Meituan's LongCat Has Just Made China AI Independent (288)

This week’s podcast is about the development and launch of Meituan's LongCat 2.0 LLM. This 1.6T parameter, frontier model was trained entirely on domestic hardware. That's a big milestone.

You can listen to this podcast here, which has the slides and graphics mentioned. Also available at iTunes and Google Podcasts.

Here is the link to the TechMoat Consulting.

Here is the link to our Tech Tours.

 ---------

I am a consultant & keynote speaker on how to increase digital growth and strengthen digital AI moats.

I am the founder of TechMoat Consulting, a consulting firm specialized in increasing digital growth and strengthening digital AI moats. Get in contact here.

I write (a lot) about digital growth and digital AI strategy (3 best selling books, +2.9M followers on LinkedIn). There is a free book and email newsletter below.

My Moats and Marathons book series is a framework for building and measuring competitive advantages in digital businesses.

This content (articles, podcasts, website info) is not investment, legal or tax advice. The information and opinions from me and any guests may be incorrect. The numbers and information may be wrong. The views expressed may no longer be relevant or accurate. This is not investment advice. Investing is risky. Do your own research.

Transcript

00:05

Welcome, welcome everybody.  My name is Jeff Tausen and this is the Tech Strategy Podcast from TechMoat Consulting.  And the topic for today, how Meituan  just kind of made history with LongCat 2.0. This was a story in the last two weeks really. It didn't get a lot of news, is, well, I mean, some people noticed, but I think the mainstream didn't notice. Kind of a huge deal that...

00:35

Effectively, China's AI tech stack is now basically independent of the US as of two weeks ago.  Not super efficiently so, but effectively, yeah, that looks like what happened. And the last piece of the puzzle turned out to come from Meituan, which I didn't see coming. Yeah, basically, they built a model that is completely  domestically trained, Chinese chips, Chinese everything.  It's open source.

01:04

mixture of experts, frontier level AI, completely homegrown. We've been seeing this happen step by step, but it was kind of the last  piece of the puzzle, pretty much. Efficiency is another question, know, but generally speaking, yeah, and it turns out it was  Meituan.  And the story behind LongCat is really awesome.  It's the coolest story. I think it's one of these ones that's going to go down in history. So I'll give you a little background on that. That'll be the topic for today.

01:35

Let's see, standard disclaimer, nothing in this podcast from my writing a website is investment advice. The numbers and information from any guests may be incorrect. If views and opinions expressed may no longer be relevant or accurate.  Overall, investing is risky. This is not investment, legal or tax advice. Do your own research. And with that, let's get into the content. Okay, before I get into Matawan, let me do sort of an update on my ongoing  rant. Last couple podcasts I've gone on,  let's call it half-baked thinking.

02:05

about  the US government, really the White House, jumping in and stopping models, what that means, anthropic sort of going for some sort of regulatory political moat uh to protect it, basically to build a duopoly, as far as I can tell. I've talked about this a little bit, but I don't think my thinking was terribly good. This is sort of off the cuff. So let me do a little bit better thinking. And a lot of this is uh from a  wage-in research, which is... uh

02:34

Pretty good research group out of China. do emails and stuff. I look at them every now and I find them pretty solid. know so a lot of this is their thing. But like the idea was, okay, know, Anthropic is doing this  scare tactic game where they're playing up the risks of AI,  you know, and basically,  in my opinion, propelling the US government to sort of stop other models and  create a license  regime of various.  And I think they're going for a duopoly.

03:04

a politically constructed duopoly. That's my own opinion. I think that's a lot of people's opinion. Okay, and they've been doing this fear monger thing and then sure enough, it was their own model, Fable 5, Mythos 5, that  out of random left field got stopped  really almost randomly without,  it looks completely random.  And that's when I started talking about, look, this is like the entity list.

03:33

Who's going to be able to use these frontier level intelligence capabilities? Is this going to be  handed out like political favors? Is it going to be only US citizens, which was what was applied to Anthropic? Which was weird because if you're not a US citizen and you work at Anthropic in the US, you can't use your own model, really? Anyways, it was kind of weird.

03:56

My opinion was like this creates a political risk that any business now has to put within the risk section of their annual report.  You cannot be too dependent on US frontier level models because they can be taken away from you at any point. So you need alternate supply chains. Fine. That was the extent of my take. And then OpenAI got hit a couple of weeks later  with GBD 5.6, I believe. Okay, interesting. And then...

