Josherich's Blog

Why it Sucks to Work in AI in China + Open Source with Kevin Xu

Kevin Xu of Interconnected Blog and Capital, as well as Jordan Schneider from China Talk here.

We’re going to workshop my take that it is way less fun to work in the Chinese AI ecosystem than the Western AI ecosystem.

And then we’re going to dive into Kevin’s opus, exploring the past 20 years of open source technology in China.

So I guess I’m here to kick it off.

The general gist of the take is that if you are at:

• OpenAI
• Anthropic
• DeepMind
• XAI
• Meta

the amount of compute you have access to to do your super cool research is an order of magnitude or two higher.

The level of business pressure that you are under is high, but the upside is just also an order of magnitude or two more positive for you as an individual contributor, as a founder, as someone playing in the broader ecosystem.

If you’re in the West, in contrast, the Chinese new entrance to the field, the 01.AIs and Minimaxes and even DeepSeeks of the world have not had anything remotely close to the type of success that OpenAI, Anthropic, or even the sort of tier two AI startups that have been founded over the past few years have been able to deliver.

So the pressure is enormous.

It’s kind of indefinite because you haven’t, even though you are making the best models in your country, it’s not that’s transforming into revenue in the tens of millions of dollars.

Kevin, we should get into the Alibaba firing resignation of Lin Junyang too then, since you mentioned the “cushy job at Alibaba”, which really only lasts about a couple of quarters.

I think that has been dealt to China tech in general, probably as soon as export control and tech sanctions and then China’s own crackdown became the norm that everybody in China has to operate in.

And AI doesn’t change that AI probably accentuates a lot of those challenges as far as lower ceiling, but also less resources with just about as much expectation for the output and the traction and the metrics.

And I think the latest personnel change over at Alibaba’s Qwen team, this just happened a few days ago, is the latest manifestation.

I think China tech or China AI, maybe, especially in the open source world, has maybe enjoyed quite a year of, I wouldn’t say euphoria, but at least a lot of really welcomed and perhaps long overdue traction, attention, not just from within China, but all over the world.

And whether that actually continues this year is a huge question mark.

And what has happened over on the Quen team may be the first sign that the economics and the expectations of all these decently well-funded labs in China, but nowhere near as well-resourced as the open AIs of the world and the open AIs of the world are still crying out for even more resources.

So you do have this dynamic of higher cost, but higher reward on the Silicon Valley side.

And the China side is just a very different set of hands that you still have to play.

I would say one thing that I mentioned near the end of my so-called opus of the history of China’s open source ecosystem, it’s 6,000 words long.

It’s probably way too long.

I don’t expect anyone to read the whole damn thing.

But near the end, I wanted to articulate one thing, which is that open source as a strategy to expand is likely one of the best, if not the only avenue in which a lot of Chinese entrepreneurs, tech entrepreneurs have at their disposal these days to really go beyond their own border.

And everybody wants to expand overseas, despite the sanctions, the geopolitical tension, the different barriers to go into different markets.

The desire and the passion and the energy has not stopped.

In fact, it probably only gotten more intense, partly because the domestic economic situation, frankly, hasn’t really improved dramatically.

So there’s only so much you can get out of the domestic pie.

They have to go overseas.

And open source is one of these ways where their own technology, at least the open source, the AI labs, can be seen and evaluated and being kicked and tested without the geopolitical stuff.

And all the toxicity of the China label as a brand for, a quarter, a quarter or two before you start to work on commercial relationships.

And so there is a very practical reason to open sourcing everything, despite the economics being initially not that great, because it’s one of the few vectors where you can break out of your market, your whole market into different markets.

Yeah, let’s do a little compare and contrast between China tech crackdown and what we just saw with Anthropic and the Department of Defense, because I think that kind of underlines another difference in the sort of ambition that you can have as a founder in this world.

I mean, Dario wants to shape the future trajectory of humankind.

