Jack Kokko - Building AlphaSense - [Invest Like the Best, EP.404]

0:00As an investor, staying ahead of the game means having the right tools, and I want to share one that's become indispensable in my team's own research. AlphaSense. It's the market intelligence platform trusted by 75 % of the world's top hedge funds and 85 % of the S &P 100 to make smarter, faster investment decisions. What sets AlphaSense apart is not just its AI-driven access to over 400 million premium sources like company filings, broker research, news, and trade journals, but also its unmatched private market insights. With their recent acquisition of Tegas, AlphaSense now holds the world's premier library of over 150 ,000 proprietary expert transcripts from 24 ,000 public and private companies.

0:37Here's the kicker. 75 % of all private market expert transcripts are on AlphaSense and 50 % of VC firms on the Midas list conduct their expert calls through the platform. That's the kind of insight that helps you uncover opportunities, navigate complexity, and make high conviction decisions with speed and confidence. Ready to see what they can do for your investment research, visit alphasense.com slash invest to get started. Trust me, it's a tool you won't want to work without. If you're attending the InvestOps Conference in Orlando this year, I'll be speaking at Ridgeline's private breakfast event on March 11th.

1:09Ridgeline gets me so excited because every investment professional knows this core challenge. You love the core work of investing, but operational complexities eat up valuable time and energy. That's where Ridgeline comes in. Ridgeline is an all-in-one operating system designed specifically for investment managers, and their momentum has been incredible. With about$350 billion now committed to the platform and a 60 % increase in customers since October, firms are flocking to Ridgeline for good reason. They've been leading the investment management tech industry and AI for over a year with 100 % of their users opting into their AI capabilities, putting them light years ahead of other vendors thanks to their single source of data.

1:46You don't have to put up with juggling multiple legacy systems and spending endless quarter ends compiling reports. Ridgeline has created a comprehensive cloud platform that handles everything in real time, from trading and portfolio management to compliance and client reporting. It's worth reaching out to Ridgeline to see what the experience can be like with a single platform. Visit ridgelineapps.com to schedule a demo.

2:12Hello and welcome, everyone. I'm Patrick O'Shaughnessy, and this is Invest Like the Best. This show is an open-ended exploration of markets, ideas, stories, and strategies that will help you better invest both your time and your money. Invest Like the Best is part of the Colossus family of podcasts, and you can access all our podcasts, including edited transcripts, show notes, and other resources to keep learning at joincolossus.com. Patrick O'Shaughnessy is the CEO of Positive Sum. All opinions expressed by Patrick and podcast guests are solely their own opinions and do not reflect the opinion of Positive Sum.

2:48This podcast is for informational purposes only and should not be relied upon as a basis for investment decisions. Clients of PositiveSum may maintain positions in the securities discussed in this podcast. To learn more, visit psum.vc. My guest today is Jack Coco. Jack is the CEO and founder of AlphaSense, where I and PositiveSum are investors, an AI-powered search engine for market intelligence. He shares how AlphaSense began by aggregating fragmented financial data sources and evolved with the advent of large language models to completely change the research experience for investors and others.

3:25He speaks to their recent acquisition of Tegas earlier this year, reshaping the business and further supporting their expansion to serve all types of companies instead of exclusively investment firms. Jack has been navigating the AI revolution from its earliest days, and you can feel his excitement when he talks about the future. We discuss building an agile platform, the importance of managing cultural integration, balancing AI capabilities with user trust and the frontier for this technology. Please enjoy my conversation with Jack Coco. So Jack, you are in a very unique position, having been one of the few entrepreneurs that was effectively building an AI product many years ago, before everyone was talking about AI, and using data and search and these tools probably for longer than just about anybody.

4:11So I think you're uniquely positioned to tell us what you've learned about applying the technology to build a great product today in late 2024. And maybe that's the perfect place to begin, which is just like a state of the union from you on what kinds of things AI enables for a product builder trying to serve a customer, what it unlocks. I'm also going to ask about what the limitations are today. But since you've been doing it for so long, just give us that felt experience and then we'll talk about the future. Well, maybe I'll start from the very beginning of what we are building with the prior generations of AI and machine learning.

