%20(150%20x%20150%20px)%20(1).png)
Invest with AI
Brett Caughran and Khe Hy lead a deep dive into AI for investing, joined by guests at the cutting edge of the field. Our goal is to be your Sherpa through a rapidly changing landscape by distilling what's working, what isn't working, and how you can leverage AI in your own process. Follow along as we tackle AI's biggest challenges and opportunities, one episode at a time.
Invest with AI
Implied CEO on the Limits and Capabilities of AI for Investing
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Ying Hua left a PM seat at Balyasny to build Implied on a contrarian bet: the big AI labs won't win finance (but it's not for the reason you'd think).
She, Brett, and Khe get into where that leaves the analyst, why Claude Code won't replace your data team, and the one part of the job she's convinced stays human.
Timestamps:
[00:00] Intro
[00:45] — She Left a Balyasny PM Seat to Build This
[01:52] — Why Bet on AI Investing in 2023?
[03:45] — Will the Foundation Labs Eat Every Vertical?
[08:30] — Is Pattern Matching Its Own Kind of Intelligence?
[11:07] — The Data Problem Nobody Talks About
[16:02] — The Alt-Data Nobody Else Will Ever Build
[20:51] — Can a Non-Coder Really Build Scrapers with Claude Code?
[25:21] — Why the Static Dashboard Is Already Dead
[32:32] — Synthesis vs. Judgment: Where the Human Stays
[37:00] — Why AI Still Can't Tell What Actually Matters
[39:24] — Solving Excel: The AI-Native Model in the Cloud
[45:32] — From Glorified Search to Cloning the Analyst
Watch & listen to every episode of Invest with AI:
https://www.fundamentedge.com/invest-with-ai
Want to actually build these workflows yourself? The AI Accelerator is Fundamental Edge's 6-month cohort for investors who want repeatable AI workflow. Learn More below:
https://www.fundamentedge.com/ai-accelerator
Follow Invest with AI
Spotify
Apple Podcasts
YouTube
Actually, personally, I think public market uh specific stocks are one of the hardest domain knowledge to grasp.
SPEAKER_01Decay's point, many of our clients are sort of building dashboards with cloud code, but this sort of topic of agent debt and the topic of token tokenomics budgets. Yeah.
SPEAKER_02When you read the surface level reports, they will have one interpretation. But when you have more domain knowledge, like depth of domain knowledge, you're going to reach a different conclusion.
SPEAKER_01Hello and welcome to Invest with AI, where we explore the intersection of fundamental investment research and artificial intelligence. We're really excited to have Ying Hua today from Implied on the podcast. Hello and welcome, Ying.
SPEAKER_02Hi, thank you so much for having me. Excited to be here.
SPEAKER_01Thanks. Uh thanks for being here. Can you uh just start sort of giving us the Ying Genesis story of sort of your quick encapsulation of your career and what you're doing now?
SPEAKER_02Yeah. My background is in fundamental investing. Uh started my career at Goldman on the equity research side, then moved to Citadel, was there for about five years, um, and then uh went to Ballyasney for about three years. Uh, and then I started implied um in 2023. Um, I was a financials and fintech PM at Ballyasny, um, but I also ran a quantum mental book. So I think I had some experience with fundamental investing as well as sort of quantitative, not quite AI, but quantitative um uh kind of tools. Uh one last interesting tidbit is in between Citadel and Ballyasne, I had a very long non-compete. Um, ended up learning Python and Gala Mastery Data Science and played it with language model back then before it became popular. So before LLMs uh became the mainstream, um this is when like BERT was still the state of art. Um so I actually had a chance to play around with it when I was at Bay M2.
SPEAKER_01And even back in sort of more um in um in style in 2026 to leave the buy side for an AI startup, but in 2023 it was it was uh more of a contrarian position. What what gave you the conviction? Sort of what did you see out of that uh your experience back then when like frankly the foundation models were pretty like not very good for my use case? What sort of gave you the conviction to to plant a flag and start an AI business?
SPEAKER_02Yeah, I think there are two uh catalysts. Number one was um I believe Chat GP GPT-4 was I remember the distinct turning point. I think um when I was initially playing around with things, a lot of things I wanted to do, it couldn't really get there. Um by GDP4, I feel like with enough prompting and enough chaining, I can get it mostly there and to kind of edit things a bit. Um, so that's number one is I definitely saw the potential of making it work and knowing full well that things are gonna improve and stuff like that, right? So that's number one. Um, number two is at the time there were a few um options out there, uh, AI tools for investing. Um a lot of them are companies that kind of started before AI and then with the ChatGPT release, kind of try to pivot. And um, I actually had a chance to play around with a lot of them. Um, I was sitting next to our AI team, so I got a chance to see a lot of things. Um, and I didn't really quite find uh a tool that really helped me. So and I think the gap that I saw, um, I thought it requires someone with a domain knowledge to solve. And hence I thought it might my back my background is interesting, that it's kind of intersection, right? I I know fundamental very well. I know enough about quantitative stuff. I'm not a best programmer, but I know enough about Kwan Sai to really fill in those gaps. Um, so that was the the reason.
SPEAKER_01That sounds great. And one of the um I think I don't think part of the reason Kay and I started this podcast, we don't really have the answers, but we know the questions. And one of the key questions we're sort of planning to explore is the sort of competition of the horizontal foundation lab versus vertical finance agents. Yeah. Uh how would you frame that competitive tension sort of today? Yeah. And how do you see that tension playing out over the next nine to 18 to 36 months?