I don’t think that’s really underplaying the vision that he’s put out in his first essays now that have turned into little novellas with the latest one that came in the new funding round. And kind of the idea that first, I think a Chinese founder could be that kind of loud and ambitious and discordant with the sort of general narrative about what the government thinks AI should be used for.

And then the, look, we talk about America has military civil fusion as well on certain dimensions, but the idea of a Chinese lab saying no in a dramatic and public way like this would just not happen.

People have learned that lesson: this is not the way to sort of have what would have been the equivalent of Dario writing that memo and it coming out in the public? It would just be someone crashes out on WeChat Moments and it gets screenshotted and sent to things.

And then three days later, all their social media disappears and they’re probably gone somewhere and maybe not coming back, ever, or at least for a good while, which is kind of what famously happens with Zhang Yiming and the crackdown on some of the ByteDance products, which weren’t even trying to reshape the face of the society’s relationship.

Just low brow humor was what Zhang Yiming’s initial wire that got tripped.

They just had fights and nudity and that’s all it, that’s all it took. Not saying no to doing autonomous weapons.

So, that’s less fun. You can’t make as much money. You can’t play machine God. And I think it’s an interesting thing.

It’s an interesting point, Kevin, about how doing open source is maybe one of the few ways you can express yourself ideologically in the Chinese technology ecosystem that by doing, but by working on these projects and releasing your technology to the world, that comes with a sort of ideological valence, which as you pointed out, starting in 2021 is something that the government sort of saw.

What you end up having to do is actually a very classically liberal democratic governance process, where you get a bunch of people who have a lot of stake in the project’s future. So, you decide on a set of priorities, you decide on a set of priorities, you vote on a lot of things, you usually have a technical committee, you can just look up any of your favorite open source foundations, whether it’s a Linux foundation, whether it’s a cloud native computing foundation, which is a sub, a subsidiary of the Linux foundation, the Apache software foundation.

• Linux foundation
• cloud native computing foundation, which is a sub, a subsidiary of the Linux foundation
• the Apache software foundation

All these open source bodies have these democratic processes and transparency and openness built into it that a lot of the AI generation of founders in China have likely lived with and grew up in for a long time as they study and level up their own technical chops.

Because everybody learned from open source in the universities, no one buys a proprietary piece of Oracle software to learn about databases, right?

That’s just ridiculous.

And now they’re expressing it in different ways. And it’s funny, you mentioned that as another vector, right? To express their own identity, not just as a way to go abroad from a business perspective.

The first generation of founders, this is like the Zhang Yiming generation of founders. There are several people who also work in open source world and databases as well. They’re very liberal. They’ve all had traces on Weibo of expressions and opinions expressed about the state of their country that has nothing to do with the technology that they’re building. And now all of that has disappeared, right?

And I think the newer generation of AI founders, one thing I paid a lot of close attention to is not so much the fact that they open source anything at all, but the way they are doing it is increasingly more sophisticated and more exposed to different ways that Western communities are being built.

One example that I cited is Moonshot’s founder, Yang Zhilin, who in their shop makes a Kimi series of models. And it’s really interesting to see him coming out of his engineer shell and engage with the Reddit community during one of Kimi’s launches directly to talk with the community. It’s very Sam Altman-like as far as how exposed and accessible he is, right? To whoever wants to talk about Kimi or give him crap about Kimi. You have to have a pretty strong, thick skin, but also a good sense of humor to do Reddit really well. That is the kind of thing that the previous generation of founders, the Zhang Yiming generation, hasn’t quite cracked. But Yang Zhilin’s generation, I think, is cracking right now. He also did this, direct-to-camera release of Kimi k2.5, I think, which is very different from a lot of how Chinese models or Chinese products in general are being released. Manus did similar things for their product as well. All these things built on top of what I wrote in my piece, which is two decades plus of open source history that evolved from taking and consuming and just using free software for their own needs. And there’s still a lot of that going on and not enough giving back, unfortunately. But over time, the maturity has really evolved as far as contributing, but also engaging with a wider community to actually take in feedback, to prioritize feedback, letting people vote on certain elements of the roadmap, which is super, super common in the open source world. And the Chinese founders who engage in that world are frankly not that different.

