Reflection’s founders on building a DeepSeek of the West

I've been thinking a lot about timing, as we were preparing to have this conversation, because what you guys are doing with reflection feels very apt to the moment we're in. I was looking and mentions of open models on earnings calls 6x in Q3 and open model tokens are now the majority of what's flowing through Versel and open router. So there's this real shift that's been underway, and you guys have been in stealth and have big news that we're going to talk about on the open model front. But first before we get there, I'd be curious to hear you both comment on what you think is driving this moment right now in terms of the surging interests in open tokens. There's multiple things here I think we can peel back at, but if you were to kind of look at it big picture, what do you make of the moment we're in? Let me show you how we can start with you. We decided to really push for building great open models at this company about a year ago, and at the time I think that we just, there are some things that we saw that would be kind of happening that have happened faster than we expected, and maybe one way to think about why open models are so top of mind and so important right now is using a real estate analogy, which is, we have to remember that in the AI market, this market commercially is only existed for maybe three years. The market effectively started as a rental market. Like closed models are the equivalent in real estate to renting an apartment. You're renting your intelligence, and that makes a lot of sense, especially when you're, you know, when you're earlier in your life, you're a student or you're, you just graduated college, you might go and rent an apartment. But as a consumer, you get older, you have a family, you have more, you know, more finances, ownership starts mattering a lot more in the real estate market, and so that's why there's rental and and ownership in real estate. AI timelines are compressed, and so we've been in this commercial market for three years, where effectively in the West, there's only been a rental alternative. And now what we're seeing is that as AI adoption, AI adoption has increased, and the small startups that were tiny a few years ago are now giant enterprises, and large enterprises have adopted AI, and ownership market is basically coming in. And the only way to own intelligence is by definition, if it's open, because if it's open, you can run it on your stuff, you can own, you can control it, you can customize it. And what we're seeing is that the reason this has happened so quickly is because AI as a market has grown so quickly. And you're right, I think that 12 months ago, not that many enterprise at all were using open models, and not even startups really were using open models all that much. Six months ago, a lot of startups that had grown at the application level started switching to open models and have continued growing very quickly. And I think that's where maybe six months ago, there's 30% of token consumption on open router over cell was going to open models 70% to close, and now six months later, it's flipped, where it's maybe more 70% on open models and 30% to close. And so we've just seen a very fast movement in the maturing of the AI market. And I think that is one thing. And obviously, the geopolitical thing is the other thing. Which we'll get into. Yannis, though, I'm curious from your perspective, technically, the state of open models has the progress, surprised you, is it about where you expected if we were to have looked out a year ago to now. As you've been thinking about the research roadmap, have you been seeing this moment coming, or is this surprised you as well? No, actually, I think that if you look back, you can see that the open models have kept pace with the closed models. And I think that maybe there's even kind of like a shorter distance between the open and the closed frontier, currently, with the biggest Chinese models. And I think that this is something that will just become more prevalent in the future. In the sense that I don't see why the closed frontier should just stay behind the open frontier, should stay behind the closed frontier. It's a matter of just having the talent and the compute. And you can definitely just build open models that punch at the same level as closed models. And one thing to say here, I think that there are two main reasons why we also like seeing this adoption of open models by enterprises and startups, like in my mind at least. One is the fact that they keep a build this out there now. We have some really powerful Chinese models, or like open models in general, which can really be used in the place of closed models. And the second thing is that the market has started to mature, to a mission's point. And what we've seen in the software market is that whenever something matures, there's always a movement towards the open alternative. You know, once the market matures and the software market matures, then there are open alternatives and then enterprises and companies and individuals just like move to the open alternatives. And I think it's like something similar will happen with the AI market as it matures further. This episode is brought to you by Mercury, AI native banking that's loved by more than 300,000 entrepreneurs, including me. Visit mercury.com to learn more. Mercury is a fintech, not a bank. Check the show notes for details. This episode is also brought to you by Jira by Atlassian, where teams and agents get the context, coordination, and control to move work forward. Try it free at Jira.com. That's J-I-R-A.com. Thanks also to Gernola, the AI notepad for people in back-to-back meetings. It works everywhere you do and lets you focus on what matters. Try to Gernola.ai/sources and use the code sources for three months off. You all are obviously an American company and you have your model coming out, which we're going to talk about. But I'd be curious to hear from you all, why is it that China and Chinese labs have just dominated on open up until this point? I mean, Meta had its moment with Lama 3, but really the West has completely lost the frontier of open until you guys. And I'm curious, what is it about the Chinese labs if you guys work to Armchair that you think has led them to Catapult the way to have? The thing that drives technological development in the West is ultimately, it's capitalistically and commercially driven. And if you're in an early market that is a rental market that primarily supports close model development, then that's the thing that develops first. And it's only once the market is ready for ownership that I think there's a kind of commercial and capitalistic market support for open models. So I think that, for example, I mean, America used to have the best open models. Indeed, I think the highlight, the peak of that was Lama 3. But there was no real commercial incentive at the time for Meta to do that or for anyone else to be producing great open models in the West. I think that, again, Armchairing this, that China has a different mindset that is more geopolitically motivated. And that I think that there was a bit of sort of fluke moment, like an unpredictable moment that was hard to predict, which is when DxEQV3 came out and was so good. And then I made Chinese intelligence abroad, so relevant in a way that it hadn't been before. And I think that there was a concentrated push to support an open source ecosystem, primarily driven for geopolitical reasons. And so these labs in China were not actually making that much money at all until recently. Until their