SPEAKER_02Yeah. I think uh especially with a recent agency improvement, I think a lot of people are in the sense that oh, model company will eat the world. Um, obviously, I'm starting company, so my answer is going to be biased. I don't think that's going to be the case longer term. So I'll give you like a um an analogy I usually use with uh our clients, and I'll give you another kind of more in-depth uh reasoning for that. Number one is I think of language model very much like universities, right? They they produce intelligence that are um capable of reasoning, basic level reasoning. Uh, they produce intelligence that's capable using tools, Excel, PowerPoint, whatever it is nowadays, right? Um it produce like kind of basic level that you can do a lot of it. It's kind of general, so you can do a lot of things. But when you dig deeper, it's kind of similar to when you hire someone very smart from college, when you bring them to your buy side, there are a lot of training you need to do on your own team. How does your PM think about things? How are things run beyond even just like a the stock level, just process by it? There's a lot of things to learn. And then on the stock level, that's even harder, right? I actually personally think public market uh specific stocks are one of the hardest domain knowledge to grasp because unlike some other field, think of medical law. A la is a law is a law, it rarely changes, right? Um, and a medical, a disease of disease, it really changes. Fundamental investing is tough. One is it's always changing. Whatever you knew as of yesterday might change based on one single news. That's number one. Number two, it's it's an interconnectivity. I used to say, like, if you can figure out the knowledge of a whole market, you can really figure out the model of a whole world because any concept in this world has some sort of representation in the market. So that interconnectivity thing, I think it's very unique to this um industry as well. And lastly, is um it requires an understanding that flows between qualitative and quantitative. It's not just qualitative. You need to think about quantitative stuff. So I think those three characteristics make it really hard to get domain knowledge. And personally, I think domain knowledge is a piece that um the horizontal model companies are not gonna do. Like, do you think they can actually figure out a way to ingest ever-changing knowledge for fine use for the market, right? Um, so that's number one. Um, I think that's tough. Uh it's mostly just as a dirty work, too. I just don't think they're interested in doing that dirty work. Um, but the second thing is I think from first principle, one needs to understand the mutation of language model, right? Language language model, at least the the bargain architecture today, um, uh, you know, it's it's next token prediction. It means um language is really a coding too. It's a coding language. It's like we put information in our language. And based on that information, you try to predict what is the next piece of information that can likely come through. It is, by definition, backward looking. It's by definition pattern matching, it's by definition not true understanding, uh, meaning it's it's just kind of fielding in the pattern. The analogy I will do is think about quantum investing. All right, quality investing, it's kind of like similar to that, it's just like information coded in numbers. You're trying to predict what's the next number that comes up. In this case, it's information putting encoded in language. Do you figure out what's the next word that comes up, right? So when you have that kind of market model architecture, by definition, you're gonna run into a few things. Number one is it's not true understanding. Um, it's not true reasoning either. It's more just how do you uh give it enough information in your language in that context window so it can do that prediction more accurately. That's all it is, right? That's number one. Um, number two is it can never figure out things that hasn't happened before. It's a pattern matching, quants, backward looking, right? It's always backward looking. When regime change happens, when pattern changes, it cannot identify it. Same thing with LLM. If information you give up till that point, if something changes, it's not gonna be able to predict that. So those are limitations. I think based on that, um, I think there's still a lot of holes to fill from each vertical, uh, and I think that's kind of where we come in.
SPEAKER_00Can I jump in here uh on that, on the pattern matching question, Yane? Yeah. Um I completely agree with you that there's tons of backwards-looking pattern matching. Isn't there an argument to be made, though, that the sheer volume of patterns that are able to be extracted is in a way its own set of knowledge, is in a way its own set of understanding? Yes. How would you think about that?
SPEAKER_02Yes, I think that's correct. Uh, but I think it depends on the input of those pattern matching. So let me give an example. How do you think language model train based on uh earnings results? It remember like early in the days, like even when Bloomberg was training as a model, like we were talking about like all the financial training, right? The idea of training, you just shove it a lot of documents. 10k, 10k, transcript, blah, blah, blah. Okay. Um I think uh if you guys ever read the earnings release, okay, management will always say, we had a great quarter. We had a great quarter. It doesn't matter how good, actually good or bad a quarter is, we had a great quarter. When you feed that transcript into LM, LM will tell you they had a great quarter. Right? So though, so my point is, number one is I agree. Uh if you can, if you can derive the right pattern, then the pattern itself is enough. But I think there is a limitation on how these models are trained because they are fed in raw data. And the other question is like, uh I don't know, I I covered a financial, so I'm much more acute to this, but um every time I read like a Wall Street Journal read about how the AIG do, it's really never the right read because they're looking at the wrong, wrong metrics. It's there's something that they didn't know. It's because everyone's kind of like someone who doesn't understand industry, when you read the surface level reports, they will have one interpretation. But when you have more domain knowledge, like depth of domain knowledge, you're gonna reach a different conclusion. So my view is I don't think these models were ever trained in the proper way of ingesting knowledge, other than it's just the raw data. So that's kind of my my view as well. But I agree with you. If we can figure out a way to train it in a in the right way, like causal way, right? Like um, you know, Snowflake is up 35% today because blah, blah, blah, blah, blah, versus expectation. Then if you train model uh with this, I think they can actually mimic the right type of reasoning. Um, meaning it can it next time it sees something similar, be like, okay, I think it's gonna react this way. But models are not trained that way. Models are trained purely on basic uh data.
