Y Combinator Startup Podcast
00:03:24 · Chaz Englander, Gustaf Alstromer
[00:03:24] S0: Yeah, so we sold our first two companies. And after we sold our second company, we did a bunch of investing ourselves, which I think on the whole, we were pretty bad at. But we became very interested in automating as much of those processes as you possibly could, right? And so it started whereby we literally were, when we would receive an opportunity via email, and having sold a company, you tend to receive a bunch of opportunities every day, it would then enter into this agencik system that, to be honest, we were just building for fun and produced a one pager for us. But the interesting thing about the one pager is, whilst we received some information via email, what was in the one page, probably 90% of that information was unrelated to what we would have received via email. In other words, let's say we got an opportunity to invest in a startup, This thing will go off and look at their LinkedIn and their background and then look at comparable companies on Crunchbase or S and P and so on, and kind of produce this one pager that you would probably go and do as a human as your first port of call. Little things like if it was a consumer company, would go and look at review websites, just as an example. Yeah, we thought it was pretty cool, then people became interested in it. And then we kind of agreed that we were quite bad investors and okay at building stuff. And that's how it started. [00:04:46] S2: And that then turned into you selling this product or a new version of that product to the top financial firms in the world, basically.
00:04:46 · Gustaf Alstromer, Chaz Englander
[00:04:46] S2: And that then turned into you selling this product or a new version of that product to the top financial firms in the world, basically. [00:04:53] S0: Yeah. So we can say now that we're about 10% of the largest private equity firms and investment banks in the world use our product. We have other customers, asset managers, sovereign wealth funds, a bunch of venture firms. But yeah, it's been pretty exciting. [00:05:11] S2: And how were they doing their work before you started using ModelML? [00:05:15] S0: So as crazy as it sounds now, right, so let me use an example whereby, let's say you're tracking a public company, right? And every quarter, every time there's a release, maybe as part of your job as an analyst or associate, you'll go to the release, the filings, and you'll put together what's called an earnings summary. So think about it like a beautifully present slide, effectively a single slide, with a bunch of information. But all that information is coming from the same piece of information. So for example, you might always get certain numbers from the filing itself. You might get other numbers like consensus or similar from a fact set and so on, and they're put into these slides. And these things take ages. We're talking probably like days [00:05:57] S2: to a
00:05:57 · Gustaf Alstromer, Chaz Englander
[00:05:57] S2: to a [00:05:58] S0: company, just to pull together these slides, right? Because everything has to be linked back to source and checked and so on. And what you can do in Model ML is to think of Model ML as being used as the building blocks for this. So you have an Excel spreadsheet that's already connected into these data sources, and you might want to export that into designs, right? And so rather than having to go and gather that information, once that release happens, this one pager effect, it was actually three pages, a cover page, your main page, a legal page, just appears in your SharePoint or Google Drive, right? And it's kind of 90%, 95% of the way there. Wow. And we think, in some cases, more accurate because actually, trawling through these filings as a human, because you can also pull from multiple data sets for the same figure, yeah, can be more accurate. [00:06:43] S2: I'm assuming that you built something in the beginning, and there was a model that could do some amount of this work. Did YC how long ago? [00:06:51] S0: A year ago. [00:06:52] S2: A year ago. So the models have probably progressed a lot in that year. What were things the models could do a year ago that were impressive, and what can they do today, and what are they gonna be able to do in the future?
00:06:52 · Gustaf Alstromer, Chaz Englander
[00:06:52] S2: A year ago. So the models have probably progressed a lot in that year. What were things the models could do a year ago that were impressive, and what can they do today, and what are they gonna be able to do in the future? [00:07:01] S0: The whole industry or the industry we're in right now, really from January, is going straight line up. And I think the main difference that we've seen this year versus last year is last year was a year of testing, right? So everyone wanted to make sure they had some sort of exposure, as in they were trialing something or looking at something, but it was still testing, right? And us having ran consumer companies, we don't really care about revenue figures. We've always been obsessed with usage and retention. [00:07:30] S2: Right, right. That's good. That's good. The most valuable thing is want bring into B2B. Like, B2B people don't always know the most important stuff.
00:07:30 · Gustaf Alstromer, Chaz Englander, Arnie Englander
[00:07:30] S2: Right, right. That's good. That's good. The most valuable thing is want bring into B2B. Like, B2B people don't always know the most important stuff. [00:07:37] S0: Percent, which we think is madness. But this year, everyone kind of... The whole world went from testing to using. And there was a fundamental shift. Think these eugenics systems, there were small things like the improvement in things like function calling with some of the newer models, but everything has become considerably better by itself. And that's the interesting thing, right? It's like, I think, even if we do nothing, theoretically, our product will improve, which is a pretty interesting world to be right? And so when you're constantly working on something and you're pushing the boundaries, month to month what's possible just wasn't possible the previous month. [00:08:12] S1: I also think the vision models. I mean, when they first started coming out and improving, we were sitting next to each other. We were going crazy, weren't we? [00:08:21] S0: It was like the stuff that we [00:08:23] S1: could do with the vision models, if you think analysing files as an example, you know, OCR, okay, was amazing. You combine that with vision and its ability to read information from tables and charts, it really just changed the game.
00:08:23 · Arnie Englander, Chaz Englander, Gustaf Alstromer
[00:08:23] S1: could do with the vision models, if you think analysing files as an example, you know, OCR, okay, was amazing. You combine that with vision and its ability to read information from tables and charts, it really just changed the game. [00:08:37] S0: Yeah. And also, I think it's a misconception. So I think the industry still thinks that AI, or a large proportion of it, is kind of where it was even six months or so ago. We're seeing today is if you look at some of these data providers, they've got humans reading information from, say, public filings and putting that in structured format. The bulk of the work that we're doing and what we're seeing is models are already more accurate than humans in those sorts of tasks. And so I think that will take a little bit of time in terms of confidence, but I think some of the more lower level, more kind of data gathering and presenting type tasks are fully already being automated at the top firms. [00:09:17] S2: This is interesting. I'll tell you, at YC, the evolution went through at the same time. So two years ago, we would say, oh, you can't really sell the software to investment banks or private equity funds. They don't really buy software. Or if they do, they may buy one piece of software every ten years. The same was true for lawyers. But something happened last year. They all got very AI curious. [00:09:35] S0: They were [00:09:36] S2: curious about AI. They were like, let's do pilots to play with the software and try it out. And you were saying that this is the year when that turns into actual contracts.