So, in my view, those are the most, quote, Western part, the most, quote, Western portion of the Chinese tech community.

They’re also ones that are, of course, super aligned with the government. And we can talk about those as well. That’s kind of a different cloth. Well, I mean, Yang Zhilin’s an interesting one. He’s yet to agree to come on China Talk yet. TBD on that, guys. But he did a PhD in the U.S., right? And I don’t think that, I don’t know if there is any other, well, I guess the Pinduoduo CEO has an American degree from Duke or something. But aside from that, I don’t think any one of these top people have English as strong as, well, I mean, Jack Ma’s kind of an exception. But the fact that he lived here or worked at Meta and Google for a while is sort of illustrative of that. And, but, and yeah, his doing it, doing a YouTube video in English was definitely a step.

Kevin, why don’t you talk a little bit about the need for, or, or why these, kind of, everyone below Alibaba feels like they need to have kind of global adoption of their models. I think even Alibaba needs it. And I think that’s one of the reasons why the personnel change. But to more directly answer your question, the Chinese technique, the Chinese IT ecosystem as a business environment has been piss poor forever and continues to be piss poor. And what I mean by that is people just don’t spend enough money buying software solutions when they can hire five people to hack something on their own.

The business environments between the software makers, whether it’s a SaaS solution or AI model and sell it to would be buyers, whether it’s an SOE or a tech company or local government or a bank or whatever. That exchange in value has never grown to the level where you just could actually make a decent profit margin off of that. It’s incredibly, incredibly hard to make money off of code. So for any Chinese technology company where the core of your product is software, you have to go abroad at some point. You look for markets where there is a stronger penchant to pay for technology versus paying for people to hack your technology or to just jigger your technology into something you want. You never want to pay enough or you’re just going to pay some service contract, but you never want to pay a good subscription, all this sort of stuff.

So with that context in mind, the latest generation of AI labs from China, at the end of the day, they’re selling a model that’s packaged either in an API or packaged in some sort of a vertical chatbot or agent solution, TBD, what the form factor would be yet. And the only way they can really make good money is to go to a market:

• primarily the U.S.
• but also Western Europe
• maybe a little bit of Japan
• pockets of other parts of the world

where buying software is just not as scrutinized, where you can really scale that model, right? So they really don’t have a good option. It’s a very tough hand dealt when you’re a tech founder in China trying to make technology products to be sold and make a big company out of it.

And open source is this vector where you can at least get people to try yourself without, again, the toxicity I said, and kind of see it on the merit. And the playbook of packaging open source things to commercial software that you will subscribe to or pay for, that playbook has been played quite a few times in the U.S. market, especially in the cloud, the hyperscalers, packaging open source databases to be consumed on AWS. And most of the Chinese vendors and companies have really internalized how that playbook is run as well, which is why, if you want to turn it to present day, why Kimi and Minimax and now Tencent Cloud and every other cloud under the sun in China is packaging Llama almost immediately as a hosted solution to be consumed by the masses or whatnot.

“And that’s just a very simple open source to commercial playbook.”

And it’s more surprising to me that less U.S. players are doing that. We can speculate on why. But that playbook is very well run and well internalized on the Chinese ecosystem.

Yeah, China never figured out SaaS or was never interested in SaaS. And the sort of companies that ended up selling software ended up kind of being almost all forward deployed engineers without back ends of a thing to leverage.

It was just, OK, you’re just going to hire some software engineers and we’re not going to pay them a lot because in China, you don’t get paid a lot as a software engineer versus an entire economy being used to buying software, which then OpenAI and Anthropic can sort of leverage into and turn into — I mean, it’s three quarters of their business now is enterprise as opposed to consumer.

So not really having that as the thing that’s going to power your business means that you have to go to these B tier and C tier business models as opposed to just making the thing and selling enormous amounts of credits for it.

And then Anthropic now having Anthropic and OpenAI having forward deployed engineers is a nice thing, a nice extension as opposed to the core business, which is a lot more straightforward: keep building better models and keep selling tokens and try to get them cheaper and faster.

