The Open Source AI Model For The West - EP 44 Misha Laskin

I think the mission of the company is far broader than that. It is to build a frontier of open intelligence and make it accessible. Meaning when we say frontier, there's no reason why open models should be any worse than closed models. Our intention is to build a new frontier lab that drives the frontier of knowledge in artificial intelligence, but doing it in an open way, in an open scientific way. So it's funny that we have to say open science. It's just science.— but in an open scientific way and release the models, enable the community of researchers and developers to engage in how they develop and build a frontier lab with open intelligence at its core.

Misha Laskin, how are you?

I'm doing well, Ashlee Vance. Thanks for having me.

Thank you for coming. You, well, you've been in the news, I guess not immediate, but it's in the last month or two, isn't it, that you guys did your fundraising or?

Yes. So we, it was announced last week.

It was announced last week.

Yeah.

Okay. It is very recent. So you're, you're, co-founder and CEO of Reflection AI. You guys raised $2 billion. That's not shabby.

Yeah, definitely. You know, this thing that's happening, like, in the progress of AI, right? If you would have asked me 5 years ago or 10 years ago, I was a physics student, that this is where AI would be, I would not have believed it.

Yeah, well, I mean, we'll get into it. We'll definitely get into all that. But I mean, just to like set some context, I mean, you guys, the big, the big part of the big message here is that you guys have raised a ton of money. You have a team with an incredible pedigree, some folks from all over the place, but a lot from, from DeepMind. You were at DeepMind. Your co-founder worked on AlphaGo quite, quite intimately, crucially. On that project. And so sort of have this all-star team. I figured that's why you were able to raise quite a lot of money. And then you have, you've got a bit of a different pitch, a different frame around what you're doing than, um, definitely than, than the OpenAIs of the world. You, you guys have come out and said you want to make this open model sort of in the open source vein, um, and, and to, to publish the weights of, of your models. And so, and you're kind of making the pitch that America needs— we've had Chinese companies, Deepseek got all this attention for doing an open model and that we need something that's competitive. Facebook's made open models, OpenAI kind of sort of is. And then I think does— and Mistral does open models?

Yeah, Mistral is definitely an open model player.

Okay, okay, okay. But this is— am I right in this is some of the heart of what you guys are saying is— makes you, makes you different.

I think so. Certainly it's become really important to produce great open models of American provenance that are competing or competitive with the Chinese ones. But I think the mission of the company is far broader than that. It is to build a frontier open intelligence and make it accessible. Meaning when we say frontier, there's no reason why open models should be any worse than closed models. Our intention is to build a new frontier lab that drives the frontier of knowledge in artificial intelligence, but doing it in an open way, in an open scientific way. So it's funny that we have to say open science. It's just science. Yeah, but open in an open scientific way and release the models enable the community of researchers and developers to engage in how they develop and build a frontier lab with open intelligence at its core.

And I want to nerd out, but I also want to make this accessible to as many people as possible. So when you, when even just some of these basics that we're talking about, when we're talking about an open model versus a closed model, when companies, when we talk about weights, I mean, it's almost like these mathematical calculations, these numbers that are attributed to the neural network and how it will behave. And to some extent, this is like the secret sauce. You do all the training and you come out with these weights that make your model hopefully behave the way you want it to behave and do incredible things. And so some companies keep those proprietary into themselves. And then you guys are talking about being able to be at the cutting edge of what these well-funded companies who have access to all this computer doing, but doing it in this more open way. Why is it even, um, you know, I come from, I covered open source software for years and it was this very like ideological thing, especially in the late '90s, early 2000s. Microsoft was rising and, and there was a whole faction of people that were, you know, they saw code like, um, like writing, like knowledge. This should not be, this should not be contained within a company. You should be able to alter your computers as you see fit. This knowledge should be shareable. You know, and they, they saw Microsoft as, as really like, it was almost a religious debate, right? Like they were evil for doing proprietary code. And man, I mean, I lived through that. It was like emotional and people really cared. I haven't felt that same level of like enthusiasm and rhetoric and everything around open models. You brought up science, you know, scientists tend to lean more in the open direction in my experience. All right, what do we do at Core Memory? We cover innovative, fast-moving, forward-thinking companies, which is why Core Memory is sponsored by Brex because Brex is the intelligent finance platform for many of these companies. 30,000 companies from startups to the world's largest corporations. Rely on Brex's technology for their finances. They've got smart corporate cards, high-yield business banking, and expense automation tools that are fantastic. I hate doing my expenses, and Brex's AIs and software run right through those expenses, figure out where we're spending money, and take care of so much stuff for you so you don't have to waste your time on it yourself. Go to brex.com/corememory to learn more and just, you know, get with the program. Let's get going. Let's get out of this archaic finance software and move toward the future. Core Memory and Brex. How do you come about it? Like, why is it, why is it intellectually, philosophically, technologically important to have these open weights? And did I, did I, did I, is it apples to apples? Have I even started in the right place?

No, I think it's definitely, I think it's definitely in the right place. I think the observation that In the buildout of software, that there were these effectively closed proprietary software buildouts, and a lot of developers pushed back because they believe that a lot of the code that powers core infrastructure should be in the public domain and it should be accessible and any developer should jump in and contribute into it, I think mimics the ideas of how science evolves. The reason we are here where we are today with AI is because of science that was done openly. When we think about the fundamental breakthroughs that power the models that are being trained today, it's self-attention, well, it's deep neural networks, self-attention, next token prediction, and reinforcement learning. And obviously there are a lot of papers within each, but if those ideas were not in the public domain, then we probably would not be experiencing the boom that we are in today. So the fundamental progress in science and technology happens because things are done openly. And when we think about from an open source perspective, what operating system do most computers run on in the world? And it's Linux. I mean, yes, our consumer devices maybe run on iOS or Microsoft, but when you look I mean, if you just look at all the computers in the world, all the servers, and what are they running? They're mostly running Linux.

Yeah, the infrastructure, not even just the servers, but like all your consumer electronics devices, your cars, you know, most things that make the world function are probably running on some version of Linux.

Right. And for mobile devices, it's Android. There are more Android devices in the world than iOS devices. Openness spreads and is accessible because everyone can build on top of it. The internet was an open protocol that spread around the world. And there is obviously a desire from a company that's commercial to build things out in a way that's incentivized with them kind of effectively winning commercially. And the current commercial rollouts have basically incentivized closed model development. So it's not irrational that AI development in the West specifically, or especially, has been done in a closed way, but it is not the fundamentally best way to build out what is the most impactful technology of our time. So openness is, I think, fundamental to this technology being both built out in a way where everyone can contribute to ideas of how you get new capabilities into it, But also from a safety perspective, I think you don't want safety research to just be concentrated among a few closed labs and 30 people within each that really think about safety. You want everyone to participate in that discourse. Now the question is, how do you set up an entity that enables you to do that at scale, at frontier scale? So you're not looking at small open models, you're looking at open models that are just as good as the best closed ones and sustain itself so it can keep doing this? I think that's the big question.

