a16z's Martin Casado: China dominates 50% of global robotics deployments — America needs to act now

Let's first tell you about graphite. dev code review for the age of AI. Graphite helps teams on GitHub ship higher quality software faster. You can get started for free. And our next guest is uh Martin Casado from Andre Horowitz. Uh we have a bunch of breaking news. We're very excited to talk to you. How are you doing?

Welcome, Martin. Good to see you. Um can we adjust the levels? Make sure that we're hearing him. Okay. Uh how is your day? Just tell us how you're doing. I'm doing very well, thank you. It's actually pretty chill morning cuz I had a couple meetings canceled which is kind of the best gift you can give me. There we go.

Paradise. Cancelled meetings. Hop on TVPN. What more could you want in the Silicon Valley life? Exactly. How's uh how's your how's your Q4 been? Uh pretty mellow. Uh most of my companies most of my companies are still in Q3 because we shift by a month. But pretty good. Thank you. There you go. Wait, why why a month?

Is that is that standard now? Is this something it is? Well, it is very standard.

I mean listen so I did you know enterprise sales for my company for 10 years and I don't remember a Christmas or a New Year's where I could be home because uh we were like deep enterprise sales so 50% of the year came in in the last month and so everybody realized that you know if you're pushing to the end of the quarter it's probably better to do it when people are still working and so most will use uh the end of January and so that's the majority of the companies I work with by a month.

Yeah. When how often are you uh investing in a company and and they haven't made the shift yet? Are you going that early? Honestly, if they don't, I I recommend them to them. Yeah, I imagine that's a first. You like you like the holidays at the end of the year. You like spending time with your family.

I mean, I would say 2008 to 2012 of my life. Maybe 2016 was me on the phone like in a beach somewhere trying to close a deal. So, and it paid off and here you are uh leading the infrastructure practice.

Uh, I want to get to robotics in China, but uh, first I'd love just an update on uh, how you're thinking about infrastructure at Andre. You have 1. 25 billion under management for that fund. And infrastructure, I mean, we're seeing deals for energy. Sam Alman's doing stuff with SKH and HBM.

There's, you know, the the Neoclouds, there's the hyperscalers. There's so much in infrastructure. And then there's also just like dev tools, which I think could maybe be in there. So, how do you define it? What's interesting? Where are the key themes these days?

Yeah, so we do primarily software infrastructure and things that directly relate to software infrastructure. So everything from like they say chips to AI models to to your point dev tools to the traditional comput, network and storage.

Um, a very interesting thing about this AI wave is like it's very hard to distinguish between like a like a deep infrastructure model company and like say an application company. And so we're having kind of a renaissance in the field where everything has to be redone.

like you have to redo networking, you have to redo compute, and you have to like build new dev tools. And so, you know, we're just trying to keep up with all of this change that is coming from from AI. How are you thinking about uh competition from the hyperscalers when you're looking at a startup?

We talked to a bunch of founders who they're going up against AWS and Andy Jassey's not asleep at the wheel, but at the same time there's this meme of like, you know, a VC would always tell you what if Google does this and then we have, you know, Cloudflare and like, you know, GitLab and there's like a million companies that we could name off that are, you know, decacorns, public companies that successfully went up against the the, you know, the the public clouds and won or at least saw, you know, massive businesses built.

So, is there is it all about the founder? Is it about the particular market or the technology or lock in or being programmer record or being a you know some sort of system of record like how are you thinking about it?

Um so listen I've been an infrastructure investor for 10 years and um uh listen in that entire time the public cloud has been incredibly dominant right and so I would say after reinvent which is the AWS show I have to like move into therapy mode like I've got actually I'm actually on the other side of this I'm never the one that actually worries about the public clouds I'm always telling like the founders and I I have this line my my line is is like can you name one company that AWS has put out of business one just one good point like they literally have copied everything.

They've hosted everything and like and so the reality is is that if you have an entire if if the market is large enough and you have an entire company focused on it, you're going to win and that's just the reality.

And AI in particular is a new behavior which you know you know companies like incumbent companies are terrible at new behavior. So like this is definitely an advantage to startups like it's like a new buyer. It's a new use case. It's a new way of thinking about things.

