WEBVTT 1 00:00:10.128 --> 00:00:18.501 Hello and welcome to the Macro Brief from HSBC Global Investment Research, the podcast that looks at the big topics driving financial markets across the world. 2 00:00:19.222 --> 00:00:20.824 I'm your host, Piers Butler in London. 3 00:00:21.265 --> 00:00:25.511 And on this week's podcast, we're catching up on the latest developments in artificial intelligence. 4 00:00:26.172 --> 00:00:30.434 A sell-off of chip stocks saw billions of dollars wiped off stock markets this week. 5 00:00:30.914 --> 00:00:41.660 And with the so-called Magnificent Seven Stocks in the thick of their reporting season, we're asking how investors and companies are feeling about AI given competing narratives and market jitters. 6 00:00:42.220 --> 00:00:47.061 To discuss this and more, I'm joined in the studio by Mark McDonald, head of AI and data science. 7 00:00:47.141 --> 00:00:48.781 Mark, great to have you back on The Macrobrief. 8 00:00:49.201 --> 00:00:49.761 Thanks, Piers. 9 00:00:49.821 --> 00:00:50.682 It's great to be back. 10 00:00:51.062 --> 00:00:52.642 So I was looking it up. 11 00:00:52.702 --> 00:00:54.922 We haven't spoken since December of last year. 12 00:00:54.942 --> 00:00:58.603 So in the world of AI, quite a lot has gone on. 13 00:00:59.443 --> 00:01:08.705 And in fact, you've just co-authored a report with Alice Sapinda, who's our head of global and EM equity strategy on AI narrative shifts. 14 00:01:09.265 --> 00:01:13.966 So I thought it'd be great to talk about these narrative shifts and you identify five of them. 15 00:01:14.786 --> 00:01:23.228 The first one is AI euphoria, which I guess encapsulates this idea that AI changes everything and we are at the foothills of a singularity. 16 00:01:23.408 --> 00:01:25.069 Are we at the foothills of a singularity? 17 00:01:25.489 --> 00:01:30.710 I think this is one of the things that at the moment, it definitely doesn't feel like markets are trading like that. 18 00:01:31.530 --> 00:01:38.932 I think when people envisage what they mean by a singularity, sometimes people let their imaginations run away with themselves. 19 00:01:39.512 --> 00:01:45.278 The market sometimes fully buys into that narrative and at the moment seems quite sceptical of it. 20 00:01:46.359 --> 00:01:52.484 So the piece which we're referring to here kind of tracks the prevalence of these different narratives through time. 21 00:01:53.105 --> 00:01:56.308 That one right now is not particularly dominant as you might expect. 22 00:01:56.788 --> 00:01:58.428 The next one is AI disruption. 23 00:01:58.648 --> 00:02:02.049 There's been a lot of discussions about the impact of AI on jobs. 24 00:02:02.489 --> 00:02:06.310 It feels like that narrative has evolved quite a lot, even since we last spoke. 25 00:02:06.650 --> 00:02:07.491 Yes, it has. 26 00:02:07.851 --> 00:02:14.972 I think, you know, with any new technology, there's a tendency for people to focus on what might be taken away from them. 27 00:02:15.432 --> 00:02:26.255 And so it's very easy to see, as AI progressively does more and more tasks well, that people begin to worry that it will begin to take over more and more jobs. 28 00:02:27.035 --> 00:02:36.282 Often this tends to underestimate the degree to which people are flexible and to which the global economy is actually a complex system that will react. 29 00:02:37.323 --> 00:02:51.713 And historical evidence suggests that new technologies generally, they're good for productivity, they're good for growth, and they generally generate more new jobs associated with the technology than get replaced by the technology. 30 00:02:51.733 --> 00:02:51.833 Yeah. 31 00:02:52.253 --> 00:03:03.417 So we generally think that this fear of large-scale AI-driven unemployment is an overblown fear. 32 00:03:04.217 --> 00:03:12.420 And often I think what will happen is people's roles will change, but also who's going to get the productivity benefit. 33 00:03:12.500 --> 00:03:17.362 Often it ends up being the consumer that gets the productivity benefit through a higher quality product. 34 00:03:18.002 --> 00:03:21.265 And AI kind of just raises the floor for what's acceptable. 35 00:03:21.746 --> 00:03:25.809 And so it might not end up replacing as many jobs as people expect. 36 00:03:26.069 --> 00:03:33.036 But there is this paradox, isn't there, that there was this fear when Excel spreadsheets were introduced that accountants would be put out of jobs. 37 00:03:33.436 --> 00:03:35.218 In fact, it allowed them to do a lot more. 38 00:03:35.438 --> 00:03:36.179 Well, exactly. 