SPEAKER_01Can you can you talk to us about just the the data challenge? I mean, part of the reason that I didn't really conclude that LLMs were institutional great in 2025 was so much of the uh of the sort of data that informed outputs was from uh from the open web. And so the training corpus is sort of fundamentally a sort of compressed web scrape.
SPEAKER_03Yeah.
SPEAKER_01And I think you know, what should the open web may be great for you know coding um sort of tools, but in terms of like institutional grade finance workflows and data, uh not not so much. Uh how have you sort of tried to crack that problem? And how have you seen the ecosystem evolve to sort of handle that institutional grade data problem?
SPEAKER_02Yeah, I think data is an easier problem to solve now, right? Because it it's interesting. Like I I remember we started the company in 2023, and within a few months we just saw a bunch of companies starting. And then when my co-founder and I were like, it's interesting, a lot of company, the dollar founders don't really have investment background. Why would they start a company like that? And then we they hit us like, oh, a C C document. There is a large corpus of free data out there that you can just put a put feed into LLM and then it spits out some answer. Like it's it's a very natural reaction to go. So number one is I think source of data. Um, you you want to go to a source, you go to SEC, go to company websites, you go on to go to um transcripts. Um, and that's number one. And number two is we have taken a slightly different approach. Like we don't buy a lot of data, we actually process all the data ourselves. We do all the pre-processing ourselves. And then um, to give you an example, um, I think every competitor out there, more or less, buys transcript data. We actually do our own live transcription. And we got a lot of questions like, why would you do that? By the way, it's not not even that, it's not that easy to build. There's a lot of uh um uh like busy work, right? Uh that a lot of uh nocil glorious work that we needed to do um in that. But the reason why we do it is number one, is we get to control how data is processed. Number two, we get to control when the data comes in. So, like for example, our live transcript, it's live, it's sub one-second delay. As soon as the earnings hit, it can uh uh come through. Number three, it allows us to gather raw data. We're doing that raw data. So if you talk to most of quant that buys data, they are like, oh, um, how is sentiment on this topic of our management, right? Um, but they're using clean transcript. How can you tell sentiment on a topic if you don't include all the filler words, the raw transcript, the ums, the uhs, the spaces in between sentences? You can't really do that. So we wanted to find a way to have the most raw data that could be processed as quickly as possible and control the full pre-processing step. So that's number one, I think, something that we did. Number two is very early on, we thought that like the qualitative data sets is more or less obvious to people, but there's a lot of quantitative data sets that people are not including in there. Um, and that has twofolds. Number one is a lot of questions should not be pulled from qualitative data. Uh so the very simple, uh, this is like a very common trial question, right? Pull Apple's China sale over the last 20 quarters. Like uh before, like when a lot of people are doing this as a sample question. That question should not be pulled from 20 different transcripts. That should be a structured data set. That should be a Bloomberg pull. Like I have joked about that's a terrible question because a lot of people can pull that data a lot faster using Bloomberg code than you typing out that whole sentence about like, you know, what you pull. That's number one. So I think some data should be structured to begin with. And I think that goes back to some of the data quality stuff, um, doing models and stuff like that. That should be structured. Uh, second thing is um institutional grade is also, we try to find data that people would have wanted to get, but has a lot of hurdle in getting. Um, so what do I mean by that? Some of the alternative data. Uh, not alternative data like a credit card reports or whatever stuff everyone, every hedge fund uses. There are a lot of data out there that only works for one or two companies. No one really is scraping that data because it's not a viable business to scrape that data. Who are you gonna sell to? It's only for work for two company. Quant's not gonna buy it, right? It's not enough data to backtest. So we very early on have helped clients say, hey, you can actually use us a tool to script those data. We can help you script those data and everything's integrated. So we fill in the data layer by one, going to the sources. We try to do a lot of pre-processing ourselves. Two, um, we're filling a lot of gap with the quant uh with um a structure data set to one improve the quality, but also give more scope uh to the availability of our data sets.
SPEAKER_01What would what would be some like one or two examples of that sort of sort of data?
SPEAKER_02I'll give you an abortion one because I think no one actually out there is actually caring about it. Um so I don't disclose anything that our client tells us, but I used to cover insurance companies, um, car insurance companies. Um and car insurance company, uh, you you estimate a loss ratio, a lot of times kind of fingering the wind, you know, like, oh, the weather is about the same, a little colder, so I'm gonna bump up a few basis points in loss ratio. Um, when I was a band, one of the quantum mental things we did was we um actually uh built a tracker uh or built a scrape for seven of the toppest like highway patrol um websites. And we pulled um data around accidents. So for every major city, major like biggest uh seven of the top 10 markets, which I think is more than like 70% of Norwegian share, um, we actually know exactly how many car accidents is happening. And then we look at the duration of the car accident, when did it clean up as a proxy for severity? Then we actually have an index, it just runs in the background. We have an index on is this um uh is this uh should be a higher uh higher uh loss ratio or not. And then we combine that with domain knowledge, right? Think about um we know that every third quarter, uh, insurance companies do a reserve study. Do we think they're gonna release reverse reserve? If they do, we're gonna add that to whatever the data we have. So that's how I think about um kind of uh data that's only, by the way, there are only like three auto insurance companies that trade that data for. But that's that's my point. Like who's gonna build that data? It's it's it's not uh it's not an easy data to build, who's gonna build it, right? So that's an example of uh, I think a useful data set. I think um that's gonna become increasingly more important because those are data sets that fall in the crack of quantum touch it, and there's not enough data set back test to do it, right? You have uh you may have 30 years of data, four quarters, that's 120 data points times three. Like that's not enough data to do a back test. But fundamental people can use that as part of the mosaic.