So they don't have the incumbent institutional momentum. And so I think that like you know we're free of of a lot of the trappings of the cloud and I think the future's very bright for startups. So I I I don't spend a lot of time thinking about the incumbents.

That said, I have to say one thing is listen anthropic and open AI are are startups really in the classic sense. Um and they are incredibly dominant and so I would say I would put them in a different class. There is there is a class of these new types of incumbents. Think like Stripe.

Y Figma where the founders are still there. They're still hungry. They're executing very well. I think those are much more dangerous than you know Google or Microsoft. Yeah.

I was talking to a founder who was uh going up against an AWS point solution that was rolled into stuff and I was like you you are going up against Angie Jasse but you're really going up against the particular product manager who's on that particular product and that product manager might not be taking calls on a beach in on December 25th, you know, like they might be taking a vacation and so if you're going to outwork them, it's possible you could beat them.

Totally. And and like this is a company that, you know, is famous for bragging about two pizza box teams, right? Okay. Yeah. So fine, you've got, you know, the 12 people on this product that's dying on the vine. Like if you can't beat that, you kind of deserve to lose. Yeah. Yeah. Yeah. It's a great point.

Um well, I I'd love to go into uh robotics in China. Uh you recently published a Wall Street Journal op-ed, which I love. I read the journal every day. Before we dive in there, one question around uh when you're evaluating uh how you think about uh how you think about revenue ramps specifically.

You know, there's always this question if a company's ramping revenue just just uh at at an ungodly uh rate. There's a question of like how how durable that revenue is.

Is it is is it going to be possible for another company to come into the category and sort of pull off the same type of um pull off the same type of growth?

What are you evaluating for these software infrastructure companies on whether they have actually done the hard work to figure out develop something that's unique and proprietary or they're just capital or they're just like a we talk about it's like a a barnacle on the side of a whale like opener.

We're into the barnacle economy. We love the barnacle economy where you just naturally grow at the growth rate of anthropic and you're like it's a miracle. Yeah. Yeah.

Um I mean listen so there's always this concern of easy come easy go which is kind of like what you're asking about like you know you got this revenue so quickly how hard could it be?

Um but you know listen we're 3 years into this and we've got a lot of data points that this stuff is actually pretty durable right I mean midjourney bootstrap to actual scale and it still is very much a leader in image um you know open AAI um roared into dominance um in text and it's still the leader anthropic roared into this I mean cursor roared into this and I think that these revenue ramps are much more of an indication of the size and the growth of the market which I mean these things have just been so large and growing so fast and of course like the leader that becomes the brand monopoly will follow brand effects and then people will know the name and then they'll go ahead and adopt it.

So I think that you know like there is no indication just because the revenue ramp was quick that it's not durable. That said I don't know of any endemic defensibility in AI. I think companies have to figure out traditional modes and there's a lot of traditional modes we know about, right?

Integration modes, two-sided marketplace, brand modes, like whatever the moat is, they've got to get to that. So, when we evaluate the company just because it grew quickly, we're not like, oh, easy come, easy go, right? I mean, that's kind of like intellectually lazy.

On the other hand, we do look for like what actually is the mode, right?

If you're basically just, you know, reselling TPM, you know, tokens per minute from Anthropic, that's probably not as durable as if you've got an a workflow and a bunch of users that use it and, you know, they save configuration in this and you've got good retention numbers, etc.

So, we do kind of full-fledged investment like we always do, but I don't think it's correct to just say just because the growth is high, it must not be durable. How quickly can a brand moat emerge in in infrastructure? Yeah, I mean it's a good question.

I mean the last time I remember brand modes like this was the internet.

I think when markets grow incredibly quickly the like education is always kind of um out of date and so you default to brand like remember Netscape was largely a consumer phenomenon and very quickly everybody was talking about Netscape um very early on and we're seeing a very similar thing.

So I think it's been probably 25 years since we've seen brand effects dominate adoption especially for highly technical projects. I think the last time was probably the internet and then before that like maybe the PC like Windows 95 again was very much kind of a brand name.

