China Expert: AI Race, Key Policy Decisions & Unpacking Geopolitical Chip Tension

Could you just walk us through what's happened in the Chinese AI ecosystem?

>> The Chinese government I think postreform era hardware over software.

>> What what's the general attitude of like Chinese founders and builders?

>> The main focus is a kind of competitive pressure one to stay alive. China's

never stopped, right?

>> I remember this whole debate of like even make a good model given the type of censorship you'd have to do.

>> If you can figure out how to make a model not racist, you can figure out how to have it not say politically controversial things. So, what role is

controversial things. So, what role is the Chinese government playing?

>> The story of the Chinese government with respect to AI has been um >> Jordan Schneider is one of the pre-minent thinkers around China, hosting the popular blog and podcast China Talk. I'm Jacob Efron and today on

China Talk. I'm Jacob Efron and today on Unsupervised Learning, I sat down with Jordan, who's also been a close friend for about 15 years. We had a really interesting discussion. We hit on the

interesting discussion. We hit on the Chinese AI ecosystem overall, some of the key players and trends and what's going on. We talked about the Chinese

going on. We talked about the Chinese government and their attitude toward that ecosystem and how they're supporting it. And we talked about US

supporting it. And we talked about US policies toward China. We got Jordan's take on a bunch of the key questions around chips being debated right now and like any good conversation with Jordan.

We had about 15 tangents in between.

Overall, a really fun episode. I think

folks will really enjoy this. Without

further ado, here's Jordan.

>> Jordan, thanks so much for uh coming on the podcast.

>> What a pleasure, Jacob. Long time

coming. 15 years in the making. Here we

are.

>> Who would have thought, you know, six, seven years ago when I went into venture and you started like this China newsletter that we would have a joint podcast.

>> I did.

>> You you saw this day.

>> Of course. I've been waiting for it ever since we we revived a failing college politics magazine together.

>> That that is true. Still still a career highlight.

>> Indeed.

>> Maybe we can get the Y politic to cover this uh this episode.

>> You think they'll sponsor us?

>> Yeah. Yeah. We get a little ad or something.

>> Sure.

>> Hi guys.

>> Well, very much looking forward to having you on. Uh we have not done a bunch of China stuff on this podcast to date and so I think you will be our uh our resident expert for our audience.

I'm going to kick things off in like the most broad way possible. Could you just walk us through what's happened in the Chinese AI ecosystem pre-hat GPT and post chat GPT?

>> Pre-Chat GPT um we had a Chinese AI ecosystem which was doing AI not in the sort of like LLM scaling laws paradigm.

there was a lot of sort of vision stuff which was world class. Um and you know the the inklings of beginnings on robotics models but the sort of you know

scaling is all you need transformer architecture um paradigm which brought us open AAI and anthropic and chatbt and the world that we've been living in for

the past 3 years was not something that really anyone in a Chinese technology company was investing a lot in. So um it

was a real wakeup call for uh the Chinese technology ecosystem with uh sort of lots of hand ringing but also kind of energy and excitement and

funding that this sort of broader technology ecosystem which was really desiccated um in the wake of uh COVID and the tech lash and crackdown that you

saw really from like 2018 2019 all the way through uh the fall of 2022. 2 um on the one hand was very freaked out that like America had this incredible cool

thing that China didn't but also um uh there was enough capital talent like uh compute energy in order for lots of

experiments to be started. So over the course of 2023 um and 2024 you had a number of labs spin up. I think h probably like 12 or

spin up. I think h probably like 12 or 15 if we really want to like expand the bubble of like who was credible >> and how much of a coordinated effort was that like top down versus like hey

there's just a lot of talent. So the

interesting thing about the Chinese government and she in part she in particular but the Chinese government I think really over you know post-reform era hardware over software um so

software um uh innovation software companies were not conceived by the central government as like contributing to national power and national greatness

and the sort of reason that you had the tech lash and uh all this drama with uh Ant Financial and DD getting smacked down and whatnot was because at the end of the day the Chinese government did

not see um softwaredeled companies as like really doing all that much fundamental to um make China flourish in the way that um the party envisioned it doing so.

>> Seems like they thought it was actually bad for society.

>> Yeah. that yeah that that these these great engineers working on algorithms and models are wasting their time rotting our children's brains when they could be you know designing cars and

weapon systems and what have you. So the

2023 uh sort of expansion was a fascinating one because it was almost entirely sort

of private money driven um where uh you had these VCs who like were really just sad. there all these like like

sad. there all these like like tearjerker articles about oh we have nothing to invest in we can't raise our next round this like okay we have a new story to tell um and because the Chinese

government very early on banned um uh western closed models from uh getting uh deployed within China you have this protected market which is a dynamic we

we've seen in many and like many other uh sort of uh big corporate stories over the past 30 years in China so I would say that um there has been an enormous

amount amount of Chinese uh subsidization and support on the hardware side trying to build a kind of separate uh and resilient ecosystem away

from the sort of like TSMC and you know Tokyo Electron and ASMLs of the world into something that China wouldn't be vulnerable to from a export controls perspective to the tunes of hundreds and

hundreds of billions of dollars over the past 15 years. But on the sort of software side of things like very little and it's it from a financial perspective

it started to be um on the margins but like deepseek is deepseek um and Alibaba Quen and Miniaax

and Kimmy um are not um you know Chinese government innovation stories. They are

Chinese engineering, Chinese private capital, Chinese like big uh software companies that got hated on for the past 5 years stories just being able to do

their thing and ultimately deliver models which are um you know competing globally.

>> Were these efforts like venture funded because obviously you had these big companies that like did you know uh you know that that produced these models and then my impression of Deepseek was was like some hedge fund that like started doing these things. What is the story

behind a lot of these like major Chinese players that we hear about?

>> So, it's all of the above. Um, Deepseek

is this weird exception of this guy who ran a hedge fund who just thought models were cool. You know, he was sort of like

were cool. You know, he was sort of like Zuckerberg and he just like bought a lot of Nvidia chips before they became impossible to buy. Um, and he had a team that was sort of working on this and

then they ended up, you know, going from spending time making like high frequency trading models to um to large language models. Um, Alibaba, China's biggest

models. Um, Alibaba, China's biggest cloud provider, uh, China's biggest, uh, owns China, the largest kind of e-commerce platform in China, also a big payments business. Um, a big sort of

payments business. Um, a big sort of like, uh, finance business as well on the side of it. Um, they were working, you know, had had enough computer

science talent, brought in a lot more.

um other big companies like bite dance um which you know you got you guys know what Tik Tok is um uh you know had talent and capital and compute um there are a handful of smaller models which

did raise VC funding I think having um there were a handful of folks who had a pedigree from you know having worked in in western labs primarily there's a big

kind of USChina compar uh uh knowledge interaction transfer which we can get into but yeah you had you saw raises of um you know not like OpenAI and

anthropic raises but um you know eight figures like low n figures to try to spin up something with like business model TBD but like let's just go train some stuff we'll and we'll figure it out on the back side.