And it’s interesting. I think there was a moment in the Chinese tech ecosystem where buying software could become more scalable. It was trending that way all the way up until the tech crackdown—not so much on the Jack Ma side of things, but cracking down on education tech companies, cracking down gaming for a little while there.

Before that, there was a period where engineers were actually getting expensive in China enough where paying for the solution as opposed to paying for the body to make the solution might have been worth it economically. And then the crackdown just released a bunch of engineers into the market who are unemployed overnight and they need jobs.

So all the companies who weren’t going to pay because of the labor supply, let’s just say, just got a big boost of supply. So, oh, never mind. We’ve got all these engineers we can hire for cheap. Forget about paying a subscription for your stupid little CRM or something like that.

And so it’s funny if you want to replay that history, if the crackdown hasn’t happened, what would have happened to the software industry in China is another interesting branch that we will never get to really live, but we can speculate.

“That’s fascinating.”

Kevin, can you talk a little bit about the consumer fight domestically in China?

I mean, I’m a lot less close to the consumer side, to be honest with you. My investment is in the infrastructure side, the B2B side, the enterprise side. I would say most of the revenue is coming from consumer apps or APIs that power other consumer apps.

I think the consumer landscape in China is always a bit more—you can call it vibrant, you can call it fast moving—you can call it there might even be a higher willingness to pay on the consumer front. We were joking earlier in our live stream about gifting on live streaming platforms and whatnot. That’s a very well-baked business platform that is very to consumer in China for a very long time.

So I think the fight is there, but again, the market is just walled off. If you make the best Chinese speaking or the best chatbot, whatever consumer app for the Chinese market, the way you can really go abroad with that is hard, not impossible. TikTok has shown some ability to scale beyond that. If you find some commonality or common denominator to scale, right?

But as far as who’s up, who’s down, obviously Doubao from ByteDance seems to be the one. It’s a ChatGPT equivalent as far as the consumer level of both awareness, but also usage. Alibaba is actually quite behind, even though there might be nominally number two with whatever, the Qwen branded chat app that they’re releasing just very recently.

Another reason why maybe they’re much more commercially focused is because they’re losing that market share. While the cloud infrastructure B2B side is still going to be a very tough business for them to grow both domestically and abroad.

One of the interesting things about how this is going to evolve is we’ve recently got news that chat GPT is spending less effort on the shopping side of things. They had a big launch six months ago. I used it once or twice. It really wasn’t great. So maybe it deserved to die.

ByteDance and Alibaba sell a lot of stuff, sell outrageous amounts of things.

• OpenAI
• Anthropic
• ByteDance
• Alibaba
• Doubao
• Jack Ma
• Kevin “And if there are any two companies in the world that are going to figure out how to kind of get that sort of chat bot shopping experience, right?”

I have faith in the kind of Taobao and Douyin teams to make that happen.

And I think the WeChat, Tencent team too, for sure.

“They’re doing that as well, right?”

So it’s just it’ll be interesting because the kind of paths of highest revenue of just selling API access to enterprise isn’t really open to China.

That the sort of engineering hours that would have pursued those sorts of tier two things, as you said, the kind of open source bundling model or this shopping stuff.

We’re just going to get we’re going to get stuff.

Even if the Chinese labs aren’t making the frontier pushing models, we will still have interesting applications that the Chinese ecosystem is going to pursue sooner and with more energy than in the West.

And I think it is not a technical limitation on the models that is not letting ChatGPT or Claude help me pick out my new nail clipper or whatever.

“It’s just that it’s not enough.”

It’s not a high enough ROI decision to spend kind of management engineering hours on.

And it may be for that or other applications in the Chinese ecosystem.

Yeah, my speculation right now is that no matter how which direction the model game between the two sides evolve in the next year or two years or three years, the business environment or the trajectory of tech business in general will more or less stay the same in the sense that Alibaba or ByteDance or WeChat, Tencent will probably figure out agentic commerce, let’s just say, before Google and OpenAI does.

