Cursor's $60B Deal, DeepSeek V4 & the Death of the AI Moat | This Week in AI E11

27:54in understanding. And then we drive to just as much a people-oriented solution as a software-oriented solution to come up with the outcome. It's almost a very different kind of business. And that requires people hours, the latest in software and custom software and leveraging the latest in these coding agents. And it requires really deep domain expertise where maybe forward deployed engineers matter less than forward deployed bankers and forward deployed investors that we've pioneered. And like that idea of like forward deployed expertise being maybe someone that's not even technical, but can leverage these other technologies to drive to that business outcome as the future of, you know, this category before, you know, everyone just has a neo firm, AI investment bank and AI hedge fund and all the like.

28:43It does seem, Matan, that if we were going to define what would be the moat in future startups, a way to think about it is this promise to your customers, to your partners, that you will relentlessly iterate and not give up. Because if anybody can build anything or anybody can copy anything at a pretty brisk pace, and AI software is obviously deflationary, cost of things goes down, well, then what's left? what's left is that you will keep iterating and you will quit on me and you won't let your software deprecate and become brittle it's almost the promise of the company to its partner yeah totally totally and i think this also reminds me of uh i think this is a carl leschenbach quote where he said at the end of the day people buy software right it's not just like oh you checked all the boxes now instantly you know we're going to switch to something else it's like But this is why, as you know, we're in a very technical space, you know, software development agents.

29:46One of the reasons we're beating all of our competitors is that we have an absolutely killer enterprise sales team. And we treat them as first class. Like, I think it's also, you know, growing up in the Bay Area, there's a very common fallacy that it's like research and product are like the most important. And then, oh, sales and marketing, you know, if you have the best product, it'll handle itself. And that is just so naive because the reality is your product is the whole journey from the very first time they hear your name till their 10th renewal after a decade of being a happy customer. Obviously, the software is a big part of that.

30:20But the reason why some of these large banks are going with us is because we're going to go sit down with them and deeply understand, you know, the use cases that they have. Not just sit on our high horse, like, oh, we're the best engineers in the world. We built the best product. You know, manifestly, you must use us because, you know, we're the best. And I think that kind of shapes the way we build our team as well, where it's kind of cool where no longer is the boat just, we built this software and it's so hard to build and everyone else is too lazy to build it. So you're going to use this forever.

30:50And it's like any feature we release, a competitor could release within two weeks. So the differentiator is not, you know, our ability to create certain features. is it's now like, what is our DNA? What is our product philosophy that determines what we decide to actually ship? And when, how do we engage with our customers when we do that? How do we help them change their behaviors? And I think it's fun because it's a much more interesting problem. Like before, basically the moat was just, we'll do the boring stuff that you guys are too lazy to do. And we have such a headstart that now you can never do it.

31:23Like that actually leads to shittier software. Because if you just have that headstart, then it's like, whatever, people need hundreds of engineers to catch up. Now it kind of holds us all to a much higher bar. I think it's also like the operational excellence on the runtime side too, right? It's like, you know, you might be able to to vibe code like a HubSpot replacement, but then are you going to also like have a DevOps team that like operates it and make sure it's doing the right thing all the time and then adds new features to it? It's like, why do you want to do that when it's not even your core business?

31:50It just doesn't make any sense. I've had this experience now, Russ, many times where my OpenClaw agent says, hey, I can build you that piece of software. Would you like to replace Slack? Because Slack has some limitations. And I'm like, the first thing I think of is, well, I'm paying$6 ,000 a year for Slack for 25 handles or whatever it is. What is my maintenance? And when I lose something, and it's, wow, Slack's a bargain at$10K a year for my little venture firm. I get more than$10 ,000 worth of value. Now, if it was$1 ,000 a seat per year instead of$250 or whatever it is per year, I would have a much different thought about, well, then maybe it is worth it.

32:35If I was paying $250 ,000, I could have one developer on staff doing that 10 % of the time. So it really is deflationary, but you also have to pick your battles as a business. Then we are really talking about brand and communication and go to market, like you're saying, Matan, if I said to you, what's the cheapest phone you could buy and why aren't you using an HTC phone instead of, I'm assuming we're all on the latest iPhone because we're CEOs in Silicon Valley or in the industry, well, we'd probably all have either the most recent Pixel, most recent Samsung, or most recent Apple, because those brands mean something to us.