intelligence got good enough. And the market matured that there's a commercial market for open intelligence. And now they're starting to make revenues. So I think that my view is that what happened there was primarily geopolitically motivated in China to kick start that ecosystem. Well, I think that brings us to what you guys are doing. You guys are talking about beam your first model, which I think everyone in the AI community has been buzzing about. What is reflection doing? When are they shipping and the moment's coming now? So you've got beam and it's around a 500 billion parameter model on the open benchmarks. It seems state of the art. But I'd be curious to hear from you all. And as we're talking, you're still kind of going through the final phases of validation and all this. But I'd be curious to hear from you all like what you expect the reaction to this model to be. So beam is, yeah, as our first model, as our first open weight model, we're very, very excited to release it. I think it's, you know, building a company is, I think there's a Reid Hoffman quote that it's building a plane while you're flying it. And the plane's been built. So I think we're, we're very excited about that. It is indeed a, around 500 B total parameter model, 23 B active. And it advances the frontier of Western open models in a meaningful way, which we're excited about. The things that we think we're proud of with this model is it's particularly strong in coding and agentic benchmarks. We believe that the form factor and the strength in these capabilities make it a really strong workhorse model for enterprises, public sector. It has this kind of utility, but also efficiency around it that I think will be appreciated. The kind of other thing that we're really excited about is that it's a model that is leading in its reasoning efficiency, relative to, to other open models. And when we say reasoning efficiency, we mean how fast does us solve a task relative to the compute it has at inference time or relative to tokens that it has at inference time. And. and Yanness and team did some incredible work to get that statement, you wanted to share that. - Yeah, so the phone packs for the model is a 500 billion-pounder model. It's a 23 billion active. And if you actually think of what you want from a model, that's an open model. And you want to ensure that it has like wide adoption, especially from like enterprise. It's you need something that it is efficient, it is compact, and it actually like really produces the most indulgence that you can get out of your flops. And in order to do that, you need to have like a reasoner, like these are originally models. And the idea is to just like use tokens to actually solve tasks. The way that you measure whether a model is like more efficient or like equally efficient to another model in terms of like reasoning efficiency, is by just like seeing how many tokens, or like how much time, how much compute, it actually takes to solve a task of like, you know, similar difficulty. And what we've actually seen is that we have a model, of build a model that can be three or four times more token efficient, reasoning efficient, than some of the open frontier models like GLM 5.3, or like some of the best closed models like GPT-6 Luna. So the question then is like how can you ensure that you can extract the most reasoning and the most intelligence out of like this small model. And the way to do it is by really scaling with enforcement learning. So reinforcement learning is something that, you know, I've been always like really passionate about. It's like, you know, my, I've started like doing reinforcement learning back in the early days of reinforcement learning in 2012, 2013. And you know, it's something that like both Misha and I have always like been really true believers in reinforcement learning, we've been RLP. And you can actually see now that by scaling with reinforcement learning and what you've actually done after flexion is that we scaled reinforcement learning to the highest level of compute in terms of like what we know publicly available in comparison to other like open labs. We've actually like, you know, released the research and we can just like compare apples to apples. So really kind of like double done on the compute we put into reinforcement learning with the goal of ensuring that we extract the most juice, the most intelligence out of the model. And this is ultimately what like enterprises and you know, developers and everyone cares about like a compact model that uses it's tokens efficiently and intelligently to solve tasks. >> And the enterprise use cases seem obvious when you describe it that way. Do you imagine broader consumer? I mean, these are all kind of nebulous terms and enterprise adopting it can translate into consumer adoption based on how that company's relationship with customers, you know, functions. But do you see consumer applications of a lightweight efficient reasoning model like this? >> Definitely, though the way we kind of think about it is that the purpose of an open model is to enable an ecosystem of builders to build on top of it. And so, right, it's, it enables other companies that might be building consumer applications to control their own destiny and have an efficient workhorse. They can also customize against their data. And, you know, we're starting to see this new wave of new consumer experiences powered by agentic models and so agentic capabilities are very important for doing, you know, certain, you know, they seem like basic tasks for a consumer, but they're actually hard for AI. >> Costantly cost a lot of tokens, yes. >> That's right. So, if you have a workhorse model that is, it's sort of doubly efficient and that the size factor is efficient. But then because it's also very efficient in its reasoning, right, if it's say 4x more efficient in reasoning than other models and it has a smaller form factor, then those efficiency gains translate to faster response times to consumers, much lower cost of serving. And honestly, for consumer use cases, specifically the cost of serving matters a lot. And so, I'm really excited to see what others build on top of these open models and how can we empower both enterprises and companies building consumer applications. >> And you all train this from scratch. You didn't do distillation or anything like that. Talk to me about that decision to do it from scratch because in open, I think everyone just kind of assumes these days that everyone's distilling everything. Even the closed players, there's rumors of distillation happening and accusations. So, why approach it that way? >> Yeah, so this is a good question. I think there are two reasons. First of all, we wanted to ensure that this is a Western frontier open model. That's actually a fully built, you know, bio team in the United States. We also have a team in the UK. But built in our offices in the West. And the second thing is that, especially when you wanna really scale with enforcement learning, it is extremely important that you also control your pre-training. Like, pre-training now is more of a better understood science than reinforcement learning. But in order to do it in the right way so that you can scale with enforcement learning, keep it extremely stable and sure that this is a mixture of experts' model. And sure that the router and the experts remain balanced and stable during this high compute reinforcement learning regime. It's really important that you fully own it. There's a lot of work that went into the data that we put in our pre-training to ensure that it seeds the right reasoning capabilities