For sure, it will live within their own wall garden, but they will figure out that experience before while the U.S. side or the Western world will still race ahead when it comes to enterprise adoption because of all the dynamic that we just talked about, which is exactly the dynamic that we had 10 years ago or 15 years ago when it comes to which side is better at what.

And one last thing I would say about open source, which is that it isn’t just a B2B packaging play that is fairly natural.

The thing with open source is that it touches the developer more directly right away.

There is a vector in which it could touch the other kinds of consumer application or just landed application, practical application a little bit quicker than closed source kind of beautifully packaged but inaccessible models.

I kind of name check unitary in my long piece as a robotics firm that also open source their own VLA model.

So the hardware OEM side of things in China are also in the open source game to get their software packaged out there, to get more people interested in their platform if their vector into the market is actually to make the hardware really, really good, whether it’s quality or affordability or whatnot.

So open source could play in the consumer side, but it typically doesn’t play as well in the consumer tech side, but it could play well in the hardware side.

And that could be another thing that you see more offshoots in the Chinese ecosystem from the hardware vendors of the world.

What else do you want to talk about, Kevin?

Do you want to do it?

Do we talk about are you done with your conjecture about China?

Not less fun.

I thought it would be more personal than just the industry.

Oh, I mean, I think it’s a personal, I don’t know, personal.

It’s just, did you have a more personal take?

No, it was just, I mean, it’s tough, right? It’s a hard hand.

Well, I mean, like you just have, it’s, it’s, you have less money, more competition. It’s more intense. Upside is lower. Kind of political walls are way tighter closing in. You can’t be, you can’t aspire to become like a master of the universe in the way that you can in the West.

I mean, what, what, what, what, let’s, let’s talk to the other sides. What’s the upside?

You get to distill models. Everybody can do anything about it. Everybody gets to distill models.

I, I don’t know. What else is an upset? I guess you’re.

Yeah.

Right. There’s no upside. If you, if you have the ovarian lottery choice to be born 10 years ago and you’re a tech entrepreneur in your genes, which world you want to be born in. I think it’s 10 out of 10. You want to be born in the United States. There’s no question about it.

That that is a fork in a row. And it’s really unfortunate, right? Given the quality of, I don’t know, innovation or just grit, I think, to have to work within that much constraint and to still put out what you can put out to the world.

I think it’s admirable, but admorbidity only gets you so far when it comes to market cap.

And, you know, we’ll see how that goes.

And yeah, I don’t know if you want to talk about the whole Alibaba fiasco. I think it’s TBD where that’s going. But again, it’s just another episode, I think, of the whole difficult. Yeah, let’s do it. Well, let’s just, I mean, just to, just to kind of level set: Tencent market cap, $600 billion. Alibaba market cap is somewhere around that is 325 billion dollars. So those firms have been at it a long time.

If they were, I mean, the closest, the closest comps are in the multi trillion dollar range in the U.S. and little OpenAI and Anthropic have already, or on the, or on pace to surpass. Both of them having existed for three and let’s say 10 years, but we don’t really count the opening after 10 years.

So it’s not fair at some level that you can have just as sharp engineers doing, working just as hard, probably harder to be honest. And not, not having the same kind of ability to technical outcomes or overall compensation, global impact.

Yeah. Yeah. Should we address viewer questions? I feel that’s part of the job of a live streamer. Thoughts on XAI implosion from Daniel.

I stay fairly close to early stage investors on both sides of the aisle or both sides of the, I don’t know, ocean. And frankly, XAI’s financials and trajectory as a business is pretty terrible. And has been terrible since the get go, regardless of how you feel about the models and what it can do.

So really folding XAI within SpaceX and you kind of lump Twitter X in there too, as a conglomerate concern is one, the best way for the initial investors to have a good story or a good ending out of the XAI investment, which has always been pretty aggressive from the get go, given how much they’ve built.

