We saw this geopolitical tension coming, where if it's only one country that's producing open models, then the rest of the world kind of becomes dependent on those models and on the infrastructure of that country. And it's extremely important for America as a country to be competitive, and not just competitive, but have the best open-source software and models when it comes to AI, and ensure that our allies, that the world, and domestically, we all build on our own open intelligence.
Support for the show comes from Engine. Every year, $1.6 trillion moves through travel. Most of it still runs on technology built decades ago. It works. It's just not intelligent. Engine spent the last 8 years building the foundation that intelligent travel requires: robust inventory, hotel relationships, pricing intelligence, and payment rails. The unglamorous infrastructure work that simply can't be bought. Today, that foundation runs travel for nearly 40,000 businesses and 2 million travelers, and it's growing 60% a year. Intelligent travel. The future of business travel runs on Engine. Learn more at engine.com/propg.
Welcome to the PropG Markets Founder Series. I'm Edson. There's no shortage of concerns surrounding the AI buildout. Communities are pushing back against data centers. Workers are worried about being replaced. And some fear the technology could pose an existential threat. But one of the most immediate challenges for the US might be the race against China, where open-weight AI models are gaining ground by offering a cheaper, faster alternative to the most powerful proprietary systems. Our next guest is building a company designed to give the US a stronger foothold in that race. The startup recently released its first model, Beam, which it says outperforms other Western open models while being cheaper and faster to run. As the competition with China intensifies, the company is increasingly seen as a strategically important player in America's AI buildout, and investors are taking notice. Backed by firms including Sequoia, Citigroup, and Nvidia, the company closed its June funding round at a $25 billion pre-money valuation. So, can this new model give American AI a competitive edge against China? And what would it take for the US to win the race in open-source AI? This is my conversation with Misha Laskin, co-founder and CEO of Reflection. Misha, thank you for joining me. We'll get into open weights, which is really what I want to focus on in this conversation because it's such an important topic that I think a lot of people don't fully understand. But before we get to that, could you tell us a little bit about your background? I know that you have a background in theoretical physics. You worked at Google. Tell us a little bit about the story leading up to starting this company, Reflection.
The reason I got interested in physics, it was, you know, I moved to the United States as a kid. We moved to a rural part of Washington State and had a lot of time on my hands and basically developed an interest in physics at the time. It might be maybe part of it was that, you know, this area was one of the places where the sites for the Manhattan Project, it was the Hanford site where the plutonium was enriched. So it kind of had a rich history of physics in it. But what I really liked about physics was that it was a fundamental science. It was that had based on simple principles that could explain the majority of things that we observe in the world and was at the root of the technologies that we are using every day. So I kind of saw physics as the kind of the fundamental science of the 20th century. The first computer program was brought forth by physicists. It was the ENIAC program at the Institute of Advanced Study. This was during the World War II era when computers are first kind of being used for calculating ballistic missiles coordinates. Then, you know, things like the transistor were invented by theoretical physicists at Bell Labs. Systems like GPS only work because of special relativity. You have to account for the speed of light in order to have accurate GPS. So it kind of seemed to me that the fundamental science that was at the root of all the technologies we had at our fingertips was physics. And my switch into AI happened when I noticed that something similar was happening in AI, where it went from a field that had a lot of, I would say, diversity, but there was no real simple way to scale intelligence really until deep neural networks were discovered and scaled, starting 2012 with the ImageNet moment. But when I saw AlphaGo, which was a system that played the game of Go, which is out of the traditional board games like chess and so forth, Go is considered the most complicated one because it has the highest what's called branching factor, meaning that the amount of moves, the amount of different kinds of moves that every particular board position or configuration can afford is just very high. So it's a very difficult and intellectually challenging game for people to play. And a neural network was trained to play it at a superhuman level. And in fact, that system had never stopped learning. You could just pour more compute into it, and it would have just gotten better. And what I saw happening there was that a science that I thought was as fundamental as physics was being discovered in real time that would power the majority of technological advances coming into the 21st century. And it kind of seemed like, you know, some big discoveries are made in AI that make it kind of the physics of our time. It's kind of the physics of the 21st century. And that's kind of what got me into AI as a field in around the mid-2010s.