Which was, you know, it was always a big question in Linux land, open source software land. It was like, who's funding the— I mean, it's a little bit different, but, you know, much of that software was these supposed— it always got pitched as kind of like volunteers who are working on Linux. And some of that was true, but it really turned out companies like Google and IBM and you name it, Sun Microsystems, whoever, were paying decent salaries to people to work on this software and contribute it back. And so they had good jobs and this was important to a lot of companies. And then it was always harder for some giants. I mean, Red Hat managed to create this model where they made money off supporting their software and services around it, but they were really like the biggest, you know, flagship example of somebody who succeeded off this model. So what's the, what is the cha— what's the equivalent in open source land? Because you've got, you've got what, I mean, you're up against 5 or 6 American foundational model companies that, that, that are selling all kinds of the, their software on these subscription services. I mean, do you do the same thing or is it, is there, um, sort of like there was with open source software, slightly different play?

It's a really good question. And this is where the analogy to previous waves of open source software breaks, because previously open source was cheap to develop. I mean, it just required developers really. And for that reason, you can, you didn't necessarily need very large commercial entities around a lot of these projects. It's not that expensive for Google and some other companies to front the bill for some headcount. That builds out Linux or some other piece of open source software. Well, of course, there's a long tail of contributors who are independently contributing to this stuff as well. The difference now is that models, open or closed, take a lot of money to train. And so you can't do this in a way that is not commercial. But the important thing is to ensure your commercial interests are aligned with the mission of the company. So I think the big challenge is that if you don't think about those things in tandem, then it's a tail wagging the dog sort of situation where you'll start open, but then your product or your commercial business will take you in whatever direction. You need to go. And so far, each time that's happened, it's been closed. Like OpenAI started with open, you know, open artificial intelligence.

We were, we were kind of here not too long ago. I mean, I mean, at the beginning, OpenAI was meant to be this open counter to, to Google and DeepMind and then whoever else was coming. And then, but also to your earlier point, even every company that was working on AI in a meaningful way were publishing tons of papers. It was a lot of academics that had joined the labs. And it's just, it's just in the last like 4, 3, 4 years where this started to close up a lot.

Yeah.

Yeah.

And I think it's because the commercial interests pulled the companies where they needed to go. And I, I can understand, you know, where they're coming from. When OpenAI was founded, it took however long, half a decade to get to any point where the commercial structure was clear, probably 7 years from its founding to ChatGPT. Of course, you can't predict what's going to happen 7 years from now. So you can start off with a certain intention, but then you strike gold and your product takes you, right? Where your company follows the product. And I think that that's every single large attempt at open intelligence in the field of AI, or I guess, yeah, in the field of AI, has either not made it in the sense that it was just not resourced well enough to really punch at the frontier, or the commercial interests of the company took it down a closed path because there's no— there has to be a reason why open intelligence matters for you commercially or for your product. And that's why it's really important to kind of set up both. So the way I kind of think of evolution of commercialization of artificial intelligence is almost like a Maslow's hierarchy. Of needs that need to be met, starting with consumer. That's with ChatGPT and the initial kind of set of consumer products. Actually, before ChatGPT, there was Character AI, which was probably the first viral, viral product built on top of a large language model. And that category has clearly been defined by a set of companies and prominently OpenAI. And so the consumer need has been met. I think moving further upmarket, we have cloud APIs. So where Anthropic Gemini and the GPT APIs compete, these are closed model APIs that developers build products on top of. And I would say the primary consumer of that market is the startup. And the startups have now grown to be very big. So now these startups are Some of them are large enterprises, but I kind of think about that being like the Stripe for intelligence in the same way that Stripe grew with Lyft and Uber. These APIs are growing with the startups that are building on top of them. And that brings us to the last category in Maslow's hierarchy of artificial intelligence commercial needs, and that is large enterprise. So once you move up market and there's basically nothing else beyond there because, I mean, large enterprise, what's bigger? Very large enterprise.

Yeah.

So you have large enterprise and at that point that's when there's a transition point where open models start to become important from a commercial perspective because once artificial intelligence becomes a very large line item for you and is a core part of your IP, you start thinking about how do you optimize it? How do you bring it in-house? How do you run on your infrastructure? How do you keep your data from going to the closed model providers? How do you customize the model? You want to be able to predict your costs and you want to be able to have ownership and customize the model. And so I think in part what's happened is that it took a while for the arc of commercialization of artificial intelligence to hit the point where large enterprises were ready for its adoption, because in a sense, the slowest to move, right? Consumer happens first, then startups, and then large enterprises. And so I think now is the moment in time when there is a commercial structure that— where open intelligence is fully aligned with it.

And you guys would help those companies set this up and manage it and tweak it and sort of like in the Red Hat model, or?

I think that that's roughly right, but much bigger because what you're providing is effectively the intelligence core that's going to power the company. And so if you are embedded as the intelligence core and the key provider that they, in the same way that there's like the sort of infrastructure AI core, like GPUs and the infrastructure around them, you're now part of that as the intelligence core. There are a lot more things you need to do to make an open model useful to an end customer. I mean, the very basic of it is that you serve an open model on their infrastructure via an API, but there are so many more things that you need to do to make it more universally useful. You need to help customize it. They want to probably build some internal agents around it, so you need to provide software for that. And there's the surface area of value that you can then capture and go monetize with the open model at its core.

Okay. Okay. So you're— the OpenAIs, the Anthropics, the Googles, they're kind of keeping all this stuff within their walls and just renting it out as people want. And you guys would be helping companies actually build on top of AI themselves, and your technology would be at the heart of that.

That's right. So It's not everyone needs to themselves have access to the models and build, right? Build the intelligence itself or own it or customize it. And so I think the closed model strategy is great for startups when you're just kind of getting started and you don't want to deal with all the infrastructure complexity. But once you're a very big consumer of artificial intelligence, then you actually want to own it.

Okay.

Once it becomes a very large line item, and you're spending a lot of money on it, then you want— it's very similar to how many enterprises have repatriated their cloud workloads. They first built on top of a cloud provider, but then the bill became so large that it made sense for them to bring things back in-house.

Yeah.

Okay.

In sort of slightly changing gears for a second, I mean, It's been— when did Deepseek hit? Was that like 5 or 6 months ago, or is it longer than that?

Well, V3 landed in December 2024.

Okay. And that was like this big moment where people really woke up both to sort of the state of AI in China and then, and then the techniques that they had used to make it work. And then, and then I guess the openness of it all. I remember in the weeks, you know, the first couple of weeks when it came out, there was all this panic and everything. And then I think it was Dario at Anthropic that wrote a paper. I think it was Dario. And if I remember right, he was arguing that, well, it wasn't quite as cheap as everybody thought because they'd been, you know, accessing GPUs they probably weren't supposed to have and probably spent a lot more on their infrastructure because part of the thing was like, oh my God, they've learned so quickly from the US cutting-edge techniques, but they— and they've improved them and they've done it at a fraction of the cost and I mean, do we still— now in hindsight, like, what is— what does the world think of Deepseek? Because at least that, that moment, that month of like people panicking and freaking out seems to have passed, at least in my mind, to some degree. Um, so yeah, I mean, like, now in hindsight, what do we think about Deepseek?

