AI Agents Will Upend Investing Even More Than Coding - Frank Long, Coatue

AI Engineer · 14 min · 166 sentences · from YouTube's caption track

Each timecode opens YouTube at the start of that sentence. Line anchors (#s42) are the cue ids in the WebVTT, and every line carries its start and end seconds. All transcripts has every talk, and the whole corpus as one file.

1. 00:01[music]
2. 00:12My name's Frank.
3. 00:13I work at Coatue, which is a technology investment firm across public and private markets.
4. 00:17So, in our business, what you really need to do is be able to take high conviction bets on big themes and watch those play out.
5. 00:26I think honestly a career is the same way.
6. 00:29And last year when I joined Coatue, the bet that I wanted to make was I think that agents will have an even bigger impact in investing than they did at coding.
7. 00:38I know that's a big statement.
8. 00:39We're all here because of the transformat- transformative impact that AI has had in coding, but I really believe it strongly and today I want to talk a little bit about
9. 00:48the framework that we use to actually realize that impact.
10. 00:52So, at a high level, you can think about a hedge fund as a function that takes in two parameters, context external and context internal, and spits out one tuple,
11. 01:04selection, size, timing.
12. 01:06What stock?
13. 01:08How much?
14. 01:09When?
15. 01:10Right?
16. 01:10And I think this is a good framework to actually to ground our work, which is we're trying to solve this compression problem of how do you take all the information in the world and spit out this one simple tuple
17. 01:24and the product that we that we provide is essentially a time series of this tuple that constructs a portfolio.
18. 01:31So, let's start from the top, which is the inputs.
19. 01:34Very simple philosophy.
20. 01:36We use our screens, we look at stuff, we do work.
21. 01:39Every single thing that a human can see in pixels on the screen, an agent in your building should be able to see via API.
22. 01:46Uh the analogy I use is imagine you brought the smartest people ever, the greatest investors alive, and you brought them into your company, but you gave them a computer without logins to 75% of your services.
23. 01:59It it's almost how smart the system is if it doesn't have the tools to actually action that.
24. 02:05So, the two parts, external context, right?
25. 02:08No surprises here.
26. 02:09Think the news, prices, filings, all of that should be easily accessible and not via dragging and dropping, copy and pasting, like directly via API.
27. 02:23Internal context is just as important, right?
28. 02:27So, this is your own firm situation.
29. 02:29So, that might be our current positions, potential liquidity and risk requirements we have around certain names, and past case studies or um you know, history with a certain sector
30. 02:41or stock or management team.
31. 02:43That stuff's all super important.
32. 02:44So, this is the foundation and I think uh no credit to me, KKR has been an amazing leader on this for over a decade.
33. 02:52It's both technological and cultural.
34. 02:54We have an amazing data science team, amazing, amazing processes, amazing, amazing systems that bring all this in.
35. 03:00And just as importantly, we have a culture that is super, super rigorous and built from the ground up to capture all this insight.
36. 03:08For decades, we've filed every single note about every single trade, about every single decision we made or didn't make, about what went well and what didn't go well, and that's poised us to expose us to AI.
37. 03:20So, if you think about it, I probably can't just take all of those things, jam them into a context window, and get my magic three numbers and go home, right?
38. 03:31It can't be that easy.
39. 03:32You probably have to break down this function a bit further, right?
40. 03:37So, maybe the first step is I take in that external and internal context, and I gather it, right?
41. 03:44I want to gather it, I want to filter out what's relevant and not relevant.
42. 03:48Okay, well, I pulled some news articles about the NBA.
43. 03:51That probably isn't material to stocks, but oh, well, this is actually related to some local permitting that might actually impact data centers, like there's a lot of nuance in this gather step.
44. 04:02Um then you'll see probably next step is I need to analyze that information and form a view.
45. 04:08Right?
46. 04:09I actually need to think about what that means for the markets, public and private.
47. 04:15Uh and then I need to take that view and turn it into a decision.
48. 04:20A decision object has these three parameters, right?
49. 04:24These three sort of sub sub pieces.
50. 04:26So, you can see every single step has an input and an output and the input for the subsequent step is the output that the previous function generated.
51. 04:37I think as an engineer, this is a way for me to conceptualize the problem in a way that makes a lot of sense.
