Transcript 0:00 We want to build computers that can think and do things like humans. But we raised more than two billion dollars to actually have, like, access to the compute we need. Building these models is like building a rocket. 0:08 The secrecy and, like, the time that you're taking, do you think you will be able to match the closed labs models? 0:15 [upbeat music] If we go back to AlphaGo in 2016, that moment for the outside world felt like miracle. 0:28 For you as an engineer who works day and night in the trenches, what is the biggest change in how the progress actually happens now? 0:34 Well, it seems like, uh, the fundamentals, uh, for the most part, uh, have stayed the same. When we were building AlphaGo, it was, uh, an extremely challenging engineering-wise, uh, project. 0:45 So it kind of like required scaling our models and scaling our, our training recipes to massive runs that, uh, weren't the case back in 2016. 0:53 I think that, like, most people had a couple of GPUs, and they were just, like, training things back then. 0:57 Now it's the norm that, like, uh, anyone who wants to train these big models, either with, uh, pre-training or reinforcement learning, they use, like, um, hundreds, thousands of, like, GPUs. 1:05 In that sense, we were a, a bit ahead of the curve. At the same time, you know, it was kind of a collaborative project. There were, like, many of us working on it. 1:13 You know, again, back then, it was more common for, like, a few researchers to kind of like work together and just, like, build something together rather than, like, big teams with, uh, project managers and just, like, you know, deliverables and everything, which again, is kind of like more of the case right now. 1:25 So I'd say that, like, in many ways, AlphaGo and how we did our work back then is like a preview of, uh, what was to come and, uh, how things, uh, work now. We even had, like, different phases. 1:35 You know, AlphaGo was trained in different phases where you had, uh, you know, just training on human data, which is kind of equivalent to pre-training in a way, and then reinforcement learning, which is kind of like what we call more post-training now. 1:45 And we had, like, human testers to just, like, tell us, like, what are the mistakes that the model was making, which is, again, kind of quite similar to how things are done now. So, uh, and lots of compute. 1:56 So I'd say that, like, um, if something many of, uh, the ways that we structure ourselves or, like, the ways that we went about doing our research is actually more similar to how things are done now than how research used to happen back in 2016. 2:09 I guess, I mean, are we still in the era of, like, breakthrough moments, or is it mostly messy operational work now? It's actually a good question, and it kind of like boils down to what you consider a breakthrough. 2:20 I think that, like, uh, for different people, different, uh, levels of discovery constitute, uh, a breakthrough. 2:26 I think that, like, we have all the ingredients or, like, the most, uh, most of the ingredients in order to build really powerful agents that can do almost anything that a human can do on a computer, and that's a form of AGI. 2:36 In that sense, it's actually more of a matter of, uh, you know, executing, uh, finding the right methods, finding how everything fits together and just, like, doing, uh, you know, a lot of engineering and, uh, research. 2:47 For, like, some other things, we might need to innovate and, uh, come up with, like, you know, breakthroughs, like different architectures or, like, different learning algorithms or whatever it is. 2:56 It kind of like really comes down to what you want to build and, uh, 3:01 you know, what's your definition of AGI, and different labs and different people have, like, different definitions and, uh, what you think might be missing. 3:07 What is your definition of AGI and superintelligence, for that matter? Yeah. For me, I feel like, uh, it's something simpler. 3:13 It's like literally an agent that interacts with, uh, software in a computer and can really do what a human can do. So in a way, it's kind of like, uh, like, you know, do many more things on, with our computers. 3:24 And in that sense, I think that, like, we don't really need any massive breakthroughs. We just, like, need, you know, better engineering, better methods, kind of better combination of existing methods. 3:34 But, like, uh, not anything that's kind of, you know, really a game changer or anything like that. 3:38 I think 2025 literally became the year of coding acceleration, and I want to read this from March 2025 when Lightspeed co-led your Series A, and it said, "Reflection AI is leveraging its deep expertise in reinforcement learning and large language models to solve autonomous coding and, more broadly, unlock the path to superintelligence." 