00:43:15,879 Speaker 1: earlier when we were talking about bottlenecks, and it's like, sure, 748 00:43:15,960 --> 00:43:19,719 Speaker 1: everybody wants more data, everyone wants more compute, No one 749 00:43:19,719 --> 00:43:22,919 Speaker 1: would complain about these things. But where the rubber meets 750 00:43:22,960 --> 00:43:26,960 Speaker 1: the road is like, does the institution have the capacity 751 00:43:27,080 --> 00:43:30,400 Speaker 1: to actually maybe from a cultural standpoint, a sort of 752 00:43:30,719 --> 00:43:35,200 Speaker 1: hierarchy standpoint, to actually get the most out of these tools. 753 00:43:35,520 --> 00:43:37,920 Speaker 1: And this is clearly a very hot area. And so 754 00:43:38,080 --> 00:43:42,120 Speaker 1: for example, just last week, Microsoft announced a new thing 755 00:43:42,160 --> 00:43:44,840 Speaker 1: which they're calling the Frontier Company, which is basically a 756 00:43:44,960 --> 00:43:47,759 Speaker 1: new like sort of subdivision that seems to want to 757 00:43:47,800 --> 00:43:51,920 Speaker 1: specifically solve this problem, go into an organization and figure 758 00:43:51,960 --> 00:43:55,600 Speaker 1: out this optimal structure. And you know, arguably this is 759 00:43:55,640 --> 00:43:58,279 Speaker 1: even like what a company like Pallenteer is trying to do, 760 00:43:58,560 --> 00:44:01,959 Speaker 1: which is, you know, there deployed engineers. We know about 761 00:44:02,320 --> 00:44:07,840 Speaker 1: Claude sending or Anthropics sending engineers inside Goldman Sachs to 762 00:44:07,920 --> 00:44:10,800 Speaker 1: really leverage I hate using that word because it's so cliche, 763 00:44:10,920 --> 00:44:14,359 Speaker 1: but yes, leverage these tools. What specifically are you seeing 764 00:44:14,440 --> 00:44:17,840 Speaker 1: happening on that front. Do you have third party companies 765 00:44:18,280 --> 00:44:21,120 Speaker 1: who are coming to you and saying, look, we can 766 00:44:21,960 --> 00:44:25,400 Speaker 1: work with you to find what is the org structure 767 00:44:25,440 --> 00:44:28,720 Speaker 1: of the future for man group such that it's getting 768 00:44:28,760 --> 00:44:30,919 Speaker 1: the most out of these tools for a long time. 769 00:44:32,719 --> 00:44:36,320 Speaker 5: Yeah, I did smile at the multi billion dollar for 770 00:44:36,640 --> 00:44:41,360 Speaker 5: deployed engineer and division that you just mentioned, partly because 771 00:44:41,880 --> 00:44:44,200 Speaker 5: I mean that's the way we've been like set up 772 00:44:44,520 --> 00:44:46,040 Speaker 5: internally here for about. 773 00:44:45,880 --> 00:44:47,719 Speaker 6: I think fifteen years. Same. 774 00:44:47,960 --> 00:44:50,680 Speaker 5: We're very big on We're very big on platforms, and 775 00:44:50,760 --> 00:44:53,719 Speaker 5: we've got a bunch of teams and Tashara runs one 776 00:44:53,760 --> 00:44:57,880 Speaker 5: of those that build out these big cross cutting elements 777 00:44:57,880 --> 00:45:01,200 Speaker 5: of platform technology, in his case the Data iplatform, but 778 00:45:01,360 --> 00:45:04,920 Speaker 5: a big part of the tech team are already and 779 00:45:05,000 --> 00:45:07,920 Speaker 5: having for a decade and a half forward deployed engineers 780 00:45:08,360 --> 00:45:12,080 Speaker 5: sitting with our quants, sitting with our discretionary managers. And 781 00:45:12,560 --> 00:45:15,480 Speaker 5: one of my jokes often the people I'd be interfering 782 00:45:15,560 --> 00:45:17,400 Speaker 5: to come join the team is all right, let's go 783 00:45:17,560 --> 00:45:19,919 Speaker 5: walk across the fifth floor here in the bank House, 784 00:45:20,200 --> 00:45:23,200 Speaker 5: and I want you to tell me who the engineers 785 00:45:23,200 --> 00:45:25,759 Speaker 5: are and who the quants are. And I bet you're 786 00:45:25,760 --> 00:45:27,960 Speaker 5: going to get it wrong, because