**A** (0:00):
When I went on and I tried to like understand this question, I think a lot of the answers usually were answering more like why China wants their eye, which is like very obvious answers that they want national sovereignty and like domestic control. They don't want to be dependent on like you, the US at any layer of the stack, whether it's like chips, models, you know, software, anything partially due to export controls. You know, this is like dreadnought moment. Dreadnought is like this like period in World War I where like the British invented a brand new ship that like rendered all old ships effectively useless. And like, it restarted this like arms race where it actually gave Germany an opportunity to be like on the same playing field and build a navy that was like somewhat equivalent to the British navy and like, kind of like restarted this like arms race. And so, you know, I think China really does see this as like a dreadnought moment, particularly for software which like, they have been sort of behind the US on for the last 20, 30 years. And they want AI that's better designed for China. You know, Chinese language is a big thing. Like, my girlfriend is Taiwanese and she thinks that, she says like, yeah, deep Zeke is much better at writing like Chinese like prose than like, than like chat or Gemini is. And obviously, you know, socialist values. They want information control requirements. They care a lot about making sure it's like home controls your. But none of this answered my question about why open source, they could do all of these things while like keeping a much more closed source model. Similar to like, we have the big labs in the US doing so before I kind of jump into a couple of the reasons why it's kind of important to understand how Chinese companies work. You know, you know, the socialism with Chinese characteristics. It's like important to realize that like a lot of these companies are not operating in like a purely capitalistic way that we might expect like the US like US Companies to do. They have to like, you know, they are capitalistic in some sense where they do have a profit motive, but they have to balance this with political alignment. Because a lot, you know, a lot of these actually labs do have some level of direct or indirect state ownership. Zai was like, kind of incubated out of Sinhua University and has like a heavy level of like state control. Internal M is one of the big robotics labs. They're like, they are a government initiative, basically. Huawei, which is the big chip manufacturer, is. I mean, it's controversial, but they're effectively state controlled. And you know, oftentimes in other ways States can like force a lot of their will on these companies, that a lot of their companies need good licenses to publish any sort of AI products. So they control, you know, they have to follow all these rules around data labeling, training, compliance. There's always like compliance requirements that the government kind of like can push on these companies. And the big thing is they set, they signal through the biggest way that the government kind of enforces its will upon these companies is by like setting national priorities and strategic initiatives and then using a combination of like carrots and sticks to like benefit and punish companies for complying. So like carrots can include things like subsidized compute, preferential procurement. So like, you know, who gets contracts from the government for different initiatives, tax breaks, credit from state banks, access, national data sets, political goodwill. They have sticks that include things like regulatory investigations which can, you know, spread into some other parts of their businesses as well. So some of the big labs like Alibaba or like, you know, Tencent, they have other big businesses that, you know, they can find that if they're not doing things that the government wants, they can might, you know, they're not going to stick to that just AI side. They have content moderation, mandate, kind of like convention, a lot of licensing requirements, security reviews, heavy fines, block IPOs and like disappearing as happened. So yeah, we have to like, when we think about like what, what a lot of these labs is into incentives are, we kind of have to keep in mind that they're balancing their personal incentives as well as like the state's incentives. So why open source? Well, I would say first thing is always like the very traditional benefits of open source, which is like, it's a force multiplier. It accelerates collective development. The innovations on silos. You know, you have a lot of different companies and researchers and everyone like kind of working together to contribute to the progress. And that helps a lot because, you know, China is kind of coming from behind. They, you know, they were behind on open source on like a lot of AI development and you know, having this force multiplier of like there are multiple labs sort of contributing on top of each other works is kind of what helped them like catch up to where they are today. And it also helped them deal with a lot of the export controls in ways that we'll get into in a little bit. And then the other, you know, from the labs, personal perspective, you know, similar for a lot of reasons why people do open source development is there's a prestige aspect to helps attract talent. A lot of the best AI researchers want to, you know, see their work, be used and actually be used in public. And so having it be open source helps attract a lot of top talent, attract investors, attract awareness, media attention, government supports like Deep Seek. Even though it's open source, you know, most of their usage is still through their API and they kind of became a household name, especially within China. They're the most popular AI product in China right now. So, okay, there's all those like normal benefits of open source. Something kind of, you know, are very common in the US as well. I want to focus now on like what are the benefits like from more like Chinese specific characteristics. So one of the big things is national development. So this is what called technology and the rise of great powers. And, and the premise of this book is basically like what's important for, like how for societies and civilizations is not who can invent technology the fastest, but it's really about who can get that technology to diffuse through their society and through their economy the most efficiently. So you know, through. In the, you know, in the second industrial revolution, even though a lot of those innovations are actually invented in Britain, the US is one of the most like, adept at like actually co proliferating those