3:59 So it's almost a year since then. Again, what did last 12 months force to rewrite in term of you building autonomous coding, how you thought about it back then and how you think about it now? 4:11 I mean, when we started, we really believed that, uh, as I said, like, reinforcement learning is the next frontier, and, uh, we just, like, wanted to focus on reinforcement learning. 4:19 But, uh, if you want to do really kind of reinforcement learning at the frontier, you need to have, like, an extremely powerful base model that you can just, like, post-train. 4:26 You also, like, need to have a model that you can then deploy safely. 4:30 What we saw is that, especially with, like, you know, Llama Four not being a particularly good model, is that the whole Western ecosystem was missing a powerful base model that we could even use to do reinforcement learning at scale. 4:43 One thing that changed is that, uh, we understood, realized this, uh, limitation, that we needed to have access to, like, more powerful open base model coming from a Western lab. 4:52 And we also realized that, uh, we were the only ones, uh, who could actually build it based on, like, how everything was, like, set up, so we set out to just do that. 5:00 And this is why Reflection is now building, um, its frontier models from the ground up. We're doing both, uh, pre-training and, uh, reinforcement learning. How is it coming along? 5:09 Lots of, uh, incredible things we'll have to share this year. Any ideas where and how will it look like? You, you have to wait for that. Well, let's talk about Asimov then. 5:19 You started with that as a autonomous agent for enterprise. What changed there in terms of all these different coding agents and systems being launched? You know, Claude Code just celebrating one year. 5:32 OpenAI also kind of refocused to being more coding system, and Codex performing pretty well. What's your agent? What are you working now on? Yeah. 5:41 So we're actually, like, building, uh, again, we're building frontier open weight models, like frontier agentic models. 5:47 So you can think of, uh, an equivalent of, uh, you know, the US DeepSeek is, like, how some people call it, or like, uh, you know, agentic model's powerful, like, uh, how Claude is or, like, you know, GPT 5 point, uh, two or three. 5:58 This is kind of like what the focus of the company is. 6:00 In terms of, like, coding, uh, we realized that, I still think that, like, this is, uh, true, that, uh, one of the main, uh, limitations of, uh, coding agents is that, like, they don't have access to all of the context that, like, uh, your engineers have access to. 6:15 So that's kind of like the, the thing there. 6:16 At the same time, you know, we've actually embarked on, like, a, an extremely ambitious goal of, like, building these models, not just with reinforcement learning, but, like, uh, from scratch with both pre-training and post-training. 6:27 And, uh, we just, like, really decided to focus on that, so this is what we're doing. What are the main bottlenecks that you, uh-... coming across? 6:34 I mean, if you want to build frontier models, you need to first of all, uh, attract the right people to do that and really grow out the team. You know, we've actually attracted, uh, a world-class team. 6:44 We have, uh, people from, you know, Gemini, people from OpenAI, people from Meta, Apple, like, you know, all of the best labs and the data providers. These are, like, people who've done it before. 6:53 They've built the frontier models, and they've actually joined us because of our mission of building, uh, open intelligence. 6:59 Uh, and again, we are the only place that's actually committed to that, so this, uh, this is something that, like, resonates with, uh, many of our scientists. 7:05 That's one thing, you know, getting the resources you need in order to have the compute you need in order to build these models. 7:10 That was, uh, also required us to have, uh, a significant, uh, Series B in the summer, actually in October, uh, where we raised more than $2 billion to actually have, like, access to the compute we need. 7:21 Scaling in terms of engineering, building everything from the ground up. Even if you've done it, even if you know what to do, it's never easy. Building these models is like building a rocket. 7:30 There are many ways, uh, you, things can go wrong, so you need your best rocket scientists to come in and just, like, uh, to do it and have all the room, the agency to actually do it. 7:41 And we are extremely lucky that, uh, we've assembled this team, and, uh, we are building these models. What AI tools do you use yourself to accelerate? 7:48 Oh, we're using, like, all of the different AI tools, and I think that, like, different ones are, like, good, different things. 7:53 You know, we're using Cloud Code, we're using Cursor, we're using, uh, many internal ones that we've actually built. 