what you'll see is 787 00:45:28,239 --> 00:45:31,440 Speaker 5: very similar stuff on the screen, and that holds just 788 00:45:31,520 --> 00:45:33,640 Speaker 5: as true today as it did decade ago. 789 00:45:34,239 --> 00:45:37,759 Speaker 1: One last question, but another thing I'm curious about in 790 00:45:37,880 --> 00:45:42,560 Speaker 1: the investment context. So we know that AI works when 791 00:45:42,640 --> 00:45:44,920 Speaker 1: there's a big pool of data, and then it can 792 00:45:45,000 --> 00:45:48,000 Speaker 1: pull in the unstructured data and the structured data, et cetera, 793 00:45:48,200 --> 00:45:53,240 Speaker 1: and communicate across them. In an entity such as man Group, 794 00:45:53,960 --> 00:45:58,200 Speaker 1: are there any alignment issues in which you know, if 795 00:45:58,320 --> 00:46:03,759 Speaker 1: I have a subject better expertise that generates alpha, this 796 00:46:04,040 --> 00:46:07,000 Speaker 1: might be why I have a seat at an organization, 797 00:46:07,280 --> 00:46:09,520 Speaker 1: or I might be a rain maker. You know you 798 00:46:09,640 --> 00:46:12,719 Speaker 1: hear this at all kinds of different firms where compensation 799 00:46:13,040 --> 00:46:17,240 Speaker 1: is linked to someone's like deep expertise in some area, 800 00:46:17,960 --> 00:46:21,040 Speaker 1: do you think about alignment so that the firm, the 801 00:46:21,160 --> 00:46:26,640 Speaker 1: franchise Man Group is actually capturing some of the expertise 802 00:46:26,800 --> 00:46:30,759 Speaker 1: and data of the superstar. How do you get them 803 00:46:30,880 --> 00:46:34,359 Speaker 1: to sort of, I guess, contribute as much information as 804 00:46:34,440 --> 00:46:38,319 Speaker 1: possible to this thing that requires a lot of data 805 00:46:38,360 --> 00:46:38,960 Speaker 1: and information. 806 00:46:40,000 --> 00:46:42,920 Speaker 4: I think in some areas that's still a little bit 807 00:46:42,960 --> 00:46:44,280 Speaker 4: of a work in progress. 808 00:46:44,360 --> 00:46:49,279 Speaker 5: That's saying in some areas because in others, notably the systematic, 809 00:46:49,320 --> 00:46:53,720 Speaker 5: the quant areas of the business. This highly collaborative approach 810 00:46:53,840 --> 00:46:56,640 Speaker 5: and shad code basis. That's just been the way that 811 00:46:56,760 --> 00:46:59,799 Speaker 5: those areas have worked for a long time. 812 00:47:00,520 --> 00:47:00,640 Speaker 4: Now. 813 00:47:01,000 --> 00:47:04,160 Speaker 5: I'm not saying that you walk across the discretionary part 814 00:47:04,239 --> 00:47:07,520 Speaker 5: of the flaw and all of the fundamental investors are 815 00:47:07,600 --> 00:47:10,000 Speaker 5: going to be quite so free and open. I've talked 816 00:47:10,040 --> 00:47:12,279 Speaker 5: to a bunch of them about their processes, and you know, 817 00:47:12,320 --> 00:47:15,520 Speaker 5: they're open with ninety percent, but there's the ten percent. Well, 818 00:47:15,560 --> 00:47:17,919 Speaker 5: you know, this is where my personal value add lives. 819 00:47:18,000 --> 00:47:21,760 Speaker 5: I'm not so comfortable talking about about that. But even 820 00:47:21,960 --> 00:47:26,840 Speaker 5: that said, there's a very decent and genuine amounts of 821 00:47:27,200 --> 00:47:30,359 Speaker 5: collaboration there as well. And it's only when you get 822 00:47:30,400 --> 00:47:34,319 Speaker 5: perhaps to the very sensitive areas that people might be. 823 00:47:34,320 --> 00:47:36,480 Speaker 6: A little bit more reluctant to talk. 824 00:47:37,120 --> 00:47:40,120 Speaker 7: Yeah, I think the high level there's a huge amount 825 00:47:40,200 --> 00:47:42,800 Speaker 7: of shared workflows. If you think about the way that 826 00:47:42,880 --> 00:47:45,440 Speaker 7: we baptest, the way that you read an investment report, 827 00:47:45,520 --> 00:47:49,560 Speaker 7: these are all workflows that are