4:48We wanted to build a semantic search engine that would understand financial and business content, read it line by line, almost like a human analyst would, and understand all these billions of data points, millions of documents, and connect the dots between them, then map that to a user's search query and deliver the right data points, the right insights to them. And this was really hard to do because all this content was siloed in thousands of silos and sort of paywalled data sources. It wasn't like the internet where Google and others had billions of web pages with links between them. And those links would tell you about the authority of those pages and had billions of consumers clicking on pages that also told you what's good content, what's not so good content, and what's most relevant to a particular query.

5:36None of this existed in these fragmented separate databases that we needed to go and access and then semantically index and help users search. And this was a really hard problem. Even Google and others in that space had tried to build enterprise search and had retreated from that market because it's so hard to do this without those benefits that you have in the public web. So we used earlier generations of machine learning to categorize information and semantically try to understand the context of words and sentences and then organize that information so we could then map that to each user query.

6:12Later on, you got the simpler, earlier language models like Google's BERT and where level of understanding expanded. You were no longer categorizing things or looking at small local contexts. You started to understand longer sentences and passages better and really be more accurate about understanding topics and KPIs and what is a company saying? Is it bullish? Is it bearish? Like all of this kept on improving the categorization, in the linkages between conversations. But the true big breakthrough, of course, was with large language models and generative models where now you had massive context and you could have a model that reads even up to a whole book, whatever you give it to it.

6:56Obviously, different models have different capabilities, but the contextual understanding was incredible. And the model now had world knowledge and it was able to make its own conclusions statistically, but still apparently to us humans in a pretty smart way. So that now enabled you to do a couple of things that wasn't possible before. You could take a natural language query from a user and actually understand it, take a sentence and really, really solid, good, precise understanding of what the user is looking for, which now tells the machine a lot more about what the true intent is that you're looking for in a way that couldn't be expressed in those keywords.

7:34And then similarly, we had a much greater understanding of all the underlying text because the system understood the context at the really broad level. So it took the capabilities another huge leap forward, enabled even better mapping of what a user wants and finding the right contextual insights from these vast volumes of information. So it's a really exciting space to be in. And making sense of information has just become so much more easy and powerful with LLMs now. And of course, that's evolving so rapidly that the answer changes in terms of what's possible almost every month or two. Just for context for those listening, just give us a general sense of the sort of data sets, document sets that AlphaSense had pre-ChatGPT, pre-LLMs, pre-GPT models.

8:25What were the key things that your users came to you to find information within those data sets? It was basically everything that an investment analyst would use in their research process or a corporate strategy or competitive intelligence professional or a financial professional would use in their research about their competitive landscape, launching new products and things like that. So it was basically company information, company filings, earnings call and other conference transcripts. Anything a company would publish mandated by the SEC, companies put a lot of information out there. So collecting all that and then collecting everything that the sell side says.

9:03So you've got a vast number of banks and their research departments producing a lot of high quality research. But that's in the past been fragmented. So bringing all that together and making it available based on who's entitled to what. So kind of figuring all that out and bringing that to users. Back in those days, that was the data set, mostly information that you might get access to from financial data terminal that you might be using as an investment analyst or as a corporate typically would have less access to some of these more private proprietary sources like the bank research. But still, those were the types of content.

9:39And since then, we've obviously added a lot more really insightful proprietary content on top of that. Yeah. And so if I think about oversimplifying, I'm sure, but conceptually, the early version of the product is all this relevant data that is in disparate places, disparate silos, and the best possible search tool to get me the pieces of information I want based on what I'm asking for as quickly as possible. Right. It was information that's hiding in plain sight. How do you find the needle in a haystack? Everybody had the haystack, more or less. But it was just a lot of work, so much work that you often didn't have the time to invest in it.