04:25

Anthropic sort of got released from jail. Well, half of it did. Fable 5's out of jail. Anyone can use that anywhere now.  But it's got pretty severe safety rails built into it. Mythos 5 doesn't have the safety rail. So you can use it for things like cybersecurity,  biology, and that still needs a basically an approval from the company and some sort of consortium and the Department of Commerce. OK, is that the new regime? I don't know.

04:55

OpenAI put in some pretty severe guardrails. Those are now open for everybody as far as I can tell.  That was kind of the state of it. And then out of left field last week, a report came out that the Chinese government was looking at its frontier level models for any sort of these sorts of concerns.

05:17

And so a lot of stuff was bouncing around the internet. Oh, China is going to do the same thing. So you can't count on open weight Chinese models anymore. Like, let's say DeepSeek.  And the report that came out was, you the government basically met with, I think it was Minimax, Alibaba, Quinn, and DeepSeek. Who knows? These are reports. Now, I don't think they actually made any decision. I think it was a discussion. But the rumor went everywhere. Oh my God, you can't trust these either.

05:46

All right,  so what is all this? I here's a pretty good breakdown, which I was kind of reading. A lot of this is from Weijin. I was looking at this through a pretty shallow little lens, which is like, can you build a business on  supply chains  that are unreliable and are prone to  political uncertainty? That makes life very difficult, and you can't do that responsibly as a CEO. Okay, but there was like...

06:15

If you think about the risks of these things and this  do you need a quasi government licensure situation. Okay, here's a breakdown which I read from Weijin about how to think about the risks associated with these frontier level things. And there's one, two, three, five. Basically, yeah, five different risks that they mentioned, which I thought, okay, that opened my eyes a bit. Number one is the one everyone talks about, which is cybersecurity.

06:45

You know, if you have Mythos V, you know, it was, you can have this thing just running all the time hunting for vulnerabilities. Okay, that's a big problem, because it's kind of a white hat versus black hat. So we got to give the white hat people some time to use this and not let every single gang out there in the world start having this next week. Okay, so, you know, it was a productivity tool, hunt for vulnerabilities.

07:12

But then it really quickly becomes an exploit tool because you can turn these models into agents very quickly. So it's not so much,  do the research. No, now we want you to test the exploits. Now  we want to do automated attacks. OK, I can  get the fact that that's pretty dangerous.  And  if you sort of put a delay in here and only release, you know,

07:40

lower grade sort of tools for a while. You give the black hats an advantage. Okay, I kind of get the cyber security one. And the other thing is even when you release these things out in the wild, you can't really control what people do with them. They can hook them up with other tools. They can hook them up with databases. They can put them as the backend of agents. You know, these models are powerful. Okay, so cyber security, I kind of get it.  Number two, biological and chemical risks.

08:09

Yeah, mean, biological weapons, biological harmful materials, not just weapons, but anything harmful, biological chemical. You know, there's a nice healthy barrier to entry for most organizations and especially individuals in this field. You can't just scale up a lab and start testing, you know, various viruses. You need a government institute for that, apparently.

08:36

Okay, this lowers the barrier for non-experts and it lowers the barriers for individuals who want to scale up because you can have these things modeled out. Yep,  that could be a problem.  We know the history there. Fine, I sort of buy number one, I buy number two. And those, you could probably safety rail those pretty confidently. Bucket three is where I get, this is more in my area, bucket three would be agentic capabilities.

09:07

uh For GBT 5.6, they have ultra mode, which is basically one agent runs a bunch of other sub agents as far as I understand. So, you got teams of agents, hundreds of these things working together. All right.  Well, what are they doing? Are they coding? Well, coding could be a problem.  Coding is a pretty powerful tool. uh Cyber security,  as mentioned, that's a problem.

09:37

Long horizon tasks,  which is pretty much, this is an area a lot of people focus on right now.  You kind of look at like where the smartest people are placing all their chips. And for the last year, as far as I can tell, the smartest people were placing their chips in coding. Now the thing you hear over and over is sort of  multi-agent collaboration, uh long horizon tasks, which are, know, give, this is most of enterprise workflow. You give,

10:07

Let's say you're going to have a robot in your kitchen and you say,  clean the kitchen and make me dinner. That's a long horizon task. Instead of being able to break those tasks up into parallel processing,  they have to be sort of done serially. One, you have to clean it, then you have to cut the vegetables, all of that. So it has to be done serially.  So you're talking about giving  an AI or an agent  an objective that could take days or weeks.

