DeepSeek’s fundraising, unique A-share features, and what’s ahead with The Information

Hello everyone, I've been on the road, so I'll be delaying this week’s written post and next week’s podcast episode. Apologies in advance.

In this episode, I’m joined by Asia Bureau Chief Jing Yang and Senior Reporter Juro Osawa from The Information to unpack DeepSeek’s surprising shift toward external capital, its unique investor structure, and what its rapid revenue growth reveals about the economics of AI. We also explore why China’s AI labs remain fiercely competitive despite a far smaller capital market than the US.

The conversation moves from DeepSeek and Huawei to accusations of distillation, ByteDance, Alibaba, and Tencent, and the growing race to turn AI models into real businesses. We also look at China’s emerging advantage in robotics, from its dense hardware supply chain to the vast amounts of real-world data being generated through deployment.

Finally, we discuss the next frontier: world models. As the race moves beyond language models, can China’s hardware and data advantages translate into an edge in embodied AI? And in a provocative final take, Jing argues that Chinese frontier models may never fully overtake their US counterparts.

The AI Proem Podcast is part of the AI Proem newsletter, which has ~13k followers globally. To learn more about China AI, the business of AI, and how AI is impacting businesses, please check out the newsletter here and more insightful conversations here.

Chapters

00:52 The Rise of Chinese AI Models

04:06 DeepSeek’s Unique Position in the Market

12:31 Capital Structure and Investor Dynamics

16:12 The Landscape of Chinese AI Labs

19:17 DeepSeek and Huawei: A Strategic Partnership

25:43 The Competitive Landscape of AI Labs

33:51 Challenges in the Chinese Capital Market

36:44 ByteDance’s Unique Approach to Model Training

41:32 The Future of BAT Companies in AI

48:22 China’s Robotics Landscape and Advantages

53:51 Impact of US Regulations on Robotics

54:43 Unitree’s IPO and Market Dynamics

Grace Shao (00:00)

Hey Jing, hey Juro thank you so much for joining AI Proem today.

Juro (00:03)

Hey

Jing Yang (00:03)

Yes.

Grace Shao (00:04)

to start, you know, let’s start with Ox alpha. That was such a teaser. It turns out to be ZAI again, but I think it was very much expected by people who do watch space closely. It was no surprise in that sense. But I think what’s shocked a lot of people is that, you know, the inference was running on Chinese domestic ships and potentially GLM five point three flash.

The price, it’s priced at about like one fortieth of Opus 4.8. That’s quite significant of a gap, right? Tell us about what you think about that, the implication on these continued Chinese open weight models that are coming out strong but cheap, and how that’s impacting US frontier models.

Juro (00:43)

Yeah, maybe I can start on this one, but I think those flash models are coming out, like Deep Seek had the V4 flash, which is a smaller size one, and that became very popular. And then this one comes out. Alibaba just had a three point eight flash as well. So those are smaller size, you know, lower cost models that can still handle a lot of like AI Asian type of tasks really well. So

know, this is like a sort of sweet spot for demand, right? So a lot of Chinese companies with open source models are coming out. So that’s one, you know, one thing about this, you know, that’s part of that broader trend. But I think as you pointed out, the chip part, you know, it kind of shows how far the Chinese chips have come in terms of inference, right? And

They have become more capable, especially Huawei chips, being used, you know, more and more for inference. And you know, Deep Seek came out with V4 and you know it was also optimized for you know to run on Huawei chips as well. And we’re gonna see more and more of this for sure. But I mean to what to keep in mind that when it comes to training, it’s not the same story yet, not quite yet. So, you know, a lot of Chinese companies are using, still using the most advanced

NVIDIA chips and trying to gain access to that and which we also wrote about recently.

Jing Yang (02:03)

Yeah, I mean the only thing I I’ll add to that is I think the bigger picture is that there is a shortage of inference trips in both the US and China, but in China the shortage is much more runs much deeper and you know

A part of that is because of government policies, right? We still have not seen H two hundred actually officially being allowed to export into China. So the right course for all the major leading Chinese AI labs is they have to adapt the inference to domestic trips and to you know this concept of homogeneous compute, which has been discussed in Silicon Valley, but as well as in China.

And then the hemogeneous part in the Chinese context is very different. It’s about being able to run your inference on a combination of NVIDIA and and and say five different other Chinese GPUs. Right. So we saw that Kimi K3 had to suspend our new customer subscription shortly after it was released, and that was directly a result of not having enough inference chips.

to to to you know to to meet the surging demand.

Grace Shao (03:14)

I think Kimi wasn’t the only one that faced this kind of issue. And Deep Seek and ZAI throughout last year, I think, had to like, you know, manage their demand as well. But I wanna bring the conversation to Deep Seek. You know, you guys broke the story on well, actually Juro just broke the story on Deep Seek’s revenue. And I believe Jing Yu wrote the story about Deep Seek’s fundraising. It’s quite fascinating because DeepSeek, quite secretive, you know, their investor letter or investor conversation was leaked.

Apparently the founder was not happy that it was leaked, But, you know, they are fundraising now. And that’s shocked a lot of people because for the longest time they didn’t want to take any external capital. Jing, do you want to take that first and just talk about, you know, their capital structure, their fundraising, how unique it was, and then maybe we’ll pass it to Juro to talk about, you know, his recent findings about the money they’re making.

Jing Yang (04:03)

Sure. Yeah, so both Juro and I and our colleague Qianer as well, we’ve been reporting Deep Seek very closely since the beginning of last year. So I think one thing I like to sort of take the credit for is that I think most media allies sort of moved on after the initial buzzy period around Deep Seek early last year, but we stayed. I think we immediately recognize that this company is gonna be a very unique

like sort of existence in China AI landscape for the time being. and we reported, I mean I was just as shocked as anyone when we got a tip when we had our very first fundraising story. Because knowing the company the way we knew it for the last eighteen months, we were I don’t know if you were about you, but I was quite shocked. I think I edited it, I wrote that story in like huge disbelief. Like it’s a funny moment for our

journalism because a lot of times you kinda expect things will happen in a certain way or the direction of travel, at least you have a sentence on the post, but this one I think it completely surprised me. And then so in the two months that we’ve been chasing every step of the way of the first ever fundraising, one question that always lingered in my head is what

Prompted this change of this dramatic change of attitude toward external money. So eventually we were able to do a sort of the deeper story in which we revealed that it was Anthropic’s mythos preview that contributed a lot to

CEO Liang Wenfeng thinking, because I think there was a period of time, if you remember, like in the second half of last year, people were in AI research community, people were doubting or casting skepticism on whether the scaling law still exists, if if it’s still going to yield kind of you know progress. and then I think mythos showed that scaling law still exists. And then that’s when Liang realized that okay,

even if we have done innovation when it comes to you know in cr improving model efficiency. To really get to the next level, like Mythos demonstrated, we need to fully embrace the competition and we fully in not the competition, but fully embrace the game of, you know, you know, gathering resources. We need a lot more data and a lot more compute and therefore you need money. And and it’s a

money at a scale at a magnitude that Leon himself cannot fully fund anymore. That’s essentially the reason. And I think the broader takeaway while having covered that story is it is kinda like, you know, how the the peer pressure and the the the competition really makes it very difficult for any any lab

to just you know be stay on the sidelines because you know up until April when DeepSeek D came out with the first funding pitch, they were I think I believe the only holdout right around the world’s major AL labs to not have done any funding, fundraising.