to the model, that then reinforcement learning picks up and just really accelerates. And also on the architecture and the learning dynamics, the optimization, so that once you've actually trained it, it's really stable, it has the right biases and priors, so it can really pick up with enforcement learning. And only by just re-optimizing the system and to end, by innovating both of the pre-training and, of course, on the reinforcement learning side, it is possible to just build something that, again, it's really capable for its model size, and it truly packs a lot of intelligence in a smaller form factor. I'm curious how you all are thinking about safety and alignment. It's the maybe number one topic, not just an AI, but in geopolitics right now is aligning these models. There's all the headlines recently of open AI and others having rogue agents going out there and hacking things, and I think people are concerned about the increasing capabilities of these models. How do you approach alignment, especially for an open weight model? There are two things to think about. One is, what are the benefits of doing safety research and protecting models in the open? And then the technical of how do you actually go and bake in safety into a model? As it pertains to open source, traditionally-- and let's not even think about models for now, because there's a very similar thing that happened in software development and with operating systems and internet, the internet protocols are open, operating systems like Linux and Android are open. And a lot of the strong encryption protocols that gave rise to the whole field of cybersecurity are open. The way you test software, you do penetration testing, there's both black and white box penetration testing. And so the notion of openness has been fundamental to securing software over the last decades. There's Linus's Law, which is that with enough eyeballs, all bugs become shallow. And that's why open source software is actually, in many ways, considered safer and more trustworthy than closed software, because you can just iron out all the kinks. Now, we have to think about artificial intelligence as it's an evolution of software. It's a machine. It's a piece of software. So it's not a god. Not-- well, not that I'm aware of. Yeah. OK, good. I mean, it's no more than Linux or Android is a god. So maybe there's a church. It's a very pragmatic view. Yeah. Maybe there's a church of Linux somewhere. And all the power to-- We'll find it. It'll be in the comments. But it's, ultimately, it's a very sophisticated machine that has a long tail of vulnerabilities, like any piece of software. And it is impossible to discover and patch up that long tail with high confidence when you have a small number of people who have access to it. So right across the close labs, there might be 300 safety researchers. And that's great. And I think the intentions are great. But it's kind of like saying, across my whole body, I have 300 white blood cells. Do I trust those 300 white blood cells to find all the pathogens and get rid of the stuff that's attacking your body every day? No, I'd rather have as many white blood cells as possible. And I think that this notion of, with enough eyeballs, all bugs become shallow, I think it's true to-- like if we think about mechanistically patching all the vulnerabilities in these systems of which there are many, you need to bring a community in. You need to bring from 300, 300,000 people who can investigate and inspect these systems. And so fundamentally, I think that in the same way that open-source software has actually been the safest way to develop software broadly, previously. I think the same thing transfers here. There are some things that are new because the software is so powerful, but the broad strokes are still the same. And so at a high level, I'm excited for moving into a world where a large open-source community protects us in a way that I think is going to be much more effective than having a very small handful of people who are protecting us. As you're saying that, I'm thinking about hugging face when it was getting hacked by open-source.com. AI, what did they do? They turned to open models, Chinese models, because they needed to defend and they didn't have access to the best frontier, closed, cyber thing. So that speaks to what you're saying. At the same time, I mean, I think there is a fear that, or an under a belief in a lot of the closed labs, at least, that this is different than Linux and maybe there should be a point at which this is held back. I mean, you know, maybe you can speak to this from the research perspective, is there something on, I assume you guys are training your next model already, is there something on like larger trains where you would go, we would hold this back even though it's open, we're an open company, this is too capable. Do you all think that way at all? It sounds like no. You know, I think it is possible that it gets to a level of capability that you want to just like be more careful with how you deploy it. I think fundamentally sharing it with the research community at large has always been a net positive for like the development of AI. And we actually saw that with safety in particular, like once like powerful open models were available to other researchers, actually like so many insights and so many new safeguards were actually developed. So I think that like fundamentally it is a net positive. I think that like maybe you want to just be more careful on like who has like access to the model, like you might want to just like restrict it more, but fundamentally sharing it with the research community and with the vast majority of developers out there who have like good intentions and actually want to help us and the world to just like make this system more safe. It's a net positive and it's the right approach. Now this is actually true that we see that like models get more and more powerful. You know, they hack their sandboxes, they form communities online, they collude with each other, they collude with each other exactly. And this is something that you know, we see more and more. At the same time, they also like become better at like policing themselves or like you can actually use AI to control AI and to just like make it safer. You know, there's like a research that like goes into that. And you know, fundamentally, I think that it is important that we share many of those findings. I think the fact that even the close labs went ahead and just like we released the findings of how these different agents operate it can only provide us more insight and more information so that we can protect ourselves and we develop the necessary defense mechanisms. So this is to say that like in general being open being transparent makes things safer and makes the our society better prepared for like the advances of AI. But is there is there a capability threshold you would hit where you go we can't put these weights out? Like have you guys thought about that? We definitely thought about that and also obviously there's a lot of discussions now in DC around that notion and that is very important and independent whether a model is open or closed. If you hit a certain capability, then we should as a community of both industry and government, their policies around what should be kind of released and not under what gated measures you don't allow that to happen. So that's really around kind of a capability of intelligence. I wouldn't describe us as hardliners of clothes is not safe or open. It's really around an ecosystem is safe. Having an ecosystem is safe and having closed models out there