So you were working at Google, and then my understanding is that you start Reflection in 2024. Why did you start this company? What did you think was happening in AI, or what wasn't happening in AI that you could contribute to the world?
When we, it's funny, it's 2024, this is only a bit more than two years ago, but it seems a lot longer on AI timelines, and the field looked very different then. So I'll kind of rewind. What did the field look like in early 2024? GPT-4 had shipped, right? And ChatGPT had taken off. The first Gemini models were shipped, Gemini 1.0 and 1.5. These are models that my co-founder Yiannis and I contributed to, especially on the reinforcement learning side. But at the time, intelligence was really only good at chat. The notion of coding, agents, all this stuff, reasoning, none of that existed, and it didn't work. Like these models were just good as chatbots. They were not really good at any sort of cognitive work that needed to be done autonomously. And our bet at the time was actually twofold. We saw the beginnings of an open-source ecosystem. So some open models had started coming out that were pretty good, you know, the models from Meta and others. And so we thought that those models would give us a good base to build on top of, instead of having to, right, sink in our own large R&D money into pre-training these models. But then the big bet that we made as a company initially was that we thought that the paradigm of reinforcement learning, the same paradigm that created AlphaGo, my co-founder was one of the key contributors to AlphaGo, so the same paradigm of teaching AI systems to learn through trial and error could be scaled on top of language models and create systems that were not just good at chat but good at cognitive agentic work. Meaning, I mean, and we actually went in with coding as the entry point, that models that are good at coding, that can build software, that can do other agentic tasks on a computer, legal, finance, and so forth. And right, we thought that with building on top of open models, scaling up reinforcement learning would allow us to unlock this kind of capability in these systems. Now, what happened very quickly, actually, in the first year of the company, is that we were able to get, you know, proof of concept of these systems of reinforcement learning working at small scale on small models. But at the same time, the field had actually discovered how to scale up reinforcement learning. So OpenAI's first models started coming out, the O-series of models that had introduced reasoning into the kind of public market. And so what happened was that reinforcement learning as a problem was solved actually faster than we anticipated. You know, we thought it would take years. It took less than a year, both, you know, within external labs and even within our own experiments. And then we saw another thing happening, which is that the Western open models started significantly lagging behind China. Chinese open models started coming out that were very strong. I think we all kind of remember the DeepSeek V3 moment that happened roughly around the same time as the reasoning models were coming out. And then a whole ecosystem of Chinese models came out that was very strong. And about a year ago, a bit over a year ago, we made the decision that the most impactful thing we can do as a company is to take on the open intelligence mission because we saw a couple things coming that, again, we thought would take a longer time to play out, but then play out a lot faster. One was that we saw the geopolitical tensions coming, that it cannot be the reason open models are so important is because they are kind of the thing that the world, open source is the thing that the world adopts broadly. When you think about what is the most widely adopted operating system, it's, you know, it's Linux for computers and Android for mobile. These are open-source operating systems. The internet protocols are open, and these were originated in America and adopted by the rest of the world. So open technology is the thing that becomes kind of the foundational thing that everyone adopts, and with it, it brings in the infrastructure and kind of software of whatever nation is originating, right, the open-source software. So we saw this geopolitical tension coming, where if it's only, you know, one country that's producing open models, then the rest of the world kind of becomes dependent on those models and on the infrastructure of that country. And it's extremely important for America as a country to be competitive, and not just competitive, but have the best open-source software and models when it comes to AI, and ensure that our allies, that the world, and domestically, we all build on our own open intelligence. So we saw that coming.
Can I just interrupt for one moment? What actually is, what does open source mean? What does it mean to be an open model? When you describe Linux, which is open source, what is the difference between open source versus closed source?