I think that We need to give credit where it is due and that it was an incredible accomplishment. And because what happened is that a group of basically high-frequency traders, so quants in China, who did not have the storied large language model background like some of the people who built out these labs in the US, found a way to replicate a lot of what was known internally in big labs or some version of it. I think the details are always different, but roughly these are the kinds of tricks that one would uncover at a frontier lab. And not only replicated without, because these labs aren't open, so it's not like that information was really out there. It's Oftentimes there are papers in public where there's some information, but you need to know which, which ones to look for. And so they replicated it at a great degree of excellence and shipped something with maybe not as few resources as they claimed, but still significantly less resources than the US labs that was not that far from the frontier. So I think what they did was incredible and it was a wake-up call. And I think that the reaction to it in terms of the panic was justified. And I think the only thing that's happened is that now we're living in the world where open intelligence is being dominated by Chinese labs instead of DeepSeq. There's DeepSeq and then 5 more, and we've just gotten used to it. So I actually think the panic is not high enough. It's kind of, okay, now the world is the way it is, and we're importing Chinese models.

So we, we got complacent or, or, uh, aren't thinking about this deeply enough?

I think so. Uh, however, I would say that the— but we do see, I think, panic in the right places. Or maybe— and, and the way I kind of think about this is, for example, in July, the White House released an AI action plan., which was very much about this problem of what do we do about the fact that we're importing Chinese models into the US and building on top of that technology instead of exporting it to the rest of the world, right? What do we, what do we do about this? That was kind of a, you know, the, the plan was sort of a proposal for fixing this issue. And so I think the fact that the administration is taking this as a serious issue of national security is the right way to be thinking about it. So I think that maybe while in the general public the panic has faded, I do think that we are seeing the right, I would say, strategic questions being asked.

So, and I never paid enough attention, shamefully, to what Meta was doing on the open front. I mean, was Deepseek that much better than the models Meta— because Meta had spent tons of money, right, and, and resources and, and for a while I remember was keeping up in a lot of ways with, with the proprietary, uh, Frontier models and was doing well on a lot of benchmarks. Yeah. So it, we, I mean, it was Deepseek that much ahead. Is, is Meta still, even now that they've, they've kind of changed strategy, is, is that's taken a backseat or?

So Deepseek was way better. I think Deepseek was, you know, a kind of qualitative leap. Away from anything that was being produced outside of China, Meta included, and apparently at a fraction of the cost. So I think that was— I think what was eye-opening was that, I mean, Meta has been and will be putting in just a crazy amount of CapEx into getting GPUs for training these models. And DeepSeq was able to get away with a much more powerful model for certainly something that's a fraction of the cost. Like, we never know how much, but there are export controls. The chips that Chinese labs get are, are not as good. Uh, they're obviously— they have their internal chip economy that they're trying to build up, uh, but they're, they're kind of playing— they have, um, a disadvantage on the chip front, and despite that disadvantage, uh, they're able to rise to the occasion.

Yeah, I mean, it's probably to some degree helpful. You have to figure out how to do things with less resources, right? And be more creative and innovative sometimes.

Yeah.

Yeah.

Yeah. So I think that what was surprising is that because American labs are mostly closed, people don't have a sense for what it takes to build these models. And DeepSeq was the first time that I can think of where a powerful model was released with the details of how it was trained. And the details were surprising because oftentimes, I mean, something that surprises scientists that get into this is that when you're trained as a scientist, you're trained that it's all about breakthrough kind of ideas that are oftentimes theoretical or mathematical. I mean, I was a theoretical physicist and a lot of it was around how do you creatively find these completely new ideas to approach tough problems. And that is true for the development of these large language models as well. But what surprises people is that the majority of those ideas are at the infrastructure engineering level. They're not so much at the, I need to invent a new theory of artificial intelligence or something like this. They're very much at the infrastructure optimization level. DeepSeq, when the paper came out, half the innovations were how they coupled their algorithms to fit perfectly with the hardware that they had. And I think it opened people's minds in terms of this set of problems that matters. And the fact that that happened then and is not really happening outside of it, that's exactly the issue of why do we need an open science around this? So that the community as a whole can be calibrated to the ideas that matter and be directed in their exploration, right? So we can develop this technology together.

And then part of the fear would be not only that U.S. companies are importing DeepSeq and working off it, but I mean, it would go back to this more fundamental thing that we talked about before is, um, you know, if you look over the last 30, 40 years, the U.S. crown jewels are Windows and Oracle and, and these software layers that have spread through the entire world. And, and once you're on these things, you're, you're sort of on them for forever in some ways. And, um, you know, it's really powered the U.S. economy in a lot of ways. So the— I assume another fear would be not just the, the U.S. is importing these things, but, but other countries who have not spent the resources or can't on, on all these massive data centers and don't have the scientists to train these models, that some Chinese model becomes this like de facto layer of AI to some degree.

That's exactly it. That's actually the bigger fear. I think that's the bigger fear to importing this technology, because if we want to, we can always figure out kind of ways of being kind of domestically independent. But if a technology spreads around the world, then you can't really stop that. Right. And so exactly kind of you're saying that the protocol for the internet was developed here and spread around the world. Linux and Android were developed here and spread around the world. And technology is very much one of our greatest exports and national treasures. And now we're saying the arguably the most impactful technology of our time— I mean, I think certainly of our time and arguably of the last century or something like this— is being developed right now. And we are losing the— we're losing ground in exporting it around the world. Right. And so if we don't set the standard now and export it, then someone else will. And right. And that game will be lost. So it's a very real problem that's happening right now.

Okay. Since day one, the Core Memory Podcast has been supported by the fine people at E1 Ventures. They are a young and ambitious VC firm in Silicon Valley investing in young and ambitious companies and people. Thank you so much to E1 Ventures for all your support. Maybe I'm just like not running in the right circles. Going back to some of the stuff I was talking about, you know, I used to go see Bill Gates speak and there would be protesters outside. I mean, all the free software people, the open source people. I mean, this idea, you know, Scott McNeely even, I mean, hardcore capitalists were up there like calling, you know, um, well, I think Ballmer, Steve Ballmer called Linux a cancer. I mean, all this, this stuff was very, was so heightened. Um, Clearly you've created a company to try to combat this, but why is, why is this not more of a, why is this not more of like a religious and intellectual war this time around? Or am I, am I just missing what's going on and people really are quite animated about this?

I think certainly in the research community, people are quite animated around this. It's a good question, but One thing is that AI research at scale is only accessible to a very small number of people. And so that might be part of it. I think that in the previous iteration of open source, a developer could quickly take something, fork something, and start building. In this instance, that's true if you do things at very small scale, but at large scale, it's not just because of the table stakes are so high. Yeah, but it's a good question why, at least in the research community, it is very much the case that when we talk to candidates or our friends, that there's a real real need for open intelligence and open science. I think a lot of people miss the days of what Google Brain used to look like about a decade ago when self-attention was being invented and all these things were being done in the open. And I think maybe there's a sort of, well, it is what it is kind of thing, because, well, we understand these things take a lot of money to fund and you have to make money, you have to generate revenue. So it makes sense why it would be closed because for the consumer and kind of cloud API products, there's, you know, opening these up only puts you at a competitive disadvantage. So that might be part of it. But I don't understand why— there are some people who are very vocal, I'd say, particularly from within the government. They have people close to AI policy are very vocal. Obviously there are some members kind of in technology that are vocal, but by and large, I think that you're right. For some reason, it— it— I wish there was more of kind of community banding around it vocally because that would help us kind of change the course of things.

