Sequoia's Dean Meyer on America's open model paradox: US frontier AI is being distilled into Chinese open-source models

Speaker 1: Let me tell you about MongoDB. What's the only thing faster than the AI market? Your business on MongoDB. Don't just build AI. Own the data platform that powers it. Up next, we have Dean Meyer from Sequoia Capital here to talk about America's open model paradox. I love a good paradox. Also, fun fact, I heard somebody debunking Jevan's paradox. It's not a paradox. It's just an just an effect. It's just an economic effect. Is it a real paradox? Breakdown America's open model paradox. Feel free to get a little bit of an intro. What's up, guys? But how you doing? What's up? I'm good. Can you can you can you guys hear me? Oh, yeah. We're we can hear you. I can hear you. Let's go. We can hear you.

Speaker 8: What's up, guys? Great great to be on the show. Thanks for having me. It's fine. You know, my my myself and and and my partner Constantine at Sequoia, you know, we've been riffing on Sovereign AI for a while. Yeah. And I think kind of, you know, we wrote this post that went a little a little viral, and kind of the the core core concept was, you know, US is creating this frontier capability Yep. That kinda gets leaked into Chinese and white models. Mhmm. And then you have the entirety of the American kind of builder ecosystem, which is Neolabs, you know, companies that are somewhat dependent on these Chinese models. And so, you know, today there's like, you you know, if you're an American lab and you can legally learn from a Chinese model, but you can't learn from a kind of frontier US model, which is what they're doing, Then the question is how do we build kind of the, you know, the legal path almost for this direct route as opposed to just having this dependence on Chinese subway models. And so we kind of address that and and share a few ideas, which I'm happy to talk about. But that's kind of like the overall the overall thing. Yeah. I mean, there's so much there's so much news about this particular topic. The essay was written at the perfect time. Just just before Kimi k three, man. It was Yeah. It was good timing. We we had no intel. We had no intel. But, I mean, were you but were you a believer in

Speaker 1: in just the fact that the open source market generally lags the frontier by six months and that's not changing? Because there was this graph that was going around for a while that was showing that maybe the closed source was scaling on a slightly steeper curve than open source and this was driving sort of about a couple months ago when we were in between open source releases. We were people were starting to say, oh, maybe they're they will plateau because the training cost will be so high that you'll have sort of like what happened at Meta. That will just happen in China because China will just say, yeah. Like, if we're going to spend a billion dollars on this training run, it's not worth it to open source it. But it feels like they found an economic model that makes sense. Kimi's license is out today, and they're gonna take a portion of revenue if you're really selling a lot of Kimi tokens. And so Mhmm. Were were you were you was this built on an idea that that the open source and closed source frontiers would always be on the basically the same track?

Speaker 8: Yeah. Well, I I I think I think, look, that that is that is the question. Yeah. Yeah. And so, look, our view is I I mean, first of all, like, the the the gap is closed a lot. And so kind of the key conversation online over the past week was, you know, is distillation the reason for this? Or is there some other kind of mysterious reason? And the truth is, like, you know, there are two camps, like everything. You know, people are are kind of extremists in in you know, when when there's conversation. Everyone there's a group that are like, hey, distillation does nothing, and there's a group that's like distillation everything. Mhmm. The truth is from the conversations we've had, it's somewhere in the middle. Right? Like, it's 20%, thirty, forty, it doesn't really matter as much. Mhmm. It is meaningful that we know. We know from the labs, you know, there's tens of millions of of of queries through these, like, large scale proxy distillation accounts. We also know, on the other hand, if you speak to anyone who's, like, leading in synthetic data gen on pre or host training at the labs, like, they will tell you, even if you do, like, hardcore KYC and then, like, let's say you get access to these kind of, like, you know, you don't have all the traces, you don't have Yeah. Kind of, you know, like, the chain of thought. There's a lot of RL that happens. Right? And you just get the output. Like, let's assume you just get an output from a query or a prompt. It is still extremely valuable. Like, that is 100% true. And so we know distillation is real. And then the question becomes, well, if distillation's real, you know, how do we address the asymmetry that exists between, you know if you're a Western, you know, US company or, you know, just a Western company, you're not going to, like, go and set up, you know, 35,000 proxy accounts with, like, different IPs around the world and go into philanthropic for the fear of being sued and being called out and and even being cut ax you know, shutting off access. And so there is this structural asymmetry,

Speaker 1: and and and think it's something we have to think about. So you have to create the meme of token, Max. Yeah. Go ahead. You have to create the meme of token, Max, the token leader would rather Maybe. Do you think that Anthropic or anyone else, let's just say a a closed source frontier lab in America, if they believe that they've been distilled by an international company, should they just sue? Even if there's no even if there's no way that on any reasonable time scale that they're gonna win a settlement or block anything, would it just throw down the gauntlet to say, we're willing to put this into the the the official record under threat of perjury that we believe this is what happened, and there's, a signed affidavit of their belief that there is wrongdoing?