>> Sounds familiar. I mean who like obviously in the US the version of these has been folks that were in the big labs you know a lot out of like deep mind or you know uh or meta or open or anthropic and they spun out and they start these

new labs like the set of people that are at the forefront of the model development ecosystem in China a lot of them have western experience is it just like hey once you know it became clear to scale transformers like there's certainly plenty of great engineers over

there that have been messing with this stuff like who are a lot of these people >> both so we have folks with PhDs at Carnegie Melon and folks who spent time in Microsoft and Google. Um, Deepseek is

an interesting exception where um the sort of lore is that his English wasn't even good enough to like get into the top grad schools in China. So like yeah,

he wasn't about to move to the um to the US. And um sort of made a point of only

US. And um sort of made a point of only hiring domestic talent because he thought it was kind of underpriced in the marketplace. But um you know if you

the marketplace. But um you know if you go through the background of a lot of these uh top researchers like the way they establish their credentials um both

to the large uh kind of established tech firms that want to hire them as well as the funders who they need to raise money from is by having some sort of uh kind

of western big tech or western academic credential. though this is changing and

credential. though this is changing and I think sort of one of the remarkable things of like you know you've had a lot of young researchers on your podcast is this isn't necessarily a field where you

need to have spent 20 years in it to be on the frontier. This was one of the um uh the open questions that some people but not China talk readers had over the course of the past 5 years is like can

Chinese engineers do it? Are they

creative enough? Are they lateral thinking enough? And it's like yeah they

thinking enough? And it's like yeah they can. It's a big country. They got a lot

can. It's a big country. They got a lot of really smart folks. And by the way, you know, if you look at a lot of these western labs, like a third of the last names are Chinese. And those Chinese uh

of that third, I would say 2/3 or 3/4ers of them grew up in mainland China. So

there's just an enormous amount of cultural exchange, intellectual exchange, academic exchange which is happening before and continues to happen even um or despite all of the kind of

geopolitical tension that's being uh you know spun up because of all this >> that like research collaboration. Do you

expect that to continue? What what are what's in the tea leaves there?

>> We'll see. Man, I'm going to Nurups next week which is in San Diego. Um, this

year is the first year they've had to spin up a second off-site one in Mexico basically because Chinese researchers struggled to get visas to the most

important AI research conference in the world. And that I think is like

world. And that I think is like emblematic of a lot of the sort of larger challenges that these two ecosystems are u you know have today and will continue to have to like play nice.

I mean we are not in a sort of like per perfect open free market. this is a this is a dual use technology. This is the you know the story of uh the the decade

when it comes to economic growth.

>> I guess you going back to your timeline um it felt like you know with Deep Seek and then afterwards there's been this collective movement in China of open sourcing models. Um what's up with that?

sourcing models. Um what's up with that?

>> Hey everybody, I'm Ann Maris Walton and I'm the new producer of unsupervised learning. Today, I'm super excited to

learning. Today, I'm super excited to introduce a new concept to the show where we open up the floor to you. If

you have any pressing questions about the AI ecosystem that you'd want to ask Jacob and an upcoming guest, please submit them using the Google form that's listed in the show notes below. We're

super curious to hear about what you have to say and very appreciative of your support of the show.

>> There's been this collective movement in China of open sourcing models. Um,

what's up with that?

It's a good question. Um because as as you can tell better, you know, as you know better than I do, it's hard to make money when you're not selling a thing,

right? So, um I think from the beginning

right? So, um I think from the beginning it was a sort of market adoption play and a bit of a race to the bottom play where like if there was a really good

model that was open source and free, it would be really hard to sell a closed model which was not all that much better. And in the US, you did um over

better. And in the US, you did um over the past 3 years really have a considerable gap where if you wanted to have access to the best um capabilities, you could pay for it.

>> Yeah. I mean, Meta tried the strategy and then backed out, right? It just

didn't work.

>> I think a few things happened. So, like

the technology ended up working out that you could get really good, pretty close stuff without having to spend tens of you know, billions of dollars on training runs. um which is something

training runs. um which is something that um made it easier for the Chinese players to um uh to to take this to be able to sell stuff which was like pretty good and still and still usable in this

paradigm. You know, all the firms are

paradigm. You know, all the firms are trying to make money. Like we had the one year of fun research, let's just let's just do stuff. And now there's starting to be more of a crunch. Our our

15 AI labs that we started with are now down to like really five or six. And uh

most of them uh are sort of either like Deepseeka is is an exception but most of them are sort of like aligned with either Tencent or Alibaba and then the sort of the applications and use cases

are much more straightforward like yeah if you are Amazon um you want to have a model um so chat GPT doesn't take over like the whole buying ecosystem that's

sort of the same thing um from the Alibaba perspective and open AI has not really been born In China, it's still like a few small players who are fussing around and making pretty good models and

a handful of giant companies who can pay for this because like they have other businesses which make enormous amounts of money and spin enough spin up enough cash for them to you know stay relatively close to the frontier, right?

But I mean to your earlier point it seems like to the extent that similar use cases to the US will exist in China and the fact that these closed source models are banned in China there feels like a large like do you expect the

Chinese LM market to kind of shake out similarly to the way the US one has?

>> That's a really good question. I mean I think it hasn't because we don't have AI labs that are strong enough to stand on their own. And I think that is sort of

their own. And I think that is sort of like this comes back to like a fundamental uh uh difference between the sort of broader western ecosystem and

the Chinese one is there's just more VC money slloshing around in order to scale up an open AI or an anthropic to be able to compete with the likes of a with the

likes of a Google, right? And what we've now seen in China is like, you know, we'll see. Maybe some of these smaller

we'll see. Maybe some of these smaller labs will try to go public and raise a lot of money, but um the sort of relative um the relative independence that I mean, not really anthropic

because they're kind of like, you know, beholden to a lot of these other players, but but Open AI has been able to develop over the past three years is not something that's happened to any of the um sort of players who've entered uh

over the past few years in China.

>> Right. I mean, well, does the Chinese government just, you know, sit on the sidelines and be like, "Hey, let this develop as it develops." Like, I imagine there's a huge incentive to concentrate compute, concentrate talent, like really

get something state-of-the-art or what's their attitude toward this stuff been to these?

>> Yes and no. I mean, I think it's working pretty good. Um, if you're if you're if

pretty good. Um, if you're if you're if you're on the Chinese side, like look, pre-training is nice. Um, but by the way, the Chinese firms can still access

compute abroad. um and the best uh you

compute abroad. um and the best uh you know and and sort of use the best Nvidia chips. I mean we have this famous Oracle

chips. I mean we have this famous Oracle uh basically powering all of bite dance in America which is maybe the the the story that's gotten the most headlines.

But but um the sort of the compute constraint is as much one of capital for now as it is as it is like a coordination problem. And you know,

coordination problem. And you know, there's lots of downsides to like putting your thumb on the scale and saying you're the you're the lab that we want to have win it all. And and really

what we've seen over the course of uh how the Chinese government has um sort of shephered the growth of other emerging technologies and industries is they're fine to have companies fight it

out for a while, right? Like BYD was not um uh uh anointed. Um Huawei was not anointed. they had to like battle for um

anointed. they had to like battle for um a dec you know I mean in Huawei's case like multiple decades before they sort of rose to um national champion status.

So we're still in the early innings on that and that phase and I don't think the technology today is quite the sort

of um uh you know AI 2027 like one one you know one run to rule them all and if you don't get it you're going to be lost forever. It's a much more dynamic

forever. It's a much more dynamic ecosystem is a repeated game. We're

still figuring out the business models even um much less like whether or not AGI will be achieved once I collect enough chips in a basement.