33:16It's going to have the best camera it's going to have the best ecosystem it's going to be solved and we're actually spending five times as much as we need to spend conservatively four or five times right my time but yeah and i think and i think what's interesting there is also like we're used to a world where the reason you buy things is because you cannot build it yourself now the reality is you can build anything but do you really want to spend 10 of your time maintaining slack like is that is the jason calacanis court competency that you're really good at maintaining slack probably not and so now there's this laziness calculation which is like okay well now you have some defense where if slack is going to be really predatory on their pricing now you can be like f you i'm going to go build it myself because you're raising the prices significantly but if they're like just below the threshold such that you're too lazy to actually do it yourself then they'll like you know stay there because at the end of the day you probably don't want to be maintaining slack I would imagine.

34:12It is a really interesting thing about the economy. It really speaks to free markets. In a free market, we are now going through an incredible transition, George, that I don't think any of us have seen in our lifetime, which is buy or build is actually a valid question for almost everything. I live on a ranch. When you live on a ranch, one of the things I learned leaving a city, and I lived in New York, LA, and the Bay Area, people come and take your garbage. Okay, great. When you live on a ranch and it's a mile to the front gate, how does the garbage get from your house to the front gate in a garbage can?

34:52And I asked the guys, and they're like, well, we pick it up at the streets. You have to hire a service to pay on the street. So there is another service that I use that I pay to take the garbage bins from the ranch house and the barn to the front of the gate. Now, if that person said to me, I'm 10xing my price, you know what I'd say? Okay, I'm going to drag them there. I'm going to buy a pickup truck and put the garbage bins in the pickup truck, drive it to the gate myself, because that makes no sense. But I am not. If he doubles his price, I hope he's not listening. I'm still going to pay it because I don't want to do that.

35:25It is now really what we're getting at. And this is going to be exacerbated. And it leads to two other really important discussions that I think my guests today for the audience are uniquely qualified to answer, which is in this really important discussion, where does the value, where does this training go? Because agents, as all of us know, and the public does not writ large, are now learning. They're now learning. And what they're learning is the value. So in George banking, what those banks train, your software to do, what you train your software to do on their behalf, that's where the value actually resides, is that knowledge and that continuous learning, which then has made me, big wind up here, start to wonder for startups and for my company, a venture firm, how do I capture that value?

36:17Maybe I need to make SLMs, small language models, VSLMs, verticalized small language models, because I don't want to give that knowledge to OpenAI. I don't want to give it to Anthropic. So how do you think about capturing the value that your agents and your software are discovering along this journey? And do you worry about educating other large language models? So we have a very unique approach to how we kind of try to solve this problem. And it's by building what we call deterministic agents. And to kind of grok that, like just as a small thought experiment, if everyone had an open claw at a 10 ,000 person organization, that organization immediately goes to chaos.

37:03Like it is actually like the worst case scenario if you are the CEO of that institution. And the reason for that is because, you know, everyone has their own preferences. Everyone has every, you know, their own, you know, desire to do things certain ways. Someone might want to spin up a Replic clone. Another person might want might want a Slack clone. And the coordination and the issue of just managing those people and then managing that set of 10 ,000 agents on top of that ends up actually creating pure chaos in the organization. What you're getting at, I think, is the idea that it's more important than ever to manage agents or to actually embed in them the expertise or the firm-specific way of doing things or our specific way of doing things, which is especially important as organizations scale, whether they're, you know, just human, human agent, hybrid, or just agent.

37:53That idea of like managing, yes, you can do that with training and with, you know, encoding your preferences in a system prompt or a set of skills, and then having your open claw or your clawed or your GPT go and explore and exploit and do whatever. But increasingly, what we're seeing in the world's largest institutions is that you don't even want these always on long running agents that are doing arbitrary things. Actually, quite to the contrary, what you want is a defined task of what we call a deterministic agent that goes through a, you know, 10 step or 50 step or 85 step specific process, the firm's way of doing it.

38:32And once you actually have encoded that, maybe there's like skills for one piece, another piece is code or executing code, another piece is writing code. But once you get all the through all of the 85 or 100 steps, that becomes load-bearing institutional-grade software that you can then scale to 1 ,000 to 10 ,000 agents to 10 ,000 employees, et cetera. And that's a different layer on top. It's almost a different, as kind of Matan mentioned earlier, model agnostic, almost task agnostic layer, where you're prompting without any sort of understanding the overarching task, single steps one by one to get to the right institutional outcome.