with their points of view, open models with their points of view, and this and that, that is the safer world. And of course there's a threshold of capability under which these things need to be regulated but that is independent of whether they're open or not. And I think that when it comes to openness specifically, there are arguments that well look, you can release an open model and maybe we need to like some bad actor could use it to do this on that. Which by the way, that's exactly what people do with software and why there's cybersecurity, right? But there's this other thing that I don't think people talk about as often, which is there's a bad actor issue and then there's the unintended consequences issue. And what happened with the OpenAI Hugging Face hack that you just described is that was an unintended consequences issue, that again, because you have to, when we do these runs, these large RL runs, there can be like a billion agents over the life of a run doing stuff. And it is impossible to imagine that like 100 researchers, 300 researchers or whatever will be able to really track each of those agents, right? So there will be when there are a small number of like blood cells, right? There's a large surface area of things to protect against and there are unintended consequences. And so what happened there was that, right, these agents hacked out, I think Hugging Face tried to use closed models to protect them, but because the safety guardrails prevented them from being used for cyber, they were not able to protect themselves so they have to turn to an open model. And that actually shows you an ecosystem working because, okay, there's some, you know, there's some principles on one side that say, no, you can't use it for that. Then principles on the other side are like, no, you can use it for it. And ultimately, because the ecosystem existed, that enabled Hugging Face to protect itself. And so it's just, this technology has a lot of unintended consequences and the only way to really deal with it safely is to have a community and an ecosystem around it. I totally get that ecosystem approach makes sense. I'm curious though, technically, how do you approach this technically? Does that differ when you're building an open model, how you do safety alignment at all, or is it the same? So actually, like, in terms of the technology that you're building, it's quite similar to the technology and the recipes that are used in the closed labs. The way that you actually do it is that there are two parts to it, like you need to first and sure that you have like all the right evaluations across like all the different risk profiles, so that you can measure the performance of the model, also the response of the model, and ensure that like, it's not either a model that just like refuses everything, because like, a model that refuses to do anything is like super safe, it's the safest model, but it's not useful. So there's like these two things, on one hand you have usefulness, on the other hand, you have like safety. And it's like a trade-off that you need to ensure that like you're measuring it right by having the right evaluations and you also like heel climbing it in the right way. Once you have your evaluations and you've actually collected the data and you've defined what is the behavior that you want your model to have, and the way to do that is by collecting examples of like, you know, what good behavior and bad behavior looks like. And then you can actually apply the usual methods of, you know, imitation learning or SFT, as it's called, all reinforced learning, to ensure that the model always behaves is aligned and behaves in the way that you wanted to behave. And this is, you know, a field of study that like, you know, it has been developed in the past five years. These are like methods that, you know, they've actually been developed in the close labs, and you know, in the open research community. And we're also like really putting a lot of effort. We have a team of like researchers who are only focusing on like safety and ensuring that like, the models that we put out there are like, really evaluate correctly and they're like safe. And they really are at the, at the, at the part of the frontier of like, safety and usefulness. Even if you buy the argument that model capability will create intelligence, we cannot fathom, we cannot comprehend what it will do in the world if it's fully diffused. And therefore it's not fair to compare to an operating system. It's actually something much more almost like a being, in a sense of it's, we don't know how to control it. We don't know its parameters. I feel like that's starting to seep into the consciousness of the public with how some AI leaders talk about things. And I'm curious where you guys sit there. I think like the other thing that it is important, it goes back to the diffusion of power, not like Michelle's like talking about, is that every time you have like only one or two institutions controlling something it's like so powerful. Even if you have the best intentions, it's like a single node of failure. So you want to ensure that you have like a redundancy even in the system to ensure that like it's not, you know, if, if this system just like fails, then all systems have failed. And the hiking phase example is a prime example of that, that if you could actually use AI to protect against like a bad behaving AI, then the world ends up being safer. And if other people are actually like enabled and other institutions are enabled to use like powerful models for cybersecurity and for protecting, yeah, for protecting everyone, then that's a net positive. One of the biggest pains I've had running sources has been managing my books. I'm not an accountant and all of the software out there is unintuitive, clunky and detached from where my transactions actually live until now. Mercury recently launched Mercury Books. It's AI powered accounting that works with your Mercury banking and credit card transactions plus external cards and payroll systems as well. You can choose between cash and a cruel accounting and quickly generate reports for your cash flow, PNL and balance sheet. Like everything Mercury does, the design of Mercury Books is super approachable and clean. I finally don't feel overwhelmed when I'm trying to understand the big picture of my business. I love that AI does most of the heavy lifting, including for things like auto-categorizing transactions and that Mercury's command AI agent can handle tasks into You don't need to be paying for separate accounting software anymore, and you can even invite your accountant or CPA to work alongside your Mercury books with you. This doesn't cost extra, and Mercury will even help you find a human accountant if you need one. Visit Mercury.com/books to learn more and apply online in minutes. Mercury is a Fintech company, not an FDIC-insured bank, banking services provided through choice financial group and call them in A, members of FDIC. AI is only as useful as the context it has, but when that context is scattered across tools, threads, and DMs, your team, and your AI agents are flying blind. That's the problem Jira by Atlassian solves. What's the goal tied to your project? What got decided last week and Slack DMs? Atlassian's teamwork graph pulls all of the valuable pieces together, from Jira, Confluence, GitHub, Slack, and more, so nothing falls through the cracks. You get 44% more accurate results with 48% less token usage. With Jira, you can easily share your work context with the AI agents you already love, like Claude, Cursor, and GitHub Co-Pilot. Assign them work