>> What role is the Chinese government playing? You alluded to like obviously

playing? You alluded to like obviously very focused on the hardware side, super involved in chips and and reducing dependencies there. Anything else

dependencies there. Anything else they're kind of like focused on? People

probably heard of America's AI action plan. Like China had a version of it.

plan. Like China had a version of it.

It's all kind of like fine, not super dramatic. like, "Yeah, we're going to do

dramatic. like, "Yeah, we're going to do more AI in education. We're going to do more AI in healthcare. We think open models are cool."

>> They also have like relatively meaningless platitudes in their uh government docks.

>> Oh, yeah. 70% adoption in all companies by 2035. Like, okay, cool. That seems

by 2035. Like, okay, cool. That seems

that's that seems like a nice idea. Um,

you know, one of the sort of like risk vectors that people were talking about was that the Chinese government would be really scared of what these models would produce and then kind of like crack down

on uh, you know, folks access to it because they're worried they'd be making >> I remember this whole debate of like if you could even make a good model given the type of censorship you'd have to do.

>> Yeah. It's like I don't know. They

figured out how to censor other stuff.

How racist is Claude not that racist? If

you can figure out how to make a model not racist, you can figure out how to have it not say politically controversial things, which by the way are like 0.001% of queries. People are asking what

of queries. People are asking what basketball shoes to buy, not um you know what happened in their country in 1989 primarily. So um so yeah it's it's been

primarily. So um so yeah it's it's been much the the sort of the story of the Chinese government with respect to AI has been um more about how they I think

have like um maybe learned some lessons from uh the heavy hand that they put on the semiconductor ecosystem. I mean that was very rocky. At one point they jailed everyone who was running the most

important sort of industrial policy big funds >> uh 2021 2022 or something. Yeah. So so

they had like a hundred billion dollar fund and the top seven people have now been disappeared for corruption. Right.

The reporting is that that was downstream of she being frustrated about the relative amount of progress that the indigen indigenization has been made.

And like look, if the VC firms are going to put up the money, you know, you can spend it on other things. So yeah, it it'll like there's there's more downside

than upside risk, I think, for the like eye of Sauron to come on this. And for

whatever reason, it's it's been um pretty low touch for now.

>> From published documents or anything else you can get a sense of like how top of like how worried is the administration about this?

>> It's really funny. So, so the the Chinese uh the Chinese system, I mean really since Mao has done these things called like study study sessions where they literally like sit around, bring in

the top experts and academics and like quote unquote learn about a thing, you know, like have a lecture series like ask questions. They had an AI one in

ask questions. They had an AI one in May. We wrote it up. We'll link it in

May. We wrote it up. We'll link it in the show notes. you can kind of triangulate what exactly they're focused on by like which which guests they bring

in to um uh to talk about them. You

know, pretty anodine people, not not very crazy things. AI is important. We

need to care about it. Not a ton of AI dumerism. I think there's like

dumerism. I think there's like >> is there like a Chinese alazer or like you know do they have the that >> no I mean the thing the thing about like the right lens to understand the Chinese

the Chinese government's approach is you know what is it going to do to national national power and what is it going to do to political stability and I think

there is a sort of um a broad understanding that the sort of social dislocation that um this technology could potentially produce use with job

displacement is like an opportunity and a risk and like we'll see how it plays out. It's not hitting um you know as

out. It's not hitting um you know as hard as what um sort of reforms in the early 90s with stateowned enterprises did for instance. So I think they're

relatively confident in their systems ability to kind of absorb that um uh discoation that will come from the productivity growths and on the like AI

doomer safety side not a lot um from the from the government perspective you know a line here or there uh you see um uh

the sort of level of pledness of various uh founders is interesting. one of the top AI people at Alibaba. He just talked about AGI for 30 minutes at their like

big annual conference. It wasn't

anything novel or interesting compared to what you would see in the like it's like very very analogous to what you would see in the discourse in San Francisco. Um but it was remarkable that

Francisco. Um but it was remarkable that it happened from such an established player of like talking about you know grand vistas of like new scientific opportunity uh and whatnot. So there

there is a real interplay in the in the discourse I think much more from the sort of private sector founder perspective. You know they all listen to

perspective. You know they all listen to Dwarash and uh Jacob show.

>> I do I do I'm always there someone will occasionally send me like a Chinese translation of one of our episodes and like it does some real numbers. Like I'm

shocked at how much people in China seem to follow along with like the discourse and like media here around AI uh in the US. And this is one of the like um uh

US. And this is one of the like um uh perennially frustrating things for me is in the other direction like there's like

10 of us really uh whose job it is is to like do the translation in the other direction and follow the Chinese AI ecosystem and explain it to the US.

Whereas in China like oh I'm glad I'm glad that's happened but like yeah there's entire media organizations that are following what's happening in Silicon Valley like translating the

interviews, writing lessons. um uh uh and that just doesn't happen in the US for um various structural reasons which we can delve into or not. The looking

glass is very much one is uh one one directional.

>> Yeah. What what's the general attitude you know of like Chinese founders and and builders toward like I mean as they're consuming that content is there this feeling of like a race or is there a feeling of like excitement and we're all collectively trying to build this

technology? I'm sure it's not as simple

technology? I'm sure it's not as simple as any of that. Look, I think the the main focus is a kind of competitive

pressure one to stay alive in this like incredibly hardcore like technology and commercial competition both um in the big companies as well as in the small

ones like you know it's been funny seeing this mirroring of like all the AI labs they're now working 996 like China's never stopped right and I think every time there is a big opportunity

the competition is just even more fierce. Um, and there's less breathing

fierce. Um, and there's less breathing room for these uh for these labs to survive. So, um, that is probably

survive. So, um, that is probably comprises like 90% of the day-to-day worries. I think there is a broader

worries. I think there is a broader sense and you can read this in some of the Deep Sea Founders interviews that he does feel like he's on this like humanistic quest to like you know create

a new technology and open like new vistas of of of of science and and um innovation uh which is which is nice but there's also this very frustrating

backdrop um which I think is probably best exemplified by a um uh a Johnny John anecdote. So, he was the bite dance

John anecdote. So, he was the bite dance um the bite dance founder. Um his old social media posts from like the late 2000s, early 2010s made clear he's liberal, like he's frustrated with

censorship and whatnot. Um and he was kind of running, you know, basically running a media company like he had he had like a Buzzfeed equivalent of Buzzfeed like was big and good and tech

enabled. um as uh Tik Tok in China, Doy

enabled. um as uh Tik Tok in China, Doy was growing. Um uh he and his uh

was growing. Um uh he and his uh leadership team like disappeared for a few days and then at like 512 in the

morning on like a Monday um he posts a letter which is basically an apology saying we did a bad job of uh uh moderating our content. We need to like

serve the party um better. we're not

upho upholding, you know, socialist values and this sucks. So there's this there's this like really tragic like

central tension is that the um you know a lot of the leading lights of the Chinese technology ecosystem do have this global liberal perspective

oftentimes have worked u worked in the west or in western countries aren't you know really all about like making the CCP great again um and you know have

kind of like you know globalist just in the nice cosmopolitan, let's say, um uh uh visions of what what they produce can do, but they're living under a system and in a sort of like global state of

things where it doesn't matter what they think. So um you know as this technology

think. So um you know as this technology becomes more and more relevant from a from a national power perspective uh it

seems inevitable to me you know if uh sort of the Chinese government continues to be run in the way that she has run it for the past 15 years that the sort of operating space for this type of um you

know positive sum interaction is going to is going to shrink. It feels like there's a real financing gap today in the Chinese ecosystem. And in the US, you've got tons of venture investors, rich capital markets, folks like Nvidia that are, you know, would love to have

as many potential model companies uh that can exist. Um and it seems like, you know, there's one version of the world where the Chinese government is just happy to let everyone club each other to death, see who the few winners are. Uh obviously, we've seen scaling

are. Uh obviously, we've seen scaling laws going to seem to continue to hold up. It feels like this is going to be a

up. It feels like this is going to be a capital intensive game.

