Super Intelligence | Dot’s Launch | Dario’s PR Team | Legora & H-Company CEO Join

Yeah, it took us 18 months to grow from 1 to 100 and then 6 months to go from 100 to 200 and now we need to break the record for 200 to 300 as well.

Like every quarter I feel like it's a new company. So I need to re-qualify for the job, but so does every executive in the company, right? So you kind of need to step up to the challenge. I think for me, at the start of this year, it was about being in the arena for every single marketing release, every single deal, every single feature that we're shipping. Now we've gotten to the stage where we're over a thousand people. And so setting the direction and sort of being in the field to impact that direction is more important than being involved in every individual deal personally. Now as we're scaling this quickly, what we're finding is that the most important job is actually to maintain the culture and maintain the velocity and the thing that got us here even as we're getting bigger. And so I'm spending 200 days on the plane. I'm in every single office. We are just doing everything faster than what has historically been possible.

I'm 27. I feel like I have a few more years on the road. And I think it's important that the whole company sees that we're willing to do whatever it takes to win. And for us, I mean, I started the company in Sweden together with August and Sig. And when you start a company in Sweden, you need to be ready to get on the plane a lot. Now, it's unfortunate that SAS got sold to Copenhagen and I need to always make my way via Denmark in order to get somewhere, but it's been fun.

Yeah, I pay taxes in Sweden.

This is completely true.

So, let me explain how I did it. Well, in Sweden, you actually get paid to go to university. So it's a really great system and it's even better. That also means that it's quite complicated to take up two seats. And I enrolled in the technical university of KTH and I did my degree in computer science. Now after about a year, that didn't really feel like a full-time job. So I got a full-time job as an engineer and then that didn't really feel like a full-time job. So I also enrolled in the business and economics university called Stockholm School of Economics. Now, in order to enroll into both at the same time, I called KTH and said, 'Yes, I want to enroll in this year's semester.' And then I called them and said, 'Never mind. I regret it.' Called the other school, 'Hey, I want to get in.' And then I called KTH back and said, 'Hey, I messed up. Can you please let me back in?' And then I was admitted into both without them knowing, you know?

And actually two of my best friends, three of my best friends all did both of them at the same time and so did my girlfriend. So we've become our own little mafia. And it worked out exceptionally well during COVID because there were no real presence requirements and so a couple of times the exams would clash and that was a problem because you couldn't physically get on the bicycle and go to both tests at the same time. But if I were doing both of them digitally, it was no problem. So I would just bring out both tests, a big chunk of paper. I would get two phones. So one filming me for the one at university and the other one filming me for the other. And then I'd just, you know, half the time for the first test, half the time for the second test.

Pretty well. I graduated one of the degrees and I still have my master's thesis to write at the other. So when I talk to my, you know, maybe potential future in-laws, I finished my degree. When I speak to investors, I'm a dropout.

Yeah. Well, they came at it from a different angle. I think it was like blood and Legora CEO, but I want to know, like, is this still true today and what does that actually mean for someone on the outside looking in?

So, we have taken that misprint or misquoted article and we have made it into a cultural phenomenon inside of the company. And so Q4 in Legora right now is what we call Operation Blue Smoke, which in Swedish means Operation Taste of Blood. And that obviously doesn't mean that we're going around drinking red juices. We are locked in. We are recognizing the opportunity to really seize the category that we're playing in and off the back of that build a generational technology company having been founded in Europe. And that means leaning into an insane extent. We've given everybody additional waivers to take care of stuff at home, like make sure you get home cleaning, like just get ready to lock in for Q4 because this is the time to do it. And we're feeling that in the end of Q3 right now. We're expecting to do over 10 million of ARR in the next 24 hours and we're just so in it. And the opportunity space for our vertical is so large. There are so many software opportunities and tools that haven't been built because legal has been such an undigitized market until LLMs and until we could really work with language.

Yeah. Yeah. Yeah. Okay. There you go. Tell me about why it's probably time.

Let's go back to 2023. We founded the company and the current working thesis at the time was you need to build your own large language model and you need to fine-tune it for every use case that you wanted to work on. And you had a Bloomberg spend, you know, millions and millions of dollars building a Bloomberg law model and then you got GPT-4 and it just smashed it. And we continued from the very beginning to say, well, look, if these labs are going to invest billions and billions of dollars into R&D, we want to ride that tide. We don't want to be building sandcastles. We want to build boats that increase as the tide rises. And in the beginning, it was really easy. We had to route between one model, GPT-3.5. And then more models got added to the bench. You got the Anthropic models, you got the Gemini models, and now you actually have the Meta and the Grok models, which are pretty good. And you're starting to get open source that is really working as well. And so the smorgasbord, to use a Swedish word, of models available to us is growing exponentially fast. And what we're seeing at the frontier is different models are good at different tasks. And so in order to optimize across all the tasks that can be contained around legal work, you need this very jagged frontier. So you need all the different models. And we've invested so much IP in our model routing layer and in our own benchmarking and eval data sets and processes that what we're seeing is let's ride Moore's law. Let's get better models from the labs all the time. Let's apply them in our ecosystem and then let's ride the cost benefits that we get that we can then propel back to our customers. And there are opportunities to fine-tune models for very specialized tasks. If you're doing post-training on, for instance, citation extraction of American case law, that's something we've done there. It makes a lot of sense to post-train a model because you're going to run it on 400 million cases and it needs to be very, very cheap. But the idea of fine-tuning a model or post-training a model for general legal knowledge, it just doesn't work. It's proven back in '24 that it didn't work and so trying to do it now again, I just find to be a very strange strategy.

