This video explores the global AI race, focusing on a groundbreaking release from a small Chinese research organization, Deep Seek. The discussion positions this event as an inflection point reminiscent of the early Internet era, highlighting the acceleration of AI innovation outside traditional tech powerhouses. The guests analyze how Deep Seek’s swift progress and open-source strategy are prompting comparisons to the Space Race and Sputnik, but argue that the true analogy is the Internet: a technological revolution shaped by distributed creativity and minimal regulatory barriers rather than isolated government efforts.

The speakers argue that the Deep Seek release is more significant as a wake-up call for U.S. policymakers than as a direct threat to companies like OpenAI, Nvidia, or Anthropic. They critique regulatory approaches aimed at restricting AI advancements and emphasize that true global competition and diffusion are enabled by open access, powerful research, and accessible infrastructure—factors which outpace attempts at restrictive control. Deep Seek’s permissive licensing, release of reasoning steps, and cost-effective engineering demonstrate the benefits of openness and collaboration, spurring the proliferation of specialized models capable of running on personal devices and fueling a new era of AI-driven applications.

Key takeaways from the video:

Please remember to turn on the CC button to view the subtitles.

Key Vocabularies and Common Phrases:

1. asymptote [ˈæsɪmˌtoʊt] - (noun) - A line or value that a mathematical function gradually approaches but never reaches; in general usage, it means reaching a limit or plateau after which progress is minimal. - Synonyms: (limit, plateau, boundary)

So there has been this view that the traditional one shot LLMs, you know, we're starting to maybe asymptote like GPT4

2. proliferation [prəˌlɪfəˈreɪʃən] - (noun) - A rapid and often excessive spread or increase. - Synonyms: (expansion, spread, multiplication)

And so this is definitely is going to result in a lot of proliferation.

3. distilling [dɪˈstɪlɪŋ] - noun / verb (distilling, to distill) - The process of extracting the essential elements of something, often used in machine learning to refer to teaching a smaller model using the outputs of a larger model. - Synonyms: (refining, extracting, concentrating)

Now we don't know why they didn't do it, but it just turns out that that chain of thought, if you have access to that, it allows you to train smaller models very quickly and very cheaply. And that's called distilling.

4. vertical integration [ˈvɜːrtɪkəl ˌɪntəˌɡreɪʃən] - (noun phrase) - A business strategy where a company controls multiple stages of production, often both producing materials and creating end products. - Synonyms: (consolidation, unification, amalgamation)

But it could be the case that if you're building an app, you need to vertically integrate into the model.

5. zero sum [ˈzɪəroʊ sʌm] - (adjective / phrase) - Describing a situation in which one participant's gain or loss is exactly balanced by the losses or gains of others. - Synonyms: (win-lose, competitive, fixed-pie)

And the problem, zero sum thinking is so dangerous.

6. permissive [pərˈmɪsɪv] - (adjective) - Allowing or characterized by great freedom of behavior; in software, a permissive license allows broad freedom in how the software is used and modified. - Synonyms: (liberal, lenient, unrestricted)

We haven't seen a license this permissive recently for a soda model.

7. capital expenditure [ˈkæpɪtl ɪkˌspɛndɪtʃər] - (noun) - Money spent by an organization to acquire or upgrade physical assets such as buildings or technology. - Synonyms: (investment, outlay, spending)

So on one hand you could be like, oh, we're going to see all of this kind of capital expenditure and all this capital expenditure is going to go into physical infrastructure and therefore we're going to have another, you know, fiber glut equivalent, but a data center glut.

8. glut [ɡlʌt] - (noun) - An excessively abundant supply of something. - Synonyms: (surplus, oversupply, excess)

So on one hand you could be like, oh, we're going to see all of this kind of capital expenditure and all this capital expenditure is going to go into physical infrastructure and therefore we're going to have another, you know, fiber glut equivalent, but a data center glut.

9. inflection point [ɪnˈflɛkʃən pɔɪnt] - (noun phrase) - A moment of significant change or a turning point. - Synonyms: (turning point, watershed, crossroads)

The discussion positions this event as an inflection point reminiscent of the early Internet era, highlighting the acceleration of AI innovation outside traditional tech powerhouses.

10. hyperscaler [ˈhaɪpərˌskeɪlər] - (noun) - A company that can scale computing resources at a massive level, often referring to large cloud service or data center providers. - Synonyms: (cloud giant, scale-up platform, infrastructure provider)

Which is this crazy, which is to my view, this kind of crazy hyperscaler view of the world, which is we need need more compute and more data.

11. arbitrage [ˈɑːrbɪtrɑːʒ] - (noun / verb) - The practice of taking advantage of a price difference between two or more markets; here metaphorically used for leveraging differences in data or expertise. - Synonyms: (exploitation, leveraging, speculation)

If you want to look at a place to arbitrage really smart, educated people, relatively low cost, it's hard to beat China globally.

12. futility [fjʊˈtɪlɪti] - (noun) - Pointlessness or uselessness of an action. - Synonyms: (uselessness, ineffectiveness, fruitlessness)

I do think that there is now a wake up call. Like, I think that the futility of the past four or five years of this kind of stuff is now very, very clear.

13. diffusion [dɪˈfjuːʒən] - (noun) - The spreading of something more widely. - Synonyms: (dispersal, dissemination, distribution)

It's not about war. It's literally just about technology diffusion.