>> Do you think that that is how it persists or is the Chinese government step in and try and solve this in some way? Okay. So, let's do a bit of a like

way? Okay. So, let's do a bit of a like net assessment of uh love a good net assessment.

>> Yeah, you got to do it of the of the Chinese and and I'm sorry to get you a whiteboard to uh to to really do it.

>> So, you know what are the inputs to like have uh great models that you can deploy at scale. You need talent. China's got

at scale. You need talent. China's got

as much talent as any other country in the world. Um you need energy. They

the world. Um you need energy. They

probably have uh enough energy to to run that as far as we're going to need to over the next few years. the real

constraint is going to be compute. And

um right now it seems like the Trump administration has decided to continue the Biden administration policies of not giving the Chinese government open the sort of the the Chinese mainland like as

much access to compute as they would want um in the interim. So then what we're having a a sort of battle on is

like whether the uh sort of western finance plus um TSMC plus like the entire western supply chain for

semiconductor manufacturing equipment can uh will be able to sort of outscale with an enormous like a lead of probably

10x where Huawei is right now, maybe 15.

um the domestic Chinese ability to sort of copy um and and scale up the same thing. And you can look at the examples

thing. And you can look at the examples of um uh electric vehicles and the solar industry as moderately difficult emerging technologies that China really

has been able to scale up in a way that dwarfs the rest of the um the rest of the world. But I do think that

the world. But I do think that semiconductors are different. Um, we

have basically had an open tab when it comes to the Chinese government allowing uh their hardware ecosystem to spend as much as it takes to build it up and they are where they are today. The Chinese

government can spend a lot of money but global capitalism can spend a whole lot more. I mean the fact that Alibaba is

more. I mean the fact that Alibaba is worth like $200 billion and Nvidia is worth 4 trillion 5 trillion. all the

sort of like swing sources of capital seem to be coming America's way. So

there isn't really like cavalry going going to come in from a sort of like uh like a technological defector of like oh is is is Japan going to switch sides? I

don't really think so. such that I I do feel kind of confident that the money that the west is going to be able to the money and technology the west is going to be able to deploy to scale up on the

compute side is going to be able to uh swamp uh the Chinese players for a long time to come >> end of next year is there going to be like a greater or smaller difference between u you know American and Chinese models

>> so the models themselves who knows um but you know we've kind of learned this it's not just the model it's the ability to serve it, right? And that is really where um you start to see these

constraints. I mean like like if you

constraints. I mean like like if you think the rate limits on claude and open AAI are bad like just try using a Chinese model, right? And so that both takes compute and it takes money. Um,

and look, maybe once the business models start to click in, that starts to change. But as long as you do have this

change. But as long as you do have this um relatively um uh uh substantive wall, not allowing uh China to buy the best

chips in the world from the place who makes 10 times as many as what the Chinese competitor can make, then you you've got a big edge. the the that you know doesn't necessarily mean that the

gap between closed uh you know western models and Chinese models is going to be all that big but like in order to reap the sort of gains from a business

perspective and from a broader like national productivity growth perspective like you need to have >> like how many tokens does each person get for you know >> exactly and that's the thing that I think is domestically going to be

constrained relative to the west uh and China for a long time to come.

>> Yeah. What's interesting though, it sounds like the Chinese government is focusing on the exact right area, which is that like basically, you know, opening up capital markets more or putting more money to work without solving this hardware problem is not actually all that helpful. It's like

that's that's the thing that, you know, needs to be figured out. You kind of alluded to to some of the US government policy, you know, Trump administration continuing some of the stuff the Biden administration's done. Can you just give

administration's done. Can you just give our listeners like a walkthrough of like USI policy toward China over the last five, six years and kind of where we are today?

>> We're starting uh 2015, 2016. So late

Obama, you have this consensus start to shift in America from the more we engage with them uh the more liberal and and comfortable we'll be with China's uh

rise in the world to this um understanding that the direction that she is taking the Chinese government is not one that the US is going to be comfortable with anytime soon. So

starting late Obama then in Trump won and and into the Biden administration, you had this sense that China is a strategic competitor. Um the US just

strategic competitor. Um the US just can't let them do whatever they want and gain in in every front that we consider to be of like national strategic

importance. So there are um places that

importance. So there are um places that we're going to have to start um uh controlling. So in towards the end of

controlling. So in towards the end of the Trump administration, there was a big focus on telecommunications and Huawei. And when the Biden team came in,

Huawei. And when the Biden team came in, um, uh, thankfully there were a handful of like pretty pill folks, um, in the White House in 2021 and 2022 that saw,

um, just how important, uh, compute was going to be to this sort of new paradigm of competition. So September 2022,

of competition. So September 2022, >> who are these people? I feel like before Chad GBT, it was not like, you know, I I wouldn't expected so many people in the administration to know that.

>> Yeah. I mean, it was it was a really interesting happen stance. So there are a handful of folks at CES set. Um Jason

Matheni, uh Ben Buchanan, uh Chris Maguire, uh Safe Khan, you should have all these guys on. Great for your show.

Um you know, had been writing about AI for years and >> they were reading all the research.

>> They're reading the papers and they're just like, "All right, this could be a big one." So in October 2022, you have

big one." So in October 2022, you have so >> a month before chat GPT >> a month before Chat GPT, you have this very dramatic move to to start to control both exports of AI chips as well

as semiconductor manufacturing equipment. And um then a month later uh

equipment. And um then a month later uh we have chat GPT and everyone's like oh like like at first we're like oh this is kind of weird kind of interesting and they're like oh okay like I get it

scaling laws like here's what we're trying to stop. Um so there has been a sort of there was a dance over the course of the Biden administration where on the one hand Nvidia wants to sell

chips so they make something that's like right under the limit and then the government is like ah no that's not cool. The parameters of which

cool. The parameters of which manufacturing equipment also got twisted over time and then they wanted to make sure that the the Dutch and the and the Japanese did it at the same time. So, it

was kind of delayed and then you had this big scandal of uh TSMC basically selling a ton of wafers to uh a Huawei cutout that then they ended up getting smacked for. So, there was this

smacked for. So, there was this understanding that like look, we want to win on compute and one of the ways we're going to win on compute is by sort of throwing some gears in the sand of the Chinese computer ecosystem and not

letting them just like buy up everything. and you know not great uh

everything. and you know not great uh you know not perfect not terrible but like on the margin definitely slowed down um the development of the Chinese semiconductor and and broader AI

ecosystem. So in comes the Trump

ecosystem. So in comes the Trump administration. The first few months you

administration. The first few months you had a team that seemed to be very aligned with this view that controls were important. Enter David Sachs exit

were important. Enter David Sachs exit um uh the China hawks thanks to Laura Loomer and in comes uh uh Jensen Hong into into Trump's life. And we've had

this real head spinning drama of will he or won't he allow the best American chips to sell to China. Um an

interesting wrinkle in this is that the H20s which are good not great chip uh you know like one generation behind kind of limited in some ways but still very

powerful. Um the Trump administration

powerful. Um the Trump administration decided to ch decided to sell into China only for the Chinese government to say screw you guys. We don't we don't want your stinking chips. Now, why that

happened was a fascinating is is like a is a big puzzle that the community is still very much um debating. On the one hand, maybe she has bad information and the Huawei lobbyists got in and told

them that we've got better chips and more chips than you really think we do.