So that’s quite striking and I’d like to keep reminding people that I’m not passing judgment, whether it’s a good thing to join this arms race, but just that, you know, when the last holder had to succumb to the broader pressure, it tells us something.

Grace Shao (07:08)

It’s

just a very, very capital intensive game that they’re playing. but I wanna kinda throw this to Juro then. Tell us about the recent story you just broke.

Juro (07:17)

Yeah, so we wrote about their finances. So their revenue you know for the first seven months of this year was about seventy million dollars. and that’s you know seems very small still, but that’s still you know, like about ten x compared to all of last year. So just considering how this company last year, you know, seemed like you know they had no interest in really making money. And so this year they’re just really getting started.

Right. And they’re starting to have this growth. And looks like a lot of the revenue also came, you know, quite recently as well, in the past few months. And so even though the number is small, also another remarkable thing is that their gross margin is quite solid, so forty four forty four point six percent. And if you look at their API sales for models, that’s like eighty three percent. And so they’re maintaining this like a very high

quite high margin, even though you know what they charge is you know they’re charging very low prices. And so this also shows what they’ve been doing to kind of lower the running cost, running costs of their models, right? And so a lot of those numbers are telling us, you know, kind of where they are. But I think this is important because they’re raising money and even you know talking about an IPO as well. So

No, investors will be looking at those, you know, this progress very, very closely.

Grace Shao (08:43)

I think that’s really interesting.

Jing Yang (08:43)

And and Grace you asked earlier

about the capital structure. I I’ll just

Grace Shao (08:48)

Mm mm.

Jing Yang (08:48)

briefly talk about this because I think it’s also quite important. so

Bear in mind that I think between Juro and I, we have probably 15 to 20 years of recovering venture capital as a journalist. And this is a unique deal for both of us. The way I like to talk about it is that if you categorize the types of investors in deep sea first around, I would categorize as four types. For number one is Leon himself. U number two is the sort of strategic, you know, in you know, the tens and the end of the net, right? And say it’s all. And then third type is the VC firms. And then the third the fourth.

type is the national air fund. so we don’t care, you know, Leon, he’s his money, he is he’s putting, you know, you know, in he’s putting his money in his in his own company. We don’t care what he used he he does with it. But for this following three types of investors, only the national air fund can invest in the SIC corporate entity directly. And then the other two types of investors have to input have to wire their money into a SPV structure.

like a LP structure that’s actually managed by Leon himself. And as part of that arrangement, they also do not have voting shares. They have to essentially make Leon the proxy for their votes. and they also have to agree to lock up their shares for five years. I guess with the exception of you know, events like an IPO. This is all very unusual, right? Because you are putting money into

not into the company or funding. And well, you are putting mu money into the controlling shareholder of the company are funding and at the same time you are giving up on any voting rights. and the reason for that sort of rigorous or strange structure is because as we reported that Leon wants to make sure that he only attracted in you know investors who are there for the long run, not for a quick exit. he is I guess in today’s popular

speech in Silicon Valley, a very AGI appealed. So he wants to keep Deep Seek open source and he does not like commercialization or generating revenue despite as Juro you know mentioned that they generated a pretty quick and revenue growth but that’s not their goal, right? so you can debate whether the end justifies the Minx.

but this is the structure that they came up with. And I know that maybe maybe a lot of us would tend to have an impression that while this is such a hot company and when they finally come to the market with the fundraising obviously investor would like to would just like, you know,

I’m swarming and then regardless of the terms, but actually that’s not the case. I know definitely investors who passed off on the deal exactly because of the structure, they feel like they could not accept those terms. and I think Deep Seek has actually taken that feedback because in the current, you know, the second funding round, they actually are loosening some of those requirements that were in the first round.

Grace Shao (11:39)

That’s

super interesting because I think, okay, just let me process both what both of you said. On one half, what Jing said is that he still wants control of the company and like he doesn’t want to run it like a commercial business, right? Like it’s a very unique setup. And then on the other hand, there’s a clear sentiment change or at least a strategic shift that is being pushed by the fact that he has to get money because everyone else has money. And if he doesn’t have the money, he can’t do the RD that is required, right? And then obviously the IPO that’s imminent, or at least, you know, in the pipeline. Now

I do want to ask one question. You know, he’s openly talked about not wanting to frankly make money for the sake of making money, but he said that I want to make money, but that use that money for continued RD. He’s kind of openly talked about, but even then, the margins, like Juro said, are still extremely high. So what does that say about the other labs?

Jing Yang (12:29)

Maybe I’ll I’ll try to take a stab and then

I just point out as we reported that they have this really high API sales margin, like 80% plus, because there is actually real renovation behind it. The industry term is that they have a very strong infar team. But basically what that means is that they have managed to bring down the inference cost of their models so that for the same number of tasks or to tokens consumed by you know, you you you need it, you need less chips.

fewer chips. and then whether that’s the way to go to get a higher margin. For example, Juro Feel determined by like, you know, we saw Minimax results just recently and then they have a much lower level margin for example, right? So

Juro (13:14)

Yeah, I think definitely Deep Seek, people have written about this too, but some researchers have pointed out that their, you know, innovation is in how they run their AI, you know, with like just chip requirements, are very low. Like memory chips, especially, a lot of companies are, you know, dealing with that cost right now. And so yeah, that’s where Deep Seek seems to be able to make a difference. Yeah.

Jing Yang (13:40)

Yeah, so so I think I’m trying to sort of like try no

channel my Wenfeng here not that I’ve never ever met him or anything but based on enough that I’ve heard about from people who do know him right he believes that AI should benefit all any advancement in technology and AI should benefit all. And then to make it benefit all the first thing you need is to make it cheap, make it affordable. And then and then therefore you need to first make the deployment of AI cheap, cheaper and more efficient. That’s why they have this, you know improvements on infra.

And then as a result, as long as you have something that’s both good and cheap, then obviously you are gonna generate a lot of revenue, right? Right, and profit eventually. I think for people who know him, they would say this is the train of logic. that not that it’s this high margin shows the company is so profit driven, but

When you are pursuing the greater good, money will naturally follow.

Grace Shao (14:40)

No, I think that’s definitely something you’re hearing more and more coming from the AI builders. Like I think it was Neil Mova that recently we just saw Patrick Shaughnessy. The goal is to make AI as cheap as possible. And this ‘cause there’s definitely a sentiment shift from like even a year ago where the narrative is really about like how much money can we make with AI or like, do we pay premium for intelligence, et cetera.

Let me bring the conversation back to the Chinese labs. like you just mentioned Jing, you covered Minimax as well, obviously Moonshot, Zai these are some of the leading labs in China. What’s their current structure? there seems to be a lot of them. Should we be expecting some consolidation? I guess this is two two questions separately, but just kind of give us a high level landscape breakdown.