I feel like those— I'm remembering people like Richard Stallman, Eben Moglen, who was this lawyer who used to defend a lot of the free software stuff. Maybe like Lawrence Lessig, another lawyer to a degree, these, these, the intellectuals who would make these arguments around free and open source software. I feel like those similar figures now are largely kind of the doomer. Like, I'm thinking of someone like Eliezer, you know, who the people who are where the energy is focused really seems where that, that passion, that intellectual energy is really on sort of is AI going to destroy us all or not? And a little bit, you know, maybe that has sucked some of the energy away from this other question that, that is like intermingled in all this a bit.

That is true, right? The notion of AI safety is a very emotionally charged thing because, well, you can imagine a whole spectrum of issues around safety, but where your mind first goes is the Terminator.

Yeah.

Instance of a superintelligence that starts manipulating you. And I think that there is a notion among some members of the safety community that the safest way to develop AI is in this very closed setting. That is certainly kind of a sentiment that exists. And I think it comes from— what I disagree with in that sentiment is that typically It's never the person outside of the closed lab that is saying we need to develop, you know, the only way to do safety is in this closed way. It's always the researcher within the closed lab. And effectively it's saying, I will figure out safety for you, that that's— I'll be one of the 100 people in the world who has a say on how to develop these safe models and we'll figure it out and you're welcome.

Yeah.

And I find that to be very frightening. Because safety should be something that is in public discourse and that should not be gated. Every research institution should be able to contribute to safety discourse.

Yeah.

And the only way to do that is if you do the science and put the models in the open. Uh, that's the only way to ensure that there's a diverse ecosystem around safety research. Now, the argument from in a closed model lab about why you need to do safety in a closed way is that, well, the models will get powerful enough where if you release the weights, then someone will take them and wreak havoc with them. I think that that argument has weight to it. That is something that I fear as well, that once these models become capable, Reflections models or a Chinese model, and more likely it's going to be a model outside of this company, uh, ends up getting, right, leaked. And then you have these powerful open weights that an adversary can take and start—

yeah, it's like you just released exactly how to make a nuclear bomb, right?

Now the thing is that This is an inevitability that I believe will happen with or without us. I believe that there are 5 or 6 Chinese labs releasing great open models. It is inevitable that someone releases a powerful open model, and that should be part of the safety discourse. What do we do then? We can't just say we'll be the closed lab that we won't— as a closed lab will ensure safety by not releasing open models, someone else will. Even if you do very careful know-your-customer deployments of an open model, because open doesn't mean that you, for some set of models, it means you put the link on Hugging Face and they're just freely deployable. But for very powerful ones, it could mean that you have a know-your-customer access to it. In the same way that when you set up a bank, you can't just— right, you have to go through some checks before you do that. Even in the world where that happens, someone at some point is going to put a model on their thumb drive and release it somehow. So I think we need to be thinking about safety in this more holistic sense and addressing that problem directly rather than pretending it won't happen.

It won't happen. It's interesting that China went this route. I mean, is your working assumption that they— there's so many open models there because it's, it's like some asymmetric form of getting distribution because, because the US had already established such an early start in all this. ChatGPT is this big brand. And so this is a way to get people to, to pick up your product because, you know, I remember, I remember not so long ago people would argue to me like, AI is not going to develop in China nearly as quickly as the US because the Chinese government is actually like quite afraid of it because it can't really like control fully the results. It doesn't want some intelligence that's, you know, better than the government is at things, and people know that. And so that it was gonna, you know, keep very tight controls on, on how this technology developed and perhaps actually try to slow it down. Um, And then now we see, like you mentioned, I mean, this influx of, of open models there. Why do you think this took off, this approach took off in China?

I think that it was deeply contextualized in what was happening in the States. There was this narrative and I would say perception that the US was just light-years ahead of China when it comes to artificial general intelligence, that we had both the closed model frontier, but then Meta was releasing the best open models. So we also had the open model frontier and nothing even close to our technology was coming from China. And I think that there's one world where, well, Chinese companies could have built that internally and then maybe people in China would have known, but across the globe, They would not. And so I think that the open models were— I mean, part of it I think was the founder DNA. I think that with Deepseek, there was kind of a special founder and team that was already making a lot of money from their quantitative, from their quant high-frequency trading product. And so I think part of it is that they were able to do so and had You know, I think, I think it was coming from a place of sincerity, of just wanting to do something good for the world. But I also think that there are geopolitical influences around because it's not just Deepseek. There are 5, there are 5 such companies that were counter to what was happening in the US in order to exert influence on the world more, more broadly. And in some sense, if things continue the way that they do, I think we risk winning the battle but losing the war sort of outcome where, yes, we have great businesses running on closed models, but the open model standard gets set by China and exported everywhere. And then there are going to be large businesses built on top of that. Right. Which, which effectively we miss out on or don't really get to influence. So it reminds me of how during the internet era, companies used to get a lot of users before monetizing. That's kind of what's happening right now with Open Intelligence. But then once you hit scale, like Google hit scale and then started introducing ads, well, then that became one of the biggest businesses in the world.

Yeah. So, okay. You know, you guys raised $2 billion. It's a staggering sum. I was just at OpenAI this week and I won't disclose like the exact figures I was seeing in these meetings. It was around their data infrastructure buildout. But, you know, I mean, just, just incredible amounts of money that they're intending to spend on, on data centers in the coming years. I'm sure you've gotten this question before. It's just how, how do you possibly as a How do you possibly keep up with, with these companies at this point, given how much it just seems the more compute we throw these things, the better they do and you got to train them and it costs huge sums of money. Are you, are you guys arguing at Reflection that you have such a clever group of engineers that you can kind of do what Deepseek did and have these different techniques from everybody else to get the most out of these systems and do things in a different way? Is that— how do you keep up?

Yeah. I definitely will say, I mean, certainly there's an excellent team here and we're driving innovation along the frontier of techniques of training these things. But I think even for DeepSeq, that was not enough. That's just one aspect of what it takes to build a new frontier lab that will be competitive. I think the other aspect is you have to be embedded in a moment in time where it's strategic for you to exist and where you'll be able to survive the releases that are not at the overall frontier. Every single lab, when it gets founded, it releases models that are the first models. I've yet to see a company that released a model that was playing at the foundation model game that was at the overall frontier. It took Anthropic 3 generations of models to get there. It took xAI 3 generations of models to get there. And so each of these companies, when they've done this, there's been a strategic reason for them to exist. And DeepSeq included. What was launched in December of 2024 was DeepSeq V3. So it was their third iteration of the model. The reason we can exist and thrive in this specific moment in time is because of the vacuum in open intelligence that exists in the West, in the States in particular. There's so much demand for open models right now that they do have to be capable. There's a minimum level of capability, right? You have to be competitive with the Chinese models, but you don't yet have to be competitive with the overall frontier. And so you have kind of your foot in the door in that way, because what you need to build a new frontier lab is basically the team, capital and compute, right? And a strategic partner and a commercial opportunity that will enable you to sustain kind of the capitalization rounds. And you need all those three things to happen for a new frontier lab to be built.. And I think right now is the moment in time where that's possible with open intelligence.

