Uncapped with Jack Altman
· Max Junestrand, Jack Altman
It’s human users and agent users, and every new feature that we build has to be able to cater to both.” Now we’re seeing more people basically use our agent that uses the tabular grid, or our agent who uses our word editing capabilities, than humans actually going and using those features at all. [00:00:00] S0: Chetan made a cool point to me recently. We were talking about how companies that are pre AI and companies that are fully AI native just have to be built differently in various ways. Because you didn’t build a pre-AI company, I think it gives you this unshackled mind. You’re not even trying to think about some past alternative. You’re just like, given what’s in front of me, what should a company look like? You talked about how having YC founders inside the company has been helpful and I’m sure there’s a lot there. I’m curious, what are the main tenets that you’ve observed? Because now you’ve probably hired a lot of people who did work and built companies pre-AI. What do you think are the main tenets, ideas, and cultural concepts that have been important to you, just to make it work in a fully AI-native world?
· Max Junestrand, Jack Altman
[00:00:00] S1: I think this idea that Chetan brought up—you have to be willing to kill the stuff that you’ve done in the past—is very important. In more traditional software, you have to build the foundations and then you build the stuff on top of it and keep building the stack. In that world, it was also very good to have a technical architecture where one feature would rely on the same microservices as other features. But the problem is, in AI, maybe that feature now needs to scale really quickly, and the cost of writing software is so low that it’s basically better to build your own stack for each thing. Now that we hire finance professionals or even lawyers internally to Legora—we just hired our first tax person—they come with a set of ideas. “Oh, this is how I used to do it in my old company.” Everybody’s forced to relearn, and also question what their value is on top of the general model capabilities, which was very painful. [00:00:00] S0: Totally. Bret Taylor talked about this on this podcast too. Basically people are going to build something and six months later we might just kill that thing and everybody needs to be comfortable with that. Historically, that would be a lot of painful internal conversations. Do you have to change? Is that a different culture for people?
· Max Junestrand
[00:00:00] S1: I think it’s a different culture completely. The culture is that you don’t maximize for your function, you maximize for the company always. I’m very upfront with every exec who joins Legora that, in a way, you’re joining with an expiration date. You have to continuously prove that you scale out of that, because the company is scaling so exponentially. I don’t know if it was Mark Zuckerberg or somebody who talked about hiring people with high y-slopes and not high y-intercepts. I think about that a lot, mostly because I’ve had to do that. I did not join or start Legora with a lot of experience, but I’ve proven that at every new point in time I’ve scaled with the business. Other people at Legora need to do the same. I think that goes for every function. An engineering team that’s shipping the amount that we do, previously had to be 500 people, and now we can get away with being 50. There’s even a question of whether we need to be more than a 100 engineers, or is the bottleneck really knowing what to build and building it the right way, and designing an experience that works for hundreds of thousands of people that we now have on the platform. The paradigm is shifting all the time. What’s nice about our work is that engineering is a roadmap of what’s going to happen in other industries too. The general models have come the furthest in coding, but also those organizations are very quick to adopt and shift. Engineering orgs are today looking slightly different, and I think we can expect the same in legal organizations.
· Chetan Puttagunta, Max Junestrand, Jack Altman
[00:00:00] S2: Two things you brought up that it’d be great if you could dive into. One is that Legora doesn’t really have a long-term roadmap. You guys react and build today. When you first got started, you had this nearly weekly cadence. That’s how long you would roadmap to. These days it feels like you almost roadmap on a daily cadence. Things change tomorrow. You wake up and it’s like, we have to do something different. Talk about that lack of roadmap. Also the other thing that you’ve invested heavily in is just understanding model capability and the proprietary eval infrastructure you’ve built. You’ve had these conversations with the foundation model companies about how you’re able to identify latent model capabilities that they themselves are not aware of. [00:00:00] S1: On roadmap, way back, every new model just unlocked new things. When we got early access to GPT 4.5, you just realized that holy shit, now it can finally draft an end-to-end thing and we don’t need all these harnesses and things around it. That’s amazing. Let’s unleash it in a way that works. [00:00:00] S2: By the way, to do that, you need a low-ego organization. Because you build all this IP and all this software— [00:00:00] S1: And you just toss it. [00:00:00] S2: And you’re like, okay, now the model can do it. Delete it all. [00:00:00] S0: “You worked really hard for six months. We’re deleting everything.” [00:00:00] S2: It’s incredible. [00:00:00] S1: But I think a lot of the things that we have built, we know that we’re going to delete someday.
· Max Junestrand, Jack Altman
[00:00:00] S1: But I think a lot of the things that we have built, we know that we’re going to delete someday. [00:00:00] S0: I guess you need people to opt into that at the front end for that culture to really work. [00:00:00] S1: We’ve also talked about it like this. If we were here today and we started building for the future that’s way over there, that’s too far out. Our customers are not going to adopt that. They don’t understand it yet. So we need to take them on the journey. We need to take them on the path of being successful. Every iteration cycle now is shorter. Back in 2023, 2024, I think it was slightly longer. You’d have a quarter or two quarters because the models weren’t moving that fast. Every upgrade was pretty incremental. But now it flipped. Opus 4.6 flipped in capabilities. So now we have to revisit a lot of the things that we built. [00:00:00] S0: Do you know what the next flip you’re waiting for is? Is there a thing? [00:00:00] S1: It was funny, I was at the customer advisory board at Anthropic yesterday. I’m wearing my Dario shirt here. [00:00:00] S0: You look like Dario.
· Max Junestrand
[00:00:00] S1: Thank you. Most of that conversation was about how the models are now intelligent enough where they’re no longer the bottleneck. The bottleneck is all of the software around putting the models in an environment where they can execute and do work, and humans can review that work in a trustworthy way. They’re seeing that across basically every single vertical and every single company. I don’t really think that we’re waiting for new model capabilities anymore. There’s nice things to have. It’s nice to have better context windows, it allows us to do less garbage and context management. When you overflow the context in memory and so on, you have to deal with it to refresh it. So there’s nice-to-haves, but we’re at a point now where we just have so much building in front of us in terms of bringing the model capabilities into our world. That’s where all of our focus is. On discovering what the models can do, we thought very early on that evals were going to be important, both building up an exercise of building new evals, but also building out evals for all the use cases that we want to cover. Because in the beginning it was a lot of, “how good is Sonnet? How good is Gemini? How good is GPT?” We had to test them on the different evals. A lot of our customers actually contributed to this. They would give us manual tasks that they used to do, and they’d tell us, “Here’s the evals, and we’re going to call you when we can get to 100% on these evals.” I actually remember. It was a funds-related use case, an LPA key-term review report that a Danish law firm was spending three days on. Basically an associate would spend three days putting together that report. In summer of 2024, we had 60% accuracy on that task. By the end of that summer, we had 100% accuracy. Once you get to 100% accuracy, that task is done, it’s over. I’ve adopted this mentality internally that if AI can do something, it will do it. With our product, we think a lot about solving legal tasks end to end. Once a task is conquered, it’s done. We just strike it out and we’re on this path of solving more and more complex tasks. You start with NDAs, but at some point you get to full-on share purchase agreements, which are very complex. But we’re going to get there.