Jing Yang (15:20)

Yeah, I’m happy to do a micro thing and then I’ll let Juro talk about the individual labs since he knows them better than me. It’s just something that I’ve been actually discussing with my own, you know, contents and investors a lot these days. If you look at in the US, we have it’s been consolidated to like only just two to three major labs left. And in China we still have, depending on who you include or not include the list general, you have about six to seven or seven to eight.

AI model makers that are still firmly fully in the race and any kind of consolidation is not on the cards anytime soon. And this is sort of like a characteristic of the Chinese economy and the Chinese market. If you look at EV, how many years has the EV battle been happening?

And how many companies are still in the race, right? So I think there are several structural reasons. Number one is that you know, the capital markets in China are highly regulated. It’s very difficult to do once you are already a public company. It’s very difficult for a public company to acquire another public company. Or or or in an and I transaction, if any either party is public, the transaction will become

a lot harder. and then we are still sort of we’re just a few years in the aftermath of the antitrust crackdown, right? I think so that makes a lot of the tech incumbents, you know, the Alibaba and the Bidowns of the world also a bit hesitant about that. And then number two is is sort of fundamentally the the the both the private and the public markets in China lack

like the kind of breaths and depths l as in the US. And that made up that created and the lack of, you know, when you do not have enough capital, either in private or pa or public markets going around, that makes potential acquirers very hesitant or reluctant about paying a premium. Right? there were opportunities for there were discussions.

that say for example I can tell you that one of the major, you know, tech companies was seriously considering acquiring one of the smaller labs, that then by the

And then that kind of talks actually happened several times. It was involving different sort of combination of companies, about two to three years ago. And then none of that happened because even back then, the the established tech companies are not willing to pay what they consider a premium, right? that’s number two. and so I think

I think it’s gonna be really quite hard to see any real consolidation and I I think with if Deep Seek does go public and we’re also look if Deep Sea Moonshot and Stepfund, all three of them say go public next year, then it’s just gonna gonna make it even harder for any kind of consolidation to happen. And the world is gonna be in a place where companies continue to compete.

and then the fight to the competition would just have to drag on for years, for a lot more years.

Grace Shao (18:25)

Very

interesting take. Juro, do you have anything to add on that?

Juro (18:28)

Yeah, I mean, I guess it’s also quite difficult to imagine one of those founders working for another one of the founders. Or also, you know, the idea of even Alibaba or ByteDance or Tencent acquiring them because I mean they do also have their own labs and you know they brought in Tencent brought in Yao Xing Yu from OpenAI and you know ByteDance has the biggest AI team in China, so it doesn’t really have the rationale, you know, for those companies to just suddenly

you know, turn around to buy one of those, right? So

Jing Yang (19:01)

Yeah, the the third

reason actually, sorry, I forgot to mention, I gave you two before. The third one kinda touches on what Juro mentioned is talent. There’s just so much more talent in China. So it’s very easy for someone like Jiang Y Ming or or I don’t know, Polima or whoever, where they would think, Okay, if you don’t want to sell to me at a discount, then I’ll do this on my own. I’ll roll up my sleeves and do it on my own. I can’t hire people. So

Grace Shao (19:26)

Feel like, yes, I agree with you.

Jing Yang (19:27)

Yeah.

Grace Shao (19:27)

But then you look at you read the book, the Google Deep Mind story, right? It’s like no one thought Demise would go in and work for a big tech either, but eventually it happened. I mean, now that it’s clearly proven that it didn’t work out, but it was an interesting phase where I wonder if at one point like you just can’t keep on Juan right? Like there’s like you hit a limit of the Juan and then you just like you’re forced also to consolidate for resources for talent and for the like.

For the lack of compute out there. So if you believe in AGI, the some as some claim they do, and your goal is really to have the resources to pursue that scientific research and breakthrough, then isn’t it in some ways going to the company with the most resources the best outcome? I’ve heard people in the industry argue that, right? Like, I don’t know. What are your thoughts on that?

Jing Yang (20:14)

No, I think what you said is a very sensible, rational take. But the problem is for I think the three reasons I mentioned earlier in China, this kind of rational, you know, economy of scale, right? Economics one, right? Just somehow does not apply to China that people would rather trend than

Grace Shao (20:35)

They’re not rational. They’re not making rational decisions, is what you’re saying.

Juro (20:37)

Well I think there’s another

Another factor is that you know Alibaba for example has invested in so many of those AI startups, right? And some of them have gone public as well. And for them, because they have Alibaba Cloud, which has a big business with a lot of those AI labs as well. So, you know, like they would rent chips from Alibaba Cloud and you know use their infrastructure. So yeah, I mean for them to kind of decide to just buy them, I mean

when they can keep them as major customers and, you know, also hold a stake. So I guess that’s another logic for, you big cloud platforms, right?

Grace Shao (21:15)

Yeah, and they have exposure

Jing Yang (21:16)

And and the

Grace Shao (21:17)

anyway. Mm.

Jing Yang (21:18)

yeah, and then the last reason I w I would really I think also important is is the enterprise market is still

very massively underdeveloping China, which which has so far limited the the commercial, the revenue prospect of all these companies. And then when when you are projecting your revenue, that is only like 10X in the next five years instead of 100 X, then then of of course you are gonna be very you’re gonna be stingy when it comes to MA. Right. I think we can see the kind of consolidation that happened in the US. It’s it’s because they these companies actually you know can generate a lot of revenue from it.

enterprise market so they can afford to spend on MA.

Grace Shao (21:57)

Fair. Well, we’re all speculating. Let’s watch and see and what happens the next few years. Sometimes there’s a wild card thrown at us. I I do wanna ask you guys, give us some color on what happened between Deep Seek and Huawei. I mean, we started a conversation with Juro telling us about, you know, how GLM five point three Flash was tr you know, inference on domestic trips. But I think Deep Seek really took one for the team there and made the first step over. What was I guess an incentive? Was it really just because, you know

They wanted to show the ecosystem that this could be done. was there a commercial reason on the Huawei side? Did they approach Deep Seek? Was there a top-down mandate from Beijing? Like, how did that play out?

Jing Yang (22:35)

Yeah, I’ll take this one. so we and some other outlets reported throughout twenty twenty five that Deep Seek has been adapting to How Richards.

And then we’re trying to and just by the way, in order to make the inference you know adaptable to domestic chips, you have to basically reengineer your training process, right? Once you’ve done the training within video chips, then you have to then recreate that process to domestic chips. to like I’m butchering this, but to to oversimplify. and so I think

Us including many other observers, Chubash, John Warder, Elam, masses, just assumed that this is because Gypsy has been asked by the government that you now are a national champion. You have to take a lead in adapting to domestic semiconductors. and which is why you know when we reported, I think that was a story in June that we did.

That it was actually DeepSeek that took the initiative in adapting to Huawei chips. And Huawei didn’t and it wasn’t until DeepSeek spent some time tinkering with Huawei GPUs that the Huawei engineers got a wind of it. And then they approached DeepSeek and said, how can we support you? So why did DeepSeek do this?