I think it was easily the highest, wasn't it?

Okay. Yeah. After coding and...

So it's a very good technology market fit. And what I think is hard to appreciate about law sometimes is just how massive of an enterprise these big law firms are. I mean, Kirkland & Ellis is turning around $10 billion a year and you don't know how big they are because they're all privately held. There's no public law firms and so they're all owned in these partnership models and you don't see how insanely big and profitable they are. And at the top of that game, it is very competitive. I mean, the law firms are competing for the same work. And the clients, the GCs, are starting to really put a lot of pressure on how are you leveraging AI and technology to deliver me faster, better, cheaper services. And so all the law firms are in a race to adopt the technology, not only from a usage perspective but from really operationalizing it inside of their organizations, which means rethinking their business models, how they price, because a billable hour model when you can do the same work in one-tenth the time doesn't really make sense. You know, today they rely on this very leveraged model of a lot of associates per partner, but the number of associates per partner is actually shrinking. However, at the same time, we're seeing our customers actually grow. Like, they want to hire more senior lawyers. They want to hire more senior associates because the amount of work, the legal service demand currently outweighs the existing supply by so much. And once the law firm starts to adopt it, we start to see the enterprises go, 'Huh, that looks pretty cool. Why don't I get the software so I can do part of that myself?' And so, looking at the new growth that we're having, almost half of that is enterprise and legal departments. As you probably saw, I was the only CEO that didn't get framed by Marc in Dreamforce. And, you know, most founders go to Salesforce or Dreamforce to get sold on Salesforce, but we did the other way around. And now their legal department runs on Legora and so does Palo Alto Networks. So does Barclays, right? And they are spending so many dollars on both external fees but also on their internal legal operations because these are enormous enterprises and so we're seeing a big growth there.

So the reality is, back to the point of the biggest clients, think the largest private equity firms in the world, they are spending hundreds of millions if not billions of dollars on legal services and they are demanding from their law firms to really be at the frontier of this development.

A lot of what's going on right now is marketing. It's very good marketing. It's basically saying, okay, we hear you. We're going to develop some secret sauce here in-house and we're going to leverage that to give you differentiated services. Now the reality and what I really think is the technology that we have today is so good and the gap that you see is actually operationalizing that technology because maybe let's say the hypothetical model ceiling is over here. The products today are here. Like we haven't yet figured out exactly how to squeeze out the most juice, but then most organizations and people are down here. So by just getting up to where the products actually are and utilizing them to the max and then letting us continuously increase the products, like that's how you move. Going and saying, 'Hey, we're going to spend, you know, millions of dollars with Palantir and, you know, do our own model or some jazz like this.' It just doesn't really make a lot of sense. It makes for a very good headline.

It's a great marketing, but also building your own LLM model as a law firm basically means that you are embedding your know-how into a static model weight. That only really works in a particular harness. And so there's an enormous lock-in. If you instead choose to invest in building an auditable, editable context layer that you can put any model on top of, then you get to ride the wave of Moore's law. The R&D from the big labs and the increase of capabilities of our products can be applied on that context layer and you're sort of model and vendor agnostic, which is much more important for these legal organizations.

I mean, some would maybe call us a disruptor. I like to use the word transform rather than disrupting. Because what's happening in the legal industry is really a transformation and it's a change where the limiter to scale used to be number of people. And now you sort of have intelligence on tap and you just have to turn on the tap and outflows intelligence. And the best lawyers in the world are now applying that intelligence in completely new and novel ways that is growing their businesses, growing their impact in the world. And what's important to me is that they're part of that transformation. We don't want legal organizations to feel disrupted. We want them to feel like they're going on a journey, a transformation journey, and that we're holding each other's hands. And like I sometimes like to tell our customers, we don't want to be your better half. You know, hopefully you have a better half somewhere else, but we want to be each other's jetpacks. And if we are each other's jetpacks, that means we're going to make your business better and you're going to help me make my product and my software better. And together we're going to, you know, skyrocket into the sunset.

Well, I think, you know, in the SaaS era there was a lot of incremental innovation and in the legal industry there was a lot of like nice-to-have. What's happened with AI is that it reshapes the competitive arena to the extent where let's say your firm has Legora and we're competing on a deal. I don't have Legora and you're going to deliver faster, better, and higher quality and lower cost service than I am. And suddenly I'm like, you know, like how do I compete with this guy? We actually had a great example. One of the firms that we work with here in the UK is Linklaters and they're like super forward-leaning and they've been very quick to adopt all of the new stuff that we're bringing out, which I think is an impressive feat for an organization of that size. And we were talking to a couple of folks and they were working on a deal with an American counterparty who wasn't allowed to work with AI. And the associates and senior associates at Linklaters were basically throwing over and throwing back markup in like 20-minute intervals. And the poor guy in the US was just like, you know, like getting hit and hit and he just like couldn't respond. I'm sure that was like a very painful all-nighter there.

Yeah. Well, for me, Legora's job is to bring the greatest of AI, whether that's frontier, open source, or elsewhere, safely into our industry. And that's important to me because we are literally hosting some of the world's biggest secrets and most important cases. And security for us, and I've always told the team this, is priority number zero. That comes before a new feature or a velocity or accuracy improvement. The marketing from the labs and the general public right now I'm finding a little bit difficult to decipher. It's very hard to differentiate what's hype and what's truth. There's no secret that the models are improving at a very rapid pace, maybe even at an exponential pace right now. And so I think that it's sober to have very strong guardrails and evaluation mechanisms for how strong these models are. And I do think that if you put out a model into the world that or into...