10:35

Well, that's deep research. That's a lot of things. So a lot of people are placing big bets on sort long horizon tasks.  So anyways, you can see that ultra mode and these agentic capabilities,  one, you could use them  for nefarious purposes. Two, they can sort of go off on, you don't even have to have a bad purpose. They can just sort of start taking unauthorized actions on their own, which they seem to do.

11:02

So this is sort of genie out of the bottle Terminator type thinking, who knows? I'm not sure I buy that one to tell you the truth. I buy number one and two. I'm not sure I buy the agentic capability. Because I think when you look at enterprise workflows, you're mostly talking about agentic capabilities. I think that's what we're usually talking about. All right, number four, AI based self-improvement. This is why AI, I think this is why it freaks a lot of people out, freaks me out.

11:30

It's an evolutionary technology. It can evolve on its own without us. That's a little, they call it, you know, recurrent self-improvement. The AI, okay, so we built, well, we didn't, but certain people, humans built, you know, chat GPT 3.5, 4, 5.5, 5.6. Is chat GPT 7.0 going to be built by agents and not by humans? At what point does the next generation get created by non-humans?

11:59

And right now it's kind of close. Like, from what I can read about  when the coders,  how much of this are they doing themselves versus just managing AI or agents? It seems like it's a lot of it's the agents. So how many steps away are we from basically either  self-improving AI that just sort of happens recurrently,  or let's say you have AI R &D  that is now automated.

12:25

So frontier level AI R &D, not just building models that we use, but  just  R &D in the space is just happening in an automated basis. All right, yeah, that's a problem. eh And that sort of throws an interesting wrench into this whole question of we're going to have licenses by the government. Well, licenses are usually based on products. How do you license a process where the process is going to happen on its own?

12:55

There's no product release here. How do you control a process that gets out of hand? I don't know. And then the fifth bucket, which is the one I've been talking about, is basically foreign access.  Do you give Mythos 5 to ever cyber,  black hat cyber group in Romania and Albania? Can they just start ramping up Mythos 5 and doing attacks? I don't know. You want to do something fun?

13:24

Go to, if you have a webpage, every now and then go look at like where you're getting attempted lock-ins or  brute force attacks and you look at it by geography.  It's really interesting to see where they come from.  Now maybe they're bouncing them through things, but  usually they are coming from the same places.  Like I won't mention them, but if you ever check where your brute force attacks are coming from,  there is a short list of geographies that are on the top of the list like all the time.

13:53

Anyway, so you just sort of, you know, just dump it into those worlds and let them have superpowers. What about foreign militaries, intel agencies, security groups, national security groups, who knows? And I think that's kind of what I was thinking about with all of I was thinking cybersecurity and sort of foreign access as a question. Okay, yeah, it's kind of a big deal. So I think that's a little, you know, I felt bad about sort of ranting with this sort of

14:22

shallow half-baked thinking. But yeah, that's about where I am.  So  it's an interesting question.  And you know, what is frontier level today versus six months ago? You know, I haven't forgotten the whole COVID experience where literally it was November, December, they were going, this is the US government, they were going on the news.  Don't go to Christmas, all your relatives will die.

14:50

And it was just full on, this thing is risky. We've all got to be careful. And then January, one month later, the Ukraine war started and COVID disappeared as a topic entirely. It was just gone. And I'm like, what happened to these big risks and fears we had two months ago? Now they're not a problem anymore, I guess. We can already see that happening with these frontier models. This thing is super dangerous. Okay, everyone gets a license. So a year from now, is anyone going to care about

15:19

GPT 5.6 at all. So what are we talking about? Anyways, I'm skeptical but you know, I don't have to deal with government type problems. So I guess that's a blessing. Okay, that's a bit of a, I won't go into this topic again, but it's interesting. Well, for me.  Okay, let me get onto the topic here. Okay, let's talk about Long Cat. I think this is going to go down in history as one of the coolest business stories ever.  It's awesome.

15:47

A month or so ago there was news about DeepSeek v4 and it really caught my attention, caught a lot of people's attention because they basically said, look, this is our frontier level model and Deeps is obviously right on the edge.  And this model was trained on Nvidia. So people are like, well, where'd you get the chips? What happened? But the inference apparently works  equally well  on Nvidia as well as Huawei chips.

