Y Combinator Startup Podcast
00:28:12 · Max Junestrand, Gustaf Alstromer
[00:28:12] S0: So there's actually a story of a Swedish VC who was very bullish on doing the pre seed of Legora. Yeah. But they ended up not doing it because we did not have any lawyers in the team. And I met them at the airport. I met one of the partners from there at the airport quite recently, and they were very very bitter. And Legora or Legora turned out to be the lesson for them that perhaps domain expertise is not what you need to succeed. I think that you need a willingness to learn about the market, which we did very very early by, you know, exposing ourselves and spending lots of time with lawyers, like as much as we possibly could. Yeah. But I don't think you need domain expertise. Probably, you know, dependent on on the market that you're going after, like maybe if you're building in quantum computing or in fusion, that's a good idea. But legal was easy enough for us to pick up as we went along. [00:29:14] S1: I spoke to some founders yesterday, and we spoke about sort of like the next categories where the LLMs are not sufficiently good yet, but they will be. And it made me think of what you just said when you were playing around with Bard in What the early would you what give you the confidence to build a legal AI product on a model that wasn't that wasn't good enough? It was good, but not good enough. And there's many other founders probably in this audience who are thinking about the same problem but in a different domain. Maybe like, I don't know, mechanical design or some some some other part of... Somewhere where the models are just not good enough.
00:29:50 · Max Junestrand
[00:29:50] S0: So I I think the sort of tribal knowledge when we got started was that you have to fine tune a model. Right? In 2023, 2022, that was really like the name of the game. Yeah. I even remember, I think it was Bloomberg who spent, you know, millions and millions and millions of dollars billing like a Bloomberg Law model. Our view was partly because we didn't have enough money, and partly because we truly believe that to be right. You know, the models will keep improving. Right? Like, I'm sure Sam is going to come up on stage and say that later today. Yes. And so by just betting that the models will keep improving and what you should be focused on is how can you deliver that... The value that the models generate to your market, that was the crux for us. Yeah. Even with the first generation of ChatGPT, that was unusable in law. Partly because the model wasn't good enough, but partly because it actually didn't support private conversations, European data hosting. I remember the first sales I ever did, we basically said, we are like ChatGPT, but we are compliant in Europe. Like that was the pitch. That was not a very strong pitch, and know, ChatGPT would figure out how to do that later on. But at the time it was good enough, and so I think we recognized that you don't have to build for the world over here, like you need to build for the world today, and maybe like one step ahead. Yeah. And it's kind of like this ladder where you should maximize the opportunity now, make sure that you get really good relationships with your customers, and so that you can also have that conversation with them. Because if everybody's betting on this exponential uplift in model capability, the Legora platform will continue delivering more and more value. I think that we have built 1% of all the software that we will build over the company's history, and probably the the product will look very different from what it what it does today a few years down the line.
00:32:00 · Gustaf Alstromer, Max Junestrand
[00:32:00] S1: Yeah. A lot of news on models this week or last two weeks. [00:32:05] S0: It's gonna be like a great episode for whoever makes the documentary. [00:32:08] S1: I'm I'm hoping in some ways we're making a documentary right now about everything that's going on. I'm not sure if someone is, but I hope they are. Open Weight Small is very much the topic right now. And when we spoke on stage a couple months ago, you talked about that you're not attached to one specific model. You can work with many different models. Have you updated your view based on this week? But, like, how do you think about... So, like, I'm sure there was a moment, like, a year and a half ago when it seemed like it was just one or two models that were really sufficient to do... To power Legora, and now there seems like a lot more.
00:32:46 · Max Junestrand
[00:32:46] S0: Yes. Which I think will be great. Yeah. And and, I mean, you know, clearly there will be applications of different models in different use cases, and so if you take customer support, you're optimizing for, you know, speed of turnaround, latency to the customer, and cost on solving a particular ticket. I think Fin used to charge like $1 per solved customer ticket. In law, you actually want the most amount of intelligence quite often, because the fraction of token spend or software spend compared to like, you know, human expertise applied to the problem is really tiny. If you're solving complex litigation, you're gonna wanna throw the most brain that you can on that problem. And so we have really sort of two different camps, like one group of our users are asking for the ability to loop really complex models just to work on problems for a long period of time, like racking up big LLM costs, and another group of our users are saying, can we get access to open source? That is gonna be cheaper on our consumption. And I think the answer is somewhere in between. I think ultimately, one of the core IPs and muscles that I encourage as many of you as possible to build is the ability to eval new models and to eval new use cases, because that is the superpower that then allows you to route things effectively. We did a lot of that early on because we hired a lot of lawyers into Legora, and we made part of their job customer facing, but part of their job was also to build out use cases that we would throw our evals on. And the frontier of like, you know, cost intelligence and where that sort of frontier lies, you know, yet to be seen. I think there's also a compute problem where when we lock in big contracts with the labs and with our compute providers, that is actually a huge thing. We cannot have our system go down when, you know, a number of percentage points of all the lawyers in the world are working with it, and their clients rely on it.
00:35:14 · Gustaf Alstromer, Max Junestrand
[00:35:14] S1: Are there specific benchmarks you look at for how... Yeah. I'm sure you are, Ben, basically. [00:35:18] S0: Well, we we look at the Legora bench. So we actually announced and released that on Friday last And it has been that has been work for for the past three years, and we've kept that internal up until Friday. And the reason for that was it was basically OpenAI or or the Anthropic models. Mhmm. And so it wasn't like that interesting to publish benchmarks on. Yep. And we would only make those two models available in our system and the user could pick. Now, the smorgasbord of available options is a lot larger. And actually we found, to our surprise, that SpaceX and Grok was one of the best performance models, especially given the cost Yep. On our benchmarks. And we don't even have a a... Have them on our data processing agreement. And so we actually can't yet offer them to our customers, but now we will start to. [00:36:11] S1: And do you see your customers having having strong opinions about which models they wanna use? Or is it sort of like they just want the best, whatever is [00:36:18] S0: the best? Some of them have an... They have a strong... Especially like banks, you know, big law firms have a... They don't want Chinese models.
00:36:18 · Max Junestrand, Gustaf Alstromer
[00:36:18] S0: the best? Some of them have an... They have a strong... Especially like banks, you know, big law firms have a... They don't want Chinese models. [00:36:27] S1: Yeah. Makes sense. So I want to go back to just a little bit earlier, like when you decided you want to build a startup. A lot of people in the audience here. I've had two dinners in the last two days with attendees at Startup School. We talked about when should you start a startup? People are working at DeepMind, people are working at internships or jobs right now. Often, you're one or two years into it. What would you say was the trigger for you? To say, I'm just gonna do this and not do McKinsey or the other options that you So [00:36:59] S0: my my dad had two companies during the dot com boom, both of which went to zero after raising too much venture capital. Maybe some [00:37:11] S1: Has he been a good mentor for you? [00:37:13] S0: Yes. Yeah? Absolutely. [00:37:17] S1: How And... Often do you speak about... [00:37:18] S0: Oh, it's probably every day or every other day. Balagora. It's great. Yeah. About lots of stuff. I used to... Now, I'm, like, too tired to walk home often, so I have to, like, take the train or like take a cab, but I used to walk home [00:37:32] S1: Mhmm. [00:37:32] S0: Then I would call him, and that would be a lot of fun. And sometimes we call him senior adviser. We have actually... One of our biggest customer pitches, we we were stretched for resources. So he came in and did the PowerPoint for for that client. [00:37:49] S1: So he knew everything at this point.