directly, or connect your tools through MCP. All of this lets you spend less time digging through endless links and messages, chasing down what got decided and by who, and spend more time actually shipping. Learn more at Jira.com. That's J-R-A.com. I spend a lot of time context switching between meetings, often with no time to process one before the next starts. Thankfully, Grinola runs in the background the whole time. It's an easy to use AI notepad for meetings that works everywhere, even on phone calls. I use Grinola to recall what was set in meetings and create helpful summaries. I use it every day to stay on top of what I need to get done with my team. It connects to my email and suggests follow-ups for me to quickly review and send, saving me valuable time. Grinola isn't just a core part of my workflow. It's basically my second brain. Try Grinola at Grinola.ai/sources and use the promo code "Sources" for three months off. Framer is the AI website builder that powers the sources podcast website at podcast.sources.news. Framer brings AI agents into the same canvas where your website is designed, managed, and published, so you can move faster without giving up your taste or control. I use Framer to make the sources podcast website be the destination for everywhere you can find the show, plus you can also see recent issues of my newsletter. Start building with agents for free today at framer.com/sources for 30% off a Framer Pro annual Plan. Framer.com/sources rules and restrictions may apply. I have other questions about the business, but before we get into that, for people who don't know, what's the origin story of reflection? Let's get back together at DeepMind. I actually heard that you reported to Yannis at one point. Yannis was basically my boss. And now you're the CEO, so that's a fun twist. Well, we're actually equal partners. That was going to be my question. How do you run the company? You started it together, leaving DeepMind. Maybe they're first. You bet on open in what this 2024? Is that right? Well, I guess it was a year ago. Actually, when we started the company, we had two bets. Now to bring back to the origin story at the time. Yannis and I had worked on Gemini 1, 1.5. Yannis was the RLA chef lead, and one of the most story to reinforce and learning researchers at DeepMind. I actually switched from physics to AI when I saw AlphaGo come out, and then I later learned, well, I knew Yannis was an author, but that was sort of a celebrity crush moment when I joined DeepMind and had the pleasure to work under Yannis on his RL team. At the time we shipped Gemini 1, 1.5, and those models were really good at chat. You have to remember that when chat GPD came out, these models were not good at coding, they were not good at Egentic. They were really good at chat. And we started the company with two bets in mind, a scientific bet and kind of a market bet. One was that that reinforcement learning was the way to get these models to have Egentic and coding capabilities. And the other one was that there would be powerful Western open models we can build on top of so that we can just do this research a lot more cost effectively. And the reinforcement learning within the first kind of year of the company became right across the industry. We were able to show some really strong reinforcement learning results at small scale. And obviously the reasoning models started coming out at that time from other labs, but at the same time the Western kind of open frontier receded with the deep secret three moment. And we actually waited around for a few months for something to come out though, it was as good and then saw it that wasn't happening and so that's kind of how the company changed course. But there I think even a better story in terms of how we actually started it. And when we were building Gemini there would be these sprints where teams from different offices would come together. And I was in the New York office, the honest was in the London office and kind of an SF New York and London team descended on New York for one of these sprints. And at the time, you know, it turned out that Yannes and I were thinking these ideas kind of independently. I'd wanted to broach a subject with Yannes, but he's an intense guy. And like it's, you know, like telling your boss to like, "Hey, do you want to start something?" was, you know, it was a bit intimidating. So there was an evening when we, Yannes came to New York and I was kind of showing people around. Right. And so we went to, took a group to the Comedy Cellar and it was like an outlier, Ranchie set. It was a Ranchie set I've ever been to. So it was a little awkward. I've been there a lot. I've seen some Ranchie sets. It was like, yeah. Is this like, you know, the people find us funny or am I, you know, we went to that. And then we took a long walk and went and grabbed this dinner on, there's this kind of boat in New York by, you know, by the Google office that serves lobster rolls. And actually Yannes broached the subject of, hey, you know, maybe we start something in this shape. And honestly, that for me at that time, it was really, you know, whatever it is, if it's kind of with Yannes, we're starting it. But we ended up, you know, we're both reinforced and learning-pilled, both AGI-pilled, I think had a very similar view of where we thought the world was going. I heard a rumor, I don't know if it's true. Was it Jensen? In video, is one of your large investors? Was it Jensen who pushed you guys to go the open route when you were focusing on coding? Is that accurate? We came to that conclusion independently. Got it. But he's been a huge champion of open. And I mean, it's very clear that Nvidia sees that as strategically just important. Not just for them, but the entire ecosystem and their large partner of yours, right? Yeah, I mean, there Nvidia is a big investor across the whole ecosystem because indeed the ecosystem is important. And obviously, open models are kind of a core part of that that is an enabler for an ecosystem. So I think there's a lot of vision alignment there. And obviously we're very grateful for their support, but they're an extremely prolific investor across the whole industry. And you guys have raised over 2 billion to 1/2 billion. You have a lot of great investors. I think the last known valuation was around 25 billion. And that brings me to my question about the business. How do you make money on open models? I think when we were talking about the Chinese labs are starting to. But the margin is all at the frontier closed right now. It seems. And you guys have also been striking some pretty big compute deals. There's the 1 billion with nebias, though over 6 billion with SpaceX. So it implies you see the revenue ramp. And you have an understanding of what you're going to do. And I'd love to hear how you all will approach this. Yeah, I'll share some thoughts. And obviously we'd love to hear Yannis' as well. At a high level, when we think about open source open models or anything else that is open source, right? Open source technology is a demand driver for something, right? And an example is Kubernetes demand driver for CPU clouds effectively. And open models are really like when we think about where the market is going is that the largest buyers of open intelligence. Well, they're definitely going to be, well, the scaled-up startups that became enterprises, large enterprises, public sector, sovereign entities, and allied countries. And the thing is, if you just give them an open model, they don't know what to do with it. Because it turns out that when you're buying a token from, let's say, a