Maybe it's a negotiating ploy to get Blackwells um uh you know, instead of this like gimpy knockoff um model. And

then there was this big drama uh in the leadup to uh Trump's summit with Shei in uh Seoul where he got on the plane, the stock went up 10% of the expectation

that he was going to decide to sell them and Lutnik and Rubio and Besset uh and Davidson Greer like all huddled in Air Force One being like Trump don't do this. And he comes off the plane doesn't

this. And he comes off the plane doesn't he doesn't sell the chips. But now we have another story that they're considering selling a you know a version of it. So the this administration I

of it. So the this administration I think is of is very or is definitely of two minds of on the one hand this is an edge that we want to keep. On the other hand maybe there are some upsides to to keeping China buying Nvidia chips for a

while.

>> Yeah I get the argument for not selling them. How would you strongman the

them. How would you strongman the argument for selling them?

>> So the the most compelling case that Jensen has is that AI accelerators are not just their hardware they're the software as well. And the software keeps

getting better partially because um you know Nvidia and AMD uh and Huawei themselves invest in their own architectures and also because there is this broader um kind of global community

of developers which is which is contributing to um the ability for CUDA to just like be better than a random chip you get off the shelf that doesn't have the software optimizations. And you

know there are thousands of very very good Chinese engineers um who wouldn't you know who given their brothers would love to stay on Nvidia because they are the best chips and by not selling chips

into these Nvidia chips into China um there is a risk that that engineering energy instead gets spent on improving the sort of Huawei software ecosystem. I

feel like people have also written about this like ideal goldilock scenario where like you give you know Chinese companies enough of chips that like they don't spend too much time like building their own. Like I feel like Ben Thompson's

own. Like I feel like Ben Thompson's written about that and Dylan from Semi analysis. Is there like a way to stick

analysis. Is there like a way to stick the landing where it's like you still have most of the of of like the cutting edge chips in the US but like you've given at least enough of an incentive not to like go all in on Huawei and and and the local ecosystem.

>> Howard Lutnik said um you know we want to get them addicted to American chips.

And I think this is a story that has played out so many times over the past 30 years that the sort of the goal is indigenization and it's going to come

whether you like it or not. Whatever

your business plan, whatever your China strategy is, if this is if if what you're selling is an important is like a nationally important strategic research uh uh resource, then you are going to

have domestic energy which has been supercharged by the fact that there have been so many export controls and it h there have been choke these choke points that have been raised. People say like

oh like Huawei went from this like second source no one really wants to deal with to like what's front and center. Well, you know, the next

center. Well, you know, the next administration is only three years away.

I don't think these companies are stupid. Um, and uh you I would almost

stupid. Um, and uh you I would almost make an argument to the other side where if you're trying to make a build a lead um you know, why give them this bridge

and this breathing space um that selling a lot of chips would provide because the sort of as we talked about before on the hardware ecosystems like the hardware the Chinese hardware ecosystem is not

capital constrained, right? They have

all the energy. They have all the attention. They've had more money p

attention. They've had more money p poured into it than any, you know, it is the biggest industrial industrial policy project in human history. There's been

more money poured into it than than anything ever. So, um they're going to

anything ever. So, um they're going to try as hard as they can. And and my my takeaway is that selling Nvidia chips particularly into China, not even um

allowing Chinese firms to rent them and in Malaysia or Singapore or AWS or what have you, um the the the the drawbacks don't exceed the benefits. But look,

these are the like there are a number of unknowns which are in this is a very complex sort of uh equation to pan out like it depends on your assumptions of of how good Huawei is going to get. It

depends on your assumptions of to what extent you know Nvidia is capital constrained like are there you know is the is the limiting is what's limiting

American uh or sort of western AI buildout compute is it energy um you know like is there more um would more

demand for TSMC actually like lead them to make more chips or are they already kind of tapped out um so I don't know a lot of variables it's a complicated

thing the One illustrative point I would make to you is that I have tried, you know, I've done a lot of shows on my podcast about this. I have tried to find

an expert that has not uh that isn't sort of financially tied uh to the to the like Nvidia upside outcome of uh making the case and I haven't been able

to find that person. So, um, there seems to be a pretty strong consensus among the folks who take this seriously that the balance leans towards not selling

them the ships and definitely not selling them the semiconductor export equipment. Um, but there are legitimate

equipment. Um, but there are legitimate arguments on the other side. What is

like the status quo they have of this ban? Like is it is is it effective? I

ban? Like is it is is it effective? I

feel like the Wall Street Journal had some piece about chips getting smuggled in, you know, recently. It feels like you can, you know, rent some of these chips in in clouds maybe in other countries, you know, on the ground for these Chinese firms. like are they pretty constrained in their access?

>> Yeah, I mean that's the that's the sort of interesting thing about this is there's nothing stopping Chinese firms from renting compute abroad and the renting the best chips abroad. That

whole lock in thing I was talking about is like these companies are still training their models on Nvidia. They're

just doing it like in Malaysia or with the chips they already have or the chips that they've already smuggled because Nvidia is still a lot better than um than what than what Huawei has to offer.

So the upside of like actually legally selling the chips into China, you know, really is just like I don't know uh a couple, you know, maybe 10 2050$50

billion more of uh of revenue over a 5year horizon. But it's unclear to me

5year horizon. But it's unclear to me how much of that money will be marginal money versus money that the Chinese firms are just spending uh to access.

>> I feel like Biden diffusion rule had something around like you know >> Oh yeah. So that's gone.

>> Okay. So that's completely >> Yeah. Yeah. We're not We're not too

>> Yeah. Yeah. We're not We're not too Well, it's interesting cuz we were trying to prevent this, right?

>> It's Yeah, cuz this is this is what they were trying to do. I think the Trump administration is less bullish on it, but there's this very weird liinal thing where they like issued a recision. They

they said in an FAQ from BIS that the diffusion rule doesn't exist, but they haven't actually crossed it out on paper presumably because there is a there is still like an internal fight in the

Trump administration of whether or not um this diffusion rule is a thing that we're going to um live by or not. So

yeah, it's you're asking the right question. A lot of this stuff is TVD.

question. A lot of this stuff is TVD.

And the thing about the Trump the Trump administration is they just change their mind all the time. Like we're optimizing for vibes. We're optimizing for

for vibes. We're optimizing for headlines. The folks who try to project

headlines. The folks who try to project some grand strategic thinking on it um I think are connecting dots that that don't actually exist. Like we're like he's all about sort of internal power

struggles and you know being this like flexible player or something. So I would be shocked if we ended up uh you know given how many twists and turns we've

had over the past uh year or you know 10 years of Trump that we'd end up getting real consistency on this thing beyond him saying okay AI is important we want to like win in it.