39 00:03:36.199 --> 00:03:36.699 Yes. 40 00:03:36.919 --> 00:03:43.225 I'm sure when Excel and similar products were released that people were saying it was terrible news for accountants and financial analysts. 41 00:03:43.685 --> 00:03:53.008 But actually, there's been an explosion in jobs related to that type of calculation, because when you make something cheaper and easier to produce, you tend to get a lot more of it. 42 00:03:53.489 --> 00:03:55.509 And this is the classic Jevons paradox. 43 00:03:56.110 --> 00:04:02.872 I think also as well, like if you look at the life of, say, people in this department as research analysts, you've got sort of three aspects to their life. 44 00:04:03.452 --> 00:04:05.333 most important, which is like seeing clients. 45 00:04:05.933 --> 00:04:09.935 You've got the writing the research report and you've got the coming up with research ideas. 46 00:04:10.235 --> 00:04:14.277 Analysts are already using AI tools to help them do that research process. 47 00:04:14.638 --> 00:04:19.520 But what you're not seeing is people doing the same amount of research as before, just more quickly. 48 00:04:19.980 --> 00:04:23.422 People are doing much more thorough research because it's a competitive market. 49 00:04:23.862 --> 00:04:28.884 The critical resource of profitable investment ideas hasn't increased because of AI. 50 00:04:29.364 --> 00:04:33.066 And you've got everybody competing for that same resource with better technology. 51 00:04:33.426 --> 00:04:37.188 So people are still pouring the same amount of hours and effort into doing so. 52 00:04:37.208 --> 00:04:44.311 It's just that it allows them to research more thoroughly unless the consumer of the research gets the benefit because they get higher quality research. 53 00:04:44.811 --> 00:04:51.098 The third narrative is hyperscaler overspend, which is clearly front and center at the moment. 54 00:04:51.499 --> 00:04:56.224 So this is looking at the fact that CapEx investment numbers have grown exponentially. 55 00:04:57.425 --> 00:05:01.250 And the core question is, will all this computing actually be required? 56 00:05:01.730 --> 00:05:13.199 Yeah, I think sometimes it feels like these companies can't win because when they weren't spending all this money on capex, people were complaining that they were wasting all this money and had all this cash and they weren't doing anything useful with it. 57 00:05:13.520 --> 00:05:19.424 And as soon as they do start investing for the future, then the people worry that they're over investing and overspending on capex. 58 00:05:19.485 --> 00:05:23.588 And we go through periods like now where the market punishes them for doing so. 59 00:05:24.208 --> 00:05:28.472 Whether it turns out to be sensible investment or not, I guess time will tell. 60 00:05:29.172 --> 00:05:34.593 At the moment, it seems like there is vast potential demand for AI products and services. 61 00:05:35.714 --> 00:05:46.996 And as these tools keep getting better and better, I think companies are only just beginning to find out ways to practically embed them into their operations in ways that really move the needle. 62 00:05:47.596 --> 00:05:55.478 And as soon as companies start seeing the benefit from that, then there's going to be significant demand from the business community for spending on AI. 63 00:05:56.069 --> 00:05:58.251 So really very hard to tell at this stage. 64 00:05:58.671 --> 00:06:05.336 I guess the numbers are so huge and some of the companies are now burning through their free cash flow and actually borrowing. 65 00:06:05.436 --> 00:06:08.879 So the market is kind of saying, is this all really going to be justified? 66 00:06:09.640 --> 00:06:12.962 That is one of the narratives that's trending at the moment. 67 00:06:13.222 --> 00:06:22.690 I think despite the AI trade having been obviously fantastically successful for several years now, often the market narrative is looking for a reason to be negative on AI. 68 00:06:23.510 --> 00:06:29.413 And that direction, the directionality of that commentary is always the same. 69 00:06:29.914 --> 00:06:33.155 People are looking for reasons to be negative, but the rationale for it changes. 70 00:06:33.536 --> 00:06:39.299 And so you see these wild shifts in the driving narrative as to why the AI theme is going to fail. 71 00:06:39.839 --> 00:06:44.061 It's just that it flips rapidly from rationale to rationale. 72 00:06:44.842 --> 00:06:50.565 And often these are quite short-lived little tempests and then the market narrative changes. 73 00:06:51.025 --> 00:06:57.227 One negative, so to speak, is the fourth narrative, which is China competition concerns. 