Speaker 8: I mean, I I think they they probably would, and nothing would happen. Like, you know, I think not nothing would happen. And so anyway, like, you know, US companies are are kind of deterred from doing this. So I don't think it would change too much, unfortunately.

Speaker 1: It would be very symbolic. I I I think it would be I agree with you. I think I don't think anything would would change from the actual lawsuit. It would cost some amount of money to file this lawsuit, but at least it would it would they would have something to point to of like, hey. Look. We tried to go through the traditional path and you were able to follow along. You were able to look at our evidence, which was entered into the public record. And, oh, yeah, the the CEO of the lab that we're accusing of distilling us didn't show up to court or whatever. You know, whatever the process is, the process can Yes. Can begin. May maybe this is, like, a day away. Maybe this is, the lawsuit drops tomorrow. Who knows? But What about Yeah. It seems like that's the next logical step because right now there's lot of Yes. There's a lot of there's a lot of pushback against, like, hey, why do you have a problem with this? And they say, oh, maybe it's distilled.

Speaker 2: Well, you know, that's the next that's the next step. Maybe. Don't know. Yes. So so Kimmy came out with their their license this morning. Mhmm. And Mhmm. They they in the license, it says if if you're just running it, hosting this yourself for your own use cases, that's fine. Yeah. It's free. But if you're effectively reselling it as a neo cloud, you have to give us a revenue share. I can't imagine that Washington is gonna tolerate that. Just just the idea the idea that like a bunch of US Neo Clouds are then sending billions of dollars back to to China to support one of their leading labs. Just to me, that feels like it could provoke a response that felt like it was maybe coming last week when different people in the admin were saying, hey, we're we're we we we believe distillation is happening here. We're we're we're pro open source, but we're against how they were positioning it, IP theft. On the flip side Mhmm.

Speaker 1: I was very quickly able to find a Bottega Veneta dupe that comes from China and is effectively distilled on that brand, and the government isn't stepping in about that. You know? And that money is flowing straight to it through Amazon and through other platforms, Xian and T and just Yeah. But it's not it it doesn't have any geopolitical

Speaker 2: significance.

Speaker 1: You're not you you wouldn't go to war for And it's a a caring company. Yeah. No. That's true. That's true. Anyway, sorry. I was No. I think look. I think look.

Speaker 8: I was surprised to read that. Look. The truth is, like, I think it's not dissimilar from the database market where, like, you know, you had Redis, like, early on, there were a range of these companies release really cool open source, and, like, they just got the hyperscalers copied them effectively, and, like, that was the end. I just you know, it's impossible to enforce. I I I like, I don't think that will work for for the Chinese. But, look, the broader question here is, like, there are two major risks with, like, depending on a Chinese US eco you know, a Chinese open source ecosystem. The first one is the dependency risk. Like, the framing we have is, you know, in some sense, you know, at Sequoia, we always say, we're only as good as our next investment, which we actually believe. And I think in some sense, the analogy is, you're only as good as your next model. And so as long as you're dependent on the Chinese to give you the next, like, hit of, you know, of of of the thing, you're you're kind of you're kind of in a in a bit of a in a, you know, in a bit of a weird position where you have this dependency. And then the other thing that I don't think people are kind of well calibrated on is just the risk of backdoors. And I'm not saying like, just to be clear, I'm not saying that all Chinese open weight labs are backdoring models. Mhmm. But the truth is we know for sure, and there's, like, a lot of research on this, both from the labs and outside the labs in academia. It is, like, an NP hard problem to know if there is a backdoor in a model that we know. So it's impossible to verify. And on top of that, there's, like, a ton of crazy evidence that shows you can you can, like, time pull a model, which basically, you know, one way of saying is, like, a model, you could do all the evals you want, and suddenly, like, the you know, it it gets a certain input, and the model literally changes behavior. It will start, like, exfiltrating data and doing some crazy stuff. And, like, this this stuff's real. Like, this is not science fiction. We know and so kind of the analogy I would give you guys is, like, I think the state of AI security today is, like, very similar to that of, you know, early days of the Internet when, like you know, in in in 2001, like, Wi Fi was, like, you know, Wi Fi encryption was effectively broken, like, publicly broken. Yeah. And still, like, no one did anything. Yeah. And I don't even think we're at that moment yet. Yeah. And it's gonna it's, like, it's gonna get really bad if we, like, don't better understand how to secure these things, and two, like, build our own kind of independent ecosystem. And so the the yeah. Those would be, like, some some thoughts as well.

Speaker 1: There's so much more we could go into here. We gotta have you back on the show and talk more because, like, there's Cool. There's so much going on. But thank you for writing the essay. We'll let the audience go and read the full piece. Yeah. Great to have you on and come back soon. I'm looking forward to the next one because it feels like the the conversation has already enhanced so many different directions. We didn't even get into the the the the politics of anyone having a cyber weapon in their pocket. There's all these different scenarios to play out beyond what you just mentioned, which are a bunch of other considerations. So it's a fun one for Salon. We'd love to talk Thank you so much. Yeah. Yeah. Was awesome. To everyone. We'll see you soon. Great to hang. Talk to later. Bye.