And so I think it's more again an indication the market's very large. It's growing very fast. There's so much confusion in the market so that a few leaders end up becoming brand monopolies and then the adoption is literally because people can't keep up. So they just kind of adopt the one that they know about. Yeah.

There's speaking of brand, there's actually a comment in the chat. A6Z blows my mind. How can they be a venture fund and a brand at the same time?

And I feel like there's like like Andre really is focused on brand in many ways both the Andrees brand but also partnering with brands on the consumer side or or uh you know uh internet consumer side. Um but do you have do you have thoughts or advice for infrastructure founders on brand?

Is there is there is there an equivalent of like make sure you have your.

com or make sure that you have you know a specific brand package like do you spend any time thinking on that or is is that at all relevant or is it just something that you know kind of check the box make sure it's converting customers and think about it in more quantitative terms yeah you know so quite frankly this is a pretty new phenomenon like I am part of the nerdiest part of injury you don't really think about brand it's like not that's not the conversation we have normally it's either talking about like the mechanics of you know inbound bottoms up growth or the mechanics of enterprise sales or something like that.

That said, we've all had to all of us the founders and the investors had to become students of this because like you'll have model companies that'll be consumer companies and those are very much brandled.

You'll have um I mean you know I mentioned cursor before like you know cursor is a dev tools company but it's got so many users like brand effects are at play.

brand is really important and and X now because this stuff is proliferating so quickly has become a primary distribution channel like X has its own way of propagating information that's very very different than going to like traditional press or doing like a press release or writing content marketing or whatever like we're in a very different world and so I would say you know historically I personally while the while the firm has I personally have not been very focused on brand like the opposite but now I've I've had to learn a lot more about it used to just be make the docs dark mode.

Do you think some of the But now, yeah, now we are emerging. Sell the sell sell the thing and be technically correct. I was like that was kind of my Yeah, exactly.

Do you think public SAS is generally oversold or people too bearish or is the is the is the the the conversation we keep coming back to Brett Taylor was on the show talking about how it's easier to change the the your tech stack than your business model.

and he was referencing, you know, the dynamic between, you know, seat based, but that's but that's kind of not as relevant in in public company infrastructure companies that might be disrupted by because because a public I I I imagine that most of the infrastructure competitors, the incumbents are already on consumption based models and so you're not seeing a dramatic shift in the business model of the companies you back.

It's more of a technology shift or am I mischaracterizing that? Uh, okay. Okay.

So I I definitely think we're seeing um a pricing shift from seedbased to usage based and and listen we saw that listen when I was doing my company we saw the shift from perpetual to um to uh to recurring right to SAS recurring right and so we're seeing that type of shift and that's being accelerated by AI it was started by the cloud but it's being accelerated by AI because you know the tokens are so expensive.

I don't think that alone is depositioning anybody right I think that the the SAS companies are very good at adopting that and using that. I really think a lot of the story of AI is actually kind of new capabilities and new use cases and like that's kind of like the untold story.

I mean these things now speak English, right? They develop an emotional connection. They can create something from nothing. Like that's cool. That's net new behavior. That's net new budget. And you can see this in this dramatic growth. Now um of course budget has to shift to cover that, right?

And so I think we are seeing a slowing down of traditional SAS spend as you would expect when you have something super cool and new and whatever. But I don't think that this spells the demise of of the traditional sector. I mean it's just it's actually not even good at things like systems of record yet.

Now over time that will change and maybe it'll change the consumption layer and once you have the consumption layer you can change the back end. But I really think the story we should all be talking about is like what are these net new use cases? Like what is it enabling? Where does that budget come from? Yeah.

Um as opposed to like who is this necessarily going to kill because a zero sum thinking I think is the biggest mistake we all make in this AI wave. Totally. All right. Well, let's switch gears to robotics. Yes, there is a what's the major robotics company out of China? Unitry.

The there there's a unitry for sale on walmart. com for like $20,000. I'm hovering over the buy button. Uh how unamerican am I if I pull the trigger? Uh take me through the landscape. Yeah, buy. I think we're gonna we're maybe gonna try and hack it and do something weird. Um but uh yeah. Yeah.