7:58 So we have, like, a whole host of, like, AI tools that we've, uh, built for our own needs, and, uh, we use them heavily. 8:05 So you're currently working on the model, but al- also, is there an application yet that you plan to work on or working currently based on this model? The focus currently is to just, like, build the models. 8:15 There's gonna be applications to follow after that, but all hands on deck. Building these models is challenging and requires all of our mental focus. 8:22 That's true, but also, when we think about open weight models, they are rarely up to the speed with the closed labs models. 8:30 You, the secrecy and, like, the time that you're taking, do you think you will be able to match the closed labs models? 8:37 The main reason why people haven't managed to build, like, frontier open models is because, like, they haven't managed to find a, a commercial engine to support open models. 8:46 This is kind of, it's not, being open doesn't really stop you. If something, actually it accelerates you. 8:50 So, you know, for us, the important thing is, since this will be our first models, to just make them to be the best open models out there. 8:58 This is kind of, like, the focus, and at the same time, we think that we will have everything else in place so that, like, in the future we'll just, like, close the gap and eventually become the lab that, like, builds the best models in the world. 9:10 This is the, the ambition of the company. This is the trajectory we're on. How do you plan to evaluate it? What is the actual metric you op- optimize for? The main metric that, like, matters is, um, adoption. 9:21 That's kind of, like, my, that's the only metric that, like, really matters. You just, like, put it out there and ensure that, like, people like it and they use it. 9:27 And of course, like, internally we have, like, many evaluations. 9:29 Uh, we make sure that, like, we don't just, like, you know, follow the benchmarks or whatever it is, but we have, like, uh, unseen evaluations, really try to ensure that, like, the evaluations match, like, uh, real-world use cases. 9:40 We kind of, like, follow all the necessary steps, but once you've done that and you've built the models, you need to just, like, work closely with, uh, people who use them and ensure that, like, uh, they like them more and they use them more. 9:51 Are you still focusing on autonomous coding, or it's more like a general model that you- It's a general model. It's a general agentic model. Why did you change that? 9:57 Because we think that, uh, it is important that, like, there is a general, general open model in the West, and there isn't one, so I have to build it. 10:04 I recently talked to Minimax, and they said that they also work in a general model, but coding gives sort of much more feedback immediately, and it's just a focus for them. 10:15 I mean, coding is of course, like, an extremely important vertical for us too. This is true, but, like, the model is not just a coding model. It's, like, a general model. 10:21 Agentic means, you know, in many ways it means coding, like tool use coding, agentic reasoning. Those are all capabilities that fit into, you know, coding. So- Mm-hmm... yeah, it's a general model. 10:32 Definitely coding is a big part of it. What are the parts of research you're looking at in, um, reinforcement learning? What is the most interesting for you there? I mean, we're looking at many things. 10:41 Some I can share, some I can't at this point, but the most important thing is how do you scale, and I guess this is not, everyone's kind of, like, looking into that, but how do you do the create assignment correctly? 10:51 So how do you scale the length of the trajectory so that, like, the agent can, uh, you know, take many, many, many, many, many steps and just, like, still learn effectively from that? 10:59 Any other research directions that can be combined with reinforcement learning? I mean, we have, like, uh, we're looking into many, many things. Reinforcement learning can mean many things. 11:09 Uh, it also means, uh, the use of synthetic data. It also means, uh, ensuring that, like, uh, you do pre-training in a way so that, like, you can maximize your reinforcement learning downstream. 11:19 We actually do, uh, do research across, like, the whole stack. This is, uh, the benefit of actually owning the whole stack is that you can optimize it across the stack. 11:27 It's, you know, you can ensure that your pre-training is, like, uh, RL aware and that, uh, you know, the, the mixtures and, like, the data and the synthetic data that you use and everything that you do is actually, like, fully optimized end to end. 11:38 And, uh, we have, like, research efforts and, um, you know, projects, like, across the stack. Everything is happens just so fast in AI. How do you follow? 