>> So today these Chinese firms can access decent amount of comput abroad to train they presumably if they had popular products could run inference on those products abroad like I can't tell how much of a constraint it actually is.

Yeah, I mean there there's some kind of like data privacy law stuff with like using cloud that's outside China and into China. So that I think is one of

into China. So that I think is one of actually more of the constraint on the Chinese side as opposed to the American.

But yeah, I mean it's it's a it's a access to capital compute problem as much as it is a uh uh uh like the US government isn't letting them have access to Nvidia today. uh on the Huawei

side, like obviously there's like uh technological improvements that need to happen and then there's also just like being able to manufacture at scale, right? And have any sort of like demand

right? And have any sort of like demand for it. I realize this is probably a

for it. I realize this is probably a hard question, but like do they have a shot? Like I mean, how do you think

shot? Like I mean, how do you think about that question when you talk to other folks? Yeah. I you know if you go

other folks? Yeah. I you know if you go back to 2020 and 2021

the the chips that Huawei were designing and making um at Smick were basically as good as uh Nvidia's chips. So from a

design perspective they are not lacking in any sense of the word. So it really comes down to the sort of scaling within the Huawei fabs and within and within the smick fabs and the big constraint that they're operating under is they

don't have EUV. So that means that um their chips are sort of stuck at 7 nanometer and you know you can push it but like all the technological trees

they are going down are the one are ones that all the western fabs and the broader sort of global ecosystem decided were less efficient because using EUV uh

lithography to make this stuff was a better um you know mix of like effort and yields and like just capital and technological research allocation. So um

a lot you know one of the one of the big ironies of the Biden administration semiconductor equipment control policy

is that semicap basically doubled uh in 2020 from 2022 to 2023 once they controlled this because there was this big uh influx of capital and folks you

know the the Chinese semicap the Chinese semicer ecosystem sort of saw the writing on the wall and they're like oh we got to hoard um both chips as well as a manufacturing equipment. There also

still a lot of loopholes where like there's you can have a subsidiary which is 45% owned. There are some fabs where like one side like a fab on one side of the street is controlled and the fab on

the other side of the street isn't. Um

and there was a big lag in sort of when the controls uh came on in the US versus in the Netherlands and in um uh and in Japan. So, um, they do have a lot of

Japan. So, um, they do have a lot of equipment in the in the building, but again, it what we're focused on is like the relative amount of compute that

Huawei is going to be able to manufacture versus the rest of the world. And the fact is like um uh if you

world. And the fact is like um uh if you look at the Huawei road map versus the Nvidia road map and the and the TSMC road map like um you know I will I will

send you the like infographic I've vibe coded but it's basically like a 15 it's like a 15 to1 ratio at least uh for the next three years. There are some there are some knobs

>> ratio of production or >> uh yeah the compute that is going to be made. Um and you know there are some

made. Um and you know there are some knobs that sort of change that from 15 to 13 to 10 if you kind of make different assumptions about Chinese

access to high bandwidth memory or maybe Huawei gets better at designing this that and the other thing. But like at least for the next 3 years, I mean 10 years from now, 20 years from now, TBD,

but in the in the near to moderate horizon, like this is a real edge that the western ecosystem is going to have um compared to China's.

>> Yeah. I mean, I'm struck by about lithography like should are we not paying enough attention sucking up to the Netherlands? Well, you know, this is

the Netherlands? Well, you know, this is the interesting thing is that because there is an enormous amount of uh American technology in all of these uh

machines. I mean, you know, ASML like

machines. I mean, you know, ASML like was American, you know, lots of lens like there are lots of parts which are like fully supplied by the US and like

so so um I had Jake Sullivan on my podcast uh uh a few weeks ago and the um the main uh sort of hypothetical I got him to engage with was whether or not

the US should have done it the hard way as opposed to the nice way. and

basically saying uh sort of doing a doing a fataccom plea with these uh uh with these allies countries selling equipment saying sorry you can't sell anything because what because what ended

up happening was the Chinese uh uh even though the Chinese uh Chinese firms were not allowed to buy EUV equipment they spent tens of billions of dollars buying DUV and that's really what's powering um

uh all of this uh all of this Huawei buildout as well as the legacy um the legacy chip buildout. So, um, uh, there's like a lot of latent power that

still exists. Like they can't really say

still exists. Like they can't really say [ __ ] you on this. Um, we we we do hold the cards here, but, um, the Biden administration decided to to value the

alliances a little more, which led to all of this kind of like frustrating, uh, two step forward, one step back. You

know, there was a hope within this community that the Trump administration would put the pedal to the metal on this. And it's like, look, if the

this. And it's like, look, if the Netherlands is worried about America leaving NATO, how worried are, you know, what's what's one company's like political economy really matter in the grand scheme of things? Same with same

with the Japanese firms. But um that has not that's not how it's really played out. And because of the whole rare

out. And because of the whole rare earths um drama between the US and China, uh it seems like whatever um controls that the Trump administration

was considering to tighten on the semiconductor manufacturing equipment side are are on ice for now.

>> Yeah. Interesting. I guess if you had the year of the uh you know Trump administration and you could kind of like you know wave a wand and put one policy forward like anything you know that you'd be doing. So on the export

control side, I would say uh you know I want to take it in a slightly different direction. So the the chips and science

direction. So the the chips and science act which was passed um in uh in 2022 led to you know part of the reason why there's this been this whole kind of

domestic fab renaissance um had two parts to it. So, one is the is the um kind of uh loans, grants, tax benefit to

uh R&D tax benefit uh that has helped build all these fabs in uh in Arizona and uh and Texas. And but there's also

this R&D component um uh it was uh $13 billion to sort of re to kind of create like do basic and translational research

u that would push the frontier of semiconductor manufacturing forward. Um

the Biden administration kind of fumbled the bag on this. They were sort of slow um to to start spending the money. they

didn't end up developing a lot of um uh friends uh friend within friends with an industry and um the Trump administration a few months ago canceled it uh or

canceled that the the the iteration of the nonprofit that was going to run this that the Trump that the Biden administration had set up. Now what's

going to happen to that money is TBD. I

mean whether it goes into like just buying stakes in companies because we think that's cool now whether they try to spin up another version of uh of this whether it just like ends up being a

kind of subsidy to corporate R&D um is an open question but like it would be a bummer if um the biggest kind of injection of research and development

which if we're looking at like a 10 20 year horizon is what is going to matter um uh when it comes to this sort of like semic conductor and broader compute

competition like I would like that chunks of money to do chunks of money at that scale to do basic research do not come around very long uh you know very very often in the American political

system and I and I do have hopes that um the Trump administration will still come up with some creative and impactful ideas to spend that money.

>> I'm sure this is not your favorite part of being a a China expert, but I imagine I'm I'm contractually obligated to ask you a little bit about Taiwan.