The high-level distinction: closed, you cannot own, you can only rent; open, you can own entirely. Right. So it's kind of like the difference between the rental and the ownership real estate market. When you are renting an apartment, you don't actually own it, right? You are subject to, you know, the whims of your landlord, which could mean price hikes. It could mean, you know, an eviction. It could mean, you know, all sorts of things. But it's very convenient, right? You just pay rent, and someone else manages your property for you. You don't really have to worry about fixing the, you know, when the pipes are leaking or something like that, right? You call, you ask your landlord, and they figure it out. On the other hand, in real estate, there's the ownership market, right, which is you buy a house, but then you also inherit all the problems that come with ownership. And for various reasons, right, it makes sense to buy homes for financial reasons, for control reasons, and so forth. And you can think about open source in software and AI as really the same thing. You are either renting software or AI from someone, or you're owning it yourself. And so when we think about open source, you know, in AI, we really think about it, you know, we think about it as open models. The model, you either, right, you're either renting it from someone, or you can put it, you know, I can give it to you on a flash drive, you can own it, and you can run it on your own infrastructure and your computers. That is the fundamental difference.
My understanding, which I don't think fully aligns with the analogy, is that an element of open source and the open models is that they're free. I mean, when you buy a home and you own a home, you got to pay money to buy the home. But with open source models, my understanding is anyone can use it, anyone can build on it, customize it, and that is a sort of a crucial element. Do I have that wrong?
No, I think you're right. But it's kind of like, what would be the analogy here? The model may be free, even though some models choose to license. So you can have an open model that someone can own that has a license attached to it, right? And that's, you know, Meta's latest model, Muse, has a license attached to it, a commercial license if you're a distributor. Kimi K3 has a license attached to it if you're a distributor. So you can have open where there's a license attached to it that, you know, that can be monetized. And, you know, by the way, even Linux, like, it has, like, it's not just, it has a license attached to it where, you know, if you make certain modifications to it, those have to be open as well. You know, so it's not like there's a whole ecosystem of licenses around open source software and open models. And so it doesn't necessarily mean that open means free. Now, most licenses on open models today are free, and even our model, you know, Beam that we released, it's an Apache 2 license, meaning that it is free. But just because the model is free doesn't mean that everything around it is free. If you are a customer and or an enterprise and you want to run open models, well, you need to get the compute from somewhere, so you need to pay for the compute. You need to get the software for managing that compute and for serving the models, right? Just the technology for serving the models is complex and hard to do, so you have to pay for that. You have to, it typically is maybe embedded in some kind of agentic workflow or customization workflow, and so that typically, right, you partner with someone, you pay for that. So you can kind of think about the open model as the engine, but the engine is useless without a car, right? So you have to build, you have to have the car around it. And so open, even while the model itself could be free because the license is permissive, the model in its own right is not that useful. I can give you our open model on a flash drive, but in order to make it useful, you need to have all the stuff around it.
Would it make sense then to say that for the open models, and one thing also, what would be helpful as a clarification on the difference between open weight and open source? Open weight is the new terminology that we are hearing in the AI world, and it's not totally clear what the difference is. But would it be correct to say that for the open models, you're giving away the training for free, but you're not giving away the inference for free? You're saying, here's your model. Here's the underlying model, kind of, as you say, the engine that's free. You can have it and you can mess around with it however you want. But if you want to run that model and if you actually want to build a business around it, yes, you're going to have to pay someone to run the model, probably us. Is that the right way to frame this?