T take me through the current understanding. Um in general it feels like maybe I'm wildly wrong, but it feels like tension between China and America is calming down almost. I don't know just maybe that's just the news cycle and people have shifted to other geopolitical topics.

But I mean, Poly Market has the the the the the chance that China invades Taiwan by end of 2026 down at 14% from 25% and and and we're doing more trade deals. There's crazy news today, but walk me through how you're understanding the the the geopolitical landscape. So, listen, I'm not an expert on on geopolitics.

I will say today China put export controls on on rare earth metals that are, you know, useful in like things like lithography. So, that's clearly like I would say not a d taunt. I mean that clearly is kind of a more of an aggressive uh poster. Probably this is because Trump is meeting with Xiinping.

Probably this is just kind of for negotiation. I don't know right. But I I would you know again I am not an expert on these things from my perspective which is through the lens of technology like there is not a dant. I think that things tend to be escalating.

Um now that said you know the the uh tactical collaboration between China and the US is the strongest it's ever been. I mean, you know, uh, half the founders I work with are Chinese Chinese, half the teams I work with, half the papers I read, half the models I use, half the models the companies I work with use.

I mean, like, I think from a technical academic standpoint is the strongest I've ever seen in in the industry. And so, in that way, there's very much a positive, but politically there's clearly some jocking and I think we're trying to all understand what that means.

And and and I'll just tell you my my very quick view is like um I don't think anybody wants a generational struggle with China, right? I mean like the like the you know the hope is is that both of us collaborate together and take over the solar system, right?

These are two just phenomenally great countries with a tremendous amount of talent. But there is some level of polyiana in that belief. And so then the question is is to what extent do we want to depend on China? It's not it's not about being aggressive. It's not about being proactive.

It was like, you know, do we want to depend on them for for critical infrastructure? And I I would say my position is the answer is no. We don't want to depend on them. In which case, it's important to understand what they're capable of, what we're capable of, and what we can do to develop local capabilities.

Do you read anything into the the fact that Xiaomi was able to be a phone company that launched a car and Apple was maybe rumored to be working on a car and never got it out? Is is the is that just a market structure thing? I'm interested in understanding like where is China behind? Where are they ahead?

It feels like Tesla was way ahead in electric cars and China's catching up there, but now Tesla's playing catchup in humanoid robots. Um, and I I I'm I'm interested in the landscape of like what each country does best. Yeah. Yeah. Okay. So, I mean, listen, this is going to be very polarizing topic.

So I just want to say up front that I'll send some opinions. These are not rooted in fact. There is no agenda here. It just you know um as far as your first question on on the structure of companies like Xiaomi and the car and Apple. Um you know Asian companies just tend to do more right.

I mean think about the Korean cha balls. There was like for a long time there was five companies that did everything like everything. They did everything from the cone to the car. And so like I don't think that that is like a China versus US thing. I think that's something quite a bit different.

Um, you know, that said, we've seen China be like, you know, very very effective at a number of areas in recent memory, right? I mean, 5G, China basically won, right, which is, you know, critical infrastructure communications, right? I mean, for solar, they've been dramatically good.

And then I wrote this post on robotics where they're very, very good. And so, it's clearly a country with a ton of capability. Um, on the other hand, like listen, like the United States continues to like do two things that I think are are are um pretty distinct from China.

One, they're very good at complex software systems, right?

like like you know not building a model but building kind of like the opening eyes and the anthropics and you know and everything around they're very very good at complex software system and two this is going to sound funny but the US is actually really good at building stuff people like and I and no but I I think that I think there's this I think there's a specific reason for that which is um we're more like the rest of the world than China is like China like caters to a local market and the local market is just different behaviors different buying behaviors different political behaviors different economic behaviors and so like the stuff that they produce is just weirder.

Um, and so in some way I would think like like you know there's there there's two areas where they're just not as strong.

very complex software systems which we historically have just not seen come out especially for B2B but also like software that the rest of us can like use and adopt outside of like these viral consumerry things like about to say yeah they boo right like this consumer thing but I'm not a consumer guy man I'm like there's these cars that have like uh LED front dashboards and karaoke machines inside and they float and they're all very cool but I don't know if the American consumer actually wants that necessar Well, but I think I think you could reduce it to like, you know, listen, people are people and whatever, but like enterprise buyers are very different in China than they are in the United States.