technologies throughout the, you know, American economy. The Soviet Union is an example of a country that was actually pretty good at inventing technologies, but actually pretty poor at, you know, proliferating those technologies and like getting them used in like civilian and industry outside of like military use cases. And so China really, I think, I don't know if they've read this book or not, but I think this kind of helps understand their mindset where they're really kind of focus on who can deploy and integrate AI the fastest across our economy. So if you look, if you read some of like Chinese, China's like official policy, like documents, they categorize AI as national infrastructure in the same category as like electricity grids, railways, Internet. So they're really seeing this as like a public infrastructure thing where they want it, like, okay, this should be just subsidized for the, for society and then get this gap in as many places as possible. And you know, the goal is to drive the true economy. So you know, China really like kind of has this like disdain towards finance and software where they kind of see these as very intangible, like not real. They see like the true economy is like the tangible physical output of things. And so they're willing to subsidize that in order to grow the true economy. So open source accelerates its adoption because it Makes it cheaper for companies to adopt AI for their products. Chinese companies like cheap before deep sea. Many of the Chinese companies were using Llama because you know, they could, they could deploy it for much cheaper. And it also makes it much easier for Chinese companies to just like, instead of having to train a new model from scratch, they can easily pull an open source model and then fine tune it or just scale it for their own personal needs. This is, this also is very important when it comes to export controls because fine tuning can only cost like 1 to 2% amount to compute. So you know, I know a lot of startups in, in the Bay that are like, you know, having to build their own base models from scratch for like specialized use cases. In China there's much more of a culture of just like focusing on fine tuning because that's what they have the compute capacity to support. They don't have to deal with all, you know, API limits and all this kind of stuff. Stuff. And a big thing is you have better feedback loops between like researchers and product developers. So you know, you have, you know, at AI labs you kind of have like researchers working on these models. And yes, you know, some of them, like OpenAI do have an end user facing products. But now you have, in China you have this like system where you have, you know, product developers on all sorts of different products having like very low level access to two models and like helping push forward like research and create like very powerful feedback loops. Next we have like geopolitical sort of related reasons. And so this kind of comes down to this concept of like commoditize or complement. So commoditize your complement is this idea of like if you, you know, there's different layers where value can accrue and if you have a huge advantage in one of them, you want to commoditize layers so you can capture most of the surplus value. So an example like would be like Microsoft owned the OS, Windows OS and so they wanted to make PCs super like commoditized so they can capture more of that value into the OS layer or you know, you think of like Uber and drivers, right? Uber wants to commoditize drivers. So China knows that their competitive advantage is not in software. They want to catch up, but they, they don't really have strong belief that they're going to overstep with the US and like dominate in this. And so if they can't win, they want to at least commoditize by providing something that's close enough but for a lower price and a lot of these big open US Labs, kind of a lot of their business models and everything is dependent on like paid API calls, subscriptions. And so if China can like sort of undercut a lot of these prices, offer license free inference, they're kind of going to help undermine the SaaS based business models. And you know, this is something China has done for a lot of other industries. Like you know, they've subsidized solar to sort of undercut like all the other manufacturers and then until they get like a high like you know, dominant market share in that industry. And so you know, China knows their competitive advantage is not in software but they know they dominate the physical stack, hardware manufacturing, chips, robotics, power generation. And so they want to see value shift from software to the hardware and compute layers. An example would be like Shanghai AI Labs which is like I said, that one's the government initiative. They are probably one of the front forerunners in like open source robotics models. And that's because they want to, you know, if the software, software for open source for robotics is free but the value is really going to come down to the hardware which kind of has a very strong dominant advantage. And yeah, so yeah, and so you know, some graphs just on like you know, the annual industrial robots China is like taking over. Everyone's kind of barely on the chart power generation. China has like just vastly overseeded the US Even when it comes to chip manufacturing. Chip manufacturing is like the one place where you know, they're still not nearly as much. I'm sure Charles knows a lot more about this than me but you know, they are really sort of trying to like build up their chip making capacity and like, you know, if you look in the last like five years and they are the biggest grower in semiconductors even though they're not doing like latest generic semi semiconductors, they spent a lot of time like innovating on like, you know, how to work around it and get like, especially when it comes for like infrared and stuff like that where you don't need the latest gen systems. And so you know, China really does want to win this and you know, in case they can't, they always have their plan B which is invade Taiwan. The other big thing here, reason why they want to open source and post open source a lot is for Nvidia, one of their biggest trips, choke point is in Cuda which is like the software stack used for like, you know, runs on these Nvidia GPUs and China doesn't have access to these Nvidia GPUs. So they kind of want to open source the AI ecosystem to standardize around a non US controlled stack. So particularly can which is designed, it's an open source software ecosystem unlike Cuda, but it's especially designed to run on Huawei's like infrastructure. And so the idea is that like okay if they