39:10that's encoded by our forward deployed talent and our go-to-market team, again, kind of per what Matan was saying, is really important and really a differentiator in today's day and age. This is a fascinating picture you're painting, George, for a number of reasons. I think, Matan, if we are putting together the thread that you discussed as well, which is startups and products and services companies are about this forward deployment, then that means these large language models, whether they're frontier or they're open source, are quickly becoming commoditized. And we will not know the difference.

39:51We won't know which one we're using. Just like I couldn't tell you what Seagate hard drive I'm using, or if it's a commodity Taiwanese hard drive, or none of us could tell you if we're using Corsair RAM or what other RAM we're using, it's all been abstracted. The large language model might be the same as the memory and the storage in our computer. And then all the value is created by, wait for it, the humans and the agents, which are most analogous to humans and the harnesses, the skills, the memories that they actually build, huh, Matan? Yeah, I mean, I completely agree with that. And I think it's right because if you look at the downstream users, like in our case, at least, engineers, and I'd imagine it's similar kind of for both these gentlemen as well, is like our downstream users don't care what model it is that they are using.

40:39They care that it gets done quickly, reliably, affordably. Like that is what really matters to them. It is a means to an end. And when it is a means to an end, that is something that, you know, you want to have optimized. Now, my sense and like, you know, being very clear eyed about this, the model providers are going to very aggressively put up a fight to be as monopolistic as possible and as, you know, not amenable to model agnosticism as possible. Because on one hand, you have this collection of, let's say, four companies, OpenAI, Anthropic, Google, and XAI, that all want to have as much pricing power as possible.

41:14And then you have the users of these tools, whether it's developers, bankers, whatever the case may be, where they just want to have the optimum cost and quality and speed for their given task. Companies like us here, our job is kind of to be that intermediary that gives them leverage against the monopolistic tendencies of the model providers if you're locked in on just anthropic they can jack up the prices and you kind of can't do anything about it meanwhile if you use a model agnostic tool and they try to jack up prices you can be like screw you we're just taking all of our usage to someone else and that gives you like uh i think it just allows the free market to do its thing i think they're going to be dynamics where like game theoretically then each model might try to be spiky at different types of tasks.

42:03Like if general performance, you know, they're all loosely the same, then, you know, to have some edge, OpenAI might try to be better at testing. And then Anthropic might try to be better at documentation or reviewing or C++ or Python. And, you know, there's going to be this dynamic that emerges that at the end of the day really benefits businesses and, you know, allows them to avoid this lock-in while getting the frontier of performance here. Russ, if LLMs are going to be commoditized, just like hard drives, bandwidth, or RAM became after two decades of the internet and five decades of personal computers, just grinding those prices down and increasing the value of those commodities, then do you believe that the value will reside in the agentic level that is continuously learning and in the application layer where, you know, founders can really provide customized, bespoke, verticalized value.

43:02Yeah, I think, you know, you look at it as like a stack, right, or a layer cake. And like at the bottom of it is this LLM, right? And I think in these early days, that's where most of the value had accrued to. But as that layer gets commoditized, you know, almost like liquid, you see the value kind of spread all the way up and pool at the application layer. I remember maybe a year, year and a half ago, OpenAI published these projections through 2028 or 2029 or something like that of where they see their revenue kind of concentrating. What you saw over time was you saw the API layer actually shrinking a bit in where they expect revenue to be.

43:45Then the vast majority if it was actually in chat GPT or at the application layer. And so I think that that's kind of the dynamic that would play out in the macro environment as well is that a lot of the value, most of the value actually, you know, pools at that application layer, but there is significant value at every layer below in between the LLM and the application, right? And so for us, like we play, you know, we're kind of like all the undifferentiated deterministic infrastructure that wraps that stochastic core of the LLM. And so there's going to be significant value there. But we think that the developers of applications that build on top of us are going to see the vast majority of that value.

44:24And there is a breaking news story, DeepSeq version 4. This is the open source Chinese model. And this really relates to the discussion we're talking about where value will reside. This frontier model includes version 4 Flash, version 4 Pro variants. So you can go fast. You can go deep. And here it is. DeepSeq V4 Pro costs$3.48 per million output tokens. Quad Opus 4.6 charges 25 million per output token. And I think that's the best analogy here. I'm not sure what 4.7 costs. And so here's your comparison chart of this incredible battle that's going on. I think this is also, you know, talking about the open models is a kind of very relevant here because a lot of these enterprises, many of their use cases, you don't need Opus 4.6 level intelligence to go and write your documentation, right?