SPEAKER_01And to do that by hand for a fundamental investor, there's probably not there's probably not an R an sort of enough of an ROI to try to get five fifths more accurate, but sort of combining many of those signals together into a forecasting forecasting mechanism sort of could scale that precision and forecasting in a Yes.
SPEAKER_02And now now it's uh think about every part of that process a lot easier. Like it took us a few months to build it ourselves, but now scraping is a lot easier, doing analysis is a lot easier. I can schedule run it a lot easier. I don't even have to go like before there's also technical problems of like we run a Python program, it spits out a number. How do I pull that number into my Excel spreadsheet? My model, like even that was a hurdle, right? Because uh because uh I basically had to write a script to open my Excel sheet, put a number in there, refresh it, save it. Like a lot of things are a lot easier now. So I think I my hunch is more and more of these like alpha edge, like things that would be nice to do before. Now you actually have time to do it, and it's a lot easier to do it. I think it's gonna become increasingly more important.
SPEAKER_00Yeah. Can I just ask on that um data scraping question? Is that an AI specific skill in the sense that AI is going out to do it? Or because that's what it strikes me as, like you said, a Python deterministic uh process that you know you go find the site, you you know, you clean the data and you put it in a different place. And so can you clarify like the AI role in that? Is it that it's just easier to code that into existence, or or is there an actual like native AI activity in what you just described?
SPEAKER_02Yeah. So I'll split it in two parts. I think one, it's easier on the data collection side. So get the raw data. Okay, that's uh uh we're probably one of the earliest ones to have like kind of an agent that does scraping. So it does most of the scraping, huming the loop, we clean it, da da da. Um, but there is a second part is uh it's very similar to how everyone reads 10K a little bit differently. Everyone should process a data slightly differently. And uh from that point, it's no longer deterministic. So gathering data is just deterministic, cleaning data is deterministic. There is a right answer. But how each team uses that data, how each team um, what insight are they trying to get from that data is non-deterministic. Um, but that part is also, I think, accelerated by AI. And and this is especially important for we have a lot of even teams inside of large firms, even though they have a lot of resources technically, but there are still a lot of things you're like, uh, it it will take me too long to explain it to the data team. Or by the time I explain to the data team, the catalyst would have gone. Now they can just do it. Um, it it takes too few minutes, they can take a first crack, and does it worth more time invested than if it does, then let me do a deep dive and not it can pass. So I think that part is the AI native part.
SPEAKER_00And just on that, do you think that Brett and I have had this conversation, like the prevalence of claud code for non-coders?
SPEAKER_03Yeah.
SPEAKER_00Do you you being closer to it and us being kind of the clawed, the non-coder coder types? Is it a is it realistic that a non-coding analyst can go in and be like, okay, claude code, help me generate this scraper that pulls this deterministic data and so on? Or is that a bit of a pipe dream that you know is going to collide really harshly with reality?
SPEAKER_02Uh I'll preface say that I definitely think we're biased, right? Like I if everyone can use cloud code to build every process, then I think we don't really need to exist, right? So I'll prepare say I do think about answer is biased. Um, I don't think it fully works, and and I'll tell you why um from a few exam a few angles. Number one is data pipeline. So you need to schedule build it, right? Something is easy. You run it every day, every whatever. Sometimes it's not. Some data comes at a random interval. How do you figure out when to scrape? That's number one. Number two is you need to store that somewhere. You're not gonna run everything at uh like uh all at once, right? Like you need to store that data somewhere. Number three, if anyone has done scraping manually, you will know that website changes. Uh whatever deterministic thing is gonna break, uh, pretty much for certain. In a span of a year, every single scrape we build probably has some changes. How do you adapt to that, right? Um, and then lastly, I always tell you this, and this is, I think, um, I don't think it's fully solved, but when we first started out implied, we actually wanted to do quantitative stuff, meaning like natural language quantitative stuff. And we quickly basically decided that language model wasn't good at it. And even though language model has improved, I don't think some of the problem had been solved. I'll tell you why. Um AI generate Python code. How do you verify something? How do you verify that the logic you're building there is correct for a non-technical person who doesn't cannot read Python code? Right? Um, it's like we think there AI can help, but I don't think it solves completely. Um and the reason why that's important is I remember very interestingly when I was at Citadel, we had um uh data scientists, we put us data scientists onto our team, and we want to describe something, and they'll come back with this is this is relevant, this is not relevant. And every time I was like, I need to see your code. Uh, and I picked up Python when I was at Citadel, but I need to read the code to understand what you wrote. And then I can tell you, oh, you made a logic flaw here. It could be just like one sign difference, it can be very small. And it's just really hard for non-technology people to truly evaluate that, right? So um, I think some of that is all like the way we approach it is um you might not know the code, but you should know what the right answer is. So you can evaluate if the answer is good or bad based on certain points, like kind of sanity check. It's kind of like if your intern hands you an excel, it's not gonna check every cell, you're gonna sanity to check if you swap. They're waiting to do it. But I think a lot of especially the more complex things, I think it's really tough. So those two things I think is really really tough. It's like, I think clock co-building sum is very good for one-off things, it's uh harder for a continuous production level, self-improving, like that part is is difficult. And I think going back to just from an industry perspective, I don't think like this is actually true for like uh people's process today. Uh, like I think a lot of the work that individual team has done is duplicative work, meaning like, like, think about uh company release earnings. How many Amazon uh release earnings? How many teams right now are updating their model, right? Like that there should be one single source of truth. Uh, it should not be uh everyone doing their own thing. Same thing with data pipelines. Like, I don't think it makes sense for everyone to try to uh, you know, I cover 50 stocks, I have 50 data sources to manage all those 50 data sources with a pipeline through. Would I rather just use a company that can take care of all of that? And how I analyze it is up to me. But how the data is gathered should be just one company. So that's another thing, way to think about it.