14. endpoints [ˈɛndˌpɔɪnts] - (noun) - Devices or users at the edge of a network, as opposed to central or core infrastructure. - Synonyms: (user devices, terminals, clients)

you're going to end up breaking the problem up to the 7 billion endpoints of the world, which will have vastly more compute than you can ever squeeze into one giant nuclear power data center.

DeepSeek - AI's Sputnik Moment? Steven Sinofsky and Martin Casado Discuss

The lesson is not Sputnik. The lesson is the Internet. This is another step to basically AGI in your pocket. There are always pockets of people innovating. WorldCom and AT&T did not predict the Internet was going to come out of universities. It really is the AI race, just like we went through the space race and we need to win. There's no way this doesn't play out like the Internet.

It's been a busy few weeks. I don't know about you guys, my Twitter feed, all podcasts, everything, deep seat everywhere, maybe unsurprisingly, but what's your TLDR like? Let's just start there in terms of what came out and maybe also your take on why it blew up in the way it did. Because we've seen lots of releases in the last, let's say two years since ChatGPT. Yeah. The quick overview of course is like out of essentially nowhere, a small hedge fund quasi computer science research organization in China releases a whole model. Now, those in the know, it didn't just appear. There's been a year and a half or so of build up and, and they're really good. And they're really good. Nothing was an accident. But it, it appeared to take the whole rest of the world by surprise. And I think there were two big things about it that really caught everybody's attention. One was just like, how did they go from nothing to this thing? And it seems to be, you know, a constant factor of compatibility and capabilities with everybody else. And this number got thrown around that it only cost $5 million. Yeah, six. The number. $6 million. The number is irrelevant. Yeah. But because it turns out they wrote a paper and they said, hey, we innovated in this particular set of things on training, which even here was like, oh, well, that was pretty clever. And, and then because of the weirdness that we don't need to get into of the financial public markets and how this whole thing happened like on a Friday, the whole weekend was like everybody whipping themselves into a frenzy so they could wake up Monday morning and trade away a trillion dollars of market cap, which seems to be a complete overreaction and craziness. But that's not what we're here to talk about. To your point, there's a lot of moving parts here and there's a lot to consider. It's actually a fairly complicated situation. So there has been this view that the traditional one shot LLMs, you know, we're starting to maybe asymptote like GPT4. There hadn't been a big advancement. But then there's this can be this new breath of life, you know, and OpenAI released a reasoning model which is a one, and everybody's very excited about that. And so, you know, in this, you know, grand tapestry we're considering, you had like all this excitement about 01 and how that's going to drive compute costs and Nvidia. And then, you know, R1 comes out and it looks pretty good. And then all of a sudden they're saying, well, you know, if you can do it just as cheap, is this going to actually drive the next wave? And so forth. There's a lot of build up to O1 which lent to the R1 hype. And then I think to your point, people didn't know really what to think about it. And so this, and I agree with you, it was a total market overcorrection.

It's also worth pointing out that in addition to people saying, wow, this is a great model. There's a lot of like theories and rumor around, oh well, maybe this is, you know, the CCP doing a psyop, you know, you know, maybe it costs a lot more. Maybe this is very intentional. It was right by Chinese New Year. There's just a ton of rumors. Maybe we'll do our best to kind of dissect everything going on. Yeah, maybe let's just do that. Because to both of your points, there was a lot here, right? There was the performance element, there was these quotes around costs, there's the China element, there's the virality. It hit number one in the app Store. There's also shipping speed, I think. Martin, you, you shared that they released an image model shortly after and then you mentioned released on a Friday. So there's this huge mixture of just people reacting, some people who know what they're talking about and some people who don't, quite frankly. And so we're like 10 days or so out from this release, which by the way, as both of you said, that was the R1 release. There was the V3 release, what, two months ago or so, which was the base model. So now that we're a little bit further out, what's the signal from the noise?

So maybe I'll give you the lens of know Chinese people are smart. So. So there's one lens, the lens that I hold, which is China has great researchers. Deep Seq has actually released a number of soda models, including V3, which is actually probably a more impressive fetus. It's almost like a chat gpt4.and oh by the way, to create one of these, you know, chain of thought models, these reasoning models, you need to have a model like that, which they had done and we had known about all of the contributions that they've done have been in the public literature somewhere, just nobody really aggregated. So there's a, there's a thought that I hold which is this is a very smart team that has been executing very, very well for a long time in AI. They are some of the top researchers. The fact that they spent $6 million just in the chain of thought is actually not out of whack which what Anthropic has now said they've spent in OpenAI has said that they spent. And so this is a meaningful contribution from a good team in China. And so it means something and we should respond to it. So some of the, what the outcry is warranted. I do think that we respond to it, but I don't think for the reasons a lot of people are saying.