74 00:06:57.247 --> 00:06:59.147 That deep seek moment was not a one off. 75 00:06:59.908 --> 00:07:07.470 And I guess it's kind of encapsulated in saying these U.S. companies are spending billions of dollars and actually China can do it much more cheaply. 76 00:07:08.030 --> 00:07:15.055 Yeah, it feels that the US and the Chinese AI labs are competing on entirely different axes. 77 00:07:15.475 --> 00:07:26.622 So for the US AI labs, when they release a new model or a new tool, they need to be able to claim that it is the best, at least on some metrics. 78 00:07:27.403 --> 00:07:31.065 The Chinese AI companies have been competing more on efficiency and, 79 00:07:31.785 --> 00:07:38.811 So they're not trying to come out and say that this new model is the best at this whole range of tasks, but they do want it to be good enough and dramatically cheaper. 80 00:07:39.692 --> 00:07:42.774 And it's going to be interesting to see how this plays out. 81 00:07:43.595 --> 00:07:49.780 Often, particularly in a commodity business, being good enough but dramatically cheaper is a winning strategy. 82 00:07:51.101 --> 00:07:58.627 I think it really comes down to how much demand there is for increasing excellence in the model performance. 83 00:07:58.828 --> 00:07:59.348 Yeah. 84 00:07:59.508 --> 00:08:06.573 I can imagine a bimodal world where you get a small number of winners of the enterprise AI race. 85 00:08:06.793 --> 00:08:16.420 So whoever can win that race to become the default choice of most enterprises when they're looking for an agentic AI software tool. 86 00:08:17.060 --> 00:08:20.483 So whoever can win that race is going to be very sticky business. 87 00:08:20.923 --> 00:08:23.345 And that should be very easy to monetize and thus... 88 00:08:24.145 --> 00:08:30.189 That really does warrant spending lots and lots of money on CapEx today in order to try and achieve that prize. 89 00:08:31.170 --> 00:08:42.156 For the AI use cases where companies or people are accessing the AI tools via API calls, that is very commoditized because you've got lines of code that just periodically call an API. 90 00:08:42.517 --> 00:08:48.000 Switching that from an anthropic model to an open AI model to a Google model or to a model that's 91 00:08:48.480 --> 00:08:51.725 from one of these Chinese companies that's been much more focused on efficiency. 92 00:08:52.626 --> 00:08:53.347 That's very easy. 93 00:08:53.667 --> 00:09:02.859 And so I suspect that that part of the AI monetization race, that will be a commodity grind to quite low profitability. 94 00:09:03.299 --> 00:09:06.263 And so you're going to get this bimodal distribution, I think. 95 00:09:06.807 --> 00:09:07.528 Super interesting. 96 00:09:08.009 --> 00:09:14.917 The final narrative, which we're kind of living through at the moment, is AI positioning capitulation. 97 00:09:15.598 --> 00:09:19.203 Is that an unwinding of a momentum trade in market terms? 98 00:09:19.855 --> 00:09:21.936 I suspect it's mostly that, yes. 99 00:09:23.156 --> 00:09:43.542 Rather than something fundamental having changed, it really does feel like, as I mentioned before, there's this long-running trend for narrative and commentary to be looking for negativity with the AI theme in the face of continued profitability and strong performance in the trade itself. 100 00:09:44.562 --> 00:09:54.384 This feels like one of these short term periods of negativity, mostly like a positioning unwind rather than a fundamental change in what's happening with AI. 101 00:09:54.705 --> 00:10:02.867 So without getting too technical for our listeners, the report ascribes probabilities to these scenarios and what you call return signatures. 102 00:10:03.467 --> 00:10:04.527 How do you go about doing that? 103 00:10:04.807 --> 00:10:13.112 So it's actually very similar to our other regime models like Eccles and Cockles and Tickles that look for regimes. 104 00:10:13.132 --> 00:10:14.954 How you come up with those names, I don't know. 105 00:10:15.074 --> 00:10:19.296 Yeah, I mean, obviously coming up with the names is the most important part, so you've got to have good branding. 106 00:10:19.416 --> 00:10:24.740 But these look for regimes in equity, commodity, and rates markets, respectively. 107 00:10:25.741 --> 00:10:32.265 And what we do here is first you feed in data to a clustering model 108 00:10:33.545 --> 00:10:40.088 And this data is generally based on returns and volatility and correlations from assets in the asset class. 109 00:10:40.629 --> 00:10:44.791 And the first thing is the clustering model tries to group periods of times together. 110 00:10:45.411 --> 00:10:47.812 And since these periods of time are much more similar to each other. 111 00:10:47.852 --> 00:10:49.113 So looking for patterns, I guess. 