Okay. Okay. So you, you need to stay competitive, but, but there's just such a difference in your approach that, that, that can be effective to some degree in and of itself.

I think so. And like when we think about the way Anthropic progressed, I mean, they were a language model company, just like OpenAI is a language model company. And yes, there is a focus on coding, but that was not really discovered until later in the generations of models that were being trained. There were strategic reasons that I believe made Anthropic possible at that time, largely because it was the only other large language model startup out there. And hyperscalers are placing their bets with OpenAI, Microsoft, and than Amazon and Anthropic. And I think each time an intelligence lab has been built like this, there's just been— it's, it's not just the talent. Like, the talent is just kind of a baseline. Like, if you don't have the talent, you can't even do it. But the talent is just a baseline for getting started. And then you need these other things to come in place for something generational to come out of it.

Okay. Um, I don't know if this question's cynical or not. This is, I just, when I was doing my research for, for the episode, I mean, it seemed like you guys, you, you kind of came out of stealth mode in, in, or announced your existence in 2024. And, and the initial wave of stories and the way you seem to be presenting yourself was, oh, we're gonna go out and create agents and, and we are gonna go after coding. First that we're going to make a better coding agent. And, and then we think, we think the world, at least this is my interpretation, was, was we think there could be a variety of agents and we're going to go off and build those. And then when you announced this new funding round, there was much more of this emphasis on the open models.

Yeah.

Was this like, is this— a cynic could look back on this and sort of be like, oh, okay, they've adopted, you know, this, this rhetoric. Because it feels sexy and this is like a better sales pitch than having yet another coding agent, or was this always the master plan?

Building an open intelligence company—

Am I right on some of those secrets?

Building an open intelligence company was not the initial master plan of what we were doing. We started before, you know, when we started, we actually bet, we were betting that there would be other providers that build good open intelligence. We started early 2024, Llama 2 had been coming out then, Mistral launched, and there's like this momentum around open models. So we believed someone would build it. And our bet and mission as a company was that, well, we're going to build frontier agentic models. Like that was, and when we said coding, it was always because we believed coding was just the kind of skill that you need to build just a general agentic model because the way a model will interface with software is through code. So our mission, and this is what we told our employees when they came in and our initial investors, both through the seed and A, is that we're building super intelligent agentic models. So frontier agentic intelligence. We believe two things, that one, reinforcement learning is, as a technology, is going to be the thing that will drive that frontier of knowledge. That's how we go from chatbots to agents. And that happened. And we are one of the early companies to witness that and see those results happening internally. And the other bet that we made was that there would be powerful open models coming from the US that we'd be able to train on. That someone else would put the bill for getting these powerful pre-trained checkpoints and we'd be able to train on top of them. And that's why we naively thought we could do this with around $100 million. And what ended up happening was that, well, the reinforcement learning and coding bet turned out to be right. When we think about, we don't even really say it as much anymore because now I think it's just obvious that frontier agentic models are coding models. That's what Claude is. And so when we say frontier open intelligence, we mean frontier agentic intelligence. That's kind of where the world is going. But what changed was that over the last year, the best open models came from China and not from the US. And that results in a very concrete problem for companies building on top of that intelligence is that you can't sell it. When you go to an enterprise customer and you say, look at this capability, I've driven this capability in code, they say, that's really interesting. But then they find out that you're training on top of a Chinese model and they say, well, we'll use it for a prototype, but we can't put it in production. And so our evolution was that we realized Well, we either, you know, you either have to find the partner that will build the open intelligence or you have to become that company. But becoming that company requires immense amounts of capital to even have a ticket to play the game. So we did not start as an open intelligence company because we thought someone else would do it, but we evolved into one.

Okay, that's interesting. I mean, it's got to be I guess exhilarating and challenging and scary to kind of like go back to your existing investors and other investors and be like, we're tweaking the plan and we're going to need a lot more money, but it's going to be, it's going to be more awesome.

Well, the interesting thing is that from an internal perspective, the vast— everyone at the company was working on setting up a frontier language model training stack. And building out reinforcement learning capabilities and making these models good at coding. And so from that perspective, you know, our team just, just continues doing what they're doing and now expanding the capabilities. Right. And bringing in full-on pre- and post-training in-house. Yeah. And I think that from an investor perspective, I mean, ultimately we We are quite— I mean, we're quite transparent with our investors and the board, and we've had— we have these discussions in open, right? That's kind of— we're building this thing together. And so it didn't come as a surprise to anyone that this is what needed to happen.

Okay. And you were— you were born in Russia, grew up in Israel. Were educated in, in Canada, at least in part, right?

Uh, not in Canada.

Not in Canada. Yeah, I got that wrong, right? You ended up at Berkeley, University of Chicago.

Yeah, I kind of— I lived in so many places. Uh, I was born in Russia, moved to Israel as a kid, then grew up in Washington State.

Washington State. Sorry, that's right, that's right.

Yeah, yeah, uh, yeah, in the desert in Washington State. Uh, And then studied physics at Yale and then UChicago, did my PhD there for theoretical physics. Actually, when I saw AlphaGo, realized that I needed to get into AI and then did a postdoc at Berkeley where I sort of had a— it was my kind of 2-year boot camp into getting onboarded as a researcher into frontier AI. Yeah.

Peter Biel, he's influenced many people. And your co-founder is Greek, I think. Yes. Like, isn't the other cynical take, it's like, why, why does a Greek guy and a Russian-Israeli guy care about, you know, the US infrastructure over the next 100 years?

I actually think that one is the opposite. This country is built on by immigrants and You know, when we think about who's the biggest proponent right now of open intelligence and exporting it, it's Jensen, and he's Taiwanese. So, right, we are immigrants. And I've actually seen, you know, I became an American citizen sometime in college, and my family is the only part of the family that, you know, moved to the States. And I've seen what the life rollouts are like in Russia and Israel. These are both countries that are, you know, at war, um, completely different lives there. And I think that immigrants coming to this country are 1 or 2 standard deviations above the median in patriotism. Like, it's really incredible what this country can do for people coming in who are willing to work hard like, you know, you get to live the American dream, which is so much more incredible than the way your life can basically roll out in another country.

Yeah, yeah, yeah. No, fair enough. When you were, when you were at DeepMind and you were working on, on what was it, Gemini 1.5, I think.

Yeah, the first one, 1.5.

Yeah. I mean, what was it, what was it like internally as I remember when DeepMind was created and it was— Google already seemed like the AI powerhouse around whatever AI was at that time, in the sense that it had all these data centers. It had, um, it was the only company that was doing really interesting things around automated translation and some early, um, photo image recognition type things. And then it acquired DeepMind and it felt like DeepMind had run around the world vacuuming up all the most interesting academics. I remember sitting back kind of being like, oh well, they've, they've won. Nobody else even really seems to be taking this remotely as seriously. Obviously Elon and Sam and Greg Brockman, a bunch of other people had the same idea and they're like, we have to create some kind of counter to this. But, but you know, when you were within DeepMind and, and you're working on Gemini, but ChatGPT is taking off and OpenAI is having this sudden rise, I mean, what was like morale like inside of the company? And yeah, I was— I'm always curious what that must have felt like. It just, uh, this thing sort of came out of nowhere and all of a sudden OpenAI was the king of the AI world.