16:18

And that kind of got attention. It's like, wait, this thing can run on local AI infrastructure without Nvidia GPUs. Well, uh that's a big milestone. But obviously, the core capabilities happen at the training phase. And no, it wasn't trained locally. Well, Meituan basically just announced LongCat 2.0.  They officially released it a couple of weeks ago.  But it's been floating around.

16:46

since at least it appears like November of last year in  various forms. They had it under a pseudo name. They weren't releasing it under the name.  It was being used on Hugging Face and uh Open Router, but  I don't think people knew it was Meituan. I it was called Owl something. Owl. Anyways, the official release was recently and they basically said that this thing was trained entirely on Chinese chips. Inference and training all domestic.

17:16

Nothing foreign and that was big news. Okay, that was like the last piece of the puzzle  To get China to basically have its own independent AI tech stack that is not really subject  To let's call it political weaponization  By the West now I would say the US because they were the leader but you know the as ML machines coming out of Europe Okay, that's not us. Although they were applying pressure. So yeah

17:46

but the West, really the US. So, okay, so what happened? Now, I'm trying to get the story, but here's the story. I've read it a couple times in different places, and this is the best that I've got. Okay, so Chat GBT comes out, you know, end of 2022, basically in November. And in February, 2023, a guy named Wang Huiwen, founds a small business called Lightyear Beyond.

18:15

And on the board and one of the investors is Wan Xing, the CEO founder of Meituan and Wang Huiwen  was also  a co-founder of Meituan. I'm not sure how active he's been involved in Meituan if at all.  So, okay, the core team has set up something that looks, you know, going into the LLM space pretty quickly.  Something happened with health issues with  Wang Huiwen. don't...

18:43

I don't know what is a couple months later June,  Meituan basically acquires  this small startup.  It turns into something called GN06,  which is kind of an incubator unit  within Meituan. Okay, so they start building out their internal capabilities with this thing. So I don't know if that was the plan from the beginning to start it outside or not, but anyways, that's where it  sort of got to pretty quickly.

19:12

Nobody's really talking about this, right? It turns out Meituan has been active in LLMs almost from the first day. They were in this kind of faster than Tencent, faster than ByteDance almost.  They were starting up in that same period, but right at the very beginning,  they jumped in, which is interesting because Meituan, you know, they do on-demand services. They do food delivery. They do hotel reservations, things like that. That is not necessarily a group.

19:42

you would expect to be building frontier level AI capabilities. You would expect them to be focusing on certain use cases, but really they're going to play on the frontier that kind of appears like what they're doing. Okay, so over the next year, 2024, 2025, they're kind of moving along. start, now why are they different? Why are they interesting?

20:08

It looks like, as far as I can tell, from the very beginning or early on, they were focusing on a very specific question, which is, how can you get  MOE, mixture of expert, high level performance,  using domestic hardware? How can you do that while at the same time  keeping your inference fast enough to be useful in a real time situation? Right, you can see there's sort of three,

20:37

factors there they're trying to do. One, we need performance, obviously.  We need it with an MOE type structure, which is pretty much what DeepSeek did in the others. But we want to use hardware clusters that are domestic. But we got to have the real time. We can't have latency.  That's a very specific type of question.  I'm not sure anyone else, I haven't read of any other of the major Chinese companies  really laser focused on, are getting off Western hardware.

21:07

that early on in the game as a model. Now Huawei obviously was, they had to, but anyways, as far as I can tell, that was a very specific question, which turned out to be the right question, or at least it turned out to be an important question. Okay, so 2024, 2025, there's this sort of iteration approach where they sort of release lightweight models and it looks like a testing phase.

21:34

I've been reading through sort of the technical stuff they were doing, some of which I understand, some of which I don't.  Whatever the mathematics they were using to get these MOE models to work on basically 7 nanometer chips. I mean, that's kind of the question. How do you get MOE to work at a high performance frontier level  using 7 nanometer chips? Because  if you're stuck on 7 nanometers because you're using DUV instead of EUV lithography, because you can't get the ASML machines,

22:03

How going to do it? And as far as I understand, you can't get too far beyond seven nanometers  without EUV, unless you jump into what Huawei's doing, which is their new logic folding architecture. Who knows? We'll see if they pull that off the next couple of years. Anyway, so they start iterating with these small lightweight models, LongCat, Lite, Flash,  68 billion parameters, testing performance,  lots of token routing, trying to, you know,

22:32

route your tokens to the right expert levels. Cash is 256K. Okay,  that seems to be the first wave of experimentation. That sort of leads to a focus on more specialized reasoning models.  So you got sort of this lightweight architecture, you're testing, you're experimenting,  then you move into specialized expertise, because ultimately you're going to be building an MOE. So they're building out reasoning  specialties in  parallel.