10:14So you had to guess and hope for the best that you had covered enough. And you were always fearing that you might be missing something because it was really so inefficient to come through this information. So this allows me to ask a question I'm so interested in, which is, okay, so you had been doing this. You had been using machine learning models and I'm sure investing tons of brainpower and dollars and people into an amazing fast search experience. Take me back to the day that ChatGPT came out or GPT-3 came out or whatever the relevant milestone was for you all. how that then changed your world?

10:46What changed in the product? How did you take in that new capability? Did it displace a lot of the advantage that you had built up? Do you feel as though these general models make it easier for other people to do, let's say, great semantic search in their document database for enterprise search? How did your world change and how did you react to that first technology and then the subsequent ones that have come out? So again, it was going from these smaller language models with more predictive or language understanding capabilities, but without generative capabilities. And so the large generative models gave a lot more capability, a lot more general purpose abilities that you could deploy.

11:23For us, that was just an incredible playground and opportunity to go and just apply these brains in a box as I like to think about them in lots of different ways. There are really such general purpose tools that we used to spend the first few months thinking about the first very obvious things to build, like summarization of documents and summarization of insights across an industry and so forth, like letting users get access to cross-company insights in a way that would have been too tedious in the past and suddenly became very doable and figuring out how to control these LLMs and how to control, build guardrails to prevent hallucination and retain some of the capabilities of our search engine, where some of the beauty was really in the user experience of being able to really see the individual sentences and snippets of text where the insights came from and quickly validate and get to the source context and read the underlying document, really know where all the pieces of information came from, figuring out how do we deploy this in a way that retains all that greatness when this new world where you're now able to use algorithms to do it in a smarter way.

12:32Now, did that give the rest of the market new tools? Yes. I think that it's in some ways given deceptively interesting tools where you even see, I mentioned earlier, how even Google had an enterprise search product that they pulled out of the market because it wasn't feasible. Now you see some large corporations and large financial firms tinkering around. Their IT teams have been given a budget by a CEO to say, hey, go figure out something to do with Gen AI. And the most obvious thing that seems to always come up is that, hey, we've got all this internal content hiding in all kinds of disparate places.

13:10And we've done all this work to produce all these insights and memos and so forth. But when we need it, we don't find it. So can we go solve this? And suddenly Gen AI gave some tools to those IT teams that allowed them to try to do and create some very interesting demos that then got funding often. And so people started to try to build these things that even Google had backed away from doing in the past. So yes, it certainly gave capabilities that weren't there before. But as people started calling this RAG, retrieval augmented generation, and started to look at all the components of creating a robust software system, the reality has still been that, okay, you've now made some new leaps, but you also have some new difficulties.

13:54is the leaps that this technology enables. We feel like we're a company that has spent the last decade and some years on top to really get really good at this. And so we made those leaps quickly. And you've got the large frontier model companies offering their solutions. And those are advancing quickly and staying on top of this technology and developing something great that users have now come to expect by using something like a JetGPT. it's actually now become in some ways even harder because of the really fast pace of development. Just when you think you've built something great, some new great model advances in its capabilities and that's available to consumers.

14:34And now those consumers expect that you have that also in your professional solution. And so I feel like that actually gives us even more advantage as a startup that can move quickly and deploy these capabilities and has a really smart search engine that feeds the Gen.ai on top that is doing now the conversation with the user on top of this really powerful search engine that semantically kind of finds all the right information so that it can feed it to the Gen.ai to summarize. So I feel like this actually has been a massive accelerator for our roadmap and created a ton of even higher demand from the market because we've gotten closer to the original vision we had of a machine you could talk to and get answers from.

15:15I always thought it was much further away. There was when you could actually deliver that. And now we're suddenly able to, I could pick up my phone and pick up the Alphacense app and ask a question, any business question, get the answer. And I thought that was decades away, realistically, but here we are and that's possible. So it's a wonderland for a company like us. What have you learned so far from user behavior? Everyone can sort of imagine all these cool ideas and then you put them in front of people and maybe you never know exactly what they're going to use. Is there anything that surprised you about what they use or don't use that's newly possible because of these models, whether that be synthesis or summarization or condensing or just better search?