112 00:10:49.133 --> 00:10:49.673 Exactly. 113 00:10:49.773 --> 00:10:51.094 So we're not telling it anything at all. 114 00:10:51.114 --> 00:10:55.396 We're just giving it the data and saying, you tell us what the natural groups are in this data. 115 00:10:56.716 --> 00:11:09.185 And then from those groups that the clustering model produces, we then train a classification model that given data today can allocate it to one of the same clusters. 116 00:11:10.406 --> 00:11:20.613 And how you name the clusters and identify what's really going on is partly by looking at the returns of assets in each of these clusters, so these return signatures, and 117 00:11:21.180 --> 00:11:24.182 Partly looking at the drivers from this classification model. 118 00:11:24.542 --> 00:11:33.967 So the classification model, we can use modern explainability techniques like SHAP, which are kind of a cornerstone of many of our data science pieces of research. 119 00:11:34.748 --> 00:11:38.270 And this allows you to understand why a model is doing what it's doing. 120 00:11:38.670 --> 00:11:48.658 And so you can say for this cluster, the sorts of characteristics that make the model confident that it's in this regime are X, Y and Z. 121 00:11:48.838 --> 00:11:57.425 And from that, you can then get a good understanding of what really is the natural hallmark of that particular cluster. 122 00:11:57.465 --> 00:11:58.586 And that allows you to make these regimes. 123 00:11:58.606 --> 00:12:00.948 So a cluster would be, for example, in terms of what we've been discussing earlier, 124 00:12:01.568 --> 00:12:02.729 Hyperscalar overspend. 125 00:12:02.889 --> 00:12:03.329 Exactly. 126 00:12:03.529 --> 00:12:10.173 What's striking in looking at the report is how frequently these regimes change. 127 00:12:10.253 --> 00:12:13.215 Is that likely to continue or will it settle over time? 128 00:12:13.996 --> 00:12:16.577 For this, I think it's likely to continue. 129 00:12:19.019 --> 00:12:23.802 The AI theme has been such a successful trade ever since ChatGPT came out. 130 00:12:24.622 --> 00:12:26.923 that a lot of people are worried about it. 131 00:12:27.023 --> 00:12:27.584 It's new. 132 00:12:28.144 --> 00:12:31.886 The space, the drivers in the space are changing all the time. 133 00:12:32.926 --> 00:12:44.112 It feels like this volatility in narrative and explanation is just an inherent part of this space. 134 00:12:44.312 --> 00:12:49.335 When it comes to our other regime models, like equity regimes or commodity regimes, those are much stickier. 135 00:12:49.935 --> 00:12:57.640 Here, it really does feel like the market narrative as to what's driving things in AI can change at the drop of a hat. 136 00:12:57.860 --> 00:12:59.181 Because it's still so new, I guess. 137 00:12:59.741 --> 00:13:01.182 It's so new and it's so scary. 138 00:13:01.362 --> 00:13:10.028 And so much of the valuation of many of the companies involved in it are based in long-term expectations. 139 00:13:10.548 --> 00:13:16.552 And with long-term expectations in a rapidly changing area, you don't need to change your expectations that much in order to have a significant move. 140 00:13:17.057 --> 00:13:24.680 Closer to home, I wanted to ask you the question which I asked you last time is how is it impacting your work and that of other analysts? 141 00:13:25.200 --> 00:13:36.845 I saw that you had published recently a report looking at whether AI can update equity models, which to me seemed very far fetched for the clues in the name, large language models. 142 00:13:37.005 --> 00:13:37.645 How does that work? 143 00:13:38.847 --> 00:13:42.788 Large language models for some time now have been very good at coding. 144 00:13:43.028 --> 00:13:48.329 They find coding easy and they find spreadsheets hard, which is the opposite of most humans. 145 00:13:48.389 --> 00:13:51.590 Most humans find spreadsheets dead easy and coding quite difficult. 146 00:13:52.350 --> 00:14:06.494 And so, I mean, obviously this experiment that we were doing, we were looking specifically at how well or otherwise different AI tools can update the big complicated spreadsheets that single stock analysts use to model their companies. 147 00:14:07.014 --> 00:14:14.396 But I think the broader, bigger picture question is, you know, to what degree can AI just safely and reliably do spreadsheet related tasks? 148 00:14:15.176 --> 00:14:17.477 Surprisingly mixed results, to be honest. 