Production that does crazy stuff like in some way you are accountable for that. Definitely, in the same way that a legal professional is accountable for the work that they produce with the help of Legora. Yeah. And so I think there's a lot of, you know, growing up that needs to happen, and I think that leaders of the AI and tech sector can do an even better job talking about this sensibly and publicly.

Oh, totally.

So it's fun. We've actually seen a trend where a lot of organizations who got on alternative legal AI solutions back in '23 and '24, when we were still called Leia and started scaling, they have, as they've scaled with that provider, received broad but really shallow adoption. Because when you started to work with these tools back in '23 and '24, they weren't that powerful, and they would make mistakes. And in the legal sector, you only get one shot to do it right. If you deliver a piece of work product to a partner or to a GC and it's incorrect, that trust is lost. So we've seen a huge movement of churn actually from other legal AI solutions over to Legora, where they only got that sort of first level of adoption. So I actually think that to some extent there's been something to be gained by waiting a little bit. And now I find sort of post-OPOS 4.5, like after Christmas, it was such a step change in capability that we've sort of had that moment where it's more productive to work with the AI than to not work with AI.

That's a good start. That's a great... It's a great, and it's working well for us because we are investing in being not only the software vendor but really be the change partner and the, not implementation even, but like the business partner to these organizations as they go through quite a lot of change.

So what I think will play out is I think it will be very hard to drive a... I think the frontier models will be a fantastic industry and a pretty shitty business. Yeah. It's like the car industry. It's an amazing industry, but it's kind of a shitty business and like very few organizations actually make money. Yeah. I think the margin there really gets competed away, right? It's getting competed away really quickly. And so I think a lot of the end value will be created in the application layer. And, you know, that's not to say if we had a frontier European model that was as awesome as the other models, I'd love to use it. But today that's not an option. And I'm sure our customers would insist on using it here as well. But so in the absence of that, you sort of go, well, should we stick around and wait? No. Like, you got to go. Yeah. And it's become such a competitive imperative that you nail the implementation of AI in all your different functions, legal being one of them, that organizations can't afford to wait.

And we'll see. Like I think all of those companies are going to be extraordinary companies and they have, you know, collected second to Legora the most talented group of people in the world. And I do think that there's a real market for frontier intelligence. Yeah. If I give you an example from our business, bet the company litigation is you're very price insensitive. And so if you can spin up multiple frontier models to go and think about that problem for a long time and come up with a better case strategy that's going to make a better settlement or make you even win the case, of course you're going to do that. But to your earlier analogy, you don't need to take the Ferrari to the grocery store and so you don't need the most powerful frontier model for every type of task, especially not marking up an NDA. Yeah. I think what I'm still excited for is as you start to implement these systems broader and broader and as we also start to work with law schools and we start to educate new lawyers entering the market about you don't have to spend your first two years formatting documents and searching for change of control clauses. You're actually going to be doing interesting stuff. Yeah. That to me is like the most exciting update in the field.

Oh, well, if you have a working formula, don't change it. But we absolutely need to scale it, right? The most important thing to me is that, you know, when I go to a university and do a keynote or something and you get a line of 300 people who want to come up and say hi and they all say, Max, Legora is my dream place to work at after I graduate. How can I get a job? Like that for me is what I want to maintain. And I think you only do that by as you scale, you just keep adding A-players.

Well, it started as a pretty funny concept actually from an agency we work with who came with an idea that was what if we take legal movies and we sort of make like a, you know, before Legora and then after Legora and instead of the legal movie taking, you know, two and a half hours it would have just taken like a minute and a half. And you take all these like very famous scenes with lawyers and you like recreate them and the type of impact Legora would have had on them. And that got us thinking about movies and actors. And somebody went, well, what if we did AI-powered law and we got Jude Law to do it? Wouldn't that be like amazing? And we, yeah, that's actually really funny. Let's like see what he says. And we reached out and he said, and his agent said, absolutely no. Like no, no chance. And here is where I have learned that nagging is an underappreciated form of negotiation. And so we were very persistent. We kept coming back month after month and we said, hey, now we're doing a little bit better. We've just done this. And what's important for him and for frankly many other actors too is AI has kind of a bad rap in Hollywood right now. And so associating your personal brand with an AI company is actually a quite large big ask. And so he really wanted to make sure that we're serious. And once he said yes, we had written a script and we were like really excited about it. And he went, I want to pick my own scriptwriter and I want to pick my own cinematographer. And we went, okay, like that's cool. And he went, I want the cinematographer from Oppenheimer and I want the scriptwriter from SNL. Well, turns out like actually cinematography and like, you know, touching up stuff like AI is actually getting pretty good at that. So it wasn't crazy expensive. It was manageable. And they went into the scene and they did sort of like 20, 30 takes and it wasn't like that funny. Okay. And then they just clicked. And you really see if you look at the videos like his facial expressions and way of like leaning into what the emotion and what the feeling and the vibe is, I think just world class. And it sort of like came out just like so funny. And this one scene in particular where he's like recreating when we're calling from Sweden. I don't know if you've seen that one. It's the best one I think. And he just comes up with this like really funny voice that doesn't sound Swedish at all, but he's like, ring, ring, ring, ring. Hello, it's Max from Legora from Sweden. And he's like, hello, it's Jude. Right. Like it's so... It's just so fun. It had an amazing impact and I think what it did was it actually brought us out from the lens of just legal onto the, you know, kitchen table.