15:54Or what have you learned about the revealed preference of your user applying this technology? Mostly, it's been fairly expected. It's just making these better connections, summarizing across companies, industries, having a better natural language conversation with the user instead of keyword-based semantics. Maybe what has surprised me is more on what you see on the consumer side where people use a chat GPT or perplexity or Gemini, et cetera. And they, in their daily lives, in my mind, trust these LLMs too much. And I find myself doing that too. You don't have infinite time. And if you don't have citations, you just have a credible looking LLM-based answer that you can read and feels true.

16:45and let me run with that and check against my own recollection if it makes sense. It's too much work. They don't make it easy to go and validate, even if that was possible. Whatever the alarm pulls from its vast memory, it doesn't even know where it came from. That consumer world, there is a lot of trust in that. And we see what surprises me a little bit is that we occasionally see and hear about professionals using those same tools and defaulting to trusting them, in my mind, way too much. Most people don't, but that's been a bit of a surprise where we've always worked with serious researchers that do deep dive work.

17:21And our function is to take them to the right content, but they want to really read it and get the full context and see where it came from. What's the analyst that wrote it and really get to the bottom of something before trusting it, before making seven figure bigger bets based on this information. But maybe the biggest surprise has been that you also do see professional users fairly often trusting some of these public LLMs too much. If you think about the things that differentiate and just the general categories of the tools and the data, two simple things. You mentioned that you've migrated a lot towards proprietary content that just you have.

17:58you were an early aggregator of proprietary content sell-side research that you could get directly from lots of banks, but not aggregated in one place. How much do you think the separation between the winners and the losers in, let's call it enterprise financial search, will be determined by tooling versus data? Because you see perplexity, now you can search over financial documents and perplexity, SEC documents or whatever. I'm curious how you think the battle shakes out between between providers trying to solve this problem for the customer and how much of it is tools versus how much of it is more and more useful data?

18:32That is a great question. I think for us, the fundamental problem has always been about fragmentation. I experienced this as an investment banking analyst where I first came up with the understanding of the problem that needed to be solved in this industry, where I was logging into dozens of different systems and asking the library to go to even more places to try to gather information. So the fragmentation was always a huge pain point. Back then, it was so fragmented that you were keyword searching a single PDF, one keyword at a time, and doing that across lots of PDFs and trying to aggregate this information.

19:11Even with more aggregation, it's a dissatisfying solution unless you have something where you can ask one question from one system. So in that sense, it is both a technology battle of having a really great solution layer on top, the smartest LLMs applied in the smartest way and sequenced in the right way that you can create a trustworthy answer, but also then the user experience that helps users still get comfortable and trust, but verify and validate and get to the context and read. but the underlying content you want it to be all in one place you don't want to go and do multiple searches in different systems and learn the intricacies of each one if you could do everything in one place so our strategy from day one has been like hey we'd love to be one place where one search and done you can really expect that if you didn't find it in our platform it doesn't exist and we often hear clients telling us that's how they think about it that it gives them the comfort that they didn't have to go to 20 different systems to get that comfort, they could go into one system.

20:14Yeah, I'm curious also, if you think about planning a business like yours for technology landscape that just seems to be explosively evolving, how do you do that? If you don't know what GPT-6 is going to be like or something like that, how do you do effective planning, not knowing what the technology itself will allow on the product tooling side? Well, you have to try to, in my mind, organize your team and the org structure to be agile, because you just have to expect the unexpected. It's going to happen and it's going to happen very frequently. So you have to be able to react to new developments and you don't know what they're going to be.

20:54Nobody knows really. These LLM developers, even the frontier model developers, don't know what the next breakthrough might be when you cross some scaling threshold or new algorithmic development. For us, the solution has been to first try to create a system that is pretty agnostic, that can use any LLM that's out there and use different systems to accelerate their performance or build the methods where you can very quickly pivot and change what model you use for what task if there's suddenly some model is better at multimodal or better at a bigger context window is better at doing those things much faster and so forth.