And this is not to comment on the Colossus data centers, which I do think is quite a feat of physical accomplishment that only the Elons of the world can really push through to make it happen. Happens to be three hours from where I live at that. So very cool to see that happen near me, but as a business, right? As a concern, it will be very hard to survive on its own without more or less hiding in a giant conglomerate and kind of do what it does and see what they can do.

Right. I mean, we can ask the same thing for Gemini as well. We don’t know the P and L for Gemini as a, the vision. So in a sense, the independent model makers always have a little bit of a tougher time, maybe justifying the worth, but if they’re really doing well, I do think they probably deserve the valuation that they’re getting, which, again, and you can’t decide whether that is the right number or not.

Yeah. I don’t think the Gemini comparison is fair. The sort of the corporate thing that Gemini can plug into to boost growth is pretty enormous on its own. Compared to X, it would be pretty embarrassing if you were Google and you didn’t figure out how to use your giant models to make less, more money. And as, yeah, as you said, okay, we’re plugging it into X. I mean.

And what? And, well, I guess Tesla for the robots question mark, right? That’s a much more speculative, speculative thing than make the ads come up better on Google searches and YouTube. So. Yeah. Yeah.

Anyways, how is Minimax worth more than Baidu? Baidu’s got terrible vibes. I think they’ve just got terrible vibes for decades. Baidu, so yeah, this is maybe less about if Minimax is overvalued versus Baidu’s undervalued, right? You can kind of do both sides. I don’t know how you feel about it, Jordan, but Baidu’s probably undervalued from a pure financial perspective, but it just has kind of whiffed on almost everything up to this point.

Like literally everything, even including, I would say self-driving, which you’re fairly early into, from a platform perspective or robo taxi perspective, the model they were, I guess you can say early into, but it just isn’t the kind of shop that can deliver, unfortunately, based on the promises.

And Minimax, I do think it’s probably slightly overvalued in my opinion. I think there just aren’t that many ways to bet on a pure AI lab and both Minimax and Zhipu are the only ones out there in the public market right now, listed in Hong Kong.

So there’s probably a little bit of froth there as well, because if you look at the end of the day, Minimax gross margin has improved from 12% to 25% from its very first earnings report.

It is a doubling of improvement, but the base is quite low and the current state is still lower.

And that’s the thing on all these labs. It’s not like the open AI gross margin at this moment, at least it’s that much better, right? It’s in the forties, 30, forties range, depending on the business. So these are all very hard business to run, actually.

For some number, they’re both around 40, 45, 45 billion USD, which is, again, a failure for an A16Z. Maybe not a failure. We’d probably be happy with that in a 10 year fund, but still. What else is there to discuss?

Oh yeah, Alibaba implosion. We did a little bit of it earlier. “Why don’t you, why don’t you close the circle on that one, Kevin?”

I think for those of you who haven’t caught the news, there was a pretty public resignation of one of the tech leaders of the Qwen model team, Lin Junyang.

And he also has quite the big reputation within Alibaba for being the youngest P10 engineer, P10 being a very high senior engineering level that you don’t reach in your early thirties, right? I think it’s roughly where he is right now. It has open AI. Sam got fired by how it played out a little bit on Twitter. Of course, while China is actually asleep, which is super fascinating. Then a few other people also resigned. There was an all-hands meeting that happened with the CEO of Alibaba, who was also the CEO of Alibaba Cloud, Eddie Wu, to address the whole situation.

I wrote a little bit about it on the interconnected newsletter, but I think how this translates into financial results is TBD. I do think this is both somewhat idiosyncratic to Alibaba’s corporate culture, which is they’re very kind of anti sort of God figure, at least in the ranking files.

And I think Lin probably got to that level of celebrity, especially in the Western world as the only person or the most accessible and visible person that anybody from the West who interacts with the Qwen team. You basically talk to him and that he’s very good at being accessible, very helpful, generally a very good open source evangelist, not to mention really technical person as well.