Well, each AI company has its arc and you always want to be, you know, in the golden era. And I would say that DeepMind's golden era was probably during AlphaGo days.

Just incredible. Yeah.

Yeah. And that was, you know, what got me into AI. I wanted to be part of that arc, the golden era arc of a frontier lab and be contributing to breakthroughs that are happening at the frontier. So when someone else drives a frontier and it happens under your nose, when everything was in your control to be that company because all of the inventions that OpenAI scaled up, most of them came from Google Brain and DeepMind. So I think that I can speak to my morale. It was one, it was very exciting to be contributing to frontier language model research, but it was deeply unsatisfying catching up. Like that. It was unsatisfying. It's one thing to catch up when the initial set of things was out of your control, but it's another thing when you could have— you should have been, right, the team and company that did this. So being in a mode where you are reactive rather than leading the field and where you're no longer really aligned with the initial, like with the initial mission of the company. Right. And there can be, you know, people can say that they're like the company has words around its mission, around its building AGI. But when you're on the ground and you're kind of doing the work, you kind of feel what mission you're actually— right, what mission you're actually going after. And the mission of the mission of replicating something that someone else has done is just not— yeah, on its own, right, was not that inspiring. Like, obviously it has to be like stepping stones, but you want to be leading the field. Yeah.

And so, and how— I mean, it feels like they've corrected to some degree and they had this big run where they were rolling out such a impressive array of products. I mean, how do you think this got corrected? I mean, it was just, it was a matter of time or there was some call to arms, you know?

Well, they're producing incredible models and I was pretty confident that is what would happen. I mean, it's still the deepest bench of AI talent in the world. I think that Many people don't realize how many talented people are at DeepMind. And so when some leave, you know, like us, there are more people that are really good. So I think that one is that the talent bench is incredible. Obviously there's a compute advantage with being fully integrated with TPUs and the leadership is research leadership. Demis still runs the company. The people who are really leads of the projects were, a lot of them were at DeepMind very early and saw it through DP Networks, AlphaGo, AlphaStar, that the people in leadership are researchers. And I think that that is really important. So I think that what DeepMind's been able to do is pretty incredible. They're producing really good models.

Is, is Temes, or, and then Sergey kind of comes back and starts working. I mean, are either of them sort of like rah rah? I don't picture them as big speechmakers, but maybe I'm wrong. I mean, you know, at some point I just pictured someone being like, wake up guys, let's go, you know, and, or maybe everyone feels this on their own.

Well, I do think that, I mean, they're both inspiring in their own ways, but there were, I think there were some news articles around, right? Sergey going in and actually being involved.

Yeah, like 6 days a week and sitting there.

Yeah, right. And that's really inspiring, right? That when the founder of one of the biggest companies in the world is coming in and spending time directly with, with the team and, and Demis is just a very strategic leader. He's a, he's, you know, he's very strategic. And I think sincere leader in his desire to build an artificial general intelligence and make it beneficial. And so I think he's much more of actually a tempered leader. When people speak about AGI timelines, Demis tends to be more conservative on his. So I think that he's been doing this for now, 2010, must be 15 years, uh, and, uh, he's a very inspiring leader.

My sense— this is not asking you to put words in his mouth, I, I'm, I'm— maybe you could check my sense of reality. I mean, my sense was always, you know, DeepMind was out to create this very sophisticated, um, based off the AlphaGo stuff. It was like, we keep doing games We do more and more sophisticated games. Our AI goes deeper and deeper, and at some point we create this very elegant machine that understands the real world and the game that is reality. And it's this beautiful form, this elegant form of AI. And where we ended up with LLMs was something that can feel much cruder, which is you hoover up, you just trillions of words and you just throw immense amounts of GPUs at them and some form of intelligence bubbles up out of it that we don't fully understand. And just the more you do this, you know, there's a brute force mechanism to it that I think felt a bit anathema to my sense of what DeepMind was always pursuing. Do— does that ring true at all, or is that like too simple of a reduction.

No, that definitely rings true. I mean, I think that it's hard. I mean, the first thing to say is that it's hard to— you have your kind of abstract notion of how you believe intelligence should develop, and then you're surprised. So some of these things are just hard to predict. I think it was very hard for anyone to predict that language models were going to be so powerful. They were one of many things that people studied. People studied language models, NLP and computer vision and reinforcement learning. And right there, they were just one of many things. And some people believed that they would be, they'd have these emergent capabilities, but not that many. So I think that first it's kind of hard to set a prescriptive vision of how you believe artificial general intelligence will develop years in advance and for that to actually be the case. But I think that there's another point that you're making, which is, I mean, one way to put it is there's this essay, which I'm sure you've read, and a lot of, I think it's kind of the essay that's defined AI, which is The Bitter Lesson. And The Bitter Lesson basically says that the methods that we should be pursuing in artificial intelligence, the ones that will basically scale or are going to be ones that maximally leverage compute over human heuristics. And the two ways of leveraging compute are learning and search. But fundamentally, it's a point against injecting human heuristics into a system. And people often use this Bitter Lesson argument to say that language models are so-called Bitter Lesson pilled. That you just throw a ton of compute, but they're actually not maximally bitter lesson-tilled. And I think that's the point that some people are making, Rich Sutton included, that you inject a lot of human heuristics by doing imitation learning on the internet and that the most bitter lesson-tilled thing to do would be to figure out a method where an agent effectively learns from scratch and has no human knowledge. It's just on its own, learns how to do things. And probably the best example of this was AlphaZero, which ended up being much more powerful than AlphaGo, but needed a lot more compute. So I still think that there's a lot of opportunity for blue sky research. It's very hard to find the areas where things that are truly bitter lesson pilled overlap with the practicalities of how much compute you have. So I think that's like AlphaZero required so much more compute than AlphaGo to achieve roughly the same results and then it became even better. But to get to the same point that AlphaGo had, AlphaZero required an inordinate amount of compute.

All right, I'm going to ask you a couple questions which are not terribly insightful. I think we're all battling with some form of these, but I'm just super curious to see where your head is at. You know, I mean, we've had this— I've been working on this OpenAI book for like 3 years, and, and, you know, in the different cycles of that, um, at one point, you know, like ChatGPT comes out, and then it's like, oh, by the time we get to GPT-5, the superintelligence will be here. It felt like the whole world kind of felt like that for a moment. And then, um, I think I'm like at this stage now where, um, a lot of this— I lived through the dot-com boom. We were talking about this a little bit before we started recording, but some of this feels familiar to me, where it's like, you know, this thing is going to change the world, but now it feels like it kind of oozes out over you over time, as opposed to just a moment of shock, total shock, and on— the world has kind of changed overnight. Um, it's feeling to me like there's tons of progress still to be made with existing techniques, applying more compute, but that we haven't had seen something quite on the order of like the transformer that pushes the field forward in a long time. I don't know. How are you feeling? I mean, this goes back to some of the stuff we were just talking about is like, are we still— do we just keep hammering away at this and we see a clear path to whatever people want to define as superintelligence? Or are we, are we now more clear that we are elegant breakthroughs away from achieving that.