45:26There are a lot of these use cases that you could get 10 times faster, 10 times cheaper with an open model, and it'll perform just as well. So, and this is something we're seeing a lot, you know, with these largest enterprises. Yeah. What do you see here when you're looking at this, George? And have you started using open source frontier models yet? Are you having your team keep an eye on them? Are you benchmarking them and making your software headless to see what output comes out? So we've got a model agnostic layer, very similar to kind of exactly what we've been discussing. We train our own models for specific tasks to refine them even more acutely to what our customers are doing.

46:06Quite often, that's based off of cutting edge, open source models, whether it's DeepSeq or any of the other ones, especially given that we are in a highly regulated regime, and can't always use some of these non-American open source AI models. But I think, you know, this is just all pointing to a larger trend in which these models will get all very much commoditized. I think, you know, we'll eventually see small language models, models that are running on device. And that will actually change the game completely for, you know, what will end up happening, you know, where OpenAI and Anthropic end up accruing revenue.

46:45I think it's actually quite likely that, you know, people will always care about having the latest and the greatest model running in the cloud, but you'll also be able to triage tasks to these smaller language models. And so you'll have this orchestration agent that is actually going in using other subsets of quote-unquote cheaper intelligence and maybe local intelligence as well to achieve a task. So I think the cost of all of this will go down with time and the large labs will continue to have to iterate. I think there's an interesting question too that I don't know the answer to, but where even assume that like Anthropic or OpenAI have the best model, the absolute best model.

47:29I think what's tricky is that these kind of foundational labs are moving into the application layer more and more. And so there's this tension between, well, I have the best model and I want to make it available in the API to developers out there, but I'm also playing at the application layer. And like the differentiation and selling a subscription at the application layer is like having the best model. And so do you continue to always provide the best model to all developers out there that maybe can like build something similar and replicate the product, but maybe not have as good of a margin?

48:06Or do you hold back that absolute best model for your own application while providing maybe the second best one to all the developers out there? I don't know how that dynamic will play out. I think the free market will always demand that you put your almost your best model forward as often as you can. At the same time, I think there's a lot of marketing back and forth about mythos and open AI's and release models and how there's security vulnerabilities. As we get closer and closer to artificial superintelligence, I do think you'll get to models that expose security flaws. I think you'll get to artificial financial superintelligence that will expose dislocations in the market or arbitrage opportunities in the market.

48:44Has that started to happen, George? I saw somebody, I just saw on the timeline, Yeah. Somebody talking about they've they've put 10K into a polymarket account and ground it up to 70K. My team will go find this. And I was like, OK, I got to double click on that, because if that's true, you would not post that. If it was true and replicable, nobody in their right mind would post it. They would add a zero and then they would try to add another zero and they go to somebody like me or Bob and say, hey, would you bankroll me for this and split the returns? And I'd be like, F yes, I would. That's like a drop shipping scam.

49:21Like, I think there's a million different accounts where like, you know, people are like, oh, pay for my$10 a month course and I'll teach you how to make a million dollars a month with AI. The hardest part is the expertise. It's kind of the larger theme of what we've been talking about today is how you get that forward deployed expertise for the, you know, the, you know, the investing specific kind of like genius that actually allows people to make money with humans and then encode that into AI. But no, I don't think that those polymarket bots are real. You can double click into them all you'd like.

49:54I have seen instances where very clever investors in the private markets and then some in the public markets have been able to connect more dots over some subset of information or over more information than a human alone could have seen. And they have filed proxy attacks. using AI to find something that they would have missed. They have exited positions. They have invested in assets that they otherwise would not have invested in. And so you are starting to see this AI augmented decision-making becoming table stakes and changing what market really is. But no, I always have this thought experiment that when we get to true ASI, there will be all of a sudden massive movements in the financial markets as a sign that humans have no idea what's happening.

