Uncapped with Jack Altman
· Jack Altman, Max Junestrand, Chetan Puttagunta
[00:00:00] S0: Max, since you’re here, can you explain what that was? [00:00:00] S1: It’s worth going back to 2023 and 2024 when part of the paradigm was that you should train your own models and the general models aren’t great and fine-tuning is going to be really important. For two reasons, we were like, fuck that. One, fine-tuning doesn’t really seem to work, at least on the scale that we were operating. To train the new generational model, you had to put billions of dollars into it. Secondly, there was so much application that you had to build on top of the models to make them useful in your environment. Back then, just solving basic data, compliance, privacy, and great file uploads and great parsing and great chunking and all of these things, that was where the value was. [00:00:00] S2: There was another part of your experience, which was that you were actually embedded in a law firm. So you were studying the shape of what data law firms had in a way that was… Bill talks about this a lot — does an entrepreneur strike you as a learn-it-all? It was clear that the early Legora team—when we invested it was five people—was just trying to learn everything they could about how the legal profession worked and they didn’t have any bias towards it. The other thing Max said, which you should share with everyone, is that because they were embedded in a law firm in a windowless conference room in Stockholm— [00:00:00] S0: Sounds nice. [00:00:00] S2: Sounds great. They had a deeper understanding of the data model of a law firm in ways that most of us didn’t.
· Chetan Puttagunta, Max Junestrand, Jack Altman
[00:00:00] S2: Sounds great. They had a deeper understanding of the data model of a law firm in ways that most of us didn’t. [00:00:00] S1: Just to take it back even further, when we started, I offered to buy a lot of lawyers lunch on LinkedIn because I wanted to learn. I would literally cold write them and say, “Hey, I’d love to meet, I’d love to talk about IP law. I’ll offer to pay you your hourly fee and lunch.” They were all too nice to make me pay for lunch, and they’re often not even paying for it. But as Chetan put it, I think it allowed us to work with customers from the very beginning. The founding team at Legora were all engineers. The first lawyer didn’t join until nine months into the journey. [00:00:00] S0: When you had a lawyer join, had you already set the plan and the goal for the company? Was that done without experts? Was that important to do without experts? [00:00:00] S1: It’s actually funny. This is a bit of the Legora untold, first revealed here. The company formation was in 2020 and there were four co-founders. [00:00:00] S0: I didn’t know that. [00:00:00] S1: And I was not one of them. [00:00:00] S0: Didn’t know that either. [00:00:00] S1: They were working on this intersection between AI and law for three years with the early BERT models and even a Swedish-trained version called SweBERT. It was impossible to work with. Not only was it not very intelligent, it was also blatantly racist because it had been trained on Swedish forums. [00:00:00] S0: Some racist data there.
· Jack Altman, Max Junestrand
[00:00:00] S0: Some racist data there. [00:00:00] S1: When the LLMs like 3.5 came, that was when the moment shifted. We turned this into a company. Two of the co-founders left, I joined. We basically said we’re going to work in the intersection between AI and law. We don’t know what that product is, but we’re going to run like hell in this direction. Funny enough, the first lawyer who joined was a soon-to-be customer of ours. He was the CIO at one of the big firms in Sweden that we wanted to sell into. He had built an early version of a GPT plus the document management system, basically an LLM that could RAG into the existing precedent and data that the firm was using. He basically said, “These guys are going to run faster than me, and if you can’t beat them, I might as well join them.” That turned out to be a good decision. [00:00:00] S0: Are you surprised by how strongly the legal market has adopted AI? If I had thought in 2023 or 2024 about what’s going to really adopt quickly, I don’t know if I personally would have seen it coming that lawyers would be near the top of the list. You’ve invested in stuff before too, so I guess this is for both of you. Has it been a surprise over the last two years, the rate of adoption?
· Jack Altman, Max Junestrand
[00:00:00] S0: Are you surprised by how strongly the legal market has adopted AI? If I had thought in 2023 or 2024 about what’s going to really adopt quickly, I don’t know if I personally would have seen it coming that lawyers would be near the top of the list. You’ve invested in stuff before too, so I guess this is for both of you. Has it been a surprise over the last two years, the rate of adoption? [00:00:00] S1: Yes. It’s been vivid. But second and maybe more importantly, the law firm market is very interesting because it’s this perfect equilibrium with, frankly, pretty low differentiation. If you need to do a VC deal here in the Valley, you could go to any of the top five firms and you’re going to get roughly the same thing. If one of them starts leveraging Legora to offer a better service at a better price point faster, all of them have to adopt it. So the equilibrium shifts down and then everybody has to move. What happened in the law firm market was that as soon as one big firm in a market adopted Legora and went public with it, everybody else had to do the same. That’s not necessarily the same in the in-house legal sector. If one big bank has it, another big bank doesn’t necessarily need it. [00:00:00] S0: But was there something about the process of the way work got done, or the structure of it, that allowed Legora’s product to drive so much value so fast in a way that it did force that sort of prisoner’s dilemma?
· Jack Altman, Max Junestrand, Chetan Puttagunta
[00:00:00] S0: But was there something about the process of the way work got done, or the structure of it, that allowed Legora’s product to drive so much value so fast in a way that it did force that sort of prisoner’s dilemma? [00:00:00] S1: I just think the legal sector was so underserved with great software for such a long time that there was a lot of built-up problems that we could easily solve with LLMs, but that were really hard to solve pre-LLMs. [00:00:00] S2: I also think you guys had a great insight early on, which was that there was a deference in respect to the customers. That lawyers are really smart. They’re extremely well-educated, they’re tech-savvy. They’re not programmers, but they’re very tech-forward. They use the latest software, they use the latest devices. So they were all going to be playing with ChatGPT and Claude. If you showed up with a legal AI product, it had to be better than the foundation models. Otherwise they were just going to say, “Why are you deserving of my dollars?” [00:00:00] S1: Microsoft Copilot rolled out very quickly. Every law firm in the world is a Microsoft shop. Everybody works with Outlook, Microsoft Word, and where they store their documents basically. [00:00:00] S0: To the point of you have to be better than the models, if you had to break down, as a vertical AI application, what have been the things that have allowed you to be so much better than the models that it’s worth the incremental investment?
· Max Junestrand
[00:00:00] S1: In the beginning there were a lot of just foundational problems with the models. You had to guardrail them very hard to make them useful. You had to build citations, you had to build good RAG systems, you had to overcome context window problems. There were a lot of rate limit issues, so you had to juggle different models for different types of tasks. There were just so many incremental, basic things to solve. As time has progressed, our product has moved further away from what the foundation models are and much more into this enterprise-wide platform where we’re going to transact billions of dollars of legal work on the platform. We’ve moved from building a lot of the agent work ourselves and we let the models rip a little bit more, like OpenClaw or ClawdBot or whatever it’s called these days. With Opus 4.5 and Opus 4.6, there was an extraordinary difference in level of intelligence and instruction-following capability. So I see our job as: let’s provide the model the right environment and the right tools and skills to leverage, then let’s build a UI and an interface with the rest of the business so that they can all leverage it comfortably and with a lot of trust. I do think that the model capabilities improving so quickly makes us run faster, because we have to be three standard deviations ahead of any general capability. That’s a very good motivator.