That's correct. So the difference between open source and open weight, it's really, I think, open source is sort of the high-level category around, you know, open source software. Open weight is kind of a specific instantiation of it, because in previous iterations of software, there was, there were no such thing as weights. Weights are a specific thing to AI and in particular to deep learning. So it's a, it's more of a specific instantiation. And there are some people who would argue, I mean, there's kind of open source purists, I would say, who would argue that it's not open source if you don't open source the data and training recipes and so forth. And our position is that, you know, we want, like, the mission of the company is to build open intelligence and make it widely accessible. And so our goal is to open source, you know, as much as possible that makes these models easy to consume and widely accessible. So when we launch Beam, we launch it with 40 different partners. We integrated into a number of different open source fine-tuning frameworks. We released our own open source agentic harness called MIR. We're, you know, also about to release a tech report that is sort of open sourcing the science effectively, right? Kind of writing up in some detail how it is that these systems were built, because we think it's important for an ecosystem of researchers to understand how the, I mean, how these things are built. This is kind of a fundamental tenet of science that, right, you want scientists to build on your work, and the only way to do that is for them to understand what it is that you did. But it doesn't mean that, for example, in order to get good models, you have to buy a lot of proprietary data, and you, in the contracts, you're not allowed to release that data to anyone, right? Because that's, that's some people's business, right? Selling data. So open source doesn't mean open sourcing absolutely everything, but we try to open source as much as possible to, you know, to make these models both useful to the scientific community, the research community, the developer community. But when it comes to then adoption within an enterprise or public sector, you know, context, then indeed the business is charging for the systems around the model that make it useful. And these systems are effectively serving inference to these enterprises, packaging up, you know, agents for those enterprises, fine-tuning services for these enterprises, and so forth.
Support for the show comes from Engine. When $1.6 trillion moves through travel every year, it's a shock to remember that most of it still runs on tech from the 80s. I mean, work travel works fine, but it certainly isn't all that intelligent. That's why businesses and travelers book directly through Engine. Engine closes the gap between wanting to be somewhere and actually being there. Developers, expense tools, and AI agents all plug into Engine's API. And as AI spreads, Engine only becomes more valuable. Every AI agent that wants to book travel needs real supply, real pricing, and real payments. With 40,000 businesses relying on Engine and growing fast, they're powering travel for the masses with partnerships like Non and the Tampa Bay Buccaneers. Strong hotel partnerships, proprietary supply, real-time rate shopping, and payment rails. That's the kind of foundation that takes years to build. 8 years in Engine's case, and it can't be bought. The future of business travel runs on Engine. Visit engine.com/propg.
We're back with Misha Laskin. The reason that I think it's so important for us to understand the difference here, and for anyone listening, why I'm so interested in hearing from you, is that this difference seems to be upending the entire AI ecosystem as we know it. Because, as you pointed out, we had the DeepSeek moment, and part of that was here is a Chinese model that is very powerful, but also part of it was here is an open model that is very powerful. And since then, we have seen token prices come way down. We've seen OpenAI reduce its prices by up to 80%. We've seen that the AI names that we are so familiar with, specifically OpenAI and Anthropic, whose models are closed, they have been forced to basically reprice everything because all of these open models are bringing the costs down. And that's why I'm interested in this idea of it being free, because it seems as though it's not that the entire thing is free, but this one element of it is free, which is resulting in significantly lower prices. I guess just one final question on the difference between open source versus closed. Does open source and open weight, does part of it also mean that anyone can go in and understand the underlying architecture of the model and go in and change it on their own time, whereas something like a model from Anthropic and OpenAI, you can't see the architecture behind it and you can't know how to build off of it? Is that a correct distinction or not?