21:33So you can actually, you've got less latency, more time to get more intelligence into the system. And so whatever those things are, you want your systems to be more modular so you can orchestrate all this in a much more flexible and agile way, and then have a more agile organization that tracks what's going on. And it's on top of those new developments. And even when they just come out as research papers, you are reading everything that comes from credible sources and try to stay on top of what's a real development, what actually could be deployed today, what's not quite there yet, but what's giving us some hints about what might come next.

22:07So I find myself spending a much larger percentage of my time now in product development and reading everything myself. I guess I get more of the filtered information from the team that reads everything and then one out of 10 things that really is impactful gets on my screen and many others do the same. Different levels of deep dive and screening of that information, but you've got to be on top of it all the time and willing and able to pivot. I know you can't tell me exactly how investors will work in five years that's different than today because the future is hard to predict, but I'm sure you can say a little bit about what might be possible for them that isn't possible today.

22:47It's a sneaky way of asking just what is most exciting to you from a product perspective of what you will enable your customers to do five years from now because of all this. Put your dreamer hat on. What do you hope you're able to do for customers five years from now that you can't do today? We're trying to take all the friction and barriers out of the way and try to make the cost of acquiring information and insights really, really low so you could suddenly scale it and do a lot more work, do a lot more diligence, and have it be organized for you and done for you by machines so that the things that used to take weeks that you didn't have the intern around to say, go research this for three weeks.

23:29But if it takes three minutes, because the machine can organize that information and ask the same five questions from 500 documents and say, this is the most important set of documents you really should read and maybe some tentative conclusions that you could make. Hard to really see how soon and how far that goes. But I like to think of it as the computer systems and the large language models within them doing a lot of the work and acting as a large team of superhuman analysts that get smarter and smarter. And maybe there is a large group of interns that now turns through a lot of the work that would have been a past required long time and didn't get done.

24:11Now it can get done. So the level of diligence, level of research gets quickly better. And then the systems get smarter and they can start to draw some tentative conclusions and you can start to evaluate out of these five that the machine generated, these other five scenarios, now you can start to play with them as an investor. Yeah. It's so exciting to imagine that you might be able to say, build me a industry primer on XYZ and have it be done in two to three minutes. That's pretty exciting. Right. Those kinds of things are very much doable today. But then what do you do with that? And maybe you let the machine take alternative views at the industry and try to simulate where could that go.

24:45And it's not going to be able to tell you where it will go, but it can tell you different paths of where it might go. What data sets are most exciting to you? One of the things about the AlphaSense story that's clear, we were a big investor in Tegas, and obviously you acquired Tegas. That's an incredibly valuable data set. I have questions about M &A and the role it plays in building businesses in a little bit. First, how do you know you're onto a data set that is interesting, different, useful? I'm sure you've looked at every data set imaginable. What are the features of data sets that make them valuable?

25:18To me, it's the net addition to the information volume that's out there already in terms of impact of those insights. Are they overlapping with what's already out there? Are they totally new perspectives on something? And you rarely get that new perspective. The expert interview content is definitely what I see as the most exciting one because it provides a whole new lens into what's going with companies and industries and private markets where those markets have been lacking transparency, have been in the dark in terms of information you could get on those companies. in the public markets, which are becoming a smaller and smaller part of the total picture.

26:00You've got SEC telling companies exactly what information they need to be putting out at what frequency. And you've got the sell side writing research on those companies. And so you've got pretty good information out there on public companies, but even that is filter. The companies try to do their best, to show their best sides of their business and get their best story out, positive story. So interviewing those experts, getting buy side interviews with investment analysts that are about to invest in a company, interviewing an expert at a company, former executive, customer, partner, competitor, and really asking them the most important questions about the business model, the competitive environment, the company demand environment, and so forth, and getting real answers from real people that don't have any skin in the game to alter that viewpoint.