What's your message for the president today when you guys... Okay, you know what? I love you guys ask these questions. They're stupid as... a good question. What do you think about AI and why do you believe it? Why is it good for America? And that's the kind of question I can answer which I with great respect because I love you guys and I like you guys too. But like it's like cops having fun. You can sit. What did I just... Okay. He's having such a good time. Do you think I'm the village idiot? I'm asking. Okay. Exactly. So I'll give you an answer. No. No. I gave you a question. That was... So this is a question I could answer. What do you believe about AI? Let's do that. What do you believe about... Okay. AI has to work for the warfighter. Has to work for enterprises. It has to be safe, has to be powerful. And there's only one country in the world that has this kind of AI stack. Both the people and the products. Palantir and other companies are leading on the battlefield and leading. And the questions Americans have are, does it work for me? Is it just a bunch of unpleasant people getting rich who ask agnostic questions and refuse to ask the questions that they're being asked? Or does it work for you guys? Do you feel like AI is working for you? Now this is... I think people were laughing at this but he literally takes the mic at one point as well. I think we have the video of that when someone's trying to ask Sundar a question and Sundar's like running away. But I do think it's important to like just take that as an example for right now. This is such a rare type of tech CEO in my opinion. Such a rare type of like, you know, globally known tech CEO. They have the courage to go in front of these individuals and face them head on. Yeah. And be themselves as well. I feel that's like a really underrated trait and I think that there's a lot of power in that. I think I also think and we've said this before the Palantir team doing an incredible job of allowing their leadership... incredible speaker says his mind, knows that that's his own ability and, you know, he's not kind of like shielded behind a comms person or someone. It's a kind of genuine human being behind a company. And, you know, I think that actually speaks to a lot more people than reading off a script. Yeah. I think Europe can take some lessons obviously, you know, send them out and tell them to be themselves. Obviously, there's probably some PR training going on. Don't, you know, go out there and say some crazy stuff, but like... crazy stuff within reason. I think you should say you have to be bold. You have to hold like hard opinions, strong opinions, you know. If you can't go out there and sit in front of a bunch of reporters, to be honest, unprepared. Yeah. Then... and take the mic off them. Yeah. You have to think internally. Think internally. Ask the question. Why is that? Yeah. Yeah. Let's put up the next video. There's another video here just to drive down the point. Yeah. Yeah. This is another Karp one. This is great. This is great. Yeah. Yeah. What you're going to tell the president today? They're trying to get Sundar. Mr. Nadella, can you give us a sense of what you're going to tell the president today? Mr. Nadella, you want me to get a question for you? Sure. Go ahead. You want to talk to Mike? I'll do this. Okay. This is how you do it. Okay. Like Sham. Okay. Sham, tell me about how AI works on the back. That's a question he'll answer. Or see... I think we found Karp's second career. That was our Aean Jabers and Palantir's Alex Karp of the White House ahead of the president's meeting with some key tech leaders. Let's bring in Aean now. Talk a little bit more about that moment. Aean. Yeah, Carl, that's right. A little bit of an unusual interaction with Alex Karp there. And we're just watching here outside the southwest gate of the White House for the arrivals of these other CEOs who are expected for lunch with the president any time now. We're just seeing some of them pulling in here. Alex Karp, as you saw from Palantir, definitely wanted to talk. He had a lot to say at one point, as you saw. He took my microphone out of my hands and began interviewing people on the street here. He wanted to talk. He can do anything. Satella of Microsoft who did not want to talk did not answer any of our questions. So we're getting a variety of responses here as these CEOs come in to meet with the president. I asked Karp and he got a little bit defensive I have to say. I asked him what his message is to the president. He said I'm not going to answer that question. What do you think I am, the village idiot? And he got a little bit confrontational just before that sound bite that you saw there. Oh, it's so good. But I think it's true. It's like when we look at these events where we give them the freedom to go and do what they want. Okay. For Sundar, yes, maybe. He still should do it, I do believe. And I take Musk as an example where he went on Joe Rogan, like smoked a load of weed and drank a load of whiskey with him. Like that's in his book. And the stock price dropped. I think it was like three and a half or maybe 5% that day when those clips started going out. Yeah. But then it came back and it also gave us one of the greatest viral moments of all time in the tech industry. I think that's probably a bad example. I think it's a bad example of getting its stock price. Yeah, but it goes up. It goes up. There's just a little inflection point where it goes down. It comes back up all the time. I tell you who with a team who's probably not doing the greatest. Let's go. Let's go did some Dario-isms. Okay. Yeah, let's just get this one up quickly. Yeah, we have time. We got 3 minutes. This, by the way, is such a hilarious clip. Yeah, this is a great video. Slow down. from it. Where are you at today? Look, I mean, I will say what I've always said, which is that AI has incredible benefits. I've talked about the medical benefits of the technology. I as the president has... trying to get Zach over as well. What god's sake starts laughing low... has very real risks and, you know, the mechanism... and then Dario one runs away... is still under discussion but we all need to work, we all need to work together to make sure that we can win and we can win safely. If we do this right, we work with the president and everyone here, we can win safely. Very loud. Wow. Oh my god. H destroyed my ears. Hold it. Hold it. CNN fake news. CNN fake news. Look at him. Look at him escape. Look. And then Trump gets Zuck in the conversation. Wait, wait, wait. uh... so good. Wait, there's one more. There's one more. That does seem to be the next one. The strategy for these leaders is... uh head down, turn your body, and walk away. Yeah, it's actually pretty good. Let's just get this last one. This is Dario needs to relax. Dario, you have to relax, man. This isn't even him. with this new technology. Yeah, it gets cramped. The necessity of maintaining our China and any other competitors that might emerge around the globe and the superh... speaking because I'm just watching Dario. I'm just watching the way that he performs. Not even, you know, center of attention. But very... some great videos coming out of that. Let me just tell you about 11 Labs. AI that communicates at human level across every channel and modality. Deploy conversational agents across every customer interaction and create on-brand content at scale in 70 plus languages. Go to 11labs.io. Legends of the game. Should we bring our next guest onto the show? CEO of H Company. By the way, thank you for coming into the studio. Thank you for coming into... Great Studio. Many are calling the arena. We appreciate... arena. Wow. Were