23:02

compute and things like that. Fine, fine, fine. I don't totally understand the tech there beyond a certain point.  Okay. The interesting thing seems to happen around 2025. they're, you know, they're two and a half years into this, August 2025. Long Cat Flash comes out. 560 billion parameter model. Okay, now it's Flash. it's, you know, it's...

23:28

It's not deep reasoning, but it's fast and it's 560 billion parameters. That was a year ago. That's kind of impressive.  Using about 18 to 30 billion tokens or billion parameters per token.  That's sort of pretty much what you hear, 8 to 10 % in there. And they have some tech that they've made this work.  And yeah, they trained this thing  on 20 trillion tokens.

23:57

over the course of about 30 days. Now I'm not sure if they used Ascend semiconductors or if they were using Nvidia at that point. Now they're targeting to run this thing and do the training on domestic. I'm not sure how much of that they trained.  I suspect they probably used mostly Ascend at that point.  I was,  for those of you who follow me, I was talking a lot about Huawei's  AI infrastructure last year. I really started back in March.

24:27

I was meeting with them, looking around at what they were doing. I wasn't allowed to talk about most of it until the end of the year. And I really wasn't most allowed to talk about the data, the super pods and the super clusters until January  when they released it all  at the Barcelona conference. But I suspect Logic Cat Flash was probably  using domestic chips. If I had to guess, I'd say, because it lines up with what Huawei was doing at that point.

24:56

They also did a couple other models that did a long cat video, which is a pretty small video generation one. They have one called long cat flash proven. basically does mathematical proofs long cat flash thinking. So a lot of parallel processing for sort of thinking models. Okay. That was kind of the first big. Stake in the ground was this, you know, 560 billion parameter model. And that tees up.

25:25

LongCat 2.0 because that was the big one  and the jump LongCat 2.0 which they've sort of been testing. here it is. It was called Owl Alpha on Open Router really I think since about December November this there was a open source model called Owl Alpha on Open Router and I think that was LongCat. So it was floating around being tested. Apparently the usage was quite good.

25:53

But I didn't hear anyone say that it was Meituan. Now maybe some people knew, but I didn't come across anyone saying that. So LongCat 2.0, they start doing a preview model, January-ish, but really the final was a couple weeks ago. 1.6 trillion parameter model. 48 billion  parameters per token. I mean, that's three times.

26:21

What the first, that's right up there at the frontier level. I mean, that's impressive. They're up in the trillions. I assume they're moving quick. It was trained on 35 trillion tokens. All of them domestic. I'm sorry, all the semiconductors were domestic, basically ASICs. So, you know, they're using accelerators. They're not using general purpose GPUs. They're basically using ASICs. And as far as I can tell, and I've been checking,

26:50

It looks like it was all  Huawei's Ascend  910c uh semiconductors. That looks like what it was, which are specialized for this stuff. So frontier level capabilities, frontier level scale, uh capable of doing  high level reasoning, capable of doing agents, capable of doing everything I just kind of talked about. Now, maybe it's not number one or number two, but it's in that top tier category of all the stuff that makes

27:20

government's nervous apparently. So let me talk about sort of what this means in terms of the China AI tech stack and then I'll talk about, what is Meituan actually using for  internally  and what is it being used for outside of Meituan? Because this is clearly not an internal tool. This is a broad release for AI globally. But, you know, what is the AI tech stack for China? Well, I think we kind of got a first pass at it.  At the foundation level,

27:49

We have Huawei, basically. This thing was using 50,000 GPUs in about 48 SuperPods. So if you want to go onto my website, you can read all about the Supercluster, the Superpod. I believe they were using the Cloud Matrix 384, which has 384 chips. You put them all together, basically.

28:18

Instead of having a nice small Nvidia rack with a relatively small number of high-powered general-purpose GPUs, Blackwell Hopper, this is  a much bigger rack of a lot less powerful GPUs that are  mostly specified to  algorithmic functions, accelerators, ASICs, and you have to build basically a whole basketball  stadium full of these things. Well...