I completely agree with that. And in fact, you also saw the people outside of that team in China sort of piling on to try to make it more intergalactic than it was. I mean, my favorite old friend of mine, Kai Fu Lee, like comes out on, on X and says something about this is why I said two years ago, Chinese engineers are better than American engineers. But to your point about that people were viewing this, that there was reaching some asymptotic level of progress. Yeah. Like the previous base models, you know, like the GPT lineage seem to have asymptoted around GPT4. Right. But I think what's super interesting about that is that asymptote was true if you looked at it through the lens of the function that everybody was optimizing. Totally. Which is this crazy, which is to my view, this kind of crazy hyperscaler view of the world, which is we need need more compute and more data. More compute, more data. And we're just on that loop. And a lot of people from the outside were like, well, you are going to run out of data. And I just, as you know, a microcomputer person was like, well, at some point you're going to end up breaking the problem up to the 7 billion endpoints of the world, which will have vastly more compute than you can ever squeeze into one giant nuclear power data center. And so a lot of what they did was sort of like a step function change, not necessarily improvement, just a change in the trajectory. Yes. And that to me is the part where the hyperscalers needed to take a deep Breath and say, okay, why did we get to where we were? Well, because you were Google and Meta and OpenAI funded by Microsoft, which all had like billions and billions of dollars. So you obviously saw the problem through the lens of capital and data. And of course you had English language data, which there's more of than anybody else, so you could keep going. And so the way I thought of it is we used to, when Microsoft was small, we used to just decide to scale things as like, is it a small problem, a medium problem or a large problem? And I remember at one point we started joking that we lost the ability to understand small and medium problems and solutions. And we only had like large, which was just trivial and then huge and like ginormous. And our default was ginormous, like because we thought, well, we could do it and no one else could. And that's a strategic advantage. And I feel like that's where the AI community in the US particularly, or the west if you will, got just a little carried away. And it was just like every startup that has too much money, the snacks get a little too good.

So I've heard two theories of why they were able to do this. One of them is this constraint one that you've said, I think, which is actually very true, which is we've just been using this blunt instrument of compute and blunt instrument of all data and we just haven't thought about a lot of like engineering under constraints. The second theory I heard, I don't know if it's true, but it's tantalizing. Which is the reason people have said that V3 is so good is actually because it has access to the Chinese Internet as well as the public Internet, which is actually somewhat of an isolated thing. So it's almost a superset. Like we don't really have access to the internal Chinese Internet and we certainly don't train from it as far as I know, which they do. So it could be the case. Both things are true. They could have had a data advantage. They definitely had like the engineering constraint one that you mentioned. Well, even on the data, their starting point is the Chinese Internet per se, those that has much more structure to it. It's a much better training set. And in as much as human annotated data is important here and for chain of thought, you do want experts saying, here's how I would reason about a problem. I mean, this is what this whole chain of thought is. It's basically what are the reasoning steps about something. If you want to look at a place to arbitrage really smart, educated people, relatively low cost, it's hard to beat China globally. Right. And so they definitely have access to a bunch of, you know, potentially highly educated annotated data which is very relevant here. So I, you know, I, I happen to be able to believe it sounds like you are too, that this did not come out of nowhere. It's not a psyop. It's not. It's like this is a great team taking advantage of what it has closely. But there are still things that are very significant about it that are worth talking about. For example, the license is very significant. The fact that they decided to release the reasoning steps is very significant. And so maybe it's worth talking through those too.

Yeah, please. Because I do feel like those are two things that you're not seeing headlines about. Right. You're seeing headlines about all the other things that we just talked about. You said the reasoning traces, those were released which using the comparable 01 were previously not. Right. And then the Open Source license. Let's talk about this. Yeah. So there's two things that are pretty remarkable about Deep SEQ R. One that have implications on adoption. We haven't seen a license this permissive recently for a soda model. It's basically MIT license, which is like one page. You can kind of do anything right is what it's like. Free isn't free beer. For real? Yeah, for real, for real. And like so I mean, listen, I think at A6 and Z we have one of the large portfolio of AI companies both at the model layer and at the app layer. And I will say any company at the app layer is using many models. Like it's not just like one model, like I have yet to see the GPT wrapper. They're all using a lot of models. They do use open source models and licenses really matter. And so this is definitely is going to result in a lot of proliferation. The second thing is so a reasoning model actually thinks through the steps of the problem and it uses those, that chain of reasoning or chain of to come up with deeper answers. You know, when OpenAI released 01, they did not release that chain of thought. Now we don't know why they didn't do it, but it just turns out that that chain of thought, if you have access to that, it allows you to train smaller models very quickly and very cheaply. And that's called distilling. So like the general term of distilling in LLM world means you have a teacher model train a student model who's much smaller. And so it turns out that you can get very, very high quality smaller models by distilling these public models. And the implications that this is just more useful for somebody using R1. But also you get a lot more models that can run on a lot smaller devices. So you just get more proliferation that way. So it's actually a very big step when it comes to the proliferation of this model. Absolutely.

There's this tendency to either be like to peg yourself at oh, it should just be open, but without really defining it, which I think is important in this case and I think because of where they came from and that they don't have a business model. That was part of what was unique about this was it was a hedge fund, like almost like the side project, but not really a side project, just a whole thing. You know, it has this effect that like, well, we're just going to give the whole thing away. And, and the, the rest of the companies are still trying to figure out their, their revenue models, which I would argue was probably premature. And it starts to look a little to me like, hey, let's charge for a web server. And it's like the business of selling, like serving HTTP, not a great business. And I think everybody just got focused on the first breakthrough, which was the LLM, which in a lot of ways if you look back at the Internet, what exactly happened was everybody got very focused on monetizing the first part of the Internet, which was HTML and HTTP, and then along came, I don't know, Microsoft and a bunch of other companies to say that's not the best layer to monetize it. In fact, there might not be any money in that layer and the real money is going to be in shopping and in plane tickets and in television and a bunch of other stuff. And you know, even other companies at&t got wound up trying to monetize even lower layer and but you know, that's not how you're going to get to 7 billion endpoints.