50:40They might be conned by AI into moving money. It almost solves the human sociology problem rather than the fundamental problem, which is what we're seeing solved right now. Super fascinating. In the poker world, online poker, it was basically ruined. Anybody who knows what they're doing will not play online poker now because of HUDs, heads-up displays. People will pop up a HUD. They'll play across 20 tables. Then they'll join the same tournaments. They have VPNs. You'll have 20 different devices joining the same tournament and hoping that two agents acting randomly will sit at the same table. And then in a PLO game, you now know four other cards on the table.

51:23Massive edge. Not massive as in you went from 50-50 versus an opponent to 90-10, but even a 1 % or 2 % advantage over many hands means you will bankrupt the other person. And then you can also study people's like how often they voluntarily put money into the pot, VPIP, I guess it's called, just ruined online poker. Anybody who plays online poker, it's a sucker's game. You're playing against bots, you're paying against bots colluding with each other, but that doesn't help the online, the in-person game where all that is gone. But online trading is kind of in the public markets, the equivalent of online poker, I think, which means the game could get rigged over time.

52:06And if your horizon is, I'm going to own this stock for 10 years, it doesn't affect that person. It doesn't affect a person who goes, I'm long, Tesla, Uber, and Google, because I believe in the self-driving thesis. I'm just going to own those stocks for 10 years. And yeah, what percentage I own of each might change, but I'm just going to bet on the category. Yeah, George? I think you're seeing almost like it's very similar to how Gen Z and gen alpha are post rationalist and a lot of the humor doesn't make sense a lot of the memes don't make sense uh because like you know uh i think you know everything can be overly explained for them i think you're starting to get to like markets which are kind of like post fundamentalist uh is like maybe just a fun thought experiment where it's like it doesn't really matter about like the underlying asset itself or like any of the underlying stuff and it's more about like the meme ability potential or like you know the ability for ai to create you know some level of interest here.

53:02And, and maybe that is one of the defining final moats is a brand. And, you know, that being one of the seven powers is just a funny thing to think about. And I very much agree with your point about like, if you had ASI for finance, it would not be doing any crazy complicated alpha strategies. It would just go and like sociologically scam people or like social engineer its way into a meme stock. It would figure out how do we make something the next meme stock, it would just get in early, make it a meme stock, make a ton of money, you pull the rug out yeah like that that would be what it looks like as opposed to finding some crazy complicated multi-levered arb whatever it's literally back to the the what we just saw on the tv show silicon valley where it's like oh you want to get rid of bugs well humans create the human developers are creating bugs and the software is where the bugs reside let's kill the human developers uh oh we can't do that there's no way to get out of this box okay let's just delete the software it's literally the path of least resistance is go on reddit and x create 100 spam accounts and then pump and dump OpenDoor or whatever companies on the floor that has some reasonable chance of becoming a story stock.

54:07And I think there's still some more nuance and people are still more real in that. But it's a real thought experiment, which is like, hey, I have all these people spending time working on financial models and these crazy, complicated human first workflows. And it's like, well, there may be other ways. And this is more the the outcomes pricing example to make a lot of money with AI. Would you rather go up against Jane Street or like people on Reddit, like probably like random Redditors and, you know, convince them of something. And I asked my team to go find me this quote for the 70K poly market.

54:39They came back with like 17 different stories, which mean fair warning. When those stories come out, if you engage them and you take their course, you're the sucker at the poker table. you're the person who doesn't realize when you go to vegas and it's a saturday and people are playing in their robes and flip-flops at the table that's the mark that's the person who's you know in between the sauna and the pool and they're just going to blow 500 real quick and the other seven people at the table play there for a living and they're playing from the same chip stack they're just in they're colluding against you don't play on saturdays they're playing a tournament or something i don't know what your best bet is but uh the odds are going going to be against you.

55:22All right, we've got a couple of topics we could end with, and I'll just see who has a strong feeling on it. China has blocked Meta's Manus acquisition. This is pretty spicy on a geopolitical basis. We can start talking about that one. OpenAI, they redid their Microsoft deal. So anybody can now host OpenAI's models on their infrastructure. Amazon and Andy Jassy said, oh, that's interesting, and they'll be hosting all the OpenAI models. I I think that was part of the tension there. And now OpenAI is owned, I think, 26%, 27 % by Microsoft, which is incredible. Elon Musk has 0 % ownership in that trial starting this week.