That's correct. So, you know, in practice, right, it's not even that we don't know the architecture of, you know, a closed model. You don't even know the size. You don't even, you know, it's like, right, we sort of know nothing about it. And that's in a sense, you know, it's kind of like, again, when you think about the car analogy, it's sort of like if it's all packaged up and you can't peek under the hood, like the hood is sealed effectively, so you can't peek under the hood, you can't tinker with it, and, you know, right, you can't really understand it, customize it, and so forth. With an open model, you can, right? You can see the architecture. You actually, right, can plug it into either your favorite, you know, open-source fine-tuning framework, load it up on your GPUs, and actually, you know, customize it for your needs. So you can, you know, you can take it apart, you can understand it, and fundamentally what that means is that you move from a world where there are maybe across the closed labs maybe 5,000 researchers who understand these systems and are working on them. A minority of them are safety researchers. So let's say out of the 5,000, maybe 500 are working on the safety of these models, which is a very small community, right? And this is such foundational technology that is so powerful that it's impossible for such a small community of researchers to really understand everything that there is to understand about these models and how to make them safe, how to align them, and so forth. With an open model ecosystem, you're inviting a much larger group of researchers, engineers, developers, right? Anyone who takes an interest can take a look and contribute to these things. And that is fundamentally, you know, how science is done, right? Science, right? It's not alchemy. Science is fundamentally open. That's the way cybersecurity works. Like the whole field of penetration testing is based on the notion that you have access to, you know, to software that you can actually see and understand. And actually, the most hardened software in the world, the safest software tends to be open-source software because it's been, that there's this saying that with enough eyeballs, all bugs become shallow. And that's why open-source software is so ubiquitous and robust because it's had hundreds of thousands of developer eyeballs on it, looking at, you know, various sorts of vulnerabilities. And while they might not patch all of them, open-source software is definitely more, tends to be much more hardened and robust than closed software. The thing that it's quite different from, though, is the traditional business of intellectual property, where if you create something that is extremely valuable, usually the number one priority of the business is to put patents on it and make sure that it is extremely closed, such that they can build gates around it and then monetize that thing.
And so that's why the open model business, for me, I'm fascinated by how this business is going to work, because I look at a company like Anthropic and OpenAI, who are, you know, number one and number two in the world of LLMs, but they are still losing billions of dollars a year. And, you know, for them, I think a big question around their business model is like, how can they raise prices? How can they keep prices up, and the open model ecosystem is bringing prices down? But I guess for a company like yours, and something I wonder about with DeepSeek, with Qwen, with all of the other Chinese open models, is how is that business model actually going to work? Because my understanding is that to build a model and to train it, that's the most expensive part of the whole thing. And that's a part that is technically being given away. So how will the business model work over the long term?
For most foundational technologies that we're aware of, and even in the software technology era, they started off as closed. But if the technology is important and foundational enough, it was followed by an open version of it that then took up the majority of consumption. And this, right, operating systems started closed. Operating systems started closed, and, you know, Linux came, you know, after. So, right, closed could be, you know, MS-DOS and Windows, right? Microsoft's closed operating system built a great business around it. Open-source version of it, Linux, came after, and now over 90% of the world's servers, I think the number is even higher, over 95%, run on, you know, some variant of an open-source operating system, most often Linux. The same thing is true for Android, right? The first really mobile OS came from Apple, was closed, was followed by an open-source version, Android, that now the majority of the world's phones actually run on Android. The internet is an open protocol, started off closed, right? Was built basically within, you know, the government and universities. So this is actually like what's happening here is a continuation of a very natural thing. And I think that the way we probably need to reposition our thinking about AI is that the models are a commodity, like electricity or water, right? When we think about electricity and water, we don't think about, oh, how does, you know, my energy company sustain its business? You know, it needs to be making 90% margin, right? That's not, we think about energy and water as commodities that there is some margin, right, that these companies sustain. That's actually low, but it's a, you know, massive service that's provided as fundamental infrastructure to the entire country, by an entire world. So I think what's happening in terms of margin compression on the model side is a healthy thing because intelligence should be viewed like, you know, as foundational fundamentals as electricity, water, it's a commodity. Now, the inference, you know, margins even on open models are actually quite high, and the reason for that is because compute is scarce. Now, the reason is compute is scarce, which is, you know, GPUs effectively, but CPUs as well, compute is scarce because the amount of, you know, intelligence you can pack sort of per watt, let's say, keeps growing, right? So the amount these models get smarter faster than our ability to stand up compute. And you're, and when you're in the intelligence market, that's kind of like the market of artificial labor, and it's a very kind of positive-sum market in the sense that how much, you know, artificial labor would I throw at really difficult problems like drug discovery or, you know, cyber defense, right? You'd probably throw as much as possible at it in order to get the best cyber defensive systems, in order to discover new drugs, right? And what I think will continue happening is that compute for the foreseeable future, I mean, I don't know how long this, you know, how long that lasts, but for the foreseeable future, intelligence is going to outpace our ability to stand up compute because compute is, you're standing up stuff in the physical world. And so the amount of intelligence you'll get per watt is going to keep increasing. And so you want to drive intelligence, the cost of intelligence down as much as possible so you can use it more, right? There's like this classic Jevons paradox kind of thing. So all this to say that intelligence is a, should not be seen like a premium product like an Apple, you know, like iPhone or something. Intelligence is more like electricity and water, and it...