The licensing model really matters because what's going to happen is that there's going to end up being some level of standardization. Now I don't know where in this stack or in what level, but there is going to be some level of standardization and, and the model, the licensing model for the different layers is going to start to matter a lot. And you know, anyone who was around during the Internet remembers the battles over the different GNU 3v3v4. The, you know, open this license. I remember very well. Right, well, you doing a dissertation, you know, you were going to have to Release it. And like it turns out, even your dissertation, which part of it and how you released it was a huge issue because it could make or break a whole approach. And I think that the US industry lost sight of that importance because they got so used to this model of open just means we're a business and we pick and choose. What we throw out there is sort of like evidence that we're an open company. And I think that that view isn't aligned with how technology has just shown to evolve in an era where there's no cost for distribution. Before, when there was a cost for distribution, it turns out all those things were sort of the free model was irrelevant because you still couldn't figure out how to get it to get people to use it. Yeah, totally.

So I do want to take the other side of this because I actually tend to agree with you. And so, like, what you just said is a. It could be the case that the model's the wrong place to focus and everybody thinks there's a lot of value in there and so they're playing all these cute games with openness as opposed to distribution. And that could very well be true. But there's another view which is actually the models really are pretty valuable. And in particular, like, the model itself isn't an app. But it could be the case that if you're building an app, you need to vertically integrate into the model. It could be the case. And therefore, like, if I'm building like, you know, the next version of ChatGPT or, you know, we just had today, you know, Deep Research launch, I'm building that. It could be that the apps actually require you to own the model. And in that case Deep Seek is less relevant because they're not building apps. And then, you know, this means that the impact to the opening AI around topics are not as great. Right. And so I do think that there's this kind of, this fork that we don't know the answer. Fork number one is maybe the models do get commoditized. You need to focus at the app layer, and then the license doesn't matter or the models really matter up the stack, in which case the whole Deep SEQ phenomenon really isn't as impactful an event as people are making it.

So I'm going to build on that just because I want to say you're right both times. No, and it's just the variable is T. Yes. And the variable is time. Because what we saw with and the Internet is such a great example because there's no way this doesn't play out like the Internet. Like it just has to. And what we saw was like, for a while building, building one app seemed like a crazy thing because you had to own Windows and you had to own Office and. But then a new app came along that didn't own any of those and it was Search. Yeah. And so that's why I think a lot of people also, because of age and what they live through, immediately jump to like, oh, these LLMs are going to replace search. But it turns out that's actually going to be really, really hard because there's a lot of things that search does that the models are bad at. Really bad. But what's going to happen is a new app is going to emerge and then when the new app emerges, that's going to get vertically integrated and the Research app is a super good example of that. And then all of a sudden other apps are going to spring up and then that's where you have like, oh, there's Google Maps and there's Search and then there Chrome and then it goes back and eats the things that, that it couldn't do before. And I really feel like that's what's what, what the trajectory we're on now. It's still a matter of where and what integrates. But I think that they're going to like. The thing is, is that the apps that ended up mattering on the Internet literally didn't exist before the Internet. And I think that's what people are losing sight of. Same with mobile, same with mobile. They're all, everybody is complete. There were no social apps. You know, okay, fine, I get it. There was GeoCities and a bunch of other stuff, but caught up on new thing, it's going to replace something. And the problem, zero sum thinking is so dangerous. There's zero sum and also just like the old way and there's the, the, you can think of everything as this spectrum and when, when something new comes along, it, it, the, the whole spectrum gets divided up differently. Which is what Google said when they bought rightly they said, you know, what are people going to do on the Internet? They're going to type stuff. And what are they going to type? Well, they're going to type it, but it's the Internet, so they're going to type it with other people. Okay, so we're actually seeing this happening now, which is what someone will come up with, does something like in a consumer space, let's say like text to image. And then it turns out that like over time people are like, oh, kind of, that's kind of like Canva. Exactly. Like the AI native version of these kind of existing apps. And the reason it's important is because it looks like Canva or you know, it looked like Word or it looked like PowerPoint or looked like Excel, but what's important is that they're actually different. Nothing is going to ever be PowerPoint again. Why? Because PowerPoint, the whole reason for existing was to be able to do like render something that couldn't ever be rendered before. And so all of the whole product, it's 3,000 different formatting commands. Like literally that's not a number I made up. Like it's 3,000 ways to Kern and nudge and squiggle and color and stuff. And like actually it turns out you don't need to do any of that in AI.

So the whole product isn't going to have any of those things. And then it turns out all those things make it really hard to make it multi user. And so then when Google comes along and starts to bundle up their competitor that's going to replace it, they're like totally focused on sharing. So Stephen, let me ask you this. You said something really interesting, which is this, I'm good, I thought I did. Which is this has to pan out like the Internet and you guys have used examples of different companies, the waves, the mobile wave, cloud era, all of those are things we can learn from. But I just want to probe you, is there something different here? And to bring it back to deep seq, I mean that means like this is very important to realize the capabilities of China. It's a very credible player. But I don't think that R1 itself as a standalone is going to have that deep of an impact but on the Internet.