SPEAKER_01So we're sort of at a moment in time where in the in the horizontal versus vertical debate, the the labs sort of are ascendant, in particular, in particular, Claude is ascendant. And Decay's point, many of our clients are sort of building dashboards with Claude code. But this sort of topic of agent debt and this topic of token tokenomics, sort of growing growing budgets. Yeah. Um, how do you sort of see these things? Because it is, it is seductive to sort of spin up a dashboard and um and sort of think that it's institutional grade, although most sort of deep AI vendors like yourselves, we talk to sort of, you know, cite that as a bad sort of bad uh sort of not great decisions. So how would you sort of frame those two buckets and what do you think is next for clients and sort in in terms of using and building dashboards?
SPEAKER_02Yeah. I I never would say dashboards bad. I I think like uh I think um, you know, um I think Brett, you've seen our product, right? We're one of the few uh vendors out there that actually have a dashboard. Like we we actually think there are certain things that you should, it shouldn't, every not every answer should be a chat answer, right? Uh some information should be serviced to you without you having to ask. It should just be there. This there's a reason why people like to stare at Bloomberg terminals, right? Like this, there's value in that. So I would never say dashboard is bad. Uh, but two things. Um, number one is I think the idea of a static dashboard is bad, even for the same company, even for the same industry. Like what the dashboard should look like today versus what you should look like two months from now for next earning season could be different, right? As this is based on domain knowledge things. So one is how often do you want to be changing that dashboard? Like, do you want, do you do analysts in PM want to spend their time deciding how to change? Or can you use a platform like ours to be like, I you have all the data already. I just want you to present data in different ways, or I want to or reorder data in different ways so you can show me what's important, what's not important. So that's number one is just like how much time you want to do it. Second thing is, um, I do think that the ultimate form of what invest in investing should look like, the process, is not gonna be mainly dashboard driven. Um, so what do I mean by that? Um uh I'll give you an example. So I used to love chart books. I don't know if you guys look at chart books. Like, you know, you do certain charts, whether it's like an S revision, stock, whatever it is, and then you have it, you know, do a PDF, you flip it through, right? Like you get a lot of information, you get a lot of intuition from it. It's great. But once you get to like chart book number six, you're never actually looking through the chart books, right? You're not actually going through, okay, what's new, what's worthwhile, right? So, dashboard, I think there is gonna be a similar dynamic. Uh, you're gonna build a lot of dashboard, everything is gonna be very cool. At some point, you're gonna have too many dashboards to manually check. Um, so what do we think is the ultimate end of like uh investing with it? I think for us, it's like you should be able to create dashboard or however many of you. Like dashboard is just like how would I want to view this particular data source, right? That that's all it is. Uh it just everyone has a different preference. What I think is important, what I'm screaming for, what you should be able to create dashboard very easily, create in a metric, in a visual very easy. That's number one. But two, you should have someone reading through it and tell you. It shouldn't just be building dashboards, it shouldn't stop there. It should be, hey, uh Brett, basically, you have a dashboard that tracks a hundred different data points. I'm gonna check the 100 data points for you to tell you, hey, Brett, you should look at this one. This one's interesting, this one has a big jump, this whatever. Like, my view is dashboard is a means to an end for you to reach some conclusion. And the purpose of AI is should be treating like uh you have an intern who reads those dashboards for you, who build the dashboard for you, the read dashboard for you, filter all things that's important, and alert back to you so that you don't have to have to mentally spend time thinking through those.
SPEAKER_01It makes a lot of sense to me. I mean, I think one of the fundamental challenges of front office front office software, investing software, has been this deep heterogeneity of investment process by style and by coverage area, right?
SPEAKER_03Yeah.
SPEAKER_01And so really like the only scaled software has been like the terminals, which are just sort of like an aggregation of data and allow for this customization in a launch pad, right? Like all of our launch pads look differently.
SPEAKER_03Yeah.
SPEAKER_01How do you think we take that same concept? Like, how do you how do you mass market scale to this degree when when you're sort of building software? How do you allow for that mass customization of dashboards for your clients?