28:48

It's about the size of a basketball court. If you do that, you can get up into the 50,000 level chips  to get this thing. So since you have low power chips, you have to put more of them. Now, the problem with that is you have to supply a lot more power. And it turns out the interconnect  is an incredible  engineering sort of challenge because when you tie lots more chips together, the connectivity time chip to chip to chip

29:18

increases dramatically.  So you have to get rid of all the switches and the routers and every chip connects to every other chip. And they have their interconnect bus, which is really the Huawei sort of key engineering breakthrough  that lets all  every chip connects with every other chip  lightning fast. So all of these 50,000 plus GPUs effectively act like one computer  and you can dynamically route  various

29:48

computational tasks to different configurations of GPUs. Well, that's what you need for an MOU model. Now you can do that in other ways, but  that becomes  sort of the big bottleneck. And that's what the Cloud Matrix 384 was. Then they built it bigger, which they released at the end of the year. That's when you get into the Superpod.  Now they're going into the Supercluster, which is going to be  just massive  number of chips put together. It's going to be like cities of these things.

30:18

That's coming out in the next year.  So Huawei's kind of building this dramatically. So it looks like, okay, if you're going to build the AI tech stack as sort of a, I don't know,  a team. Well, Huawei's core member of the team. And the other core member of the team, SMIC, who's producing all these chips.  Most of these chips are at the seven nanometer level. Okay, that's SMIC.

30:47

The big question with them is do you have advanced, you know, extreme lithography which you need at ASML for that?  Well, it's getting harder and harder for them to get those machines if they have them. How many do they have them? Everyone's debating this. Okay, so they instead at EUV, they're using DUV which gets you to about seven nanometers. And I'm not sure you can push it any further than that. Okay, so  they're ones making all these chips. Now I think the question is,

31:15

What is the efficiency level of the foundry? What is your yield? If you're using this sort of lesser technology to do the lithography, how many chips are you throwing away? Is it 50 %? Is it 70 %? Might be. Might be. Maybe their yield is only 20%. So the effectiveness is there. They've built this thing. But  at what efficiency level?

31:43

I still think that's a secondary problem. It seems to me they've overcome the bottleneck. Now it's just a matter of driving down the cost curve and driving up the yield curve, which China is very, very good at that.  So player number two would be SMIC. So Huawei,  SMIC. uh Third player on the team,  CXMT,  the Changxin Memory Tech Company. This is the memory  chip giant of China.

32:10

which is going public right now so you can pull their  annual report. I'm going to read that this week. But yeah, you need the GPUs. You need the CPUs. Both of those come from Huawei. You need the memory chips because memory is a huge bottleneck. OK, China's got the big player there too. They probably produced most of the memory for this thing, as far as I can tell.  So there's one, two, three, four. And the fourth player, which was the surprise, uh Mate One.

32:39

If you had asked me who was going to be the model that complements this team and gets China its own independent AI tech stack,  I would not have ever picked Meituan. I would have picked DeepSeek. I don't know. I would have picked Alibaba.  It turns out Meituan did it first, which is why this is such a cool story. Yeah,  they've got the model that's built to run all the domestic stuff and they've open sourced it and they're making it low cost.  And all the developers in China

33:09

and all the businesses can now start building independently,  which I assume they'll all start doing. They'll still do the other US stuff as well, but they'll build alternative ways of running things. They have to. So it turns out that's the fourth player. Fascinating. Now the fifth player on top of this would be apps. Now we're going to look for lots and lots of  core apps  that run on these models.

33:35

Suddenly you have the whole thing, but that'll be a spectrum of stuff. Those are to me are the four key players China is going to build on this immediately and I suspect a lot of companies outside the US are as well But the question is China going to cut off its frontier tech to non Chinese companies.  I suspect they're not I think they know  that  a good example of why is Huawei has a long history when Huawei got banned

34:04

when they got hit with the entity list, the biggest single question I thought was,  is this going to impact their international telecommunications business? Are the foreign carriers,  Orange, British Telecom,  Middle East, whatever, are they going to shift from Chinese telco infrastructure  because they're worried about political risk?

34:31

Because if they do that, then Huawei loses its foundation. You can't just be a telco equipment giant with only one country. You have to be global. Ericsson, Nokia, Huawei.  And after the first year, it became clear that the foreign carriers are not switching from Huawei as their core technology for most of them. So they are super sensitive to this idea of we cannot give any foreign companies  reason to doubt us.

35:00

We have to be trusted from top to bottom.  And they're doing the same playbook with open source models and with  sort of AI infrastructure, which Huawei's taking out. So is Alibaba, so is Tencent.  They are all pushing very strong. We have to be 100 % trusted  that if you put our architecture in place or our models and trace, you can rest easy and it will never be  hit.