So there's actually these parallels when it comes to actually capital build out that you see in the AI, which is it takes a lot of investment. And there's a special parallel that Marc Andreessen actually remind me of it, which people don't tend to see as well, which is in the early days of the Internet, like, like the mid to late 90s, a lot of investors, a lot of big money think like banks or sovereigns, they wanted exposure to the Internet, but they had no idea how to invest in software companies. Like what are these new software companies? Who are these people? Like they're all private companies, like you know, if you remember this time. So what did all of them do? They all invested in fiber infrastructure the whole time. So we're starting to see this thing Again, right. We see a lot of, you know, banks and big investors, like, listen, we want to build out data centers because they don't know how to invest in startups. Like, we know how to invest in startups. Right. So on one hand you could be like, oh, we're going to see all of this kind of capital expenditure and all this capital expenditure is going to go into physical infrastructure and therefore we're going to have another, you know, fiber glut equivalent, but a data center glut. So the counter to that point where I think is different is at the time of the fiber build out, you kind of had one company which happened to be cooking its numbers where it had a ton of debt to build all of this out. And then when the price of fiber dropped, that company went out of business and like, that caused a huge issue. You have a much better foundation, much, much better foundation for the AI wave. Right. Like the primary investors are the big three. Cloud companies have got hundreds of billions of dollars on the balance sheet. Even if all of this goes away, they'll be fine. Nvidia can take a price dip. Nvidia will be fine. So I don't think we're heading to like the same type of glutton crash that other people have, which is very, you know, appealing to draw parallels to the Internet for that, I don't think is there.

Oh, I am completely with you on that. That part of it is going to look like the amount that Google invested in the, in the early 2000s or the amount that Facebook invested five years later or the, you know, people forget that Microsoft poured, I don't know, 30, $40 billion into Bing. Yeah. And it's still number three or whatever. But it doesn't, it still doesn't matter. Yeah, I would bet, I don't know, this is a fact. I'll bet Meta's spending more money on VR than it is on AI right now. Yeah. You know, just to show you Apple too, right? Apple also, because Apple, whatever is bigger than gargantuan is how much they're spending. And so it really isn't about the investing profile or who is going to win or lose on financials. And I think that is a super important point that you made to really just hammer home, like there's a certainty that nobody's going to come out of this unscathed. But the scathing is not going to be at all what anybody thinks. And then they're not like what it was like the structure that people around like, like, like Wellcome. I believe had $40 billion in debt. Right. I mean it was just kind of one of these things were structurally it was. Oh, and there were companies that we've all forgotten about that went bankrupt over that era. Actually there was one in Seattle whose name I'm forgetting, but that was like 20 billion just poof, gone. Yeah, yeah, yeah. I mean, to your point, these companies have had so much cash on their balance sheet, they've been waiting for a moment to invest in the next generation, which also contributes to their, their willingness to scale up. Yes. As much as they did.

So let's talk about that. In your article you talk about the difference between scale up and scale out and the natural tend early parts of the wave to scale up when really there tends to be a shift towards software basically going to zero cost. So Stephen, what do you mean by that and are we at that kind of change in trajectory? Sure. Well, that now we'll just switch to make sure we're really talking about the technology now, not like the finances. But when you're big, you, you want to double down on being big and so you start building bigger and bigger and bigger computers that, that don't distribute the computation elsewhere. So if you're IBM, you just say the next mainframe is another mainframe that's even bigger. If you're Sun Microsystems, you just keep building bigger and bigger workstations. Then if you're digital equipment, bigger and bigger minicomputers. And then one day, and by the way, all along you're just doing more mips in the acronym sense than the previous maker for less money. And then the microcomputer comes along and not only did they do like fewer mips, they that but they cost nothing and they were going to be gazillions of them. And so you went from an era when IBM would, would lease, you know, 100 or 500 new mainframes in a year and sun might sell you know, 500,000 workstations to like, oh, let's sell 10 million computers in a quarter. And, and I think that that scale out where there's less computing but in many more endpoints is a deep architectural win as well because it gives more people more control over what happens, it reduces the cost. So today if you have to, to, you know, the most expensive mips you can get are in like a nuclear powered data center with like liquid cooling and blah, blah, blah, whereas the mips on my phone are free and readily available for use. And I think that that to me has been a blind spot with the this model developers now, they all do it. I mean, I, you know, run Llama on my Mac and it's, it's just the first time you do it, it's your, your mind is blown and, and then you start to go, well now that's just how it should happen. And then you look at Apple and their strategy, which the execution hasn't been great, but the idea that all these things will just surface as features popping up all over my phone and they're not going to cost anything, my data's not going to go anywhere. That's got to be the way that this evolves. Now, will there be some set of features that are only hyperscale cloud inference? Oh, yeah. But just like most data operations happen in the cloud now, but most databases are still on my device. So I, I'm smiling because like, this is the story from like a microcomputer guy. I'll tell the story from an Internet guy. Right? There's the perfect parallel, like, which is like. Do you remember the switching wars? Oh, yeah, yeah. So for the longest time you had the telephone networks and they were perfect. They would converge in milliseconds, they would never drop anything. You got guaranteed quality of service. And here comes five, nine, that's five nines. And then here comes the Internet. You had none of these things, like convergence was minutes, like it dropped habits all the time you couldn't enforce quality of.