you here? Were you here when Max was here as well? Did you manage to... just for four minutes, but yeah. Okay. Okay. That's good. That's good to hear. Thank you for joining us. Welcome to the European Technology Network. We've been incredibly excited about this conversation. I particularly have been incredibly excited about this conversation. I've been following H Company. I used to work for a... I don't want to say competitor because you guys were a lot bigger than us. But it was a... we were a smaller version of what you guys were doing, Convergence. We were doing browser-based agents a couple years back and it's around the time every single day we would wait to see if H Company had launched something and if they had we'd see it and then we'd scramble and we'd try and get ahead. So this is quite a close one to my heart which is quite funny. I'd love to start off by more of a personal question, a little bit about your background, but you were doing a lot of other things before you stepped in as CEO of H Company. So before scaling enterprise tech, your journey included serving as a Navy officer, teaching mathematics in a prison, starting a consulting business, and also opening a recording studio in the Caribbean. Is this true? Yeah, sadly it's true for my parents who were very scared after I started... this is so incredible. I would just love to know like why each of these things like how... Yeah, I think I mean, you know, I started my first job was in a big US bank in London. I was making these mathematical models. So I'm an engineer my training and I really liked it. It was like good money, good people, intellectually challenging, but I was sort of missing the real world. Like I wanted to see how the real world was. So that's where I started to be a teacher in a public high school in Paris. Then prison, then the Navy and that's I think got me to see the real world, how the real world was working. I also started a small company in China, basically visiting factories. So like seeing China at the time which was like sort of like between a weird workshop and a huge factory without humans. And I think that still animates me today and it sort of created also that sort of unconventional part of my life as a trader, like what's happening on the ground is not exactly what we think if you are, you know, in our startups between SF and Paris and London. Like that's something that I use a lot right now. And it's also what's behind like the move I put at H which used to be, which started as a research lab, was funded as a research lab but then the idea is really to push that technology to the real world as messy as it is today without APIs without anything, you know, like that's what we are trying... yeah really cool really cool I saw that I had to get in there I thought it was really really cool maybe we could start with the newer release that you guys mentioned the Holo 4 would love to maybe just get temperature check about this. What are we seeing? What's been rolled out and then we can get into a few more questions around it. Yeah, I know for sure. So, Holo 4 released yesterday. It's a family of models. We're super excited by the release and that's also the reason why I'm spending the week in London for like interviews and I'm giving a keynote tomorrow as fix around it like it's basically if you look at the history of H so we started to focus on that technology called computer use two years ago. Computer use was really a bet at the time was not really a mainstream technology. Now you see a lot of companies are starting to use it more and more mostly on the B2C side or easy B2B. But we started to focus on that and the first models were really about solving these tasks which were like can you book me a plane ticket and then do check in or can you these sort of like simple slightly simple tasks maybe you know like 20 30 steps and we were trying to max the benchmarks. So you know have you have a lot of benchmarks OS world verified is one of them. We are the first company in the world to reach 80% of the task being done on OS world. So we're very excited about it. But as we deployed in the real world and now we have customers in like many different sectors. We realized that the real world didn't really need maxing benchmarks. So we started with Holo 4 to focus on much longer tasks. So instead of like 30 steps on average, Holo 4 is doing 300 steps on average. Wow. Okay. So it's like I'm happy to give you examples but like it's high level examples. It's like I'm a sales person. I want to enter that big company. How do I do it? It's not like can you book me a trip and it's going to be like okay I'm going to spend like you know 100 task steps on LinkedIn. I'm going to find if I have connection I'm send emails I might subscribe I might like get into a conference of that company listen to the CEO speaking understand everything. So that's one thing. The other thing is cost because if you do B2C you never really know how much it cost you. I mean you pay 20 bucks a month but maybe it's costing like 200 bucks to the company and right now it's what's happening but if you're a business you're going to do maybe 1 million column minutes to AI and so we focused as well on that so we are like between 20 and 40 times cheaper than the big models because it's one of the thing we see on the ground you need that and then time to execution time to ROI so like is it just efficiency or is it really getting to a task being done and ROI you can measure at the end. Yeah. Yeah. I'd love to understand that. Kind of extrapolating out like the steps and kind of having a you're thinking of yeah the outcome you know you're thinking of like what is the value that I'm getting from you know this new kind of interaction with technology and I think for you know the general population we kind of speak about the AI dissolving into the general population quite a lot. And I'd love to understand like you know by gearing the product towards the value as much as possible in theory you're going to be able to kind of translate that to the general population more right like we sit in this kind of tech bubble excuse me and so like that's kind of you know you can't really translate yourself to how they think but I'd love to understand your thinking around that yeah I think it's we've seen a couple of ways of AI the first one was chat bots like we all use chat bots it was like question and answer sometimes a better because you plug it you know with rag to data of the company but it's basically answering emails or doing powerpoints like it was mostly there to do either things that were done before a bit faster. So the whole bucket of like productivity and efficiency and I think that bucket is interesting. We like to be more productive. Sure. But if it's the AI revolution, it's really disappointing because it's if it's just like I can make slides faster or I can book faster a hotel like that's really really disappointing. What my conviction is my vision for what we do at H is that the issue is more a nurse like if you think for a second like a nurse in an emergency room right now, do you know how much time she's spending in front of a computer? Like in percentage? No, I don't. No idea. It's like 80%. Wow. Because what happens is you have a patient that arrives, you love the patient on a tool, but then there is another tool that is like, oh, for diagnostics, so I need to copy and paste. Oh, I need to find a bed to that patient on like in the pediatrics department, there is no space. So, I'm making a phone call or sending an email and basically because you have these hundreds of software that exist. It's basically your day. So, you've been trained to