And like, can you name one large uh Chinese enterprise software company? No. Software. Yeah. I mean, it's impossible. I mean, we bite dance get into enterprise in China. Does Alibaba Quen like count or like does Deep Seek count in ter I know Deep Seek's getting inferenced a lot.

I mean, I'm like, well, listen, data brick, snowflake, sales for Microsoft.

like the US enterprise companies and then in China you're like well you're kind of trying to like well it was a consumer company but maybe they do enterprise like there's kind of these weird telco companies that people will mention but sure sure it just you know listen I you know I built a team you know I ran a team in Beijing I ran a team in Shanghai when I was working at VMware after they acquired my company and we sold software in China it's just different it's totally different like the consumption patterns are different how you sell it's different what they expect from the software is different and and and the market is large enough and localized enough that any local company's going to focus on that and if they focus on that they're just like through Darwinistic focus forces of acceptance are not going to build something the rest of the world um really understands and so I do think this is an area that the United States has a tremendous advantage is like you know based on your experience selling software uh enterprise software in China do you think in in our lifetimes there we will go back to American companies trying to sell enterprise software into China or Is that is that ship sail?

Because I I I briefly uh lived in China. I studied abroad there. I was at uh I I worked out of the China accelerator office in Shanghai. And I there was a lot of international companies that were trying to sell into China. This was 2016 era.

And there was a sense that the Chinese market sort of wanted to learn about what was being sold, but they didn't actually want to adopt it. And it was more like how do we learn about this so that we can just build this ourselves and utilize you know our networks locally to just kind of out out compete here. Yeah.

So listen I I think behavior follows business.

I think that there are some systems that are pretty straightforward to reverse engineer like a like a actual hardware system where you can pull out the embedded software and you can look at the chip like like you know like let's say network routers right you know classically you know was uh Huawei copied Cisco um routers early on I think for very complex software it's just you know it's like all the operational like knowhow um is just a very tough thing to copy you can look at the code all you want but what you want to know is what it does when it's actually running and that's just not something like the state space is too charge.

And so, no, I think behavior follows business. I think there's a lot of business between the two um countries. I think they're both commercial countries. They're full of commercial entities. I do think we're going to go back to like having a market between the two. Um but I I do think like the landscape is shifting.

Like I do think like listen like right now, you know, we buy, you know, the majority of you know, our kind of hardware parts for robotics from China. And I would, you know, assume, you know, we sold a lot of software in in China that a lot of the business software like they're adopting from us.

And and um, you know, that, you know, what that looks like may shift over time, but I don't think it's going to go to zero. Do you think humanoid robots are going to start as an enterprise product or consumer product? Are are how confident are you that it will start where you think it will start?

I mean, listen, I'm not an expert. So, can I tell maybe can I just talk about the piece that I wrote very quickly to put this in context because I want to show like the area where I am expert versus not expert. So, like you're a VC. You're an expert in everything. Give yourself some I know politics. Yeah.

Like I know human noise. I know. No, it's not. So, so listen, the area I'm expert is is is infrastructure software, right? I mean, that's what my PhD is in. That's what my company was in.

That's what I you know and so in order to you know know what is impacting enterprise software and software infrastructure we do these market studies right so we'll study like what's going on with energy we'll study what's going on with microchips we'll study what's going on with robotics like these are and we did this study um on robotics and it was in service to our software investing which we do even though like a lot of you know the groups in Horses do hardware investing right like I've sat on the board of hardware companies like we do it but like the purpose of this study was software investing and we We got the result of this and we're like, "Oh, damn.

" Like, China's kicking ass. We're like, "Wow. " They like got the supply chain locked up. They're doing incredibly sophisticated stuff. I think something like 50% of all deployments for the last three years were done in China for the globally.