And there was these crazy wars at the time. We're like, why are you doing this Internet stuff? It's silly. We know how to do networking. But what the switching people, the telephone people didn't get was what happens when you actually have a best effort delivery and then how it enabled the endpoints they just didn't get. They needed the value to be in the network and they couldn't think that way. And that really brought kind of the Internet. And I think the exact same thing is playing out. I actually see it a lot of the times. Like people, they look at these models like, oh, they hallucinate, you know, or, you know, oh, they're not correct at these things, but they enable an entirely new set of stuff like Creativ encoding. And it's an entirely white space and it's going to grow very quickly. And to assume that somehow they don't fit the old model is kind of irrelevant to like where it's going to go. What I do is I just s/qos to, to hallucinate. What happened was I was going to all these meetings in the 90s with all these pocket protector AT&T people who would just show up and they would yell at Bill gates like, QoS, QoS. And we had to go all look up what QoS was because. Because not only were we not using TCP IP, but the network we were using never worked because it was like a PC based network. And the IBM people Net buoy stuff. Yeah, I am talking to a networking. I should like the ping of death. I should explain. But it was just hilarious because they would just sit and then I literally. This is something I did. They were telling me about QoS, QoS, blah blah, blah. And if they just wouldn't shut up about it, I didn't know what it was. I walked him over to my office and this was like in the winter of 1994. And I'm like, oh, look, here is a video of the Lillehammer Winter Olympics playing on my Mac. Yeah, awesome. And it was like a. And literally it was a postage stamp. And like not. It was like the size of a iPhone icon, but it was. And they were like, well, that's 15 frames a second. I'm like, I know, it's usually like five. And like, where's the audio? I said, well, if I want the audio, I just call up this phone number. I'm your sister system. And then they just laughed at me. And so here we are, of course, all using Netflix on every device all over the world. And I think that they can't understand that these paradigms where the liabilities either don't matter or just become features. And of course that's what gave birth to Cisco. And they just went like, well, this is how we've been doing it. And it all worked. It only works in our crazy, weird universities and in the Defense Department, people who care. And now that's all we use. And I want to tie this back to Deep Seq because the reason we're getting so excited about this is because we've actually seen these types of things before. And we've seen things like Deep Seek come out before. And it's not zero sum. It doesn't replace the old thing, Right. It is a component of the new thing. And the new thing we still haven't even envisioned yet, right? It's like the Internet is just coming right now. And so our excitement is for the new thing to come. And so when I saw Deep Seek, I'm like, amazing. It's like, it's like this is another step to basically AGI in your pocket, right? Like these can run on small models. It shows that we're going forward. My reaction was not, oh shit, I need to like short Nvidia or whatever. I think that's actually the wrong, the wrong answer. Yeah, I know. I mean, I read the like, let's short Nvidia, you know, blog post that flew around that whole weekend and I was like, are you crazy? I'm like, A, Jensen is a genius. B, their company is filled with geniuses. What about, what about the TAM just expanded? Don't you like? Yeah, it's exactly. And so it is super exciting. And this is the scale out step just happened. And so now you could see everybody doubling down. And to your point that you made earlier, that I think is super insightful and really important is like this enabling of specialized models because that's what's going to end up being on your phone and that's what's going to enable the app layer to really exist. And to me, this is all the equivalent of the browser getting JavaScript. Yes, I did. Because once the browser got JavaScript, then all of a sudden you could do anything you needed without going to some standards body or building your own browser. And I think that's where we are right now.

One follow up there is. If you think about how this progresses to date, I feel like the benchmarks have always been like, which model has the most parameters? How's it doing on this coding test that isn't representative necessarily? Like, what device can this fit on? How much does it cost? Do we expect then a different set of benchmarks or things that we're judging these models by? Or should we just be looking at the app layer? Does there need to be some sort of shift that kind of moves us away from bigger, better, as you're saying, scale up and something that represents scale out. Of course, I thought all those benchmarks were just silly to begin with. To me they all seemed like, remember the benchmark we used to do with browsers was like how fast it could finish rendering a whole picture. And so Marc Andreessen invented the image tag in the browser and the neat thing that they did in their implementation was progressively render it. And then what that did is like empower stopwatches all over the world of magazines to, to write who finishes rendering a picture faster. And of course, here we stand today. Like, that's a thing you can measure, even that's a time and it doesn't matter. And so I think that, that those will all go away and we're just very quickly going to get to like, what does it actually do? I do think that the measure that's going to start to really matter will depend on the, the application that people are going after. Like, take this research stuff that just pounds appeared like this week, like everybody's got a research app. Well, it turns out when you're doing research, the metric that matters is truth. And all of a sudden we're back to hallucinate. But now, like, well, you're giving footnote links and you're giving sources because what's really happening under the covers is like, well, it's a little bit less of generative and a little bit more of ir and all of a sudden vector databases and looking things up and reproducing them matters. And so now we're going to get to a point where like, okay, people are going to develop, I think it's going to probably along the lines of imagenet, and they're going to start to generate thousands and thousands of routine tests that are like, is this true?