care for patients, but 20% of your time is doing that. 80% is doing all the rest. Which means you sometimes lose patience sometimes. Of course. Yeah. But like you know hopefully not so much but it's a big issue. That's true. Yeah. And the way we think about it is not oh let's make the nurse more efficient. It's more like okay what's the outcome? So emergency room one outcome could be halving the waiting time of the emergency room. That's something you can measure making a nurse more efficient. How can you measure it? Is it really important? Actually what we need to see some real world outcome. So that's always the way we approach our workflows. Yeah. Salesperson we're going to be like okay how much do you want to increase your turnover? Hospital emergency room like our goal is going to be halving that. And because it's not that easy because it's not just like you deploy your technology, people start to use it and then miracles happen. It's not really that real world. The real world is like you deploy it, it's not working or people are like, yeah but like you don't cover the exceptions and that's where you need to iterate really quickly. So what we do is we sit next to if we keep the example of the nurse next to the nurse. Yeah. We record what's happening. We set up a first agent. It takes a couple of minutes. You don't need like two days or you don't need to ask for migration of the old system to a new system or whatever like you are immediately operative. But after these two minutes you hand it to the nurse and she's going to be like yeah but no but like actually you know you didn't this or that is more complex blah blah blah. So you do another iteration. Sometimes you need like 20 30 40 iterations. Yeah. But because it's a couple of minutes because it's just like computer use record do it record do it after a day two days you have your workflow that's working and the traditional approach is usually let's spend three months to evaluate the processes oh you need to migrate that old software to a new software you need to API that software so that we can work with it so it's like already six months on paper yeah these six months on paper we know is mostly like 18 months and at the end 18 months you have done so many things that you sort of like forgot what you you're so happy that it's working technically you forgot the so that's where how we do it and that really quick iteration circles is really connected to computer use there are other elements that are important like trust safety compliance all that's super important as well all these things make it work in the real world I remember sorry you go I was going to say like the B2C and B2B realms are incredibly different and especially when you have like a horizontal agent that can do nothing. Like I remember speaking to people and saying talking to different department leads or you know various different companies and they say okay well what can it do for us and we like well it can do anything and be like well what department all departments and what seniority we're all seniority then you kind of get this thing where if it can do everything it can kind of do nothing really especially at the B2C level. How have you found that transition because you know we struggled with that back in the day and it was a great experience to like learn how the technology can be applied because I feel like we were also learning about the technology as we were building it as well. How have you found that transition from the B2C maybe early days where you're just experimenting with it and it's a fun technology to that value ecosystem because I know we talk a lot on the show like yes and you mentioned it we had the experimental phase we had like the adoption phase I think we're still a little we're at the end of it and now we're coming into this sort of ROI phase of the technology how did you find the transition because it can be incredibly hard yeah it's both incredibly hard and incredibly exciting because what happens is it's sort of a wedding like we come with a technology, we come with our people. They sit next to the users, but we are not the expert of like dealing with an emergency room in NHS trust in Scotland, you know, like that's very specific and if you just like impose your technology and say, okay, from now on you're going to, you know, that's the way you're going to work. It's never really going to work. But if you also just like try to mimic exactly what's happening, you end up doing like productivity things. So what's really interesting for deployed engineers that we are sending is that the first dimension of the deployment is basically efficiency. So like we focus on something, it takes two hours, we make it five minutes. But that's like the first part. The second part, what we call like the second dimension is really the big change that happens when you have like potentially 10,000 people that can help, you know, like scaling it. So an example could be you know a telco company we are working with, they have a customer service like any telco company and we all know how frustrating it is like oh my internet is not working so I'm calling I'm waiting for five minutes someone is like figuring it out like oh it's a hardware issue so I'm connected to the technical department so billing issue so like you are very frustrated. The number of people in the telco we work for that decide to change provider after a bad interaction with customer service is around like 18%. So it's a huge loss. So you spend money on customer service and you lose customers by doing that. Now first dimension you just make customer service more efficient. People wait a bit less. It's great but they still need to call etc. Second dimension is you put an agent behind every internet connection every night. They do all the tests that you would do if they were calling. You realize that your internet is not working at 3:00 a.m. You do all the tests, you realize it's a hardware issue. You order the hardware, you make it deliver, you send the text. When you wake up in the morning, you have that text saying, oh, your antenna is not working. You're going to get new hardware by this afternoon. And by the way, we give you 20% discount. You're unlikely to leave that provider after that interaction. And so it's not really providing more efficiency. It's like doing things completely differently. So in that case proactively so your customer service obviously you still need people at the customer service to deal with some of the things but you turn that cost center into a profit center. Yeah. And then there is the third dimension we can talk about later. I'd love to kind of speak about the journey of H Company because you guys were one of the first companies to kind of adopt this like FTE model. And I'd love to understand like what does that kind of journey look like? You know, from kind of pre-launching Holo 4 to like now what is the future looking like in terms of like that trajectory like how's that kind of journey shaped the company and what kind of that means for the future and where you guys are trying to get to. I'd love to understand. Yeah, I mean so FTE is becoming a very fashionable term right now like a lot of people use it. I think most of the time when I hear about it it's about deploying software or AI on the ground. So it's like this sort of like technical consultants deploying it which is obviously something that we are doing but it's like probably 20% of what our FTE do. The journey and the journey to Holo 4 is like in all the products that we make is that our FTEs are there on the ground. Yeah. And they see things. Yeah. And that's how we build the product. And it's very different from...