11:46 Yeah, I mean, things happen, uh, fast, and, um, AI is kind of, like, moving extremely fast and, uh, extremely excited about that. 11:53 I think that, like, it's also important to be able to tell, you know, real progress from, like, noise or, like, you know, from things, like, people just, like, share. 12:00 I've been in, uh, AI since, uh, 2011, so, you know, past 15 years and, uh, in that sense, I've actually, like, seen all the different, uh, waves of the, the progress that has actually happened in the, in the past 15 years. 12:13 So that makes me extremely bullish on, you know, how much we can achieve in the next five years because I've seen it. 12:18 At the same time also, like, uh, it keeps me focused and, uh, I understand that, like, you know, the important thing is to just do your work and, uh, you know, continue to, to sprint. 12:26 Because you have such an amazing experience and you've been in AI and working in AGI for much longer than many people out there, what would be your advice to people, young people who are just starting and specifically, like, researchers and machine learning engineers? 12:40 Yeah, they should just, like, do it. There's a lot of, uh, more work to be done. I think that, uh-... people should just, like, uh, enter the field. It's extremely exciting. And I feel like it's, uh, it's really fun. 12:49 A, a lot of more work to, to be done. There are two different kind of like paths, and I feel like they're both valid. Especially for younger people, I would say that, like, they can be more like, uh, risk takers. 12:59 I joined DeepMind, again, when it was, like, a, a small startup right out of college because, like, I was young and, um, I, I had, like, a really high risk tolerance. 13:08 I always just, like, uh, say to people that they should just, like, try and join a startup instead of just, like, finding a big company after they graduate because this is how you learn, and this is how you grow and, uh, it's, uh, it's really fun. 13:18 This can also, like, be true for more, uh, exploratory research. If they wanna just, like, do something that deviates from, like, the dominant paradigm, they, they should also, like, try that. 13:27 Maybe they care more about robotics or, like, they care more about, like, um, these world models. They should just definitely go for it. 13:32 At the same time, even in the dominant paradigm of these foundational models, there's, like, a lot of, uh, work to be done in terms of, um, how do we use our data better? Do we scale our methods better? 13:42 How do we do credit assignment better? Like, there are still many things that are, uh, unsolved, and, uh, we need, like, bright, young people to work with us. 13:49 When you joined DeepMind, what did Demis and Shane told you that hooked you up to AI so much? So they actually told me that, uh... 13:57 And I remember, like, my first conversation was actually with Shane, and Shane told me, he was like, "We're building AGI." I was like, "Okay, what's AGI?" 14:04 And he told me that, like, "Oh, you know, it's we want to build computers that can think and do things like humans." I was like, "This is insane. Let's do it. Where do I sign?" 14:12 And, uh, I really appreciated the ambition of, like, what they want to build and also the fact that, uh, they were, like, extremely, like, mission-first or mission-aligned. 14:23 They were really doing it because they really believed in it. They wanted to... I mean, these people could have done something different. You know, back then, AI had, like, no money. 14:31 You had to be a bit crazy to even try it. So the fact that, like, they were crazy enough and ambitious enough and capable enough to just kickstart this revolution, you can sense it. 14:39 It was coming from a place of, uh, really believing in what they were building. Well, there is much more money in AI right now. What do you tell your people to hook them to AI? 14:48 Really, yeah, I'm just telling them that, um, if they believe in, uh, open science, if they believe that it's important to have, uh, frontier open models, if they believe that, uh, AGI, which is, like, such a powerful technology, should be accessible to everyone and should be democratized, there's only one place that, uh, does that in earnest, and that's, uh, Reflection. 15:06 Thank you. And the last question, my usual question about a book. The book that shaped you or seriously influenced you, and it can be from your childhood or recently. There is a book, it's called The Idea Factory. 15:17 The Idea Factory: Bell Labs and the Great Age of American Innovation. I love this book. It's about the story, the, the history of, like, Bell Labs and how it created and how it actually shaped American innovation. 15:27 It teaches you a lot about, like, systems and about, like, companies and organizations and how innovation happens. So highly recommend that, uh, to our viewers. I love history. 15:36 I always think we still can learn a lot from that. Great. Thank you so much. Thank you so much, too. [outro music]