One of the predictions that I had kind of in mind when we started the company was that I believe we were 2, 3 years at the time away from—

I saw that.

Yeah. Yeah. From a superintelligence.. And I still believe that to be true. It's just that when we think about it, it's going to be— I think the way I might have mentioned at the time was a jagged superintelligence. And when we are seeing the progress with reinforcement learning starting to scale and starting to win IMO gold medals and these various coding competitions, And I think we're at the brink of these systems being able to make, contribute scientific discoveries, maybe not, you know, huge breakthroughs yet, but there is some net new knowledge discovery I think that is probably happening now. We just don't know it. And soon on the horizon. So I think within, narrow domains, we are within reach of a superintelligence. And that's maybe not too surprising. I mean, we've had superintelligence in the game of Go and chess and so forth. But when we think about superintelligence within certain problems in physics or mathematics, that's— you've significantly expanded. It's much less narrow than Go. So these are categories that are general enough to be meaningful. Now, I don't believe we'll see a full-on superintelligent mathematician. We'll see superintelligence within slivers of mathematics because each field, as we think of it, is actually a collection of many subfields. And some language models, the current training paradigm scales into them, others it doesn't scale into them yet. So I think I stand by that we're going to be seeing a jagged superintelligence. At this point, very soon. I think within, within the next year we'll be seeing drag at superintelligent systems.

Okay. Okay. So when people are— because everyone's obviously still— they're debating in this moment, are we, are we, you know, about to hit— is the bubble about to burst or is the bubble not even big enough for really what's going on? So, I mean, you're a little more— you're more on the optimistic camp of this. We're still— things are going, going quite dramatically.

I think so. I think we kind of, from inception of the company, which is, let's say, independent of the company, it's really just AI timelines. Early 2024, the field was in a place where it's, oh, we scale these up to be good at chat. We have no idea how to make good agents out of them. Or it kind of seemed like this is kind of all they're going to be good at. Most people at the time actually believed that reinforcement learning doesn't work because it's complex, there's reward hacking. And so the belief was just keep scaling pre-training. But then there were basically the next generation of large pre-training runs weren't going as well as planned. And so there's, I think, this thought that we're at a plateau. And then the reasoning models came about and that was a binary event that happened a year ago. And from the reasoning models coming out and being good at grade school math and maybe some high school math competitions, we now have these systems that are winning IMO gold medals within 12 months. And we've gone from systems that— maybe one way to also think about it is the emergence and the disappearance of the term vibe coding. Vibe coding referred to these systems that are sitting in your IDE and you are coding partners, but you're deeply in the loop as a human. That's why you're telling them what to do and they're filling out some piece of code for you. I remember that kind of term peaking at some point, and I don't really hear it as much now.

It was a quick run.

Because these systems have been replaced by agents that now do fully autonomous chunks of work. And so now it's not really vibe coding, it's more like prescriptive coding. You're giving them descriptions of what you want to do, and then you're following them and giving them a to-do list. And that's all happened in a year. And I think that the progress is actually so fast that we don't quite appreciate kind of— and also concerning how fast it is. Like, I think that it's kind of a frog boiling in water sort of situation where, right, It's happening. Intelligence is improving at an incredible rate that is both really exciting but also alarming. And there might be, you know, I'm not an expert in kind of economics or anything like this, so there might be some local kind of, you know, adjustments or something like this. But we are in a medium and long-term trend of incredible capabilities coming out of these systems.

Okay, okay. Random question. Very random. I mean, there's two— there's— I feel like there's two giant wildcards that don't get talked about quite as much because they're, they're more secretive, I guess, which is Ilya and John Carmack. And, you know, you're at the top of the AI research circles. Are people actively talking about what they may be? I feel like they're going to— one of them is just going to spring up one day and be like, surprise, motherfuckers. You know, I've built this thing. Do you guys— are people like actively trying to kind of figure out what they're up to? Do you— I don't know. I'm always fascinated that they kind of had their big thing, said, you know, we're off trying to do something and then have managed to stay underground.

I mean, it's all— I know, I know, you know, just as much as you do on this aspect. I mean, certainly I think SSI has been quite secretive. I think that, I mean, John Carmack's startup, the last thing I saw them write was basically training a reinforcement learning agent that was actually embodied and using a real life joystick to play Atari. So as opposed to, so deep Q networks, which were the first system that kind of showed deep reinforcement learning working at scale. Show that you could train the system to play Atari video games, but it was still kind of using, you know, it wasn't using like an actual joystick or something like this.

Yeah, I remember going to Rich Sutton's lab and seeing them play all these video games. So yeah, just the computerized version of it.

Yeah, right. And it's, I mean, I think it's really interesting, but also quite wild. Because it's kind of like saying, AlphaGo was also done in a purely computerized way. And so here it's kind of, you're thinking about not just the thing that thinks, but also the robot that moves the pieces. So I mean, that's the last thing I saw come out of there. And I think it's really good to have people pursuing ideas outside of this paradigm because we are— it is actually not a maximally bitter lesson-tilled paradigm. There are probably better paradigms. And I think it's great that there are some people pursuing the more blue-sky research. Probably it'd be better if there is more— if there are more opportunities for that, because I kind of think that language models sort of sucked the air out of the room for other things.

Yeah, I mean, you get kind of too focused on one. I mean, I guess that was John's sort of bet The— in like robotics, I mean, same thing. I always get excited and then I get let down. All this money has been poured into creating models now for robotics. Do you, do you have an intuitive sense on, on if you think we finally cracked the code and they really, you know, become much more sophisticated than we had expected?

Well, I was, I was a robotics AI researcher in Pieter Abbeel's lab, and I remember the reason I There was a paper where after that paper I was like, maybe I should just focus on the intelligence problem alone because robotics is kind of hardware and intelligence. It's kind of you have these two problems that you're trying to solve. But my anecdote there was that I was advising a couple of students who— this was during COVID so they took the robots back home to their apartments and they collected data. Overnight. And then during the next day and for a week or two, they were trying to understand why when they collected data and trained the robot and it works at some times but it doesn't work at other times, so what's going on? And it turned out that they collected the data at night but were evaluating the robot during the day. And so the data, like there was a data mismatch. By the way, what they decided to do with that problem is to basically become nocturnal and sleep during the day. I think that I advise them otherwise. But that experience kind of made me think about, well, it'd be much better. Like, there's so many problems to solve on just the intelligence side that is there a more— is there a way where you can kind of take all the other hard things out and just focus on intelligence? So that got me thinking about digital intelligence. Yeah, but seeing the progress from then to now in robotics, it has also been tremendous. I mean, probably not as tremendous as what happened with large language models, just because that was, I mean, I think really took the world by storm. But I'm very excited about what's happening in robotics. I kind of think that locomotion, just the ability to walk around in a generalized way, is solved, which was not at all the case then. So there's probably a pretty significant overhang of capabilities there. I think integrating them into real-world settings is tough. So I mean, that's a challenge. And I would say the last frontier is manipulation. But the amount of progress being made in robotics has actually been pretty incredible as well. Like deep learning, the field is not large language models, it's deep learning. And deep learning lifts all tides. And I've seen that everywhere.