And so, the purpose of the piece that we wrote was just to say, listen, you know, we do these studies normally. We did tons of studies like this, but like this was so dramatic, like maybe we should kind of make it a little bit more apparent to everybody.

um like you know uh you know how China is doing and so I don't listen I am not an expert on any given robotics market like if it's an enterprise industrial thing I'll have more of a sense to it where like from a humanoid standpoint I think there's huge expert on on that um uh but I will say that like when it comes to China and the United States I mean they're a very very real contender when it comes to you know the full supply chain all the way up to the hardware it's the software where I think you know things start to differentiate Like how do you characterize the opportunity for America America to catch up?

Like should we be more bullish on the hyperscalers, the mature companies, the Teslas of the world uh stepping up and taking it seriously or completely new startups from scratch building solutions and then kind of integrating creating their own ketssu if you will.

So here so here um let's let's talk about the areas where I think that like um uh the US has an advantage right I mean like I mentioned like complex software the the United States does a great job of um when it comes to like building products that people want I think the United States does a great job of um uh and so I think that right now when it comes to you know Chinese hardware I think we have to really look at ourselves and say you know do we want to like start putting like import restrictions and tariffs on those not not to penalize China but to kind of incur local uh capability building for these things and I think that would be an important thing for us to consider as a nation doing and there's another very interesting area which is um you know uh for a lot of AI you want big GPUs right and um so you can do a lot of robotics with smaller microcontrollers but you kind of want the big GPUs for the heavy processing like certainly for AV and it's pretty well known that China doesn't do anything below 9 nanometer very well you know like seven and five they can do it but the yields are pretty low And of course they'll be subsidized by China but they're well behind.

So I think at some level like we should continue our um uh export restrictions onmeme like you know ASML lithography machines like I think we should do that um just because it does slow them down. I think we should you know do import controls um uh you know and tariffs so that like we can do local capability building.

But again I like I have no interest in a generational struggle with China. I just think that we need to understand where we're weak and then we need to um you know put in the controls necessary so that the United States can catch up. And these are the areas I would say.

So I would say it's not it's not you know the big chips. I think we're pretty good about that. Um it's not the software. I think we're pretty good about that. But I think actually a lot of the hardware supply chain I do think that we should put our finger on the scale. Yeah.

To me, the interesting thing seeing uh Unitry robots selling on Walmart. com when everyone in the tech community realized probably starting about a couple years ago that letting DJI flood the American market with cheap drones. Yeah.

killed killed our killed our internal you know they they they were willing to sell products at a loss in order to uh you know subsidize these products in order to to make it impossible for American companies to compete and now we have you know millions of drones here in America that are flying around that we don't own and they're going to continue that and allowing them to do the same thing with humanoids feels like an even bigger national security risk to me and would again just set us back in an entire entirely new crickets.

It's crickets. I haven't heard the implications. Yeah. The implications enormous, right? It's like it's like healthcare. Like it's like economics. It's like industrial capacity. It's defense. I mean, it's like basically as broadbases software if you take the generalized form of robotics.

And so it's incredibly critical that we So we have a ge geopolitical rival and we're going to allow them to put a humansized robot that can presumably do human things like uh wait a movie about this once. Yeah. Yeah. They can do laundry.

If they can do laundry, they can do a lot of other things that maybe we don't want them, you know. So So anyways, I I think this is something that uh I hope I hope it gets more attention before it's too late because they're already selling again on walmart. com.

I I will say one last thing too is like we kind of sometimes hope that oh like listen this is an issue of scale and like you know we want to work with our partners and our allies and so like what about EU can they help out here but like if you actually look at some of their dominant robotics companies like Cukos I think it's a 100-year-old company like that's now owned by China as well and so I mean even on the global footprint of industrial robots China is incredibly dominant and so I think now is the time for us to map it out and figure out what we do to enable local capability building and again the point is not to penalize anybody or to have a lifelong struggle.

The point is to not depend on them and um and to what's happening uh in in India on the robotic side. Obviously, Apple is trying to shift a lot of iPhone production to India. Is that is that a an area that you're I don't I don't have any expertise on that now. What's going on in India? What about open source LLMs?

Um have you changed your thinking at all? I know that you were pro- open source generally. We had the deepseek moment. Now, OpenAI has GPTO OSS. Uh, Meta Lama is sort of, you know, declining in favor, but there might be a new generation of models for Meta. Uh, they're certainly staffing up for it.