52. 04:44But even this it
53. 04:46[snorts]
54. 04:46really can't be as simple as writing three functions, right?
55. 04:49So, let's just to up really appreciate the scope and understand the full complexity of the problem, let's look into the first step, a tiny sub piece of tiny sub piece of just gathering information
56. 05:03and let's do a case study on interpreting the news, right?
57. 05:08So it's probably no secret to you guys that you know, news from Asia and Korean Korean media outlets has been super important this year because semis, AI have been a huge huge theme in the market,
58. 05:22huge theme.
59. 05:22And I think this is an area where AI is poised to have a lot of impact is reading uh Korean language news better than our completely non-Korean speaking staff,
60. 05:34right?
61. 05:35Uh I think I have a good thesis that this might be a good use case.
62. 05:38So you can imagine the first step is I have uh articles.
63. 05:44I need to collect the articles, right?
64. 05:45So, this could be an MCP, this could be a CLI, license the data feeds, clean the data feeds.
65. 05:52There's actually a lot of ways to do this.
66. 05:54And a lot of dimensions of freedom for how you can even just write this simple piece.
67. 05:59How do I chunk it?
68. 06:00When I pass it to the LLM, do I want to pull the images as images?
69. 06:04Do I want to pull the images and then turn the images into text?
70. 06:09There's actually a lot of knobs here, and I almost think about those knobs as like hyper parameters in a traditional machine learning system, right?
71. 06:17I think about each of these knobs as something for me to tune and optimize against.
72. 06:23Obviously, filtration, we talked about it.
73. 06:26Korean They're probably Korean NBA fans, right?
74. 06:29There's probably a bunch of stuff that's not necessarily material to the the names and themes and stocks in our portfolio that really, really matter, right?
75. 06:40So, I want to make sure that I filter those out.
76. 06:45There's this like translate/extract/should I summarize?
77. 06:49Should I extract this into some sort of structured form?
78. 06:53Again, so, so many ways to do this, so many ways to experiment how to expose information to the model.
79. 06:58And one of the reasons this is hard is it keeps changing.
80. 07:02And obviously, checking.
81. 07:03Checking like our own internal notes, our own internal views, potential conversations we've had, meeting recordings against these things.
82. 07:11We actually need to cross-reference those things.
83. 07:14Actually, the idea of pulling information from disparate sources and cross-referencing them, that's super powerful.
84. 07:21So, you'll see on on the right here I have code, LLM, LLM, code, person.
85. 07:26There's many, many ways to run these functions.
86. 07:29I think about Uh oh, yeah.
87. 07:31And finally, update.
88. 07:32Potentially, we update our views.
89. 07:36So, there's many ways to run these functions.
90. 07:39I think of each line as having a runner, and uh the most obvious one is code.
91. 07:45Code is very good.
92. 07:46It's fast, it's cheap, it's reliable.
93. 07:50So, for something like if the price-to-earnings ratio is less than 15, I should just use deterministic code for that.
94. 07:59Runner number two is statistics, classical machine learning, and you can imagine fitting a line to a regression for specific pattern.
95. 08:08This is very powerful because you can model much more complex relationships than just pure deterministic code, right?
96. 08:15And for even more complex relationships, the reason we're all here, LLMs, AI agents, that's another type of runner, right?
97. 08:23So, for something like summarization, that's the state of the art, right?
98. 08:27It's probably totally impossible to build a a bunch of if statements that can summarize better than a model modern LLM.
99. 08:35And of course, uh people.
100. 08:38People are important runners in the system, right?
101. 08:40So, uh imagine a function meetings the management team of a company.
102. 08:45I probably don't want to send, I mean, I like the new Unitree robots, but I probably don't want to see it send an LLM or an AI agent in to do that.
103. 08:51I probably don't want to send an if statement into that meeting.
104. 08:54I think there's many cases where, because of judgment and the ability to navigate the real world and all of these things, a person is the best runner for a given piece of the function.
105. 09:04And I think this is like a super super important point.
106. 09:08There is no glory in automating things for the sake of automating things.
107. 09:12The goal is not to take as many things and make them LLM as possible for us.
108. 09:17The goal is, for any given piece of this function, what can I do to maximize absolute quality?
109. 09:26And if you actually think about the function, it probably looks something like this.