The reason we report it is that as I said earlier, Leon is someone who believes in AGI and believes that AI should be inclusive. And so inclusion in here means you know you need to have a diversity of technologies in both AI and the semiconductor, right? He based on the leaked investor call, we can see that that’s the case, but bear in mind I I’d like to point out that we when we reported this the

the investor call map was not leaked, right? Had it not been leaked. So we reported that, you know, he believes that NVIDIA should not be, there shouldn’t be just one dominant GPU designer. There should not just be one dominant technology. And then and then Deep Seek wants to play its part in increasing that diversity. That’s why they did it. There’s nothing sort of forced or you know political behind it.

Grace Shao (24:35)

Yeah. Juro, so this question’s for you. I think that’s a really good context and understanding why DeepSeek did what DeepSeek did, but help us understand just who plays what role in the ecosystem right now, if you had to generalize in one or two sentences about each.

Juro (24:51)

I mean each AI company in China.

Grace Shao (24:53)

AI lab at this point. We’re talking about I guess just the labs. You can include the big tech if you want, but just like the let’s just a h eight companies that keep on pushing out models.

Juro (25:04)

So yeah, I think for Moonshot, you know, Kimi, we can say that they are trying very hard right now to be the sort of anthropic of China and really focusing on the you know coding, API sales and and especially they are you know really expanding globally, right? And so they try to show that they can, they are the ones that can really compete at the frontier level.

And the frontier coding kind of you know capabilities. So and they know that they can really build a big business like Anthropic did. So that’s the kind of path that they see, I think. And I think Jupu is you know ZAI, what they call it now. so they are trying to also you know focus on that kind of market, but they also have a you know kind of

much stronger domestic background with the Xinhua University, you know, researchers that this you know that studied it. So I mean they have been trying to kind of transition from the previous business focus was you know doing a lot of work for domestic companies including like state owned companies and you know helping them deploy their AI models. And but then they want to also become this more

More globalized, you know, also selling AI models, you know, and also like a little bit like Anthropic, you know. I mean, that’s the path that they a lot of those guys see because that is the path to revenue right now, right? And I think deep seek, as we discussed, you know, stands in a you know a little bit different position, interesting. I mean because they do you know, they do seem to kind of, you know, they do keep their prices very low.

And also not just chasing like Frontier in the same way that Kimi does, but you know, how their models can be also easily deployed, right? And by many, you know, I mean there are many different requirements from you know different kinds of users. And so they are going for, you know, like they may not necessarily be the most powerful, but you know, they are kind of trying to

like you know AI for everybody kind of approach right now as at least with the pricing they have, right? And I think they also do have a lot of, you know, do handle a lot of domestic demand as well. Like so, that’s the national champion aspect. And then the minimax I think they have chased you know both video and LLM. And I think recently they have

They have kind of made a comeback with the latest video model. But I think it’s still quite difficult to see and they do seem a little bit confused to me that in terms of deciding which direction. And because it’s very resource intensive to try to chase both markets because video market you’re competing against ByteDance, which has you know just infinite amount of money and resources. So but

You know, they’re definitely not out of the race. You know, they are, you know, trying to keep up. So yeah, that’s and and the big labs, you know, they try to do everything. And Alibaba, especially with the cloud, you know, they try to, you know, be in every segment of the market. And you can say that about byte dance, but maybe we can talk about byte dance later when we talk about distillation stuff. But yeah. Yeah.

Grace Shao (28:24)

I like how you’ve set

set the tone. We’re gonna be talking about distillation stuff later. All right.

Juro (28:28)

Yeah.

Grace Shao (28:29)

No, no, that’s really good context.

Jing Yang (28:29)

There there is a popular saying now in

The AI space in China is the, I don’t know if you have a smart translation for that into English, but DeepSeek Jansen.

So basically the Deep Seek has set the survival line for everybody else. Either you go for Solta, Solta, Solta, or you go cheap, cheap, cheap. and then but you cannot go cheaper than Deep Seek. So that’s the survival line. And I like to just one more thing, I like to sort of compare, you know, Kimi and Memex a little bit because I’m sure you guys all remember, right? Kimi, I think back in 24,

you know, actually spend quite a bit of an amount of money marketing, advertising for the chatbot, right? And then I think at some point we realized there’s no way that they could compete against Dobao and by Bay Dance. So then they switch to, you know, working on SOTA models. And whereas Minimax continues to be

You know, Juro, actually I wanna ask you, like ‘cause they continue to still be dabbling in everything. Yeah, I think I think yeah.

Juro (29:26)

So actually

On that point I guess Minimax has kind of moved a little bit away from consumer apps, especially something like if you remember talkie, like that was the early sort of hit.

Jing Yang (29:38)

I remember talking.

Juro (29:40)

But I think internally Yeah, yeah. But internally that’s that’s definitely not

Grace Shao (29:40)

Yes, and High Law. Like they had so many products. Sorry, go on.

Jing Yang (29:42)

Yes.

Juro (29:46)

That’s definitely not the priority now. I think over the past year, the priority has been, you know, really the models, right? And models and API. So I think that’s definitely and you know some would argue that Kimi has kind of moved ahead more quickly. and they had big success with K three. But essentially that’s the direction that you know, all these guys want to move in. So I’m I think Minimax is definitely trying to

move in the same direction as well.

Grace Shao (30:15)

So then my question is if every single lab is somewhat doing the same thing, like it kind of goes back to what we were talking about earlier, it’s just gonna be like a price for a never-ending duet, right? Then are we seeing any interesting application innovation right now you think that is being a bit undercovered or underappreciated? Maybe not coming out of these labs.

Jing Yang (30:36)

Juro, do you want to

Juro (30:37)

I think the application side battle for consumer apps is definitely, you know, very fierce. But we just happen to pay more attention, like I guess everybody’s been paying more attention to the SOTA models. But I think Dobao from ByteTance and Qwen and Alibaba’s app, I mean that competition continues, right? And yeah, it’s not like they don’t emphasize that. So yeah.

It’s just that the people I think have been talking more about something like Chimi K three, but that domestic competition for apps is definitely there and it’s gonna just keep going for sure.

Jing Yang (31:12)

I think there’s been you know increasing appreciation for model rappers. I I remember when Manners just became rebel viral in March last year, people were like, this is just another rapper, and there’s no yeah, and then there’s yeah, there’s no modes

Grace Shao (31:24)

It used to be a dirty word. Now people are like harnesses. Like

Jing Yang (31:29)

if you are just a rapper and now yeah, now people call it a harness. but I think that is sort of on the right.