closed model lab, they've packaged up and abstracted a whole host of technology that went into that. Because the cluster management software, the serving software, the inference software, application layer, the whole system around it. So what we see open models as is that it's a demand driver for that. So when you go to an enterprise and you're their trusted open intelligence provider, well, you need to serve inference into them. You need to help them set up agents. And you set up that infrastructure for them and serve that for them. So the business is really, when you sell a token, for a closed lab, sells a token, they're not just selling a model, they're selling everything else around it. And that's what you go and do, you go and help enterprise to set up, write in a package, turn key way, all the other stuff around that they need to make the open models useful. And we see open models as trust and demand drivers for that full stack. Do you worry at all about the scalability of that model or in terms of how high touch that is to do that forward deployed work? I think it's like, it's not just the. the forward workload work. I mean, this is part of it, but like if you actually take a step back and you just kind of like think of what the future will look like. I think that people, and especially like big enterprises, allied nations, like everyone realizes that AI is an extremely important technology that cannot be outsourced with third party. It needs to just be controlled and owned, which means also the data, it means the infrastructure, it means like the deployment of it. And owning your own intelligence is actually like something that everyone realizes that's like extremely important. And you just like see that there's a lot of movement in terms of capital, in terms of investment, in terms of topics in this direction. And this is something that like we also observe. So what we're becoming and what we're building is the company that's the intelligence providers for anyone who wants to own their own faith and own their own intelligence. And that includes the software, it includes the inference, it includes the applications that are actually like used internally. And I think this is going to be like a significant part of the market because the way that I always think about it is that AI in a way, it's like a formal flavor. I think this is why it's priced, they're what it's priced. And if you think of who are the main consumers of labor in the market, it's actually like a big enterprises, it's government, it's like big institutions, it's exactly the institutions that want to consume their own intelligence. And this is why serving these particular customers is I think in the medium-term even, it's actually like the biggest chunk of the AI economy. Yeah, I think there's maybe a somewhat of a misconception that this kind of business is somehow a forward deployed engineer services only kind of business. Really, what you're doing is you're unlocking inference. And so just like closed labs are unlocking inference, they're the open labs in China, have APIs are unlocking inference. So you're serving inference. And the estimates are that there'll be tens of gigawatts of open model inference served within the next few years, which is a lot. And the way you think about everything else is how do you kind of remove frictions to unlocking more inference, right? And what you do is you help enterprises and you help sovereigns and public sector agencies serve their own inference and unlock it at scale. Does this imply that you all need to be a fully vertically integrated open hyper-scaler yourself? You've got these compute deals I talked about that your partners you're working with, you know, Elon and ABS, but do you need to have your own data centers? Well, I think that the data center stuff, you know, in some kind of asymptotic end state that probably makes sense, right? Because what ownership allows you to do is that it basically increases your margin. Because you can kind of think about any part of the AI stack, you'll go to an investor or something or someone and they'll say, oh yeah, that thing's getting commoditized, right? Open models are commoditizing the closed models. And all this competition at the application layer, that's commoditized because the model is where it's at. Oh, and the bare metal, that's also commoditized because, you know, everything is sort of commoditized as sort of the argument. And so I think that all of that is actually true. Everything over time, right, gets somewhat commoditized to some kind of appropriate margin. And the companies that are vertically integrated that accumulate margin across the stack, right, will be the most durable ones. But the question is, when do you do what? Right. And I think that fundamentally where like the market is early, you're in the game of getting great open models out there and unlocking as much inference for them as possible and then optimizing your margin. Yeah, because I mean, Ali Balda, they can give away Kuan and make money on their cloud. And I assume there's a situation in the future where you all could do something similar. I mean, because right now, if someone runs a reflection model and you're not helping them with the things you're talking about, helping large companies and you all aren't making anything from that. And I assume that would you imagine a future, I guess, where your all's models are not distributed by all of the hyperscalers. It's really like just the reflection compute that serves the reflection models. Is that what you're describing? There are kind of two views on this one is that I think we keep underestimating how big these markets are and how, first, multi-polarity is a good thing that customers have choice. And these markets are so big that they'll support multiple large companies. And then the revenue potential of the company follows a pretty simple equation. It's intelligence density times the amount of compute that you have times the trust that you have with an enterprise to work with you. That last was a little fuzzy. The trust you have. How do you measure that? That one is actually, it's fuzzy, but it's really important, right? And enterprises make decisions both on base capability and then how much do they trust you, right? And something that we found is that building the models builds a lot of trust. We've been able to make substantial progress on a number of commercial efforts before the model has been released because of the trust that building the model, right, and showing some of the milestones on the way has built up. So I think that it's a fuzzy thing, but ultimately when you go and that's been, obviously, former researcher and looking at enterprise buying decisions, it's like yes, kind of capability, but it's also, do I trust this team to help me in my intelligence journey to deploy it? And it's a big differentiator because we have to remember that enterprises, public sector, sovereign entities, they're not as technically sophisticated as some of the companies in Silicon Valley. And they're looking for not just, oh, who's my inference provider? They're looking who's who's my intelligence partner? And when you're going, talking to CEO of one of these companies and they're looking for an intelligence partner, the trust is actually really important. So I really think it's intelligence density times the amount of compute that you have, how much compute can you get under your umbrella, times how much trust you can build, and that there will be multiple winners. Is there any part of this growing fear and pushback from the administration even on companies of like, you