SPEAKER_02It's very easy now. Um, so uh let's let's let's uh um let's think about this way. So let's say you have um 50 data points you track, uh different data sources, right? Um one, you can auto-run each one of them, uh, and then you can have each one decide is this an interesting thing? Uh and uh say of the 100 I run today, like say I have a daily run of 100 uh data sources, uh, and then out of the 20 of them is interesting. AI determines interesting, could be a meaningful change, can be versus consensus, whatever it is. And then you can just have an uh an agent on top of it. It's like, hey, those things that's interesting make a dashboard for me. I only want a few things that's interesting. That's fully customized, right? That that that's definitely doable. Um, but I think there's another layer of customization you touched on that I think it's actually a harder problem. And you're talking about a different industry, domain knowledge. Go back to my domain knowledge point. I think how do you analyze a tech company is very different than how you analyze an insurance company. The metric you care about is different. So if you think about the terminals, if you think of some of the current vendors when they generate AI, a lot of it's standardized, right? As standardized uh whether it's model KPIs, those are not very useful. Those are not not useful. But that part of customization gets a little bit tougher because you the model layer, the the platform layer, needs to have domain knowledge in order to spit out things that matter. So the one simple example I I used to test this for every some of our competitors, every competitor I try, I used to ask them that, hey, give me like an earnings call summary for JP Morgan. If they ever spit it out margin revenue, I'm like, ah, this, this, this, this, this platform doesn't know anything. Like no one cares about revenue for uh JP Morgan or insurance company, right? Like that's just like a very obvious thing. Um, so that part of customization, I think, is a lot harder than what dashboard to service you. I think that part is a part, we've spent the last two and a half years building a ticker-level domain knowledge for that reason, is I think that is that is the the hardest piece of all of it is how do you connect different knowledges? How do you know which knowledge matter at what point? Uh, and how do you know when user asks the question, you need to pull that specific piece of knowledge in order to answer that question properly. So I think part of it is easy, part of it is I guess TBE.
SPEAKER_00Yeah, interesting. One question, uh one topic that comes up often with some of my investing clients is this like separation of synth of synthesis and judgment. Yes, right. And so the like I'll often hear uh someone say is like we're okay for AI to do the synthesis and the gathering, yeah, but when it comes to the judgment, that resides with the human, that resides with the analyst. Yeah. First of all, I don't think that's a very bind, that the that line is not as binary as as one would tend to believe. Yeah. But in your examples, you've said that, you know, okay, you know, you have you have these data, these discrepancies in data, and then you have an agent kind of flag the discrepancy. And so you're you're basically giving the agent some form of judgment.
SPEAKER_03Yeah.
SPEAKER_00And you've come back oftentimes to the role of domain expertise in this. But I'd be curious on your kind of what your view on that separation between like AI does synthesis, humans do judgment. And then where when the AI, when AI starts to make more judgment, like where does like where is it coming? Where is that training coming from?
SPEAKER_02Um, I think it's a great question. Um, so I think in a dashboard example, I actually, in most of the cases that we've built, I don't think we're actually relying on LLM to make judgments. So we basically will give a rubric flag me when these criteria hit. So it's still a human filter on what matters, whatnot. Obviously, there are disadvantages of that. It could be a new environment I haven't thought of, I didn't really flag that because of that. So that's that's I think, but most of that is gonna be a human uh rubric that LLM flags. Um, judgment is interesting. Um, I would say at the current state, um, given the given the fact that a few things. Number one is I don't believe models are truly reasoning or truly understanding. There is no true sense of understanding. It's like a mimic sense of understanding or mimic sense of reasoning. Given that there is no continuous learning, meaning um anything changes, it doesn't get included. And given that I don't think models are trained in the right way to solve financial um puzzle. And by the way, I think financial puzzle is also one of the hardest ones, right? I think like compared to AlphaGo, compared to poker, like this is much, much harder. Given all those limitations, I actually agree with the fact that judgment should rely on with human. So let me give you an example of how we we think about it. Um number one is very early on, we have instructed our platform to not give uh subjective things unless they're asked to do so. Um, so everything should be resided on checkable, verifiable, objective things. That's number one. Uh, number two is if you think about like the knowledge layer we built. The knowledge we built is purely facts, things that can be verified with facts. It can be synthesis, to your point, right? It can be like what is how credible is this management, right? Um, it it can give you a score, but the score is backable by true stats. Like every time, you know, 30% chance they actually missed the guidance they give. That's not a very good track record. Like that, that even though it's giving opinion, but it's a backable by track. So all of the things we focus on are these trackable factual things. Um, subjective stuff, we try to stay away from it. I I don't personally think LMM is at a point where it can do that yet. Um, we have some ideas on how it can get better. Um, we're we're we're playing with some stuff. It's kind of like how human analysts start from scratch, building tuitions, right? Uh it's a very much I form a hypothesis, I uh test it out, I look at the stock reaction, did I get it right? Did I go wrong? And I learned something new. Okay, I'm gonna try it again. I I think uh we can definitely build that, but I don't think the model is there yet. So I would generally agree they leave the judgment to human. And the other thing is I think human is just naturally better at forming patterns, especially new patterns. Um I I have a three-year-old. Uh, comparing AI with a three-year-old, you'll see like how much better a human is better at like just connecting the dots and actually like spin out a hypothesis where like, oh, actually that's a pretty reasonable versus AI is very much like I think too much backward looking, too much like pattern finding. Um, so I generally agree with that sentiment. Yeah, yeah.
SPEAKER_01What um one of the conversations we've had in the past, which has always been sort of a thorny issue for Quantz and sort of persist in AI, is sort of this issue of materiality. Even I was sort of going through SpaceX today, and like my AI agent was really convinced that that uh Starlink ARPU was like the core, like the the like the core thing that's gonna matter. I'm like, well, what about like the perceived terminal value of like space travel and you know AI? Yes. Um and like why do these systems still sort of have native like materiality uh uh sort of limitations in your mind?