This is totally an aside, but you reminded me of like a kind of a weird historical errata, which is the fact that Andreessen made the image tag. So in a way he's kind of also the grandfather to like some AI, because Clip, which is an AI model, basically will take an image and describe it the way it does it using the meta tags in image. So he created the metadata to, to do this. I, I, I will say back on the topic of, of, of the images, here's one thing I've, I've noticed working with these companies where these, these models are actually pretty magic by themselves, right? Like, like if you have a big model and you just expose it, people use them. Right. Which is very kind of different than computers. Like you just put the model out there. Yeah, yeah. The thing is, is all the other models catch up very quickly because they distill so well. So like that's not defensible in a way. And so the companies that are defensible that I've seen is they'll put out a model that's very compelling and then once the users are engaging with the model, they find ways to build an app around that that actually is retentive. Right. So it'll start converging on like PowerPoint or like, you know, start addressing. It's more stateful and requires configuration. So that tends to be very defensible. And then the applications that use models, they use lots of models and they do fine tune these models a whole bunch. And so I think that now, I would say the last two years have been like, like the story of the large model, it really has been and they've been magic. Like people use them and they, people really like them. And like, you know, the first time you're in chat GPT you're like, this is amazing. And now I think we're in the era of workflow around models which are stateful, complex systems, right? And, and also many models powering, you know, apps in, in more sophisticated. I. The many models is a great point like to build on that. Like this is what happened when with user interface. And so the whole notion of user interface that IBM put forward was just derived exactly from their, their green screens and their 3270s and they, they made, they made a shelf of rules on how 40 inches, 40 characters of like exactly how the UI should be. And this is the F10 button and this is the whatever. And then like it turns out that people were building all sorts of UI frameworks actually looks exactly like the browser today where there's a zillion frameworks on the, on the endpoint, you pick and choose, you do what you want to do. You can invent a new calendar dropdown if you want or not waste your time. It's really up to you. And I do think that that aspect of creativity is extremely important to applications. And then one other just to be kind of dull and boring that I think is going to start to matter a lot for apps to be differentiable and to also to use MBA have a moat which is that they're the apps are going to, to also embrace the enterprise and, and for better or worse, one of the lessons that we keep learning is, you know, if you want to get adoption in the enterprise, you're going to have to do a bunch of work like to turn off parts of your app or to filter parts of your app or to disable it or whatever it is. And I think the smartest entrepreneurs are going to recognize that like the need for like sign on, single sign on at the beginning to our back and SSL every time, every single time like it turns into. And because it turns out that's also a great way to price but it's not super hard. And I think that there's so much dumb stuff has been done about AI and alignment and censorship and whose point of view is it and all this other stuff that there's now a whole industry that just wants to show up and tell you all the things that they don't want out of AI. And the smartest entrepreneurs are going to actually get ahead of that and they'll be there to sell because it turns out that is actually enormously sticky in the enterprise. And I think that we're going to see the smart productivity tools embrace that immediately. And it could be even at the most granular level of like turn it off for these users or whatever. Well, we had Scott Belsky at Speedrun recently and to your point, he talked about Adobe and someone said, well you have all these licensed images, right, for Firefly, like do consumers really care about that? And he was like, honestly, not really. But you know who does care? It's the enterprise, right? So to your point, those are two different modalities and founders are going to have to figure that out. But I do want to touch on, you know, a lot of people are talking about Deep Seq as this Sputnik moment and that can be viewed in the lens of like geopolitics, us, China, but also if you think about Sputnik, that wouldn't have been a moment if Kennedy didn't do his moon landing speech, if we didn't actually get there. So in other words, if changes would weren't made. And so let's say you're in a boardroom, you're an advisor. I don't want to talk to the board, I want to talk to the US government, right?

And so like for me, actually the biggest aha of Deep Seq is nothing we've talked about right now. The biggest aha of Deep seek is how blind our policies. Our policies have been around AI, right? They've been so wrong headed. So our previous policies around AI have been we can't open source because it'll enable China, we've got to limit our big labs, we've got to put all of this regulation on top of it and the reason is for safety and all this other stuff, export controls, all the export controls. So we can't enable other countries, we can put export controls on chips, we've talked about putting export controls on software, weight limits, all of this other stuff. That was our entire policy. And for me the biggest, biggest takeaway, the whole Deep Seq thing is that's the wrong way to do policy. China has got a lot of very smart people, they're incredibly capable, they're great researchers, they can build stuff as well as we can and they can open source it. We did not enable them, they did this even with export controls on chips, right? So there's basically all of our activity has been for naught and what we should be doing is funding and investing in our research labs and we should be going as Fast as we can. And it really is the AI race, just like we went through the space race and we need to win and we have everything that we need to win. The only thing in our way is our own regulatory.