Course, I think something that I would like to write about more that I have not been able to for various reasons is that actually after the Manners Wave and then the open call of friends in China, what we have seen is actually there’s a lot of flying under the radar Chinese startups that actually have built really smart harnesses targeted at what I call small business like small enterprise customers. and then there’s actual mode.

in doing that in offering that business and there’s actual demand and they are offering this to not just a male and Chinese customers, actually a lot of them offering them to like you know overseas customers. and imagine if you are a company that has a loyal employee size of like about 50 to 100 in like a non-tech, non-digital industry, you want to use AI, you don’t know how, you don’t even know where to start. It’s not possible for these kind of companies to go come to go on like I don’t know, open

Router and say, okay, I’m gonna use the one, two, three, four, these models for this different task, right? They don’t have that know-how. and this is why the kind of you know AI, I think they call it AI employees, right? This AI employee business, which essentially harnesses built on top of you know models, it actually are our take our gaining traction actually proven to be quite useful. so

So I think that’s something that gives people, including me, a little bit of hope and the optimism for the enterprise market in China in the future. Yeah.

Grace Shao (32:59)

Yeah,

a lot of very strong vertical products coming out. I wanna direct the attention back to Jing for a few questions on, you know, the capital market. So the secondary market in the US for A Labs obviously been incredibly active. But we’re seeing nothing like that in China. And valuation obviously has it’s like digits away in terms of like the gap. How do you view that? Why is that? Obviously, kinda touched on it or alluded to already, just in general, China’s

You know, Deep Seek and P space is not as vibrant, but is there other reasons behind this phenomenon?

Jing Yang (33:33)

When you say secondary, you mean like the secondary share sale market, not the public market, right? I think the number one reason still is what I mentioned earlier, there’s just not enough money, right? You know, I think the American LPs have, I think a big part, big chunk of the American money has left, right? And then that void still has not been fully replaced and may never be. and that has

Has really suppressed reven you know, funding, therefore valuation. but then if you actually look at let’s say you know I actually wrote a column several months ago back when Drupal, Z.I.A. and Minimas just went public. Even though that was around the time that ZIA was less than 100, I think their highest point it was about $100 billion US, right, in valuation. That was before that. It was lower than that. But still, if you look at the price to

cells multiples, they are a lot, lot higher than the open A and Anthropic. Anthropic Open A in my memory was around thirty X and

mini mass and z.ai are like a hundred or two hundred X. So I think just looking at the pure valuation numbers are kinda like not kinda misleading or not really useful. You need to look at the multiples. And if you look at the multiples actually the Chinese labs, the ones that have been published are a lot more expensive.

so I don’t think at first that, you know, it’s just because it’s the order of magnitude less, right? Fifty billion versus five hundred

Grace Shao (34:56)

That’s interesting.

Jing Yang (34:57)

billion. That does not mean that Chinese labs are cheaper. So we you can easily calculate based on

the numbers reported on Deep Second to get their PS modules. And then you can see it’s also very high, right? So and then you c if you look at the P and S and then it’s very high, not because the P is high, it’s because the S is too small.

Okay, that’s why the modules are a lot higher. the multiples are actually one order of magnitude higher than the US ones. And then that’s because the S is too small and the S is too small again, going back to my earlier point, I don’t want to repeat myself, but it’s because the enterprise market has not been developed. It’s just a lot more limited mm compared to the US.

Grace Shao (35:38)

No, that’s interesting. I think usually we just only read the headlines and just think Chinese models are very undervalued, but you provide a lot of context there. Juro, I I wanna go back to you. You wanna talk about distillation, right? Let’s talk about distillation.

Juro (35:52)

yeah, so

Yeah, one thing that kind of is interesting about Byte Dance is that you know we wrote about how Jiang Yi Min, the founder, you know, said in the internal meeting about how you know ByteDance is not going to rely on this and they haven’t. And so you know, unlike you know many other Chinese labs that you know have really relied on distillation to as a shortcut, right, in their and to improve their models.

And that I think this is one of the reasons is that you know f if they really want to compete at the frontier level and you know, distillation, you know, like putting distillation as you know as your using that as your main strategy is not gonna really get you there. So, you know, you do need to kind of really properly train your model using, you know, your own data and you know, the data that you can actually, you know.

use for you know really advanced training. And that’s really the path to really kind of come up with your own frontier models that really sort of match or even surpass US models. And so that seems to be part of the thinking. Also another reason we heard is that they also are you know aware that you know byte dance is under so much scrutiny already with TikTok in the US and

You know, distillation has become such a sensitive topic between the US and China, right? I mean the US has accused a lot of Chinese labs of distilling American models. So if Python did that, you know, what kind of you know, what kind of criticism they would get in the US, right? And and would TikTok, you know, face any, you know, more challenges because of that or so those concerns also were, you know, we heard

were behind that kind of approach. But yeah, this definitely makes them quite interesting because a lot of Chinese labs do, you know, when we talk to people, they may not say publicly, but they do acknowledge that yes, distillation does help. And so yes, I mean but I mean there are indications that other labs also, I mean, we are starting to see more progress that maybe cannot be explained, you know.

solely based on distillation, right? Like Kimi K3 achieving, you know, really you know, strong performance, right? So but yes, what ByteNAS is doing is quite interesting. And whether they can really get to the sort of frontline and you know really emerge as the you know leader that way is gonna be also very interesting.

Grace Shao (38:26)

Yeah, I think I think you like hit the nail with that. It’s like distillation and genuine innovation are not mutually exclusive. Like you can get the benefits of distillation but still innovate on top of a certain engineering techniques. One thing I also heard was that, you know, with distillation is you inherit the good but also the bad. And apparently Jiang Ming is a bit like sensitive or anal about the fact that potentially he can always still inherit the bad or the

values or whatever of you know a distilled model. I mean it obviously that story went viral domestically as well. I think a lot of people were like, Jiang is coming out and saying we’re like we we we are not gonna go distill. We’re gonna do it differently. However, the flip side of the argument’s like, well buddy, you’re not gonna distill but n your seat is not actually Soto or anywhere near the last few iterations, right? So

How do you view that? Do you think at this point for a company as big as Spite Dance and as as much pressure they’re faced with and as much capex it put into this, is it better to stick to the principles right now and hope or continue to work hard on potentially coming out with something completely original, completely innovative on their own? You know, they want to go for the best of the best in the world, or you know, the other kind of strategy, let’s not name names, but maybe other labs or certain big tech.

I would say we catch up first and at least then it helps with diffusion and then AI becomes a flywheel in our existing business and we can let this money churn and then help put more money into AI and et cetera. How do you view that?

Juro (40:04)

Well, I think it would be hard for them to make that shift now. Having s you know, after Emin said that in the meeting and we wrote about it and if they do make that shift, you know, we or somebody else is gonna probably write about that shift. So then then that would make them look, you know, pretty bad, I think, if they

Jing Yang (40:22)

There there

be

a massive off ramp for that to happen, I think.

Juro (40:26)

Yeah. Right.

Grace Shao (40:27)

It’s not the first day one of the big techs decided to change a strategy.

Juro (40:31)

Mm.

Grace Shao (40:31)

I love how you also refer to him as E Ming Juro. You guys must be best buddies. to my first

Jing Yang (40:35)

Yeah.