Like, you know, traditional AI labs that are going to build, you know, a model, that are going to feed a lot of data into it, that are going to make it better than the other model, maybe bigger, and then at the end of the day they're like, okay, we raised like a couple of hundreds of billions, we need to find a way to sell it. So let's send people and try to sell it. We are doing the other way around, basically.

So that's like the main difference. Yeah.

So our FDEs in a way are like benevolent spies for, like, you know, what's happening on the ground, and saying, you know, oh, we need that, you know, access control, that's really annoying, but if we don't have that, the model is like useless, so we need to build that into it. And sometimes it's not like, you know, solving the Navier-Stokes mathematical, you know, like, but it's really about like getting what is blocking the real world. And then the last thing that our FDEs are doing is understanding the business.

It's a sort of wedding, like I started it, doesn't happen the first week, but like after a couple of months you are on site, you have coffee with people, you open doors. Oh wow, I open a door and there are 80 people doing something that I think they could like be much better at doing with our tool. So that's like how you start to transform the business into the sort of like dimension three that we could talk about, which is, you know, getting a new strategic asset, being doing things completely different from your customers.

I think there is space for every approach, like the iPhone is a good example. Nobody was really asking for the iPhone, but then like Steve came and said, okay, like this is the new tool, everybody's going to use it.

But I think it's an example, but there are also a lot of things that come from the ground, and that's where we operate. So it doesn't mean that our product teams or research teams don't have ideas, or we do, you know, we do regularly hackathons where everybody stops working and they start to do these crazy things. But what we are, I think, very good at doing is that sort of flywheel is happening on a daily basis. It's really like, oh, this morning I was at that sports team and they are losing, you know, 1 billion because of this, because of that, and the reason they can't solve the problem with traditional means software or traditional AI is because of that, and that's like happening over lunch, that's happening over, you know, coffee, that's happening over more like structured discussions.

But that's like all the information that we are getting from the ground, and then the tough job of our researchers and product team is really to arbitrate between like, oh, that sounds like very specific to a customer. Yeah. So you should definitely build it, but we're not going to build it into the product and the model. And sometimes it's like, oh, you know, the model is not very good at calendars because every country is doing calendars differently. You guys start with Monday, people start with Sunday. To move from one month to another, you need a small arrow on the back, like, let's make the model better at understanding calendars, and that's like useful for everybody. So it's that constant tension. I need sometimes to have very strong personalities in the head of FDE and the head of product and the head of research. And sometimes I need to step in and say, okay, stop fighting. Like I know you don't want to build that. You don't think it's important. I know you think it's important. You see it every day. It's your pain. Like let's decide.

Yeah. Yeah. It's an exciting and a tough job. I think what happened was the company was funded as a research lab. So we had that diamond in the rough. We had these great, you know, researchers and models being built, especially on computer use, and my role was really to make that useful for outside of the research world, basically. So making that, you know, diamond in the rough into something that you can use in the commercial world, in the government world. So the first element was to build the product team, the FDE team, the what we call implementation strategies. So like basically go to market, going out there, explaining what we do. Now it's a bit easier because computers is becoming a bit more mainstream, at least on the B2C side. But at the beginning we had to explain what it was, how different it was, and so that's what I built also, explaining, like coming to see you guys and trying to explain what we do is very important. And then it's really about that constant equilibrium between what we are building internally, what we are getting from the ground, and really what's my key motivation is not really technology, like, oh, do we have better technology than my neighbor, but it's more like what is the ROI? Like did we actually save lives if we deploy a hospital? Did we, we're starting to work with sensitive governments as well, like did we move the needle where the needle needs to be moved. And it's not only to make company more profitable. I think there is also an element of some companies are going to disappear if they don't adopt AI in the right place. If they just play with it. I was talking to a CEO the other day who was telling me, oh, you know, I put like 80 million euros in AI a year ago, and you know, when should I expect to see a return on investment?

I was like, no, I mean you have to be demanding for it, like the first month you don't see, like if you buy 80 million and expect 150 mil back or something. Is that like that easy?

Yeah. Yeah. Yeah. So I think we're moving, like I see less and less people talking to me about their metrics being how much I put in AI or adoption rate or number of models deployed or number of use cases deployed. I think that was like the metrics of the first waves of AI. Now it's like, okay, I put 10 million in H. How much do you bring me in six months? And I can tell you it's going to be like 21 million from my experience. Like, you know, your procurement is going to be done in a different way. You're going to spend less time buying things, and that's like, and actually that's the way we price. Yeah. So like if we fail, we don't make money. So it's also a good way for us to wake up and have like the good coffee and the good idea. Otherwise if you just start to say, okay, we have paid to check these 100 boxes on the technical requirements and we are playing... where's the actual ROI kind of thing, like yeah, yeah, exactly.

It's getting wider.