And so, uh, do you think that, uh, that market has resolved? Is it calcified in any ways? Is it heating up? Uh, is it something that people should be paying attention to? Yeah. Yeah.

So I you know I generally think um AI is following open source as we've historically seen it which is the closed source models get 80% of the market and open source mops up the remaining you know 20% and like that's happening now.

I think open source is incredibly important to the broader dynamics of markets you know like kind of anti- monopoly stuff. It's good for startups it's good for pricing it's good for academia it's good for research. So I maintain that position. I think the Chinese models are phenomenal. I think they're leaders.

I think we should keep to having the academic partnerships. I feel very strong about that. I think any sort of import or export restrictions on software is stupid. Like it just doesn't really work to begin with. And so my view is software kind of, you know, let it all play out. The US is dramatically ahead.

I would say end to end when it comes to capabilities and capacities for OSS. The Chinese models are fantastic. You know, we use them. It kind of will keep us honest. I think that's great. Where I think that we should be more serious is import and export restrictions on hardware. I mean, we've done them in the past.

We've done them incredibly effectively. I mean, you know, especially when it comes to critical infrastructure, and I don't say this from an economic or even from a job standpoint.

I literally say from a national security standpoint, like, you know, do we want to run Chinese hardware infrastructure for our our internet backbone? Probably not. And if you know what has happened in the last few years with the Telos in China, you can see why this is the case.

And so, I do think having like an aggressive posture when it comes to like what we run nationally, what our capabilities are. Um, so are you not worried about the national security uh risks of running an open-source model from a nearpeer competitor?

There's this idea of the uh the Manurian candidate hiding inside of a the weights of an LLM. It all seems fine when you benchmark it, but once it realizes that it's inside of the, you know, the NSA or something, it phones home. Uh Red Star Linux from North Korea does that.

phones home and there and you know there's there's rumors that Tik Tok and and saying is having our our physical infrastructure whether it's in manufacturing like robotics in a manufacturing context or or even a humanoid in somebody's home having that not be something that is a no I I I get that the hardware but like you're asking specific question I'm asking about software like like if if hardware is like a 10 out of 10 risk where is running an open source model from a nearpeer competitor that might want to sneak something in.

I don't know if technically they can, but but is it is it a one, is it two, a five? How would you think about it? Yeah, super legit question. So, first off, I would say like I'm much more sympathetic to like import restrictions than export restrictions.

I think it's it's just a net good for the rest of the world to use US technology and to have that footprint, right? And so, like when you talk about export restrictions, I think that like when it comes to software, like we should never have any.

When it comes to import restrictions, um I think that you should differentiate between what we use for like the government and critical infrastructure versus like what you have access to academia.

I think for example, if we if we did import restrictions of like the latest greatest models, we would be handicapping our researches in a very significant way. And I I don't think that that would justify whatever gains we might have given the fact that like half of what we get is from China. Yeah.

Like there's also like a free speech like math is should be legal argument here where uh yeah like the government and critical systems and maybe our fortune 500 companies should be very careful with what they run if they're running a Chinese open source model but if you're just as a libertarian you know you could think like yeah if you want to go and run this on your local hardware in the cabin in the woods like that's pretty American in my opinion.

Yeah. Yeah. Yeah. Yeah. I'm I'm I'm very very much about never do kind of export restrictions and import restrictions when it comes to software.

It's strictly listen I used to work for the the intelligence community back in 2001 to 2003 and like this is kind of the height of like concerns around Huawei and so maybe I'm just a little bit more sure sympathetic to arguments of not having um um you know these components in our critical infrastructure that was the height of the concern around Huawei and it took us how many years to actually act on it that makes me concerned about the height it was one of the many one of the many peaks of of that makes me concerned about by by 2045 we'll get around to banning unree in the United States.

Yeah, after they after they get a robot in every every home, we'll finally be like, uh, you have to pay us 10% tax on this. That'll be the resolution. As long as we're making money from it, maybe we'll be fine with this, uh, the result. Anyway, thank you so much for coming on the show. This is fantastic.

Always a good time. Always good to see.