56:01And then finally, the future of the, well, yeah. Let's pick between those two or three. Anybody have something they feel particularly compelled by, George? You seem to be thinking this one through. What's compelling to you as we wrap here? I think the geopolitical implications is obviously very interesting. I think China has a leg up in some ways in their ability to productionize these models at scale in their control right now of open source AI, again, at scale with multiple players there. And I think it makes a lot of sense for them. It's just something that, you know, the U.S. has to come back on.

56:48Like, you know, this is, again, too important of a battle of a technology to lose control of. It's an interesting supposition here. Hey, if China has the best open source models and they can deploy them at scale from a top-down basis, Russ, what does that mean for the frontier space and this huge race towards superintelligence? I think we all agree. We're either at AGI, around AGI, closing in on AGI. So let's make it a two-part question. Where are we at in achieving AGI, Russ? And then with these frontier models, who gets to AGI first at scale? Is it the frontier models? Is it the open source models coming out of China?

57:36It depends on the definition, right? I think everybody has a bit of a different definition. My definition in particular is like we're trying to build a human in silicon. I think that's my definition, right? Can we create ourselves in software? And I think if that's a definition, then I think we're still quite a ways away. Humans can do a lot of pretty amazing things and humans have emotions and things that still kind of uniquely define us, that these models, even architecturally, aren't really set up to mimic or approximate yet. but I think like and I think there are other parts of like the AGI definition around continuous learning and things like that you know if you've listened to stuff that Ilya talks about you know he speaks these kinds of things as well and and so I think it depends on the AGI definition now I think moving forward though in terms of like the race between being between China and and the U.S.

58:42I was listening to the Jensen podcast with Dorkash a few days ago. And I thought a really interesting thing that he talked about was kind of, while we have better chips in the US, China has an advantage on the energy side, right? They have tons and tons of energy. And so you can run more, worse chips, but at a larger scale because you have a bigger kind of energy supply and you can effectively like nullify the advantage, a chip advantage. But then the other thing that Jensen said, he talked about Moore's law, but he talked about how, like, if you look at the improvement, you know, that of these models, like there's a 50x improvement and it's largely attributed in their performance, largely attributed to algorithms and computer science.

59:39That's the main thing he talks about. And so, you know, I think that the race between China and the U.S., if you believe in kind of like Jensen's philosophy of the breakdown and where like the biggest gains come from in terms of progress, it's not at the chip layer, maybe not even at the energy layer. Those are two important components for scaling all of this up. But a lot of it comes at the algorithmic layer. And so I think that the question to be answered, and I don't think I'm close enough to it, maybe as the other gentleman here, but I think the question to be answered is like, are we winning that race?

1:00:16Are we winning the race in terms of algorithms and architectures for these models and the design of these models versus China? Your thoughts, Matan? One thing that's kind of, as someone who's very patriotic, that's frustrating is that because they are so constrained, it has bred a lot of these algorithmic innovations because they have no alternative. Meanwhile, we kind of get the privilege of like, oh, let's just throw more GPUs. Like, let's just, you know, it's kind of a lazy, easy solution. And it's very tempting when you don't have the constraints to just, you know, go for that and just go for scale, like more GPUs, whatever the best GPUs.

1:00:56but I think we should really you know as Americans like the United States we should be disappointed that we have fallen so far behind in open source I think it's like pretty embarrassing and I think in fact this you know as it relates to the Manus acquisition this also shows that we're in the middle of a very serious talent war because it's not like the tech from Manus was that incredible it's really the people there that I think is the thing that's being kind of protected and I think you know it's maybe a wake-up call and we should have a much more burning fire under our asses to make sure that we are at the frontier, not only of the big models that use all the GPUs, but also the resource constrained models that are smaller and making sure we have the algorithmic innovations, making sure it's not like an unwise bet to spend your career doing that in the United States, which right now it kind of is.

1:01:46There's not nearly as much funding to do things like that. I don't know. To me, it gets me very frustrated. Yeah, Jensen said in that podcast as well that a majority of the best AI researchers in the world are actually Chinese. And then in China proper is like 50 % of all AI researchers overall. And so they have a huge advantage on the talent side. I mean, maybe a final thought. I don't know. I think the greatest technologies, the greatest algorithm improvements have always been built in America, and they are American. And I think China is very good at copying those things. But if you look at the actual, like the best researchers and the best talent, it is all American.