Like some of the stuff I see is awesome. I've been covering all these robot fights, and so I've been watching tons of videos and You know, I know we used to always have the Boston Dynamics videos of the dogs doing all kinds of things, and then people would kind of question how much was scripted and all that. But, you know, I see these humanoids flipping, getting pushed over, jumping up and all that. But then I get excited and then I get really depressed because, you know, Figure comes out and it's like the robot's still folding a shirt. And we're like— and, you know, I go into these other labs and it's full of these arms, you know, opening and closing a microwave door over and over again. I was like, man, I thought we were doing that like 12 years ago. I'm like, like, this is still the frontier of what's going on. And so I, I, uh, I keep getting excited and then I see that and I'm like, oh, I don't know, maybe we're not, we haven't got there. But then kind of like you're talking about with, with China, I'm not like a huge China fearmonger. I mean, I, I, um, like you, I'm an immigrant. I appreciate the United States and Oh, the opportunity. I wish China and US, 'cause there's like kind of, um, similar things down at times. I, I feel like we've got like a lot more in common than, than not. Um, but when I see like you're talking, yeah, yeah. I took this trip to Shenzhen and I went to DJI and it was while I was there, it was DJI had decim— destroyed all the US robotics, uh, drone companies. Um,. And then, you know, I was kind of like, man, how does this, how does it like keep advancing? Like what's, they just so dominant. And then I realized, oh, they have like 5 competitors down the street from them who are like, the second DJI puts something out, it's like their Chinese competitors like steal the design, make it cheaper. And then DJI responds to that and is like, their CEO is just relentless. You know, it was just like a way of life. And so I see that now with the humanoid companies, there's like 12 or 16 Chinese companies that I think are going to sharpen each other more than like the US competition will. And then when I think about what you're talking about with these, these open models, and then it's a very long-winded sort of thought, but the thing I like about the humanoids really is that instantiation of AI in something. And it feels very powerful to me. And I don't know when I— now I'm just adding all this up in my head. I mean, you could see a scenario where China does reach some kind of escape velocity compared to everywhere else because it's able to meld the software side and the hardware side and all this.

Yeah.

Companies pushing each other in a way that I'm just not sure the US is like fully wired right now in the same, same way. I don't even know that's a question.

No, it's— well, it's There is a question within there. And I mean, one thing that is like all this stuff is interconnected and the humanoid companies, I mean, a lot of that is done with currently pretty small neural networks or sometimes it's not even deep learning necessarily, even though that's become more and more the case that deep learning is kind of taking over the kind of locomotion and robotics, but their brains are going to be powered by open models. They're going to be open, like the things that these language models and foundation models— like when I think about multimodal, right, they can see and they can think, right? They endow them with reasoning abilities. And so these are the kinds of companies that are going to be building on top of open models. And certainly if the greatest open models are coming from China, they'll have basically the full stack kind of vertical advantage. And I'm not really fearmongering around China either. I think that actually what's happened with the open models that come out of there was a net positive for the world because otherwise I think what would've happened in the States is a very likely concentration of AI development and just closed labs and no incentive to really do anything else about it. But we have to understand that we are in a world that has two poles that can emanate the most powerful technology of our time. It's America and China. I don't think that there's any other region in the world that has the talent, capital, infrastructure to really build these things. And it has to be balanced, right? It has to be— you have to have great models coming from, from both places.

It's amazing. Poor— if Russia could ever get their act together on many fronts, they could have been a— I mean, so, you know, there's so many brilliant Russian mathematicians and physicists and computer scientists. I'm sure they could have also played here.

Well, they were a powerhouse in theoretical mathematics and physics. Yeah, that's because in the— well, in the Soviet Union, all you needed was pencil and paper.

Yeah. Yeah. Just an incredible education system. Well, I want to be respectful of your time. Thank you so much for joining us. Is it like— I know, I don't think if I remember right, you guys haven't quite— it's like next year where you're coming with really debuting and showing your technology for the first time. Is that right? And what should people expect from Reflection in the coming months?

So, right, the capital enables us to set up large-scale training runs to build these models, which takes some time. We are an open intelligence and open science company. And so people should expect us to be releasing artifacts and findings and basically research that we think is interesting, that we think that it'd be useful if other people know about it as well, kind of the internals of how Frontier Lab gets built. We'll be releasing those along the way and leading up to the model.

Okay. And you guys, you— I think you've been based in New York. Are you moving? Where will the company be centered?

So the company actually has 3 locations in San Francisco, New York, and London. Okay. And the reason is that that's how when we were— when we were at DeepMind, a lot of kind of AI talent is actually distributed across these 3 areas. And so That's kind of how we set the company up. Our largest presence is actually going to be in SF because you can just kind of follow the sort of distribution of talent.

Yeah, no, it's just when I saw New York a bunch, I was like, man, this is going to be tough.

Yeah, no, I'm here. I live in SF. Yeah, we have a great team here and, you know, great team in like in both the other locations as well. So we— and we actually kind of, you know, fly in people to spend time together between the various locations.

Yeah. Super nerdy. Last question only because I'm a— I am a historic open source dork. I mean, I think DeepSeq released their weights under the MIT license. Is that what— what are you— do you guys know what you'll do? Sure. There's like 10 people that care about this or maybe more. Yeah.

Well, I think that the spirit of it is that you want to have a, you want to have a kind of, you know, for the majority of your models, right? You want to have a fully permissive open license that people can do research and use for commercial use freely. Now, whether it's MIT or Apache 2 or something else, right? There's some regulation that's coming in around, you know, various aspects of model training. And so that might, you know, maybe some new versions of like the MIT or Apache 2 style license will be been brought to fruition.

Okay.

But the intention is very much, right, you're releasing open models for research and commercial purposes, that they're fully free. Now, they're going to be a set of models that might be trained on data that is industry-specific, or there's going to be a set of models that are probably dual licensed, where it's free for research, but we need to set up a license. To kind of use them. But by and large, most of the models will be kind of fully permissive. We want people to just not think about it, take it and use it. That's our intention.

Okay. Okay. All right. Well, thank you so much for joining us. Really appreciate it. It was the first time we met. It was really cool. I look forward to following what you guys do. I think, yeah, this stuff is like— I'm an Android user. I get made fun of all the time. But I mean, I do actually do it because it's open source and, uh, there are dozens of us. I mean, you're right. There's like a billion, you know, but we get so much grief among, uh, among the Silicon Valley crowd. So, but yeah, it's always been near and dear to my heart. Got a Linux desktop. Um, uh, so anyway, I'm, I'm, I'm into what you guys are doing. Yeah.

Well, thank you for having me, Ashlee Vance. This is fun.

All right, man. Thanks. The Core Memory Podcast is hosted by me, Ashlee Vance. It is produced by David Nicholson and me. Our theme song is by James Mercer and Jon Sortland, and the show is edited by the Jon Sortland. Thanks as always to Brex, NE1 Ventures for making this possible. Please visit our Substack, YouTube, and podcast channels to get more of what Core Memory makes. Thanks, y'all.