149 00:14:18.377 --> 00:14:36.521 Like even although there was a clear ranking amongst the different AI tools that we looked at with the clawed suite of AI tools performing best, there were still many surprisingly basic errors that it would make in this process of updating these models with new earnings release data. 150 00:14:37.101 --> 00:14:48.688 And the weird thing about this is that I think any individual cell that you are asking the tools to update in the spreadsheet, all of these tools will be able to do that and have been able to do that for some time now. 151 00:14:49.368 --> 00:14:51.630 The difficulty comes from, I think, two things. 152 00:14:51.910 --> 00:14:55.012 One, you're asking it to update all of these numbers at the same time. 153 00:14:55.532 --> 00:14:56.653 So there's a lot going on. 154 00:14:57.613 --> 00:15:05.118 And the other thing that's difficult is you're asking it to understand the structure of a complicated spreadsheet whilst doing this. 155 00:15:05.558 --> 00:15:12.763 So the sorts of mistakes that the tools would make is that they had a tendency to go in and update spreadsheets in a way that would break the structural integrity. 156 00:15:13.383 --> 00:15:18.149 So it would go over and you've got like a column of cells where it's all the same formula that's been dragged down. 157 00:15:18.169 --> 00:15:22.274 And it would just like randomly overwrite some of them with hard coded values and not others. 158 00:15:22.514 --> 00:15:27.079 So kind of very dangerous errors that are hard for a human to spot. 159 00:15:27.800 --> 00:15:33.307 This feels like the sort of thing that the AI companies are going to have no problem in really fixing. 160 00:15:33.847 --> 00:15:40.155 I suspect that part of what's going on is that, you know, all the people who are building these tools, AI companies, they're geeks like me. 161 00:15:40.215 --> 00:15:43.578 And so they're kind of just assuming, like, why is everybody using spreadsheets? 162 00:15:43.599 --> 00:15:46.262 Why can't they just get with the program and write it all in code? 163 00:15:47.323 --> 00:15:49.105 And so I think it's been a bit of an oversight. 164 00:15:49.165 --> 00:15:51.648 And now these companies are really starting to realize the. 165 00:15:52.048 --> 00:15:58.950 like vast amounts of, you know, business activity that is controlled by big complicated spreadsheets in enterprises all across the globe. 166 00:15:59.750 --> 00:16:03.811 And I think now that they're focused on it, we're going to see rapidly improving performance there. 167 00:16:04.131 --> 00:16:05.131 More changes to come. 168 00:16:05.551 --> 00:16:10.752 And indeed, there's another change that is about to occur, which is that you're relocating to Hong Kong. 169 00:16:10.892 --> 00:16:11.712 I am indeed. 170 00:16:11.752 --> 00:16:14.313 How excited are you about this change? 171 00:16:14.673 --> 00:16:18.254 And what's your thoughts about how Asia is positioned in the AI revolution? 172 00:16:18.654 --> 00:16:19.575 I'm super excited. 173 00:16:19.835 --> 00:16:28.942 The interesting thing about Asia is I think over here in the West, we're so much more focused on what's happening with the AI labs in the US. 174 00:16:29.543 --> 00:16:37.309 But as we discussed earlier with one of these themes itself, is there is a very thriving AI scene in China as well. 175 00:16:38.289 --> 00:16:53.593 And I think the, you know, when I was visiting Hong Kong for the GIS earlier this year, a lot of the conversations I had with Asian investors, they were much more focused on, you know, this competition between what's happening with US AI labs and Chinese AI labs. 176 00:16:54.634 --> 00:17:00.095 So I'm very much looking forward to getting more experience of an exposure to that world. 177 00:17:00.607 --> 00:17:04.989 Well, I think that's all we've got time for today, but the best of luck with your relocation. 178 00:17:05.069 --> 00:17:05.509 Thank you. 179 00:17:05.669 --> 00:17:08.411 I hope all your boxes arrive in the right place. 180 00:17:09.111 --> 00:17:10.352 Fingers crossed. 181 00:17:11.032 --> 00:17:12.133 But thank you for joining us today. 182 00:17:12.293 --> 00:17:12.833 Thank you very much. 183 00:17:19.118 --> 00:17:22.221 Mark McDonald there on the latest developments in the AI space. 184 00:17:22.762 --> 00:17:30.749 And if you'd like to hear more on China's emergence as a force in chip making, then check out this week's edition of our sister podcast, Under the Banyan Tree. 185 00:17:31.670 --> 00:17:33.532 But that's all from us here on The Macrobrief. 186 00:17:33.712 --> 00:17:35.434 This episode was hosted by me, P.S. 187 00:17:35.454 --> 00:17:37.396 Butler, and produced by Tom Barton. 188 00:17:37.676 --> 00:17:40.299 Don't forget to like and follow wherever you get your podcasts. 189 00:17:41.199 --> 00:17:43.662 Thanks for listening, and we'll be back again next week.