>> Sure. Only for you, Jacob.

um you know I mean people often war game a Taiwan blockade Taiwan invasion like you know what how likely is this in in the next decade all right so let's start with the Trump administration and then

we'll do like a like a 10-year thing so the American policy towards Taiwan um to defending Taiwan has been this idea of strategic ambiguity that like we're not going to commit to Taiwan that we're

going to defend them so they don't you know try to reinvade um mainland China which is something Chiang Hai was considering a number of times um and we're not gonna tell China what we're going to do to sort of u you know make

it clear on one side or the other by the Biden administration and Biden himself um said a number of times yeah we'd defend Taiwan basically and and sort of

like adlibd and the Trump administration uh there were a number of people u you know there is a big debate I think in that broader like right-wing ferment of

to what extent this island is a place that matters for us I mean we had elders ki writing a whole book um basically saying America should care about nothing

else in the world but Taiwan because it is so um important they make all the chips you know this is our way to contain China what have you um and you

know we have other strains that couldn't give two shits that's halfway across the world 20 million people like let's focus on killing you know on overthrowing

Maduro and like shutting down uh drug runners why why are we worried about this country halfway across the world.

So, um you know, we talked earlier about this worry in the leadup to the um South the South Korea summit that Trump was going to kind of just like fold on uh AI

chip exports to China. There was another big concern um that he was going to fold on uh on Taiwan. And you even had articles written from folks in the White

House leaking how concerned they were that he would give something big away.

Now look, it's it's a similar dynamic to Ukraine where like just because America says we don't want you to fight anymore um doesn't mean that the Taiwanese pe

the Taiwanese people have no agency in this. But um you know a a big um like it

this. But um you know a a big um like it is a a big part of the sort of um the balance of the past uh 75 years which

has kept Taiwan free and independent has been this uh this uh commitment from the US and from other countries that um you know it would come it would come to it

would come to Taiwan's defense. So we're

seeing a big kurfuffle now. Um the the new Japanese prime minister said the quiet part moderately more out loud than past Japanese prime ministers said and like like now China banned sushi. Now

China banned like uh uh you know seafood seafood imports from from Japan. So this

is this is still a very hot topic. I

think he kind of gets it that this would not be great for the US. Like the

package that you would need to receive for China in order to sell Taiwan uh you know sell Taiwan up a creek is like I don't even know what would be on the

side of that ledger to make it exciting and acceptable um for uh like the American polity or even a president. Um

but that is I think a big concern.

Another another interesting risk risk factor um is Taiwanese uh domestic politics and how that drives thinking in

Beijing. So we've had um uh two DPP

Beijing. So we've had um uh two DPP presidencies. This is so you two major

presidencies. This is so you two major parties DPP and KMT. KMT um ironically this is Changhai Shak's party um uh you know the Taiwanese the the Taiwanese

nationalists um but they have now you know some like horseshoe thing. They're

the ones who are much more friendly with Beijing and more open to like repro um DPP more leftwing um and more kind of like not pro-independence but uh less

comfortable with like um bringing uh more integration between the mainland and and Taiwan. Um so the the concern

like the the risk factor um is that she gets impatient um uh and doesn't necess and sort of sees the writing on the wall

from an age perspective for him wants to like end it and also sort of sees domestic trend lines in Taiwan going in a way where um it seems less and less

likely that you would have a uh you know more political integration between the island and um uh and the mainland and we're going to you know we've had a relative resurgence um in KMT and like

the better the KMT does the less stressed out um people are that Beijing is going to overre react and they're going to want to play like the nicer long game than the the sharper uh short game when it comes to uh a blockade or

what have you. But um so anyways, everyone should care more about domestic Taiwanese politics. It's probably the

Taiwanese politics. It's probably the most important variable when understanding um uh uh when understanding these dynamics. But um at

the end of the day um my take is that she himself is not the type of guy that just goes willy-nilly invading places.

Like he has been in power for over a decade now. And I think that is enough

decade now. And I think that is enough um you know those are enough data points to show a revealed preference. And for

all of the sort of obnoxiousness that we've seen um with the the Philippines and the Scarboro Sholes, that weird kind of like battle like hand-to-hand battle

in the Himalayas between uh uh China and and India over COVID, um you know, the the increasing like encroachments of Chinese uh fighter jets kind of like

going dancing in and out of the uh uh of Taiwanese airspace. Like this is not

Taiwanese airspace. Like this is not what you saw from Putin before he invaded Ukraine. Before he invaded

invaded Ukraine. Before he invaded Ukraine, he like invaded Georgia. He

like invaded Syria. He invaded Ukraine once um uh and kind of got comfortable with the you with like very dramatic use

of military force. So, um, a world where China doesn't like the results of a domestic Taiwanese election and starts to, you know, flirt with or even do

something that looks like a blockade is, you know, nonzero. I would put in the single digits if we're looking at a five, you know, over a fiveyear span.

But they're not stupid and they understand the the downside risks to doing something like this. um which

could be really really severe for the um uh uh for the Chinese economy and even potentially for for for stabil the stability of the party. So um the riskreward matrix doesn't make a lot of

sense. It didn't make a lot of sense for

sense. It didn't make a lot of sense for Putin uh in February of 2022. So um you know like we'll have to see. This is not the thing that that's keeping me up at night though.

>> Yeah. As you're particularly agi pills like oh it's all TSMC and that's like the the big thing. It turns out there's many factors at play here.

>> You know, this is the Ben Thompson argument that like if America gets too far ahead then they'll want to bomb the fabs or bomb the data centers or something. And like maybe, but hope

something. And like maybe, but hope springs eternal. Uh I think there's uh

springs eternal. Uh I think there's uh at the end of the day like the US and China are each going to make up roughly a quarter of the world's GDP over the o

over the next 10 20 30 years regardless of demographics regardless of whether or not China hits a financial crisis. And

that is sort of a fact. Um, you know, maybe if we live in a real sci-fi world where where um uh AI means that America grows at 10% and China doesn't get it and they're growing at 1%, we can have

another conversation about whether that means they're going to feel like squeezed like Japan felt in, you know, summer and fall of 1941. But that seems a little far-fetched.

>> Yeah. Is there anything else that's happening either in the Chinese AI ecosystem or China at large that you feel like is really underdised in the media today? you know, folks are

media today? you know, folks are starting to get a sense of the robotics ecosystem. I think it's uh it's going to

ecosystem. I think it's uh it's going to be a real challenge for the west to scale up manufacturing on that. I mean,

we've been able to scale up manufacturing of chips. We've been able to do a a really remarkable data center buildout even faster than what China has been able to do on that front. Um, but

that's different than just like giant manufacturing. Um uh and this is a

manufacturing. Um uh and this is a playbook that China uh you know has run a number of times with um with drones with electric vehicles. Um and in so far as

manufacturing robots at scale cheaply is something that is going to unlock a lot of the 20 uh the the um the 21st century then yeah this is something that the

Chinese the the sort of broader Chinese ecosystem may end up having a real um a real legend.

>> Totally. Um, one thing I'm struck by is so much of robotics is, you know, about data collection and it seems like a lot of the stuff is actually already being deployed uh around China, which like gives you a huge advantage to start gathering data. I think everyone in the

gathering data. I think everyone in the US is trying to figure out like the right way to start getting these things out there.

>> I remember hearing this story for um for LLMs, too, right? I mean, like, okay, like scale's gone, like we have Merkore that's like doing its thing, right? But

I um uh it's been an interesting like the the the the sort of relative weight of compute versus data versus talent in

um uh the sort of like AI LLM ecosystem I think has sort of ebbed and flowed over time. Um and you know this is what

over time. Um and you know this is what we were saying with China. Uh this this was like in all the like pre-hat GPT books. Everyone was saying, "Oh, China

books. Everyone was saying, "Oh, China has so much more data than the West.