35:26

I think they're going to lean into this hard because that's what they've been doing for the past five years.  And I think the US, which is my country, is acting with a  of a flippant arrogance about this stuff. And I think a lot of countries are going to go for the Chinese model. Then they may go for the Chinese and the US. I think that'll probably be the case for a lot of Southeast Asia, a lot of Europe,  a lot of Latam. ah But if you go to somewhere like Brazil, do they trust the US system? Not really.

35:57

Do they trust China more? I don't know. But yeah, so this sort of global infrastructure and global open source model picture  is going to come down to who do you think is going to turn you off randomly one random day. It's not clear to me that the US is going to win that. And I think China is  very sensitive to that accusation. We'll see. I know Huawei definitely is. And they're the one building the infrastructure.

36:22

Okay, so that is kind of the end of that subject  in terms of,  you know, a major milestone in sort of China AI technology just happened. Pretty cool. In terms of business stuff, because that wasn't a lot of  usable business stuff,  what is Meituan  using this for? It's actually not that interesting.  Their use cases, uh the number one use case, as far as I can tell, is at

36:51

I haven't met with Meituan in several years. I need to, I used to, I haven't in a while. I need to sort of open that door up again. Cause I really like Meituan. I really like the CEO Wan Xing. Not personally, I don't know him. But as a business guy that I'm afraid of, I would be afraid to compete with this guy. He is incredibly agile and he has just got 20 years, well, 15 years.

37:19

of just bouncing into hotels and food delivery and  food delivery reviews and  hotels and ride sharing and like, he is so fast on his feet and he's fearless. Him and like Lei Jun at Xiaomi, you should really be afraid of certain CEOs. Jeff Bezos, I would have said the same thing. Elon Musk, I'd say the same thing.  Tim Cook, no, not at all. But yeah, he's a serious player. So.

37:48

The core use case, as far as I can tell from reading, is they're building a strategic resource specifically within coding. mean, ultimately they're 50 % a software company, 50 % ops and food delivery, and a lot of robots and delivery things. yeah, mean, coding is their core. Okay, this gives you a superpower in terms of coding. If you have agents doing most of your coding, sort of like, you know, Musk, all these companies have sort of...

38:16

They're building these AI models as a strategic resource that others don't have.  They're definitely doing that.  They're doing internal workflows. They're doing sort of internal productivity things. Fine. That's kind of what you'd expect. The other thing looks like what they're calling sort of customer facing  local  services. Now, that's interesting because Alibaba would say we are doing customer facing

38:46

use cases, customer facing value add like a chat bot you can talk with when you're shopping,  things like that, or my package hasn't appeared, let's talk to the chat bot, things like that.  Meituan is using this phrase local services because you know most of what they do is in the real world. You know they're the Amazon of local services. They deliver food,  they get your movie tickets, they can get you a ride to the airport.

39:15

You know, things like that. that they are almost like a real world version  of AI. Most AI tools happen online and they are largely trained by internet data. Well, Meituan is a sea of real world data, packages being delivered, people getting their food order, restaurants preparing orders. That's all data coming from the real world.

39:43

So as far as I can tell, they're building sort of a closed loop with certain scenarios that are very real world. What people are buying, shopping, doing day by day, hour by hour, data from the real world that then they're tying into  their models.  And then they're using that to provide customer facing local services  use cases. What does that mean in practice? I'm not totally sure yet.

40:11

But they talk about the sea of operating data they're getting from the 2,800 cities and counties they serve, from the 5.5 million like autonomous delivery vehicles, from their millions of delivery riders. That's all feeding them data and they're tying it into those sort of real world on the streets use cases. Okay, I think that's all interesting and that's what I would have expected them to do. And probably focus on hotels a lot because they're huge in hotels.

40:41

You know, I would have been, I would have thought, okay, that's some really interesting sort of use cases. I think they're a really cool use case. They're really cool case study in agentic commerce. I would not have guessed that they are going to go big and try and build frontier models for everybody. I they're so specialized. I was surprised they went so big in this, uh this way.  Now, maybe that was on purpose. Maybe it just kind of fell into their lap because of the co-founder building this thing.