1:02:28I think there's maybe like something cultural about that. Maybe it's because of the superfluous amount of GPUs. But maybe I'm more bullish on like American AI supremacy. I just think, you know, we just have to stop it from getting copied. I'm bullish, but we got to hold ourselves accountable. Like the fact that DeepSeek keeps releasing these models that are smaller and really good. And yes, of course, they're distilling and they're, you know, breaking some of those rules. But like, I'm very bullish. I just want to hold ourselves to a really high standard so that you can't even say, oh, but they're better at this or that.

1:03:02Like we should be across the board, like dominated. You want to win by a mile, not an inch. Yeah, and there's no reason to not win the open source race, except that capitalism is such a strong driver that if you're a developer who says, you know what, I believe in open source philosophically, and I would prefer to work on that. But Meta just made me, you know, and Claude just got into a match, and I've got a$10,$20 million RSU, you know, package here. I'm not working on open source for the next four years. I'm working on a proprietary model. And even Meta, which was behind in doing, you know, Lama open source, they just moved to a proprietary model for a reason.

1:03:46They need to give massive grants to those folks. So there is something happening here where capitalism is racing towards the cliff, perhaps. And I guess this could be our final thought is, can OpenAI's spending, and listen, I understand they're a partner, Russ, but this is a discussion that the CFO is having publicly with the CEO, apparently, at conferences, and spending hundreds of billions of dollars on infrastructure is very different than spending tens of billions. There's a very easy way to spend billions to tens of billions of dollars. If you start spending hundreds of billions of dollars, and you make those kind of commitments like OpenA has with Oracle, people are now starting to wonder, hey, the revenue is not growing at the pace the spend is growing.

1:04:36Is this, as we would say in poker, a risk of ruin? If you put your entire net worth on the table and you're betting big pots, even if you play perfectly, you could get unlucky in a hand and your aces get cracked. And now you don't have the ability to rebuy in. So, Matan, you seem to have a perspective here. Sure. Give me your perspective on this race to the cliff or putting your whole net worth on the table and playing high stakes. I think it's just the name of the game for this era of software where everyone is going through hyper growth. And right now we're negotiating some compute deals with some of the hyperscalers and some of the kind of Neo clouds.

1:05:16The thing that's really difficult is we have to make a call now about what we want at the end of 2027. And like for us, we have been doubling every month. Like that is really, really difficult to make that commitment like 16 months in advance because the variance is so, so significantly high when you're in such hyper growth. And that's for us, whatever. Our scale for now is, you know, much smaller than that of OpenAI. They have similar trajectories. And so the error bars there are pretty massive. And, you know, to your point, if you overshoot, you might be literally out of business. if you undershoot, you might be in a situation that Anthropic is now in, where now they seem like idiots because they didn't make the commitments up front.

1:06:00So it's a very, very risky game. I don't know what the, like when you're dealing with these year-long supply chains or multi-year supply chains, I'm not sure what the optimal way to... Well, and Crusoe Cloud reasonably says, hey, we've got a six-year lifespan on this device and on this hardware profile. Will you commit to you know, four years, five years. And then how much do you want? Because then they've got to go to Jensen and then Jensen says, okay, well, it's first come first serve here. So who's going to give me the money first or put the deposit down first gets delivery first. And this is incredibly high stakes for us.

1:06:34Yeah. Yeah. It's crazy. I'm glad that, uh, I'm not having to make these commitments like you are, Mithon. It's a, it's pretty, pretty nuts. George, Any final thoughts here on the high fakes game? Luckily, you are on the application layer. You don't need, I think, to build your own Colossus to provide massive value. You just need to rent some space reasonably from AWS, Azure, Google Cloud, et cetera. You're not standing up your own hardware, I would assume. No, but I do think that from an interesting thought experiment, most of the pricing of every piece of software or every type of knowledge work into the future seems to be kind of moving towards renting GPU hours.

1:07:18So, you know, that seems like the fundamental unit of what people will be spending money on. And, you know, you can underwrite things in GPU hours or the US dollar in 10 years, but it seems like those are like the fundamental kind of units of currency. Let alone if the new CEO of Apple, who I'm ready to fall in love with, I'll just put it out there right now because I wasn't even in like with Tim Cook because just not enough innovation for me. But I kind of feel like this new CEO, as an engineer, as somebody who worked in hardware, if he drops M5 Mac Studios on everybody's head with a terabyte of RAM and, you know, M6 and M7 come out on a cadence and everybody's laptop all of a sudden has 128 gigs of RAM instead of 16 or 32, we're going to see an amount of compute flood the market, not just in the data centers, but on the desktop.