people's expertise, but they're 828 00:47:50,360 --> 00:47:56,200 Speaker 7: not necessarily the ten percent that produces the returns. So 829 00:47:57,000 --> 00:48:01,719 Speaker 7: those areas are encapsulated in AI playbooks, AI skills that 830 00:48:01,840 --> 00:48:05,680 Speaker 7: we put in our knowledge platform and they can be 831 00:48:05,840 --> 00:48:10,240 Speaker 7: used across the floor, whereas some of the particulars around 832 00:48:10,680 --> 00:48:14,719 Speaker 7: how the strategy works and the investment pieces somewhat held 833 00:48:14,800 --> 00:48:16,200 Speaker 7: back in certain cases. 834 00:48:16,400 --> 00:48:18,640 Speaker 1: It makes sense, all right, Gary and Toshara, thank you 835 00:48:18,800 --> 00:48:22,120 Speaker 1: so much for coming on. Odd lad, really appreciate your 836 00:48:22,160 --> 00:48:23,960 Speaker 1: taking your time and talk about where you're at. 837 00:48:24,320 --> 00:48:24,600 Speaker 6: Thank you. 838 00:48:24,719 --> 00:48:25,960 Speaker 4: I'll pleasure pensilup. 839 00:48:38,560 --> 00:48:38,759 Speaker 6: You know what. 840 00:48:38,840 --> 00:48:41,279 Speaker 1: I think it was really interesting about their conversation and 841 00:48:41,400 --> 00:48:45,000 Speaker 1: part is like there is so much to figure out still, right. 842 00:48:45,200 --> 00:48:47,480 Speaker 3: I mean just the basic token budget and like where 843 00:48:47,520 --> 00:48:49,120 Speaker 3: stuff that's allocated. 844 00:48:48,800 --> 00:48:52,200 Speaker 1: Like eighty six x in since January is pretty crazy. 845 00:48:52,200 --> 00:48:53,600 Speaker 3: Well, this is the other thing I was thinking, like 846 00:48:54,080 --> 00:48:57,320 Speaker 3: eighty six times growth in token usage, Like at some 847 00:48:57,600 --> 00:49:00,880 Speaker 3: point that needs to show up in another on crete number, 848 00:49:01,000 --> 00:49:06,080 Speaker 3: whether it's expense reduction or revenue generation income generation. 849 00:49:06,520 --> 00:49:09,160 Speaker 2: And I don't know how much like leeway there. 850 00:49:09,160 --> 00:49:09,520 Speaker 6: Is for that. 851 00:49:09,800 --> 00:49:12,080 Speaker 1: No, and you figure, like right now, so you have 852 00:49:12,200 --> 00:49:15,399 Speaker 1: this eighty six x explosion, but as they say, there's 853 00:49:15,440 --> 00:49:18,759 Speaker 1: still in the moment where they haven't gotten to We 854 00:49:18,960 --> 00:49:22,080 Speaker 1: wanted to build like an internal router to minimize this. 855 00:49:22,200 --> 00:49:25,120 Speaker 1: So even with the eighty six they're still in the 856 00:49:25,600 --> 00:49:28,920 Speaker 1: phase where it's okay, you know, figure out what model 857 00:49:28,960 --> 00:49:32,000 Speaker 1: you want to use an experiment, et cetera, which to 858 00:49:32,120 --> 00:49:34,799 Speaker 1: my mind is actually kind of bullish if you think 859 00:49:34,800 --> 00:49:37,479 Speaker 1: about it. For token spend that you can grow eighty 860 00:49:37,560 --> 00:49:39,719 Speaker 1: six x and it still doesn't get you to the 861 00:49:39,760 --> 00:49:42,120 Speaker 1: point where like, oh, we got to really like clamp 862 00:49:42,239 --> 00:49:45,279 Speaker 1: down on this, Like maybe you know, who knows like 863 00:49:45,480 --> 00:49:47,400 Speaker 1: what the point is for a lot of firms that 864 00:49:47,480 --> 00:49:50,920 Speaker 1: are just starting out where they actually have to start 865 00:49:50,960 --> 00:49:52,480 Speaker 1: imposing some token austerity. 866 00:49:52,719 --> 00:49:55,200 Speaker 2: I guess we'll find out at some point. 867 00:49:55,360 --> 00:49:57,520 Speaker 3: But the other thing that stood out to me, you know, 868 00:49:57,600 --> 00:49:59,759 Speaker 3: you asked the question about how do you get how 869 00:49:59,800 --> 00:50:02,000 Speaker 3: do you at workers to give up their own sauce 870 00:50:02,120 --> 00:50:05,279 Speaker 3: that basically is responsible for them having a job in 871 00:50:05,320 --> 00:50:05,880 Speaker 3: the first place. 