Grace Shao (40:36)

name bases. okay, let’s talk let’s talk about the BATs though. Let’s talk about

Where

you see the BATs right now. Juro, you’ve been covering them in terms of their businesses for a long time. You’ve covered the executives. You’ve done like, you know, extensive profiles of these some of these leaders. Where do you see these big tech going? Like what is their end game here? You kind of alluded to Baba, you know, really going in for cloud, but is that enough in the AI air?

you know, is Qwen still a priority? is embedding AI into commerce still very important? How important is really Yao Shen Yi? Is he really the kind of savior to Ten Sen, Huen Yuan, for supposedly coming out very soon? How do you view that? And then of course, seed and seed dance. I mean, everyone knows seed dance is good, everyone knows seed dance has a lot of data, but

They’re very secretive, very hush hush, right? and given that they’re not public, they’ve actually had the luxury to not have disclosed what they’re doing publicly. Can you give us your views on this? This is an open ended question. Just see however you want to take this.

Juro (41:40)

Yeah, so for Alibaba I think Qwen AI being really center and front is that’s pretty obvious. And also how much of their investment is going into it as well. So they have been kind of

spending more kind of ahead of their plan. Like they had a three year plan. And so and their CapEx, you know, was really increasing at a very I think 10 billion dollars for the was it the quarter recent quarter they reported and so yeah it’s it’s growing at a very fast pace and so it it’s very clear that it’s their priority. And obviously if you look at the revenue

still the biggest part does come from e-commerce. But there are very few questions that come up during earnings call about e-commerce anymore, right? So a lot of the questions about the future and also whatever they want to emphasize is really about, you know, Qwen and especially cloud, because that’s how you know the main platform for selling the models. So yeah, they even Alibaba’s trying to be a bit like anthropic, right? In that sense that

You know, we can sell a lot of models. And and plus they have the cloud, so they, you know, they think that we’re gonna sell everything, you know, the infrastructure. And so that has definitely become yeah the priority. And that shift is pretty clear. And e-commerce, I think they are saying that you know they are incorporating AI into e-commerce, and that’s definitely another big theme. But the future, the big part of it.

depends on how successful this you know AI model push is gonna be. Yeah.

And others ten cent. So yes we’ve we’ve written about them and you know Yao Xing Yu’s role and definitely I mean general view is that they their model has improved and I think you know because the perception before he joined was pretty, you know, kind of negative, right? And a lot of there are people who are even saying that, you know, they’re kind of not part of the race anymore, you know, like in terms of you know

the leading models, but I think they have you know come back. And what they do have is the huge you know platform for applications and you know consumers with WeChat. And we have written you know previously about the WeChat agent and that was a scoop earlier. But I think something like that, you know, they could still kind of make a huge impact by

you know, rolling out something because they have the ability to really, you know, engage their users, right? And that’s still very powerful. So we still can’t, I mean and no matter how this model race goes, Tencent will be very important. And and then I guess yeah byte dance of course yeah we talked about it but they

They are still the biggest AI lab in China and I think the amount of effort that goes into it, I mean they are trying to do everything as well. And they also just like Alibaba, they are really going for, you know, their cloud business, you know, selling models. That has become the biggest sort of emphasis for the cloud business. So previously, before this AI boom, people didn’t really talk about byte dance as a cloud player, right?

But now they do because of that. The model business, you know, that’s really elevated the position in that area too. So yeah, I think all three of them, I mean, are definitely very important. And AI is really the center of what they do now.

Grace Shao (45:16)

For sure.

Jing Yang (45:16)

Just have one thing too quickly to add. I think

But I’ve been pretty much looking forward to the release of the Xiao Wei, right? With the Witch Hat or the Wasting Agent. Not everyone

Grace Shao (45:25)

Did you try it?

Jing Yang (45:26)

I’ve s what do you think?

Grace Shao (45:27)

I tried the beta and it’s like okay,

It’s on beta. I tried it honestly, cause this is I think exactly to Juro’s point. Like it’s really convenient, so you can’t see it. Okay, there’s a glare. But basically, it’s really convenient. It’s at the top of your like B chat interface, so there’s a lot of functional adjacency. That sounds like it’s really easy to get there. You don’t have to log out to another app. But like the whole

experience it’s more like what do I need it for, you know, and how good is it? At least for our jobs I would assume, you know, or desktop jobs, AI is seen as a very, very strong research tool at this point, if you’re not doing it like if you’re not running your own agents. But for the day to day SAO it just feels like a supercharged search engine. And I don’t even know if it adds that much value, frankly, if I’m just like, like

I don’t know, when does this shop open? Like that’s when I think a consumer application, you would open up a consumer application. Do you know what I It’s not I’m gonna open up Sellway, be like, tell me about you know tree’s IPO and like what’s your assessment on this? Do you know what I mean? It’s like I feel like it’s not it’s kind of sounds good, but right now they haven’t found a really clear use case. And then obviously on top of that, Selway doesn’t even use Huen Yuan, they use their own model, which is this other wild piece of the poll.

Tencent strategy. Look, guys, it’s almost like an hour in, and I still have a lot of questions for you guys. I’m so sorry. I’m gonna like redirect the conversation. We spent a lot of time on labs. I want to ask you guys about the other side of the hype or the international interest right now, which is all about the world models, the robots, the neol labs in China. Juro, you recently wrote about this topic. I actually just spoke to Many Core CFO for this podcast like two weeks ago.

And it was very interesting to hear about, you know, a spatial, like a 3D data company or 3D intelligence company is now pivoting to spatial intelligence and thus building world models. It just seems like a lot of companies are in this space. It’s very crowded. Can you give us a high level thought an overview of China’s strongest advantage in the space, China’s current landscape, you know, in the hardware space as well as the software space?

Juro (47:30)

Yeah, sure. Yeah. So I think for robotics, China definitely has a very broad, you know, supply chain and covers almost all kinds of components, right? And like motors, sensors, you know, structural components, batteries, camera, you know, everything. And this also partly because there’s a big overlap between you know, supply chains for electric cars or drones or you know, other products, right? And

And then there are also like a huge concentration of those suppliers, like in you know, Shenzhen or you know, Shanghai, you know, Yanzi River Delta kind of areas. They have a lot of different suppliers all in one place. So like when I was talking to a robotics you know founder in Shanghai, the same person Jim had lunch with, I think, today, so he was telling telling me about how his team can just, you know, like take a DD or

you know, even ride a bike to, you know, many of their suppliers. They’re just in the neighborhood, right? So this obviously speeds up everything, right? So I think that’s the major advantage. And then the bottleneck I think a lot of people point out, but it is the software part. And I think a lot of robotics founders also agree that you know securing that kind of talent, I mean because

That same talent can go work somewhere else that pays more. So, you know, in China that talent is still kind of limited for those guys, whereas the hardware talent is so plentiful, right for this area.

Jing Yang (49:05)

but there’s

I want to add about China’s advantage in robotics. I think it’s very easy to just say it’s a supply advantage. I think there’s also actually there’s more to that.

I’m going to talk about data advantage.

We have not seen real AI powered, real intelligent robots or robot robotic models is exactly because the lack of data.

lack of spatial 3D environment manipulation data. Right? And so even though so that in China when you have you know so many pilots happening in the warehouses and the fa on the and on the factory floors, when these robots get deleted and get deployed in these pilots, they are gathering a lot of data that can be then used to improve the model. And this is something that is not happening.