Yeah, exactly. I think it's really the way to think about it. And it reminds me, I went to see the Pope on Saturday. He came to France, and that's awesome. It was like I was really, yeah, and really privileged to be there. And the way he was talking about life, not very religious, but like it's just like taking 10 steps back and saying, you know, we tend to like follow the new things or try to be like never happy about what we have and try to get like new things around, but these things are not necessarily the things that make us happy, right? I think in the enterprise world, if I can make a parallel, it's like, okay, I'm in the middle of my migration towards that, or I'm APIing my stack, or 10 years ago when I was deploying like AI and data, I was like, oh, I have a data lake with all the data that is clean. Yes, but is it making your business better? Like can you tell me the before and after? No, usually no. And so if you take a couple of steps back, you start to realize that if your business is selling, you need to sell more. That's it. That's it. If you're a construction business, maybe you need to purchase cheaper and accelerate the deployment of your, I mean, the building. And every business has different ROI. And that's where we need to marry the business and understand how our technology, which is pretty horizontal, it's like a pen, you know, like I'm often asked about use cases and I'm often answering, you know, use cases is not the way to think about it. You don't really think if you don't know what a pen is. Not going to tell you, oh, it's great for writing love letters, but also like for drawing, playing games or like you're going to use it in a completely different way that I'm going to use it. But that's like you need to use it. And it just makes me think of like we do a lot of hackathons, and we did one the other day in San Francisco with Nvidia. And every time you go to a hackathon for H, so we give the technology to people and they can build whatever they want. Yeah. They build things that you never expect. You know, you had 40 teams, 40 different things that I would have never expected. And it's like the exciting part of building a horizontal technology. It's not just like, oh, we are doing AI for water cooler optimization, you know, like you exactly know how it's going to be used. That's the exciting part about it.

It's a central question. Yeah. Because what we do is not just building technology or not just deploying technology, it's really making sure AI... So like the one-word answer to the question is ROI. Like if you focus on ROI, you might take the wrong choice, you might be slow, etc., but you're always going to change the direction till you get the ROI. So like that's the first concept. If you zoom in, we have all our little like, you know, secrets to make sure like FDEs are at the right place at the right time. It's also a lot of a people problem. Like if you start to approach a company, if you think that it's the addition of a lot of processes and third parties and connection between that, like the wrong way to think about a company. A company is actually a lot of different people and has a culture, and so like you're going to deploy very differently. If we come back to customer service, like the customer service of Airbus is very different from the customer service of, you know, British Telecom. And you're going to need to be very deep. You might think, you know, customer service is customer service, but actually, you know, Airbus is going to send like one person at the other side of the world with a part in a suitcase on a plane to repair a plane, and it's going to do like very complex diagnostics on software that might be, you know, 20 years old. So that's where the personalization part comes in. Now the hard job of the FDE is to distinguish what is important for the customer, what is something that is more general that we need to bring in the model or the product, and then what is the next step. Because the thing that we never want to do is being called for workflow execute, you know, use case and leave. What we want to do is transform the company. So in concrete terms, like we have, you know, around like 20, 25 FDEs right now. They are like between the US, the UK and France. Their background, some of them are coming from companies like, you know, McKinsey, Palantir, consulting companies, but with a technical background always. Like the interview process is really, really tough. I don't know if I would still get in, but it's like both very technical. But also we are looking at people that understand a business, but also we are looking at people that can create connections with people. Sure. Have a coffee and realize that, oh, in the entertainment business, you guys spend so much time trying to find, trying to arrange schedules because like these annoying funders are like moving, you know, all the time. And that's like one thing. And maybe we started to work with you guys on making better videos, like transition videos or whatever, but like there is something happening there, and the only way to know about it is really like to talk about it, to get into the business.

Yeah, I mean, I think I used to live in the US and I came back for H. I think we have a very strong ecosystem. We have like four companies building models. We have Mistral. I was with Arthur at the cylinder concert last week. Big fans. We have Amabs with Yann LeCun building real, you know, world models. We also have like voice models being built in France by Kyutai. And so that's like, because also we have a lot of talent at H. If you look at people, people are coming from like the big US research labs, companies like, you know, I mentioned Palantir, these companies. But we also are very strong in the UK. It's actually, you quote me at an interesting moment because we just doubled the office in the UK. We doubled the number of... half of the leadership team is in the UK. So I think like that ecosystem is very strong as well. It's a very strong ecosystem in terms of talent. It's attracting a lot of international talent. Paris is great at attracting talent, but it's maybe a little bit less international. So we have the best of both worlds between the two countries. And I try to spend a lot of time in London as well. Like, you know, I saw you had Mati from ElevenLabs. They are like big partners. And it's just, and, you know, like hoping to maybe have a sighting of Count Beanface or all the great things that you guys are doing.

Thank you so much.

Oh, wow. I should have read.

Thank you very much. Really appreciate it. Thank you guys.

Slow to really produce the results needed. And yeah, so I started machine learning very simply, and then, you know, me and Alex met up, and then, you know, the vision for world models for extreme physics really landed for both of us, and that's how we're here.

Yeah. So, I mean, one of the big things that is challenging right now is you can't really explore the design space. So that means that, you know, there are hidden corners that are basically left unexplored because one, it's too expensive, or two, you know, you can't simulate it. And that basically means that, you know, your chips, you know, they are going to hit a ceiling very soon in how efficient and how computationally powerful they are, and that's all due to the fact that you cannot optimize these reactors further without, you know, a big delay in the technology. So what our models and the kind of technology we're doing really unlocks is just folks can really just accelerate the pace to the next frontier of human innovation. And we approached that using AI.

Yeah. So that's kind of really the crux of, like, why we exist, right? These problems are so, so hard. And, you know, we talk about simulations here. What that means is running this big expensive program on a supercomputer for hours or days or weeks. And the point is, like, how are you ever going to design something good if you're sort of searching around in the dark at that kind of pace? Yeah. And the reason for that is the physics is so challenging, and this is, funnily enough, exactly how all our most important industries work. So our models work a million times quicker. So in milliseconds you get the exact same result.