So nothing against him whatsoever. But my way of thinking about the situation is the old JFK line, which is that is no longer about what Alibaba can do for Qwen is about what Qwen can do. For Alibaba. And I think that that is the strategic shift that we’re going to see as far as where can Qwen plug in to lift up the rest of the entire Alibaba ecosystem when it comes to shopping, consumer app, obviously the cloud. And I think Qwen probably vastly exceeded its expectation as well from the way it started.

I cannot believe the whole team is only about 100 people before the whole personnel change happened. It’s sort of a you’re a victim of your own success in a lot of ways. And then the top just comes in and sort of grabs it.

Here’s an interesting comp: Mira leaves open AI, brings 10 of her best friends, gets to raise $2 billion at a $12 billion valuation, has basically done. Has published nothing a year later, gets one megawatt of compute, signs a deal with Jensen, and is now raising at a $50 billion valuation.

If you take the Qwen team and bring them outside and put the hat out to the Chinese domestic ecosystem.

So you don’t want to do a Manus and I guess they kind of already are in Singapore. Maybe it’s a slightly different thing, but Kevin, why don’t you play out the two paths?

• If you’re that team in China, you want to make it in China, raise money domestically.
• versus you’re in Singapore and you’ll take Bill Gurley’s money, whatever.

Or else let’s do the global thing, but as a not Chinese company.

You mean if the crew that left Qwen wants to do their version of the thinking machine as a startup or as a Neolab, that’s the new term. It’s a Neolab. If they want to, what are the pathways for Neolab success? If you have the street cred and track record of the Qwen team, but you want to make it on your own today?

“I literally have no answer for what could be the successful outcome of a Neolab.”

It’s so crazy. We see the new Yang LeCun Neolab raising a billion dollars right off the bat as well. But that’s a $3 billion valuation. I feel like you could probably pull off domestically in China. You might get enough people to take a flyer on you for that, but not at the thinking labs. No, not at all. I think the 01.AI / MiniMax exit story or the going public story is probably the best comp for any investor looking at persuading the Qwen, where is it?

I want to think of a Fairchild, the trader is, I don’t know how many left the traders for five to start their own Neolab in China. That’s probably your best get alcohol from a financial perspective, right? But I think having that reputation that he has garnered could give you a really big jumpstart. And this is probably the last check that anybody, at least within China, but maybe even globally will write into a Neolab.

I don’t think there’s a good Neolab story at the Yann LeCun level, at the Mira Murati level that could really garner both that much money, that much attention with relatively little to have shown for up to this point. And I think the Qwen team actually was going to have a Singapore team as well. That was one of the things that was maybe on the table that got shelved because of this personnel change.

So my best guess is that they’re going to decant to Singapore and do Manus 2.0, but even cleaner, and see how that goes.

Well, it’s a really interesting question: if all this team wants to do is open source AI, and they are in Singapore, so not at all China adjacent, there’s no kind of export control risk, which Manus kind of had to deal with for a hot minute. Is that a viable business? Is that something investors would be excited about if you have a team which is very deeply ideologically committed to just making open source models?

Right. I think a lot of open source founders, quite frankly, start out that way. This isn’t, again, an AI thing. This is an open source thing. I’ve seen at least three generations of pure open source founders who got incredible traction with their open source databases, observability tool, whatever you want to bake, make open source in the cloud stack. Right now we’re in the AI world. And then you raise money and the moment you raise money, you slowly, but surely come to realize that you have to build a business.

So it really comes down to how malleable — let’s just play this out in full characters — that Lin Jun Yang is going to be the CEO. He’s going to be the figurehead, right? He’s going to be the Dario of this new lab; how good he is going to be as a CEO. He will be a CEO. He will not be a tech lead of a division within a giant corporation that has his own set of problems. Will he do that? Right. And if he can, then yeah, they’ll figure out a way to commercialize it somehow.

I think I have faith in figuring that out. Again, the playbook isn’t a mystery, but the timing and where you start to jigger and address those things is going to be very, very hard. And no one has really quite played that perfectly. I don’t think even a Red Hat — at the end of the day, it got sold to IBM, fantastic. Congratulations, but that’s about it.