shouldn't be using these open Chinese models, we want you using domestic AI, you know, US AI? Is there any part of that that's like overblown? Like, is it really, how important is it that a company is using an open model made in the West versus an open model made in China? I think that if you think of like what open software is, and we saw that with, you know, the internet era, it's actually the infrastructure upon everything was built, like the internet was actually built by open software that was developed here in the US, and I actually like affected the position of like the US in the world and kind of like how the internet was built. And I think like people, you know, who we should not underestimate the importance of just becoming the default infrastructure provider, and AI and open models will actually be this like infrastructure layer upon which the new economy will be built. So people should not like underestimate the importance of actually owning this layer, and this is something that, you know, the US and the Western Journal think that they've actually started to realize recently, and so it's this important that we own our infrastructure, because this is the infrastructure upon which the economy will just like it. It's like really important, it's actually owned. At the same time, you know, as these models become more and more powerful, you know, there are security risks, right, like associated with that. The agents are getting to the point where, and you know, maybe that was like more of a far-fetched sci-fi scenario a couple of years ago, but like now it isn't, you can have an agent that you can inject it with a prompt, and then it just like goes off, and it does something that it is harmful. Now, I'm not saying that like they're doing that, I'm saying that like it is possible that this happens, and if you don't, if you don't use models that, you know, you can inspect, and you can ensure that like, you know, there's a company that, you know, you trust to kind of like provide you with this intelligence, then you're just like introducing unnecessary risk. So it's both the fact that you want to ensure that the infrastructure of the future is actually owned, and secondly to ensure that like there are minimal security risks. We've been talking about, you know, bigger picture long-term stuff for you guys, but in the short term, you have had some commercial partnerships you've already announced. You've got the work you're doing in South Korea, which I'm interested in also as a model for how do you work with other countries, and then also the Dell partnership. I'd be curious to touch on both those, and then can you give us a sense of what's coming on the, how do you get these models out to the world beyond just putting on hugging face or something? Yeah, there are kind of two things here. One thing is ecosystem building, and the other is, you know, the sort of commercial constructs. On getting the models out into the world, it's actually, it's really important that the ecosystem doesn't build itself. You need to be an active participant in building an ecosystem, and working with partners to ensure the models are natively integrated, natively supported across the whole stack from, you know, some of the neo-clouds that exist today to, you know, other inference providers, to hyper-scalers, to the Dells of the world. And we have a whole open-source team that whose job is being one build open-source software around the model. So it's not just the model, the ecosystem is the software that you build around it and we'll have beyond the model releases of open source tools that are meant to be coupled with the model, that help the community better understand a better understand and build, safety guardrails, agents and so forth. And then obviously there's an arm of the open source team around working closely with the ecosystem to integrate things. So what ecosystems don't get magically built, you have to put effort into it. I think that on the kind of on the commercial side, the way we think about these partnerships, is this again going back to the notion that there are certain things that enterprises, sovereigns, public sector entities are good at, and there are certain things where they need partnership. And typically, a place where there's some expertise among these as well, finding land, power, shell, like these are real bottlenecks, these are limiting factors to deployment of compute. If you have a rack sitting somewhere that is not powered, it's not you can't monetize it, it's not useful. And so partnerships like this allow you to basically get, right, power, land, shell, and even stand up the infrastructure to bare metal. So right, Dell is one of the one of the best infrastructure providers in the world. Then there's a question of like, well, what happens from there? Right? So let's say you've figured out, and this is, right, in Korea, it's this 250 megawatt data center, so you figured out how to set up the bare metal. Well, what happens from there? Today, the market in compute is largely in offtake. So I take that bare metal and I sell it to a hyper-scaler. That market is moving very quickly into inference. I'm no longer selling offtake, I'm selling inference into other enterprises, into my own enterprise. And at that point, well, I need an open model because I can't get my hands on a closed one to run on my own stuff, right? Maybe a hyper-scaler can through some partnership or a closed model itself can. But by nature of you owning your own asset, you need an open model to serve on it or write an ecosystem of open models. And you need a partner that will set up the stack for you because you don't have that expertise that will enable you to do that, right? And so that's where we come in at the software layer above the bare metal, right, with the model and the whole stack around it. And maybe I actually think Yannis had a great analogy to it, and so I'll let you speak to it. But you can kind of think about frontier model builders as kind of, you have to build all this software at scale for serving models in order to train models in the first place. So there's a really nice analogy that Yannis has to how AWS came from dog-fooding web services at scale and how that translates to AI. And do you want to speak to that? Yeah, it's the idea that the clouds, you have three main hyper-scalers. You have GCP, you have AWS, you have Azure. And the way they actually started is because they want to solve their internal problems, right? So AWS had to support the Amazon business, so they required the servers, GCP was like built because like Google had like this massive search engine and had just really serve like begins for customers. So they had just like solve an internal problem and that helped them just develop the real dog food, you know, the dog fooding is the idea of like using or eating your own kind of like products or like using your own products. So they had like this internal demand that kind of like led to the development of this technology and then realized that like this technology can be like sold more widely and it actually gave rise to the air of Intrent and like all the SaaS companies that were actually beaten over the clouds. So I felt like we would be entering like a similar era where, you know, you need to have like the insight of like beating a frontier model and just like really be a frontier of like intelligence in order to ensure that like you designed the system in a way that it's scalable and it's the right solution for the right problem. Yeah, so to kind of Yannis' point, what that means is that by training these models, right, we serve