I think with very few exceptions of maybe Tesla and a couple others. But in China, this is happening in every you know, major logistics you know, Korea company, warehouse company, and car EV makers. This is why you see CATL, company like CATO and JD and Metron have invested in so many robotic startups. And this will create this data flywheel.

That will eventually, hopefully, at least the people in the industry believe, right? That makes China maybe ahead in the breakthrough of the brain, the real intelligence of the robots. That’s something I think that’s less appreciated, but I should bring it.

Grace Shao (50:33)

Yeah

no, that’s really interesting. And think just kind of adding to that, what I’ve also been thinking a lot about is just like because like you said, Jing, there’s so many moving parts in this ecosystem and they’re all done in China, whether it’s data collection with the data fine-tuning or actually the understanding of the manufacturing of these robotic parts and all that. There’s also a lot of talent that’s in this space. That again is being underappreciated. I think people don’t realize that the hardware self-integration is actually the bottleneck for a lot of these products, it’s not really just the manufacturing.

Even some say you can manufacture these products like say in Vietnam or wherever. But like you know, I think Patrick McMickey wrote about it in his Apple book. It was a lot of it is just understanding how to hardware or software or like these more like a niche, like the one percent of the top hard blue collar actually is even very, very hard to replace and train up. And that’s been done through decades. So guys, let me ask you a relatively kind of sensitive question.

So we saw that DC’s been talking about banning robotic imports from China, but from all of us, I’m sure we’ve talked to a lot of robot companies in the US, like I would say like easily like 90, 90% of them have 90 to 95% of them have some kind of arm in China, be either supply chain or certain even some parts, right? So like if this ban really comes to force, what first of all, what’s driving that decision? Second of all, how practical can that be and what kind of impact?

Will it have American robots? Because American robots at this point are already priced significantly higher than these Chinese robots.

Juro (52:03)

Yeah, so what they cited when they made that announcement about the ban was the you know national security concerns, right? So the robots can, you know, gather data and for surveillance or you know, whether they could be controlled remotely from China or you know, that was the kind of concerns that were cited behind the decision. And I think the definitely the immediate impact is that a lot of US startups or

university labs by Unitry robots or you know other robots from China because you know they need to use them for software development or research, right? And so this means that they have to find it somewhere else, but it’s really hard to find, you know, anywhere else because most of today’s humanoids that are affordable and also available are really from Chinese companies. And I think that’s the most kind of obvious impact.

Grace Shao (52:58)

Right.

Jing Yang (52:59)

I I the one only thing I’ll add, I think the impact on the US robotics development would be that

the leading companies that I can produce at scale, you know, can assemble, like I should say, because you know, they can produce all the parts, right? That can assemble in the US at scale would be fine. And I think this regulation of this ban will, whether it’s intended or unintendedly, strengthen the leading players’ position. And then those who have not been able to assemble at scale will have an even harder time.

Grace Shao (53:30)

I see. Let’s talk about the Unitree IPO. Unitree at the time of recording was late August. Unitree just went public like a week ago. Jing, what are your thoughts on that? And I think you guys wrote something quite fascinating. To be honest, for people who are familiar with A Share, it was nothing too shocking, but I think you exactly pointed that out. You’re chuckling. Why don’t you tell us about it?

Jing Yang (53:51)

No, I mean, okay, that story was not something that I pitched. I was sort of asked to do a piece that sort of explains why Unitry’s first day like a debut popped so much at 416%. I did not think maybe I’m too close to this because I covered, you know, capital markets before. And so I’m like, okay, this is not

News. This is to me it’s just like another Wednesday. You know, I mean they went public on Wednesday, right? And and then

Grace Shao (54:16)

It’s just how Ashare works. It’s just how Ashare works, yeah.

Jing Yang (54:20)

yeah, and then and then then I had to then I realized okay, okay, actually maybe that’s exactly why I should write the piece. And so then my boss was my editor was initially not convinced. I told him I say, you know, this is because of the John Tin butt, the cap on the stock prices ceiling and floor.

that is only exempted on, you know, the first day or the first couple of days of IPO. And then he did not think that this was relevant. We’re talking about first day pop. I’m saying what are you talking about? Of course this is relevant. It’s because if you don’t a lot of people believe it, if you do not get in on the first day when there’s no cap on the stock movement, then you will not be able to get into the stocks for days in a row because though immediately the RP market opens, they’ll hit the ceiling.

Right? And then say, okay, then do you have data to back it up? I said sure. And then I went to Bookflow Data. And then there were a hundred and nineteen IPOs on the HR market last year. The mean of the first day share price was two hundred twenty-five percent. So let’s just put that in context. And then in the US is about twenty-five percent or so, right? So so this is why I think

for you know, for anyone who’s not familiar with the Asian market, this looks like really striking. But I think for people who are used to this it’s quite normal. And then do you know why people call you you must know this term in Mandarin called that sing, right? And then f like fight for new is what I sort of loosely translated. But why was it called that sing? Because actually in

This is a dialogue in s the you know, like southern part of China, like in Shanghai when China first had a stock market, right? So in the eighties. And the people would be there without digital training, right? People would be lining up on the street outside of the stock exchange, waiting for it to open so they can get in and then and then put place an actual paper, like a hand in their money and get the the the the the the stock order placed. so

So now that whole phenomenon still exists, it’s just not happening in the digital world. But because of that, you know, like sort of fighting tooth and nail, that sort of hardship we are willing to endure to just get into an IPO that people call it DASI. so that’s a big reason for why we see these POPs. The highest first day pop last year in Asia was 1211% and/or 12 times.

Grace Shao (56:45)

Which company

Was this?

Jing Yang (56:46)

that

That was a company called HAR being Big Bird Industrial or in Chinese Da Pong Yi. And it’s a company that does what they call high precision industrial cleaning. So they offer solu cleaning solutions to, you know, like machines that produce EV components. I mean you can debate

Grace Shao (57:03)

So they’re also part of the autonomous

robot kind of wave as well. They’re just not that sexy.

Jing Yang (57:08)

Yeah, but you you can debate

whether this is like a very high tech but but

Grace Shao (57:12)

Mm.

Jing Yang (57:13)

but definitely, you know, if you compare a company like that that had a twelve times first day pop and you and Uniture is like only a four point six, I was like, Okay, now I get it, right? So yeah, that’s my point.

Grace Shao (57:25)

No, that I

I thought that because I saw on your I think it was your WeChat moments. I was like, yes, Ding, this is like quite correct. Like and I think it’s a relevant context to people who are not familiar with Ashare.

Okay, guys, let’s talk about China robotics. I think because like Jing said, we’re so close to it. Sometimes I completely normalize like seeing robots on the streets in China. And like you have like cleaning robots, you have like hotel delivery robots, you have like and these things have been around for like I’d say at least three to five years. So when you look at the big trend right now with the robotics hype, and let’s just take a step back, go look beyond humanoids.

Where do you think the whole industry is going? Like what are some interesting use cases you’re seeing that’s really scaling robots? and that that is maybe being overlooked by the rest of the world.