SPEAKER_02That's the domain knowledge. Like, give me an example. If you ask that same question, what is most important for SpayX, but precede it by providing a lot of context. Uh, you know, this is where the market, this is what's going on in the market right now. These are the valuation of AI stock, these are all the all the twist that Elon said, these are all the things. If you give it enough context, it will know that you know, starting is probably not the most important thing. Um, but that context relies on domain knowledge. And that's the same reason why quans um historically struggle with that too, right? Kind of like uh I once have talked to a quant said they can mimic, they can make a uh automatic do the model line by line, but they still don't know which line item matters for a given quarter. So they don't they cannot bet every single line. They have to pick three, but they don't know which one. Um I think again, this is goes back to domain knowledge. And how do you form that domain knowledge? I think that's the hardest one. I think this is the part that human analyst kind of the advantage of human analysts is based on all the domain that they uh based on all the history they had, all the domain knowledge they have, they have a mental map of what's important and what's not important. That's that. The other thing is I think to be to be fair, like I think it's a hard thing. Uh, I remember towards the end of my time at BAM, uh, sometimes I would be like, hey, if you give me um the press release, earnings release the day before, even if you were to do that, I don't think my hit rate on how the stock is going to react is gonna be 100%. Like even humans have a hard time doing that. What's matters, what's factor in, right? So I think it's a hard part. So I think that's a domain knowledge issue. And it really depends on when you ask that question, what other relevant contexts that LLMN need to have a sense of what's material, what's not.
SPEAKER_01Yeah, yeah. One one other thorny issue that I see I I sort of run into in my build out is sort of the current primacy of Excel as a sort of the investor's workbench. And I'm sort of sort of learning that Excel is not very token efficient, and I sort of would historically push back on the AI people saying Excel isn't the great, isn't sort of the best quantitative dashboard, needs to move to move into Python, et cetera. Um how do we solve Excel, sort of ingress, egress, like the sort sort of the brain of the investment process today? Um how do we how do we solve that? And what's the future look like uh from a spreadsheet perspective?
SPEAKER_02Yeah, it's a great question. Um I'll I'll give you a few kind of bulletif thoughts on some of the things you mentioned. Um number one, I don't think every I don't think the functionality or the advantage of AI can be replicated, or sorry, of Excel can be replicated by dashboards or Python, whatever, completely. Um I think a lot of it can, like if you're doing time theory, you should do that in Python and Dashboard, you should not do Excel. But if you're doing a model where it's like interactive, you can change your one assumption, you want to see how everything else flows, that should stay in Excel. Like there's no better platform to do it in Excel. So that's number one. Do you think Excel is necessary? Um, number two is um I think the current methodology that most vendor out there is doing, which is let me treat Excel as a coding um uh uh tool, and uh let me just write code to write Excel has natural limitations. And there are a few limitations, some come from the LLM side, some come from the Excel side. LLM side, it's very easy. Um so let's say I want um uh a model to remember uh to update a model, right? Well, Excel to uh agent to update a model, then what AJ has to do is go read a document, it has to remember all of the numbers, and then it has to remember where to put that number, right? So there is like a remembering this, putting over here. Here becomes tricky. Language model is next to next tokenslash next word prediction. When you have a string of random numbers together, very likely it's gonna miss something because it it's no, there's no pattern. How do you how do you do next number token? I don't know if you ever run to so when we build, so sometimes when we do a even like a simple or like a temporary file that have a long string of random numbers of words, and sometimes like LLM will just miss that or mess up a letter, a number, and all of a sudden nothing works. But that's the same issue. You're doing numbers to to to to remember numbers to put into Excel. I think that's an issue. Then there's issues coming from limitation of Excel. So in my in my dream world, I think the process should be when earnings release hit, someone should update all of my models. I should be focusing on reading, doing other things, whatever. But I often have companies, 50, like five to 10 companies reporting at the same time. What when I have an Excel add-in, what am I supposed to do? Am I supposed to like let AI take control of my computer and open 10 Excels and try to update it? Like, I don't think that's very practical, right? So our vision is there should be an Excel exist. What one, it should be a better Excel, an AI native Excel that exists, um, that solved the first problem. Second thing is there should be a AI native Excel that exists that exists in the cloud so they can manual update stuff behind behind um like your move away from a computer, it can auto-update. So think about when company reports earnings, it can auto-update. When company is doing call, it should be able to update your forward number and tell you, hey, at the end of call, Brett, your estimate needs to move up 5% based on the consensus estimate. So it should be doing all of that in the cloud. Now you should be able to download it in Excel to play around with things. Um, but I think that is the ultimate form.
SPEAKER_01It was you that gave me sort of the idea of like almost a Google Sheets, right? Where it's like it could be it could be updated and updated in the cloud. Yeah. Shared. Yeah, shared. How does that architecture emerge in the Microsoft sort of Excel ecosystem?