should be commodity. The whole system around intelligence is expensive. Um, they'll have some margin, but the margins today on inference are actually very healthy for both closed and open model inference.
On the inference though, because and so I think the utility point is really interesting. Um, and I think Sam Altman has said that he thinks that there's a world where that might happen with OpenAI. It doesn't seem to be where the direction that they're actually headed in terms of what they're pursuing as a business, but the idea of AI as a utility is really interesting. And agreed, on inference margins are very healthy. But I do think that we're also, when we say that, leaving out this question of the training, which seems to be a very important and significant expense. But on utilities, generally speaking, utilities are in some way subsidized by federal, state, local governments in order to make it accessible for everyone, electricity, water, etc. Do you see that as the world or the AI economy where, you know, perhaps because training is very expensive, that there will be some level of subsidization, um, similar to what we see with utilities?
Perhaps in some future. I think right now the market is kind of too early to tell, right? Because kind of I guess what I'm describing is kind of the desirable end state for AI is that it's a utility, and right, like the internet, right? The internet is effectively a utility, right? It's a commodity that is affordable to most people, and there are a lot of projects in terms of how do you get Wi-Fi to work in rural areas as well. Uh, so it's unclear to me whether AI goes down the path of electricity or it goes down the path of, you know, the internet. Um, but nonetheless, whether, you know, one is government subsidized or not, um, both are sort of seen as large-scale commodity services, and a lot of very successful high-margin companies are built on top of them, right? A lot of companies with high margins that are built on top of the internet, even though, you know, access to the internet in its own right is fairly affordable.
So the Chinese open models are so interesting because DeepSeek comes along, and I mean we all remember what happened in the market. Suddenly everyone panics, and you see all of the US AI stocks going down. But it seems like what is happening in China, which is why you have emerged as this strategically important company for the US—and correct me if I'm wrong on this—but it seems as though China is using vast amounts of government subsidies to create these models, which they then distribute to the world for free, thus changing the entire complexion of the AI model market, bringing down prices significantly. But ultimately, from my understanding, the government, the CCP seems to be paying for it, um, at least when it comes to DeepSeek. Uh, I mean, it seems like that is what is happening here. But is that wrong? I mean, to what extent is the Chinese government involved in funding and promoting these open models?
I probably don't know, you know, much more than, uh, you do on the subject of the CCP's involvement, but I would agree with you that, um, it's very hard to explain the Chinese open model ecosystem without deep support from the government in some way, right? Um, it's just hard to explain. You know, these companies were around for actually some time. It took a few years for them to get to the point where they actually started charging, uh, you know, meaningful money for their products. And even so, you're right, there's like, you know, a large investment that needs to be put into it. Um, I think there's a big difference between, you know, the business models that these companies have in China versus Western labs and distributors. Uh, in China, it kind of—I mean, from the outside looking in, it looks in a sense like a captive market in the sense that while there are open models, uh, you know, in China, um, I don't really see too many distributors of those open models in China outside of the companies that are building them, right? Like DeepSeek has their API, and Kimi has their API, and, uh, you know, GLM or Zhipu have their API. And it seems to me, again from the outside looking in, that their business models are not dissimilar from closed models in the West, but in China, right? But in the domestic market in China.