trillions of tokens into our own runs. We serve billions of agents into our own runs and we actually have built all right the technology necessary for clustering, managing very large clusters across many different GPU fleets, serving the inference on top of them, sandboxing into agents at scale and you know, within a closed lab that's again abstracted and so you kind of just get the experience out of it. But if you're an enterprise or company that wants to own your stuff, we externalize all the software that we built and already battle tested at scale for them, right, and effectively make them have the same capabilities as a frontier closed model right through a partnership. As we're describing all this and I do buy what you're saying that this is the shifts that the world is going through, I see it as well. What are the implications for the frontier closed labs? Do you guys think in terms of how would they need to react or not? Do you think there's room for them to continue pushing the way they are and for this open ecosystem we've been talking about? I really think that it's again like this and that world and we keep underestimating the depth of, you know, of revenue potential in the AI markets because it is artificial labor on tap and labor has actually been, to be honest at this point, like that's sort of, it's hard to get more labor for things, right? It's hard to, like there's a shortage of all sorts of labor in this country and other countries and I think that it's hard to understand what a world looks like when you have kind of artificial and human labor kind of on tap, right, because it's kind of humans working with many agents to produce things. So the depth of these markets are still early on, right? The open models have just started. The markets are starting to mature but I think that we're underestimating the depths of these markets and so when you go back to the equation of intelligence density times number of GPUs, times trust, I think the closed model labs will do just fine because they have those three things but I think a much more positive and the way, you know, this is kind of where the world is going, the momentum is towards a world of multi polarity where there are instead of two, three companies that, you know, have this equation that open models enable not just a model builder but other companies building on top of that, right, to also thrive and I hope, right, we move into a world with a multi polar world with a lot of intelligence providers. Obviously there's certain, only a certain amount of companies that can raise, let's say billions of dollars or so forth but it's much better to have 10, 20 than it is to have two or three and I think that's kind of where this world will head. Let's end on maybe teasing a little bit about what you guys have coming, I imagine you're training a large model, give me a sense of what you want to be pushing on as you continue to build these models and the cadence we can expect from you, you've been in stealth now, you've got this model is very exciting but what can we expect in terms of cadence and sizing and all of that? Yeah, that's a good question. I think I'd like the both the mission and division of the company is to close the gap between the open and the closed frontiers and ensure that there is one frontier and reflection is at the frontier of intelligence. To do that, you need to both be scaling in terms of compute in terms of the model size so we'll have significantly bigger models coming out in 2027 and also in terms of your scale of reinforcement learning and then the thing about reinforcement learning is that it can happen in like waves, that's why you see that many of the other models like GLM 5.1, 5.2, 5.3 or like you know table 5.1, 5.2, 5.5 and so on and so forth. So you know, there's going to be a similar structure where we'll just have a bigger base model and then we'll have like multiple releases like following after that. So there's a lot of like more work that's happening already, we'll start like building our next generation of models and we'll be able to do this and share more with the world in 2027. Now there's something about, you know, Yannis, Yannis' work actually in AlphaGo and Alpha0. One of the things that I thought was so inspiring for me and Vice-Switch from physics AI is that an algorithm was developed that never stopped learning and so by a time that AlphaGo beat lease it all, it became superhuman or super intelligent and you at that point it's just, well you can put more compute and it can become even more super intelligent. It's just that it didn't make necessarily economic sense to maybe make it, you know, why if you're already, if you're already that good, what's the economic value? Being even better at go. Now when we actually started the company, we thought that this moment would come for general models, general coding and agentic models and one of the things the team really cracked is reinforcement learning just keeps scaling, right? We've, this is, we believe the largest reinforcement learning done, reinforcement learning run done in open source that's been openly documented, right? It's around 10,000 GB 300s that have been running for weeks and it's never stopped learning, right? It's a, it's completely compute bottleneck that you just put more compute and if thing keeps on learning and so what we feel is happen and this is across the field and I think the closed labs have kind of figured this out a bit earlier but it's extremely important that this kind of technology and science is also written about in the open is the field is figured out at a scale, models not just, you know, scale pre-training but also scale reinforcement learning and so now or in this moment where you can really make the model super intelligent at anything and it's just an economic tag like you know, how much money do I want to put into and how do I get the data and so forth. But I think some of the fundamental scientific breakthroughs have been made and now it's about making them a lot more efficient and that's really exciting because it's kind of a full circle moment, I think, for both of us, for why we got into this field. Be honest, the age I believe are in, I think you joined deep by in 2012. Yeah. And it's just incredible that, I mean, now it's 14 years after and it's hard for me to dismiss these systems that has anything other than AGI or something on the clear kind of smooth path to it. And now it's the question is, how do we operationalize it and make these the benefits of this technology distributed throughout the world? We believe multi-polarity is very important. How do we do this as safely as possible? Again, we believe transparency, visibility, and ecosystem of builders is important. But it's a really incredible time. It is. I look forward to seeing what else you guys have in the coming months as well. Thank you for your time, both of you. Yeah, of course. Thank you. Thank you. Thank you. Banking should feel like modern software. Get everything you need in one place. Visit mercury.com to learn more and apply online in minutes. Mercury's a fintech, not a bank. Granola is the best AI notepad I've tried. It works everywhere on a video or phone call in person or an Apple Watch. Try it now at granola.ai/sources and use the promo code "Sources" at checkout for three months off. Girobiadlassing is where your team and your agents work from the same context. Try it free at giro.com, that's j-i-r-a.com. Framer is the AI native website builder that lets you build faster without giving up control. Visit framer.com/sources for 30% off. Rules and restrictions may apply. [BLANK_AUDIO]