SPEAKER_02I'm not sure. I I I I I definitely think like, I don't think our opinion is consensus. I think uh we looked at everyone, I everyone more or less defaulted to let's build on topic itself because I think one is already available. I'm pretty sure every client of Like, we need you to be a plugin. We need you to index out. Like everyone's client talking about it. But I do think that you're gonna naturally run into a limitation. It's kind of like again, AI. I do think that you should you need to think about rethink about everything, single tool. What what tool what used to work for human now needs to work for agent? How do you make it easier for agent to do it? So the way we did it, we have an AI native Excel spreadsheet. Agent doesn't have to remember anything. Agent just have to remember the formula. It construct one single formula for every single thing. Like every Excel power user knows, like the whole point of Excel is you never actually manually do it a lot. You just do it for one column or one row. You drag it down. Like that's a whole purpose, right? You don't have that advantage in using code to write Excel. You don't have that. You lose that power of dragability, you know. So that's that. There's another piece, I think. Um, I don't know. We start to hear rumblings of it, but I don't know how big or whatever catch. I think Excel um add-ins have more security risk. Um so Excel ultimate is a it's a programming language. Um, and so you are letting AI write formulas in your Excel. Some can be visible, you can write non-visible formulas to you. And it can do whatever it wants. Uh so I think it's um it's a little bit tricky. So we basically have taken a view that we let our AI native Excel should be uh sit in the in the in the cloud. Um, we aim for auto-updating without interfering with your normal workflow, and then we just alert you when things are done. Uh so that's kind of the approach we've taken.
SPEAKER_01Yeah. So maybe, maybe Ying, just sort of frame um, if you would, sort of the moment in time. Like what everyone wants to know, like what are what's everyone else doing? It's sort of like the Spider-Man meme, like pointing at everyone. Um what do you think sort of the super users are doing with these tools right now? What does the sort of right into the bell curve look like uh today? How has that changed? And sort of sort of forced you to pull out your your crystal ball and happy to hear any predictions on where you think this goes in the next three, six, nine months.
SPEAKER_02Yeah. Um I think I'll give you the progression we've seen, right? I think the very early AI users, the power usage is I think of the glorified search and summary tool, right? Uh, what did management say, how things change is basically search and summarization or search and synthesis. That's basically that. Um, I think we're starting to do more complex things. So think about multi-step uh kind of like uh it used to be like um, you know, what did management say? But that could be just one section in a preview. I think we started to see people stack that. I think that that's that's uh what Agent has made available or made up easy to do. Um we definitely have some power users who are using our platform in a very interesting way. Uh I was uh at another um, I was I was talking to a client of ours, and he was like, for some companies, I don't even look at my Excel anymore. I just have a buy, not even in our native Excel spreadsheet, it's just in the in the chat basis. Um we they they use all the data he needs, all the commentary he needs, you predict number, compare consensus, that's it. He was like, I don't need everything else. So that that was very interesting. Um and then the other the power use we have seen is automation. Um so it's not enough to be like, oh, I can do a earning, I can do a six-part earnings preview, right? Like uh, yes, you can do it, but what are you gonna do? Type in, do a preview for this ticker, do a preview of that ticker, or do it 50 times if you have 50 companies. Like that is doesn't make sense. So um what we start to see people using us is basically how do you auto-schedule things that I always have to run so that I don't have to do it anymore. I do it once, I set up this is my framework for earnings preview. Now go run it three weeks before any company in my coverage reports earnings automatically. I just come in, I just see what the result is, right? Um, you can even do like um uh this is a I have a thesis about this company. Every day I'm going to go fetch all the transcript that's relevant, all the news that's relevant, and tell me what are data points that's supportive of my thesis, what are data not supportive of my thesis. Every day, give me alert and let me know if there's a meaningful change, right? So that. Um so I think we're moving towards that. Um, my view on kind of the next leg is um truly instead of a chat bot, instead of a platform, you truly try to mimic a having an intern, having a junior person, having another extra set of hand that is always there, that um learns from your your your feedback. You know, say don't ever do this again, you will not actually not do it. It it learns your process, it learns your process, and then it makes customizable. And it's just always there. It's no longer a single chat, it's a one continuous chat. Um, so that's kind of what we, our latest feature protege, that's kind of that. I think that's the direction everyone should go. Is it's basically like a human. You talk to you, come in, say, hey, these are the five things I need to get done, it's doing in the background. And then maybe you have a quick question like, why is snow up so much? You can pull up the guidance, whatever, and it answers you. Uh, and then the pre-schedule one, like for example, we have something, hey, alert me when any name in my coverage changed revenue guidance. Um, and and today there's a Bernstein conference, a bunch of companies actually give you guidance. It flacks me. It's proactive, right? So I think that is the future of what investment work looks like. So I think one person with an AI can do a lot more coverage, can be um um, can can can wind their coverage, can do deeper analysis, can focus on things that actually matter, actually alpha generating. So I think that's the the future that we expect.
SPEAKER_01What what what an exciting time to uh to be an investor or an observer or a builder in this space because I sort of totally agree that we're at the exponential now in investing and uh we're sort of graduating from glorified search into truly sort of agentic process cloning. So thank you so much, Ying, for sharing uh your thoughts on the space, and we'll sort of bring you back on in the in a few months to sort of uh see if your prediction has sort of come to uh come to fruition. When you figure out that Google sort of Google Sheets equivalent in XL, let us know too, because that sounds pretty uh pretty interesting. Uh what uh how could people learn more about imply or get get in contact if they have uh follow-up questions?
SPEAKER_02Yeah, just visit www.imply.com or email me at ying at imply.com um and then happy to show you the product, uh kind of walk through what we have, um, and yeah, excited to uh to be here and talk about this.
SPEAKER_01Sounds good. Thank you so much.
SPEAKER_02Thank you.
SPEAKER_01Thanks.