41:12

But I think the interesting question for them is, OK, what are these models, Longcat, going to be used for outside of Meituan? This will be my last point. I'll finish here. They are licensing this thing under the MIT license. uh Highly permissive. Highly permissive. They are giving this to everybody uh to the world, as far as I can tell. Now, maybe  the joke is  DeepSeek, what did DeepSeek do?

41:41

They built a very low cost, totally open-source model. That was what OpenAI was supposed to be.  DeepSeek has turned into what OpenAI was supposed to be. It's AI for everybody. The rules could change, but they don't seem to be. And if you read the founder of OpenAI, or the founder of DeepSeek, and you read about how he publishes his research,

42:06

He sounds a lot like Elon Musk in 2015, 2017, that this is for the world. Now, who knows? You always have to be skeptical. It certainly sounds that way. Meituan kind of sounds similar at this point. This is for everybody. It's for developers. It's for research labs. It's for enterprises. Very low cost.

42:29

totally permissible, use it for what you want. I assume it's downloadable. You can plug it into Claude Code. You can plug it into OpenClaw. You can use it for agents, whatever you want. their primary, I think I'm guessing here, their primary customer that they're building this for is developers, research labs, and enterprise who are hardware constrained.

42:58

they are solving the problem that kind of the US government created that a whole lot of entities in this world, especially in China, are now constrained by what hardware they can access. Well, this model is for people who are hardware constrained and particularly those doing MOE research.  anyways, pretty awesome.  I've been reading about some of the tech.  They talk a lot about N-gram embedding, which I don't really understand very well yet, but

43:28

of the math of how they're making this work. Anyways, I'll leave it as a question mark. You know, if DeepSeek is going to be sort of what OpenAI was supposed to be, is Meituan going to end up being the same thing but for those who have been sort of cut off from hardware by really politicians? Anyways, that's it for today. I hope that's helpful. I think it's a fascinating story. I think how they did it secretly or at least quietly is really cool.

43:57

I think that it came from a food delivery app is such out of left field that it's really interesting. And we'll see where this goes in a couple of years.  yeah, maybe look at what a big impact DeepSeek has had. This small team on a random street in Hangzhou changed the world. Is this going to be another one of those stories? And I'm kind of surprised that it's not getting that much attention. Some, but not that much.

44:23

Okay, anyways, that is it for me for content for today.  Yeah, it's been a good week here. I'm doing well. I'm getting back into content mode. I'm writing and sort of catching up. I fell behind a little bit. I'm heading out to Shanghai in a couple days. Going to go to the World AI Conference. I'll be there on Saturday, Sunday. If people are around, send me a note.  I'm just going to be running booth by booth. I've got a list of companies I'm digging into.

44:52

And I'm meeting with a couple of them. That’ll be  super fun. Really fascinating stuff. So let me know if you're in town. I'll be just bouncing around.  What else? TV recommendations. I've been watching a show on  Netflix called My Royal Nemesis. I watch a lot of Korean. Well, me and the girlfriend, we watch a lot of Korean stuff. This is a  story about a sort of uh an empress  or a concubine from

45:21

I don't know  16th century career or something like that who gets transported to modern day and she kind of falls in love or whatever with a CEO and anyways it's charming it is the most charming and it's all about the woman she's  like sort of acerbic and pushy because she's like an empress but she's also kind of nice anyways watch this show watch the first episode  my royal nemesis

45:50

The woman is absolutely like the most charming person I've seen on TV in years.  She's just fascinating and super fun to watch.  That's a solid recommendation, because I usually give you these terrible recommendations. I'll give you a terrible recommendation too. I'm watching Naked and Afraid Last Man Standing, which is, know, literally it's what they drop people, if you haven't watched this, this is on HBO Max. They drop people in South Africa or the Colombian jungle.

46:20

and they have to sort of survive and they're all naked.  They take away their clothes.  And it was the last man standing version is they keep eliminating people. They have to get food and learn how to survive and people get sick and run out of food and all  that. Naked and Afraid,  Last Man Standing, which is a sub-series. Look at season one because there's a dude on this,  this whole season, a guy named Jeff, who is just this evil Machiavellian dude.

46:49

who is just manipulating everyone  from episode one until the end. And I mean, it's really awful, but it's super entertaining.  yeah, watch season one episode or season one, all of it, and just watch for this dude, Jeff.  And he's a real person. Now, who knows if he's playing this up for the TV show? Probably.  But it's both like horrible and like super entertaining.  That's a bad recommendation and a good one.

47:19

Okay, that is it for me. I will talk to you next week.  Bye bye.