Juro (58:10)

Yeah, I think humanoids, a lot of the applications they’re talking about, especially like industrial applications or even like a home household, you know, chores kind of thing, there’s still a lot of you know, hurdles, right? And but as you said, there if you are talking about sort of robotics and automation broadly, there are a lot of very impressive, like a very fast growing applications and

Another one I can think of is like as those sort of logistics vans kind of thing and that are you know autonomous and so those are kind of sometimes in the robotics kind of category, but I think those are growing very quickly as well. And then what you mentioned about cleaning robots, yeah. And but I think with humanoids we’ve seen so many impressive sort of you know demos and robot performances.

And those are very impressive in terms of how robots can move so well, like what they call locomotion, like both software and hardware, right? But I think there’s still a lot of problems to be solved for like robots interacting with the environment and interacting with other objects and people, and especially if they involve like you know unpredictable situations or kind of you know, situations that cannot be really controlled.

So because a lot of real world applications contain those kinds of unpredictable, you know, situations. So when it comes to real, you know, I mean there’s still a lot of hurdles. And I think that’s where there is a little bit of a or maybe not a little bit, but there is a hype you know, as to you know, looks like those robots can do all the work tomorrow, right, in factories, but it it’s not that simple. And

A lot of people do say that the AI brain part of the robot is still where they need so much more, like so many breakthroughs are necessary for a lot of those applications to actually happen. And that’s another reason why, you know, I wrote about world models, but that’s, you know, a big part of the discussion because that’s where they need a lot of breakthroughs.

Grace Shao (1:00:21)

Who are the main players in the world models right now?

Jing Yang (1:00:22)

I think I want to

Sorry, before we talk about that, I just the only one thing I want to say ‘cause I just learned about this. I actually

Something new that I learned. I don’t know if your audience would know, but I want to share. So there’s a big difference between embodied AI and humanoids. And I think somehow these two terms have been used interchangeably. And then the difference is that humanoids are just robots that are shaped like humans. embodied AI actually are robots that have real intelligence. and if you look at the Chinese government’s fifth 15th five-year plan where they supported the sector, they actually

actually

spelt out both say, you know, support the development of embodied AA and humanoids. That level of sophistication from the Chinese policymakers, I I don’t think that’s been widely appreciated, right? I just want to point that out.

Grace Shao (1:01:11)

That’s an interesting take. Look, I wanna ask a question, what do you think is something you still want to share with the audience that, you know, you guys are focusing on in the near future? something that’s exciting you as you cover China Tech, Asia Tech?

Jing Yang (1:01:24)

Sure, do you wanna go first?

Juro (1:01:26)

yes, I think I mean just the speed of you know changes and you know I’ve written about some you know trends like for example earlier this year when open claw you know briefly became a huge you know phenomenon in China, right? And then last year there was like a AI agent manace and all kinds of you know clones of manace, right? And yeah, I think

These kinds of things will keep happening and I always find them after all these years. I’m still fascinated just how quickly they move. And all those founders jump in and you know I mean there’s so much of the domestic competition that’s also you know not fully appreciated, just how intense everything is. And you know, those guys launched something, you know, when OpenClaw came out, everybody, you know, immediately worked on something, you know, product and

because they know that everybody else is gonna do the same and if they don’t do it now, you know, they’re gonna look like they’re falling behind. And so yeah, I feel like sometimes like you know part of my beat is just like you know the new hype beat and but you know where I’m still fascinated by just how intense you know things change and you know like yeah how intense the competition is.

Grace Shao (1:02:44)

Juro, it can be very exhausting, can’t it? I feel like there were like eight models in just the summer. It’s very hard to keep up at times.

Juro (1:02:52)

Yeah. Yeah, I mean we

We were talking about world models and you know Jean was talking to also some investors and how they were talking about, you know, even founders with backgrounds that have little to do with world models also, you know, starting those, you know, launching those startups now. And yeah, just how that happens so quickly, right? And everybody kind of jumps in.

Grace Shao (1:03:17)

It’s kind of the new gold rush.

Jing Yang (1:03:17)

I made this prediction before I made

this prediction before but I’ll make it again. I think we’re gonna assume we’ll see the War of A hundred War models. we saw the War of A hundred L Ms back in twenty ninety three. And I think this year will be the year where we see the beginning of the War of A hundred.

or models and and the scary part of this is that at least LM is built on something that has like a a c a piece of technology or fundamental like architecture that has a consensus riched upon and and word model is nothing close to that. This is sort of the scary part.

Grace Shao (1:03:50)

No, I agree with that. Guys, one last question I ask every single guest that comes on and Jing you’re familiar with this. What is one difference you view you hold or something non consensus?

Jing Yang (1:03:59)

Okay, maybe I’ll say this. I hope it doesn’t get me into trouble, but I don’t think Chinese models are gonna like the frontier Chinese models are gonna really be able to catch up with the US frontier models. Is sort of my maybe differentiated view because

Grace Shao (1:04:14)

Well, if you put that out there, you

I have to elaborate now. Why?

Jing Yang (1:04:17)

If we assume that a lot of the progress or distillation contributed to a lot of the progress we’ve seen recently, then that just means if you continue to distill, then you continue to be playing catch-up. you may reach

you may reach, you know, to be very close to be on par, but you can never overtake. like a student can never be better at than a teacher, right? So I think that’s just simple logic. I mean I’m making this prediction or this sort of differentiated view based on the fact that distillation continues to be rampant and continues to be commonplace, right? Whenever that’s changed, then obviously my view will change too. So that’s just yeah what I think.

Grace Shao (1:04:56)

I think that was one of the arguments people made about why Zhangyiming was so against distillation, because he was under the impression that you can only catch up or distill as good as the model that you’re distilling from, right? Jira, what is yours?

Juro (1:05:08)

Huh. I haven’t really thought about this, but I I guess maybe this is not necessarily non consensus anymore, but a lot of the US China, you know, restrictions on tech and you know with chips or yeah, I mean all kinds of you know, especially US measures for China and yeah w

Every time just what I see is that how there are always you know ways to get around it and and you know people involved or you know Chinese companies kind of even talking about it as if those restrictions didn’t exist or so you know for example like access to Claude or you know or Blackwell chips or so yeah, I mean those would be like I’ve kind of stopped

seeing them as, you know, like when new restrictions come in, the sort of first instinct is how are they gonna get around it this time? Because I’m sure they will, right? So yeah.

Grace Shao (1:06:07)

It’s like a when there’s

a w will there’s a way kind of thing. Yes.

Juro (1:06:10)

There was a wheel there’s a way. So that’s the kind I

I mean, maybe this is well known, I don’t know. But yeah.

Grace Shao (1:06:18)

no, I appreciate both of your time. I really appreciate your insights today. We covered a lot. We’ll love to have you guys back on another time, but thank you so much.

Jing Yang (1:06:26)

Thank you.

Juro (1:06:26)

Thank you.

Grace Shao (1:06:27)

Okay.

AI Proem is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

Get full access to AI Proem at aiproem.substack.com/subscribe