872 00:50:07,280 --> 00:50:10,320 Speaker 2: Key stroke surveillance solves all of this, right. 873 00:50:10,400 --> 00:50:14,120 Speaker 1: Yeah, you know those articles I have to say, you 874 00:50:14,800 --> 00:50:16,080 Speaker 1: need them to offer it up. 875 00:50:16,200 --> 00:50:16,960 Speaker 6: You just I. 876 00:50:18,480 --> 00:50:20,560 Speaker 1: Have to say. There was some story that came out 877 00:50:20,600 --> 00:50:23,320 Speaker 1: a while back about OMETA it was going to start 878 00:50:23,480 --> 00:50:26,719 Speaker 1: using like training its models on its own, and I 879 00:50:26,840 --> 00:50:28,800 Speaker 1: was like, they weren't doing this already, Like I was 880 00:50:28,840 --> 00:50:32,680 Speaker 1: actually like really surprised that this wasn't already. 881 00:50:32,560 --> 00:50:36,120 Speaker 3: Especially in finance and investment, which is a highly regulated 882 00:50:36,239 --> 00:50:40,360 Speaker 3: industry already and I'm sure is monitoring pretty much everything anyway. 883 00:50:40,600 --> 00:50:42,960 Speaker 1: Yeah, I kind of assumed that all of these companies 884 00:50:43,000 --> 00:50:45,960 Speaker 1: are really building models, were already using all of the 885 00:50:46,280 --> 00:50:50,040 Speaker 1: data that their own employees were generating. But yeah, you 886 00:50:50,160 --> 00:50:52,520 Speaker 1: have to like, oh, does the person does the rain 887 00:50:52,600 --> 00:50:55,360 Speaker 1: maker suddenly just start writing everything down on pen and paper. 888 00:50:56,680 --> 00:51:01,080 Speaker 1: They're not implicitly uploading all of their knowledge to the AI. 889 00:51:01,520 --> 00:51:03,720 Speaker 1: I think that's a pretty interesting question in itself. 890 00:51:03,760 --> 00:51:05,920 Speaker 3: All right, clearly lots of interesting questions. 891 00:51:06,000 --> 00:51:07,799 Speaker 1: But shall we leave it there for Let's leave it there. 892 00:51:07,840 --> 00:51:08,040 Speaker 6: Okay. 893 00:51:08,239 --> 00:51:10,799 Speaker 3: This has been another episode of the Audthoughts podcast. I'm 894 00:51:10,840 --> 00:51:13,840 Speaker 3: Tracy Alloway. You can follow me at Tracy Alloway. 895 00:51:13,520 --> 00:51:16,360 Speaker 1: And I'm Joe Wisenthal. You can follow me at The Stalwart. 896 00:51:16,640 --> 00:51:19,880 Speaker 1: Follow our producers Carmen Rodriguez at Carmen Arman, dash Ol 897 00:51:19,920 --> 00:51:23,960 Speaker 1: Bennett at Dashbot, Calebrooks at Kelebrooks, and Kevin Lozano at 898 00:51:24,040 --> 00:51:26,880 Speaker 1: Kevin Lloyd Lozano. And for more Odd Loge content, go 899 00:51:27,000 --> 00:51:29,160 Speaker 1: to Bloomberg dot com, slash od lots or of a 900 00:51:29,239 --> 00:51:31,839 Speaker 1: daily newsletter and all of our episodes and you could 901 00:51:31,840 --> 00:51:33,839 Speaker 1: shout about all of these topics twenty four to seven 902 00:51:34,000 --> 00:51:37,480 Speaker 1: in our discord discord dot gg slash hodlogs. 903 00:51:37,880 --> 00:51:39,719 Speaker 3: And if you enjoy od loots, if you like it 904 00:51:39,760 --> 00:51:42,839 Speaker 3: when we talk about how finance firms are actually implementing AI, 905 00:51:43,000 --> 00:51:45,400 Speaker 3: then please leave us a positive review on your favorite 906 00:51:45,400 --> 00:51:48,719 Speaker 3: podcast platform. And remember, if you are a Bloomberg subscriber, 907 00:51:48,760 --> 00:51:51,640 Speaker 3: you can listen to all of our episodes absolutely ad free. 908 00:51:51,920 --> 00:51:53,920 Speaker 3: All you need to do is find the Bloomberg channel 909 00:51:53,960 --> 00:51:56,279 Speaker 3: on Apple Podcasts and follow the instructions there. 910 00:51:56,760 --> 00:51:57,560 Speaker 2: Thanks for listening 911 00:52:14,280 --> 00:52:15,600 Speaker 6: It in