And just to, just to build on that like the lesson is not Sputnik. The lesson is the Internet. Internet. Because what we learned from the Internet, which Al Gore famously claimed to have invented the Internet, but what he really did was invent the regulation that allowed the Internet to flourish. And they could have looked at the Internet and said, oh my God, this is a Sputnik moment, and then tried to turn it into what AT&T and WorldCom wanted. And they were there lobbying, trying to make that happen. They absolutely, and frankly AOL wanted it to happen that way too. And so they ignored that and they went with what made the Internet strong to begin with. And so what gave us this deep seat moment was the strength of the worldwide technology community. And so as much as people want to own it and be the singular provider, it's not going to work. The biggest difference, not to over analyze the analogy, I think it's a Sputnik moment in the sense that it's a wake up call for half the world. It isn't a geopolitical wake up call. It's not about war. It's literally just about technology diffusion. And we've had so many misfires since then. I mean, we had the whole encryption war where we tried to put export controls on encryption and all this. And although people thought we were being silly as an industry, when many of us would champion this. Well, you can't. It's like outlying math. Math. It turns out it is out that long math. And the fact that it used those chips. Well, the world's economy as we've seen, is very, very hard to put export controls on things. Remember when we were going to export control PlayStations?

Oh yeah, no, Xbox. Like the government came to. Like actually 2048 bit encryption and email. Yes. Because people came to and well, we can't have bad actors, that's their favorite phrase, bad actors encrypting their email. I'm like, what? Well, they're just going to encrypt the attachment themselves and then like there's nothing we can do about that for sure. But, but in, in this case, we've, we've, we've actually put export controls on GPUs before. I mean like a perfect analog. We were like, oh listen, you can do weapon simulation on these things. Like a PlayStation was the first to actually use the SGI, if you remember that we're going to export control that like, we can't let that into like Saddam Hussein's hands. Like the whole thing total failure. Because it just turns out global markets are global markets. And like we're much, much better in investing, which at the time we did, did in our own infrastructure. We did a great job of that. And I think it was great analogy with the Internet and with algorithm. We should be doing exactly that again. And some politician needs to stand up and be the Al Gore of this moment. And I think that we will get that. So I do think that there is now a wake up call. Like, I think that the futility of the past four or five years of this kind of stuff is now very, very clear. And I mean that even more broadly than you were saying. I mean like the, the, the people who wanted to control this technology at this very granular level in all these think tanks and institutes that were all aligned, I mean the number of books written, the number of academic departments started, the number of assaults on technology companies to align, I mean, you know, there were whole, whole meetings in Switzerland about aligning, you know, with the world leaders, you know, like we should. That's just not how anything evolves. And if the biggest lesson for computing starting in 1981 with the IBM PC or frankly 1977 with the Apple has been the creativity at the edge and just enabling that. And I think the problem that the regulators had was they had never faced regulating a connected world before. The other lesson from Deepseek is just, okay, the world is already connected. The world is already native in all of this stuff. So now the amount of, of actual calendar time it takes for something to diffuse is zero. Deep seat, I think is, was the number I saw this morning is like 35% of the DAUs of OpenAI and that's a giant spike because just all the same people are just trying it out because there's no friction, it takes no time. And so it's so unbelievably exciting to be part of what's going on right now and we just don't need to, to throw water on it and be party poopers.

I personally don't think this is a Crisis moment for OpenAI or Anthropic. I think apps are hard to build. I think that right now the apps that they put out are very complex. They actually know their users, they have very specific use cases. And so I mean, I think for them it's a bit of a wake up call. That they can't slouch and they got to move very quickly. But I'm still very, very bullish on our labs and I think they can stay ahead too. So again there's this view of, of deep seek as a Crisis moment for Nvidia, a crisis moment for OpenAI and Anthropic. I don't buy any of that. I think it's something it's more of like a wake up call for the regulatory environment. And then listen, we should all acknowledge that, listen, there's going to be global competition. We need to stay ahead. What we should see now the right reaction from all of these frontier folks is they should all just start be building apps because the best feedback loop to build a great platform for other people to use is to be building apps. And there's this whole concentrated conversation over competing with your partners or whatever. Our industry is coopetition through and through. It's Andy Grove's lesson. So just everybody should be prepared for these big players to compete with you. But history has shown that's no surefire success. If Tam agrees is 10x there's just a lot of room for a lot of folks. Yeah, I mean people like Microsoft spent 10 plus years like a distant number three in the applications business and it was a platform shift that all the other players ignored that caused it to win. And so I think that the TAM is going to be 100x. It's going to be every endpoint. The revenue is going to come from the apps side of it and then there'll be a developer side of it. It'll just be a different pricing model for different sets of scenarios but it's going to be there. So we're just, just everything is rising right now since it is this positive sum growing world.

Do you have any thoughts just real quick on the fact that this came from like an algorithmic hedge fund, like a quant. Is that any different to your expectation or does that actually signal that more can participate? It's a good reminder that there are always pockets of people innovating. And WorldCom and AT&T did not predict the Internet was going to come out of universities. Like they did not think that a physics lab in Switzerland was going to invent the protocols that become foundational. And they also didn't expect a failed corporate lab to develop TCP IP that became the standard. I mean it's not just any old, it wasn't like the IBM lab, it was like literally a lab that they'd all but shut down because it failed just down the street at parc. And so. So remember, like, SRI was involved these places that you don't even think about anymore, right? And so, like, all of this is gonna. Like, most of this isn't going to be even in any history that's written in five years. And I think that that is the excitement, Sam.

ARTIFICIAL INTELLIGENCE, INNOVATION, GLOBAL, DEEP SEEK, OPEN SOURCE, AI POLICY, A16Z