110. 09:30It's probably not just a linear statement of a bunch of different things.
111. 09:34There's probably all kinds of branching logic and all kinds of ways to actually fork this.
112. 09:39So, maybe I want to collect the information and then translate it, and my main goal is to extract quotes and check those against sources to make sure that these are actually the things that people said, and and that's my approach.
113. 09:50But that might be totally wrong.
114. 09:51Maybe actually what I want to do is I want to dedupe, and then I want to filter, and then I want to pull external figures, or maybe I want to do something totally different.
115. 09:59I actually want to I'm passing in a note first, and I actually just want to find other confirmatory data points, or maybe I'm trying to write team.
116. 10:06Uh there's a million different ways to configure this.
117. 10:09So, yet again, more hyper parameters within the control flow, right?
118. 10:15So, so many ways to think about just this very simple problem, a polling, filtering, summarizing, aggregating the news for LLMs.
119. 10:25And each of these boxes probably has an optimal runner, right?
120. 10:32So, uh the optimal runner might be deterministic code for collection, right?
121. 10:38I'm polling last 24 hours for a given source.
122. 10:42That's deterministic code.
123. 10:44Filtration, decision models, maybe.
124. 10:48LLMs, statistical methods, a combination, right?
125. 10:52So, this probably understates the complexity.
126. 10:55The reality is most things are multi-runner.
127. 10:57Most things have deterministic code plus LLM plus statistical method or human with LLM in conjunction.
128. 11:03Uh the reality is this it's probably combinatorics between my four buckets, and there's even more buckets than that because within the LLM category, I can tune the model.
129. 11:15I can choose big model, small model, medium model.
130. 11:18I can tune thinking effort.
131. 11:20X high, high, medium.
132. 11:22I can tune harness.
133. 11:23I can choose the harness.
134. 11:24I can choose which skills load with the harness.
135. 11:26I can choose which hooks load with the skills.
136. 11:28So, yet again, more hyper parameters within my search space.
137. 11:32I have an absolutely enormous amount of hyper parameters within my search space.
138. 11:35And this is just for this one tiny problem, right?
139. 11:38I'm just filtering the news.
140. 11:39So, the way uh to think about breaking down these problems is what can I actually compartmentalize in a way that we can understand and evaluate because that's what I'm hill climbing against, right?
141. 11:51Otherwise, I'm just tuning these things.
142. 11:53I'm okay, well, that should actually be a machine learning system instead of an LLM.
143. 11:57That should be a little model that we train.
144. 11:59Oh, no, that should actually shouldn't be a model at all.
145. 12:01That should be an if statement or Oh, man, LLMs are really, really bad at translating that to portfolio impact.
146. 12:07It should really be a human on that step, but they're using uh a deterministic system to help them search and pull the information.
147. 12:14It could be all different things, right?
148. 12:16And if you don't have a North Star to hill climb against, you're just kind of you're just fidgeting around.
149. 12:22And I think I cannot convey enough that this is just one tiny thing we do.
150. 12:27Imagine every single data source that we pull in, there's probably more than just three transformations to get them to the intermediate output.
151. 12:34So, the actual graph of Kensho and funds like Kensho, it actually probably looks Oh, not like this.
152. 12:42This is my screensaver.
153. 12:46The actual graph probably looks something like this, right?
154. 12:49It's probably massively complex.
155. 12:51In each of those nodes, we just talked about at least 10 hyper parameters with at least 20 different valid values for them.
156. 12:59So, the search space is is basically infinite.
157. 13:02And figuring out how to actually lay this out in a way that you can comprehend both from technical perspective, but also absorb culturally, organizationally, that is the great challenge
158. 13:14of applying AI in investing.
159. 13:16And beyond investing, what we're really talking about is creating a state machine that can take information from the world and understand it well enough to make coherent decisions.
160. 13:28There's use cases for this far and above wider than just investment firms.
161. 13:34In policy, in nonprofits, in drug discovery, in corporate strategy, I think one of the most exciting things for me about being in forms like this is talking to other people in dis- different spaces
162. 13:46with different problems that kind of congeal on the same problem space.
163. 13:50I'd love to keep talking about this.
164. 13:51I'll hang out in the hallway afterwards.
165. 13:53Here are my socials, and thank you everyone for your time.
166. 13:57[applause]