But we should have a little bookie market. So the odds that he shows up at Meta revives Llama. He could get hired away. Exactly. Just go up to the Man of Singapore office that has Meta badges and just call it a day. He gets all the compute he wants, hangs out with Alex Wong. That’ll be fun. I think that’s an outcome.

We can probably rule out Anthropic. I think we can, I don’t know. This guy doesn’t seem like he. Well, Anthropic never wanted to open source anything in the first place. Even I try to straddle both worlds and try to be nice to everybody. But Anthropic was just “no.”

Well, what’s funny, I do want to talk about this a little bit. We wrote a piece for China Talk, exploring the ups and downs of Anthropic’s rep within China. And we had this fan, we had this fascinating swing with on the one hand, everyone was hating on them because they released that report saying that all the Chinese firms had been doing distillation of their models.

“oh, how dare you, you guys stole all this other IP. Don’t you dare give us any crap.”

And then a week later they have this giant blow up with the department of war and they’re like, “oh, wow.” Dario standing up for humanity, stopping the Terminator from taking over the world.

Which, incidentally, is an analogy that a PLA official shouted out, saying:

“we don’t want Terminator to happen.”

I don’t know. It’s kind of a fun wrinkle. I think he was persona non grata number one, for all of his stances around export controls and whatnot.

But felt like he gained a few a few popularity points with the whole Pentagon kerfuffle domestically.

• Domestic within China, domestic within the United States.
• Within China, within China.
• Certainly in the United States, I think he’s garnering quite a bit of domestic goodwill.

You can just see it from the Claude download chart. It’s finally at number one after being at, what, number 17 for ever.

Well, everybody I know, not to say that’s indicative for a big sample size, who is part of the Chinese ecosystem. They love using Claude. Most of them are engineers, who I know. They love Claude code. They love the product that’s coming out of Anthropic. I think that is a good way to understand the practicality of people who work in the tech world or the business world in China. Is that, from an ideological perspective, you probably don’t like what the guy who made the product said about your country or what you do, or whatnot. But a good product is a good product and I’ll still use it, and it’s still very, very popular.

And it’s, I don’t even know how that usage falls in an Anthropic financial statement. Do they even have a revenue line that says PRC on it? Because they’re definitely getting usage. I just don’t know how that actually works from an account, a receivable, or accounting perspective.

Well, it’s all Asia Pacific. It’s all coming through Japan and Korea, right?

What’s funny is they’ve really tried not to, they had this whole blog post saying

“we’re going to try to really shut down all the, all the accounts.”

And both, both OpenAI and Anthropic have since published reports saying, we caught some Chinese state hackers slipping, revealing that they were trying to find exploits or whatnot using our models instead of domestic or open source ones. But I don’t know, it’s hard to regulate this sort of thing.

VPNs can be really good. The thing about VPNs is at a certain point, if it’s really that important, you can just control someone’s computer who’s in Japan. How are they going to know? I guess, are they going to ban everyone who’s using simplified Chinese to use their models? I don’t necessarily think so. And if you’re not going to do that, then to a certain extent, especially with individual users, where there’s a will, when there’s a will, there’s a way.

Maybe with the kind of, you can at some point get a hand on the distillation stuff. It seemed the way they were writing it up, it wasn’t super sophisticated and it was kind of obvious. And the question is then how sophisticated can you get on the distillation side to a bit of a bit of a kind of mouse game between the Chinese and Western model makers.

I think it all has to do with traffic pattern, right? A normal person or even an engineer using cloud from China to do some normal engineering or side project will just look different on the observability level than a distillation campaign or a run or what have you. And that’s what any cybersecurity person needs to do for something that has as much traction to figure out the difference. And not just blanket ban or blanket allow whatever happens on your platform.

So yeah. Cool. All right. We’re good. Maybe we should call it there. This was fun.

“Everyone subscribe to sale.”

If you work at a large organization, reach out. you get a big bulk subscription to all of the greatest AI content on Substack. I read sale.com read sale.com, but this was fun. Thanks everyone for joining today. Thanks everybody for joining.