1:08:13And what does it mean if there were a billion Mac studios in the world with all this extra capacity? I think we're going to flood the market with compute, with memory, just like we did in the days of the internet with storage and bandwidth. We overbuilt, we flooded the market, and then let alone Elon going to space and making unlimited data centers in a permissionless way. And if he, you know, has, as he announced his own fab, what's to stop him from just every day putting up a starship with X amount of gigawatts of compute? There's no constraint if it's in space. There's no constraint. It's just how fast I can make these satellites and orbit them.

1:09:02This is a brave new world. Yeah. I think the tricky thing is you need to, you need the applications, right? You need to justify the CapEx with utilization. And so the question becomes, what is going to drive utilization of it? I think a nice backstop there that maybe I harp on too much, but it's like there are so many problems latent in society, whether it's software, solvable problems, or healthcare, or just industrial. I think we only get to that kind of concern of what are the applications if we suddenly feel like we have no problems to be solving. And my sense is that generally the free market allocates resources accordingly to whatever are the most kind of pernicious problems of the time.

1:09:52So I'm more like even, you know, Jason describing that to me, it like puts a smile on my face thinking about all the things that we've probably assumed are like just realities of the world. we don't even think of them as problems but there are so many things that you know with all of this compute um and with all this innovation that we're going to be able to solve which i think is just it is like i i always think about um i don't know if you guys you know growing up i would always see the memes that are like born too late to explore you know the oceans are born too early to explore i think this is the best time like this is by far the exact moment that is incredible to be around because all of these problems are like being solved at such an exponential rate that's a great moment to end on.

1:10:34Stay positive. I think the P-Doom here is like, if we were to P-Doom this panel, I'm at like P-Doom, I'm like five. I have like 5 % chance this goes awry. I think it's like 95 % chance life's going to be awesome. Anybody else want to share their P-Doom after this epic episode of This Week in AI? Hey, where's your P-Doom, Matan? My P-Doom is zero because it is in our hands and we will not allow it to happen yeah russ you got a george you got a p doom i'm assuming you understand the concept of like what percentage doom you are i believe that's a simple way to think about it yeah i think at the current state of things yeah i'm like five or less love it george i'll put it there at 10 i'll be i'll be i'll still i'll still be the highest i hope i hope it's in our hands i think uh i think these models are more powerful and and deserve empathy uh but yeah i I still am a techno optimist and I'm still on acceleration.

1:11:28If you're at 10, by the way, that's the lowest score across Anthrop, this entire organization. I think literally the guy who works the dishwasher in the cafeterias, he's at P-Doom 12 % just by overhearing these conversations constantly. He's just like, yeah, I don't know how this is going to go for me. Like as a dishwasher here or working in the kitchen, I think I'm P-Doom 12. All right, everybody, another amazing episode. Russ, where can people learn more and who are you hiring for? Hiring across the board, go-to-market, engineering, pretty much everything as we start to scale up a bit. You can find out more on X, x.com forward slash live kit, and also on GitHub, github.com forward slash live kit.

1:12:13Okay. Matan, how can people find out more about your great company and who you're hiring for? Email me at matan at factory.ai. High killers only people who are looking to transform software development. High agency. Let's get it. High agency. Get some. George? High agency is important. I'll add running hard. We're hiring a lot of forward deployed investors and bankers, like an insane amount. And I think it will be the most important role in the next 10 years. So we're here to save you from financial work. FDBs. And FDIs. FDBs has a nice ring to it. It does. But also engineers, sales leaders, anyone who is, where can they find more?

1:12:54Maton puts it a killer and they can find more. Just email me, george at hebbia.ai or on hebbia.ai. We have a careers page. Thanks, guys. All right. There you go, folks. And I'm hiring six new people for our venture capital training program. This is, if you ever aspire to break into venture capital, you don't have to have gone to HBS or Stanford GSB. you can just email researchers at launch.co and we're going to hire six and it's a one-year training program. And if you make it to year two, you get a big salary bump and you get to work with me evaluating over 10 ,000 startups that apply for funding every year and try to figure out which hundred we're going to invest in.

1:13:33It's more like maybe 15 or 20 ,000 at this point. Go ahead and subscribe this week in AI.ai has all the links and we'll see you next time. Bye-bye.