can get the entire open source AI world to be like using non Cuda that they can kind of Huawei will eventually have GPU capacity and be able to be the primary compute provider for like you know this AI that proliferates around the world due to its open source nature and they can get over in on the energy production. Today most of the, most of the models are still being even within China are still being trained on Nvidia GPUs. But there's a process that I don't really quite understand, maybe someone can explain it to me. There's a process called porting which allows you to like tran you can an open weight model, you can like port it to a different architecture that still lets you run on these Huawei GPUs for like fine tuning purposes. And so you can kind of think of this like commoditize the where they want the value to go. Where like in the US a large portion of the value is in this like intelligence layer. They kind of China as ideal as they want to like shrink the intelligence layer and drive more value to the apps by apps like using it more so hardware apps like robotics and stuff and then a lot more for the infrastructure because they think that they have an advantage at the infrastructure layer. There's also soft power reasons. You know, they want to build their relative goodwill to the us get more countries sort of dependent on their infrastructure. They can bake in their own cultural values and norms rather than pro western norms and just like change global perceptions on China. Like you know here I'm giving a talk on why China is a proponent of open source. And you know it actually has had a meaningful effect on shaping these global these cultural perceptions. And then last we'll go into some internal cultural reasons for this. One big one is that China just historically has had a weak software as a service culture. So, so this is like. So it's kind of really small. You probably can't forget but this is from like McKinsey where it says that the size of China's SAS market is worth 5.2 billion compared to the US's 120 billion. And a large portion of this is that like Chinese companies are just unwilling to pay for software just like quarter Part of their like cultural mindset. And so another big thing as well is a lot of the Chinese tech giants are often platforms, you know, like WeChat, Taobao, Gaoyin Gao is like the Chinese version of TikTok basically and, or their cloud providers, right? And you kind of, you can see, you can kind of compare this for an analogy to like US Labs where like Anthropic is, needs to be an API provider but Meta is down to just do like open source, like model generation because that's not where their bread and butter is, right? They, they make money off of like owning these platforms for the same reason you don't see Amazon really pushing that hard on trying to build, you know, models because they're, you know, they're not really a software as a service provider. Amazon's a cloud provider. Meta is a, you know, a platform of social networks. They don't, their bread and butter is not going to be made from software as a service in the same way that like Google or Microsoft or Anthropics will be. You know, another interesting thing is that the product weight. So in China products are regulated but the weights are not. And so to like do, to publish an API or trap bot or some sort of product, you have to like the Chinese government has all these sorts of like things you have to follow. And you know, I'm sure you guys have read about like all the like fights within a lot of AI labs in the US about like, you know, the alignment team wants to like slow down like product releases and whatnot. Now imagine like that. But like the alignment team is the government and how slow they go. And so you know, this basically kind of allows labs to just like, you know, open weight, they can just publish them and like push forward the research. And so this allows them to just go way faster and you know, it allows the world to see their true capabilities. So you know, they kind of from that prestige aspect, you know, they behind their API, they have to like nerf it a lot to like meet the alignment requirements of the Chinese government. But they want like the research community to be able to see their true capabilities. So they want the open weights to be there so people can test against the true possibilities and it avoids a lot of the political liabilities. Right? So if you like open source these weights and they do something bad, it's like, okay, you know, that was outside of our, outside of our system. Another big thing is this like this sort of, is this like path dependent thing or I will say I think deep seq in particular. Was very like very ideologically motivated and like belief in open source. But this kind of like created this like social like stepness cascade where like, you know, a lot of these Chinese companies tend to be very mimetic because they're all kind of like get attention from the Chinese, from the Chinese government. And like when Deep Seq had this like big moment of attention from like the world and media and from the Chinese government, a lot of the other companies kind of wanted to like follow in those footsteps. And it kind of ended up turning this thing where these like big Chinese labs didn't want to get upstaged by like a small startup. And so it kind of created this like default open sourcing now because Deep Sea kind of kicked off this like mimetic effect. And then finally the last thing I would say is this always has been this like collectivist knowledge culture. There's been a long tradition of like sharing wisdom and collective advancement. China has never been a society that's like really, you know, glorifies individual inventor the same way that we do in the the US it's come from like a lot of Confucianism is about like good governance is like based off of the wide dissemination of correct principles. And so knowledge is something that's meant to be like proliferated. And they kind of see intellectual property. This kind of like shapes a lot of their views on like intellectual property. And just like they kind of see software and intellectual property and software more in this like realm of like intellectual property than like hardware us, which kind of leads to like, okay, you know, companies should have an edge by their like commanding of resources and like the physical world, not through like hoarding of intellectual knowledge and property. And so, you know, kind of makes sense. You know, it's a communist country. It kind of, you know, it kind of makes sense that you'd expect a communist country to actually be someone pro open source for, for a lot of these reasons. Yeah, that's it. That's all I got.