Like they have all these CCTV cameras.

Like they have these closed ecosystems. Like of course they're going to win." Um

so I think it's I think the sort of the the the factors of what is going to really turn robots on is still an open question. And um you know the the

question. And um you know the the relative strength of the US is still models um when it comes to the the robotics perspective. Now, like having

robotics perspective. Now, like having models and having compute means you get to like on the LLM side means you get to reap all the benefits, but like you need a lot of robots in order to like use

those models and use them at scale in a way that um is is like financially renumative for the folks manufacturing them. And if you're sort of dependent on

them. And if you're sort of dependent on on Chinese robots then TBD I mean you know there's an interesting question of like is that a comfortable steady state for the west to have a lot of Chinese

robots running around America. Um

>> I think people have watched I robot to uh to to be cool with that.

>> Yeah. I mean we I I remember I remember Jacob Hellberg who's now at the um uh uh uh State Department like tweeting Chinese humanoid robots like we have to ban them in all caps like three or four

years ago. So yeah, iRoot's like a thing

years ago. So yeah, iRoot's like a thing that now policy makers are now currently indexing on. So you're on to something,

indexing on. So you're on to something, Jacob.

>> Well, no, I feel like the argument that people like to make that the West has a chance is, you know, well, all these jobs are so much more expensive in the West, like actually from a business model perspective, how much how much are you really going to gain deploying these robots like within China relative to

what you'd gain deploying them in the US. And so the economic prize may end up

US. And so the economic prize may end up being so much larger uh here that you know, enough of an incentive to overcome some of the manufacturing uh disparities. Jordan, it's been a

disparities. Jordan, it's been a fascinating conversation. I always like

fascinating conversation. I always like to end our interviews with some quickfire questions where we stuff some overly broad things into the end. Um,

and so maybe to start, if you could clone yourself and have your clone go just spend a ton of time digging into other areas, what are some like unexplored topics that you're super excited about?

>> Dude, Medicare policy. What I thought we were going to healthare podcast.

>> I would love to have I would love to have your clone on my healthcare podcast. Um, that's a really good

podcast. Um, that's a really good question. I think there are so something

question. I think there are so something I have been spending time on um is thinking more about military applications of AI and the sort of fact

uh is that like the like what can be certain is that uh the sort of armies are not going to um or just militaries and national security establishments

more broadly are not going to be like perfectly fit for the technological moment. um whenever the next big war

moment. um whenever the next big war comes. And sort of if you just um if you

comes. And sort of if you just um if you just look at the amount of change that you've seen from the way um like battles

are fought um in Ukraine from 2022 to today. Um you've had radically different

today. Um you've had radically different um sort of like configurations and relative importance of different um weapon systems and what humans are doing and what drones are doing. And that's

not because we've had like incredible technological breakthroughs over the past three years. It's just that sort of from a doctrinal perspective, from a manufacturing and scaling perspective, from a command and controlling

perspective, there's this enormous amount of technological overhang that um you know, organizations both uh you know, military as well as like

enterprise are going to have to explore in order to like really make the most of what's out there. And you know from a con from an enterprise perspective like you have that creative destruction

happening constantly all the time, right? Um so you reach the frontier of

right? Um so you reach the frontier of what's possible much quicker and without as much death. Um but that's not the case with um with militaries that really sort of like play in the dark for a

while and all of a sudden like the lights are turned on and you're running around being like, "Oh my god, okay."

like now we're now we have 10 times more people than we had a week ago. Um you

you know we have all this energy, this talent, we have all this money. Um and

we have this competitive pressure of like us screwing up and people dying because of it. And that drives an enormous amount of uh change and churn and and innovation both at the

technological as well as the organizational level. So, I've been

organizational level. So, I've been really enjoying kind of reading the um the case histories of this cuz this has happened a lot of other times in history

where all of a sudden um 1914 like the machine gun which had been around for 50 years like was not relevant at the scale that it was um uh you know in the Western Front in World War I and you had

lots of different organizations all trying to adapt and train their people and and do procurement and like you know have all these sort of counter um reactions to this like incredible dance

that uh incredible like death dance of innovation that the the sort of two opposing forces of opposing armies had done. So um exploring that um you know

done. So um exploring that um you know it feels today like we're in this moment that's never been never happened before because none of us have lived through one of this of these like you know

changes not seen uh in a century from a technological perspective that are going to you know ripple through both um corporate commercial societal as well as

military cases. But going back um and

military cases. But going back um and you know living in that like 1940uh4 moment that 1941 moment even this sort of like Gulf War um kind of like

precision revolution moment uh has been a fun intellectual exercise. I have a book I will pitch to your audience. It's

called the social history of the machine gun. Um

gun. Um >> it's 200 pages long written by John Ellis who is this like recluse who like refused to come on my podcast.

Apparently, he hasn't talked to anyone in like 30 years or something. I asked

>> that would be a big get if you >> I know. I asked like four people and they're just like, "Jordan, it's not you, it's him. He's just very contankerous." Um, but uh he um you

contankerous." Um, but uh he um you know, he wrote sort of the story of the machine gun very concisely from a lot of different layers. So from this sort of

different layers. So from this sort of like technological advancement layer of like how you actually came up with this and manufactured it, manufactured it at scale, the acquisition layer of how these com you know who was trying to

sell it, how the buyers were thinking about it and how that changed over time.

The sort of tactical layer of like people having this big thing and like figuring out how to position it. And

then the sort of broader organizational layer of the militaries of what they're going to do with this technology and like how it's going to end up like changing like battle plans and

ultimately grand strategy. Um and uh you know it's an incredible case study because like you'd think they would have gotten it faster but there is just so

much like institutional um crud. And by the way it was not

um crud. And by the way it was not obvious. like there were people who were

obvious. like there were people who were right, but the people who were wrong like weren't total idiots and you could tell a story on the other side of it.

And and it's a very the book is great.

It's just like this very efficient way to sort of put yourself in the world in which this like giant um kind of revolution was going to come um and you

know take away like tens even like hundred thousand hundreds of thousands of lives but the world couldn't see it and you know TBD on like what that is

going to be for robots for AI for drones what have you um but uh it's it's always good to be reminded that like this is not a unique thing that has ever happened in human history like we've we've hit we've hit like technology

shock many times in the past.

>> No, I feel like the Spanish Civil War was like a famous example, right, of everyone looking around seeing what was happening there and then adapting those techniques.

>> Yeah. Looking up. Oh, bombers. Like, oh,

okay. Got to deal with that. Um, yeah.

So, you know, there there's a there's a great example of like uh uh American generals watching the Crimean War um in the 1850s and then the Civil War breaks

out and the ones who had who had been in Crimea were clo like got closer to realizing that they couldn't all line up like it was 18, you know, uh you know, 1807 and Napoleon and we needed to sort

of like spread out because the rifles had gotten that much better. So, you can get glimpses into the future. Um but not all the lessons are applicable. Um and

it's just this incredibly difficult like naughty thing which is impossible but there are some kind of uh principles which I think you can take out of it of just how confident you're going to be

what is noise what is signal like and how you sort of build to make sure you're resilient um into different potential futures that could play out. I

feel like China's one of those topics