I mean, let's be honest, software has sucked. Technology's been hard to use. Like, it's most often broken. And now we're actually in a point where this stuff could actually start working. People keep thinking of these things as software, and it's not. And I think we just keep mapping it to things like search, and it's not. Like, we're starting to realize how powerful this stuff is. And I just think that like the biggest constraint is no longer the tech. It's literally our creativity.
You recently wrote about sort of four buckets of styles of investing that might be applied in, call it venture — I don't know, big private market tech, let's call it. Can you enumerate those and just like say a sentence on perspective of an...
Yeah, yeah, yeah. Again, I feel like I'm just in this point in my career. I've been doing this for a decade. I've done hundreds of deals, and I sit in these discussions and I just realize people aren't communicating very well. So I kind of did this exercise where I'm like, okay, so what are the mindsets people bring to investing? And I think there's basically four types of mindsets. So the one that we were talking about before is like this traditional finance mindset, which is the correct proxy for the health of a business is the market it's going into and its finances. That's what it is.
Yeah. It's like, what are the SaaS metrics? What, you know, what is the size of the market? You know, and that is a proxy for the entire business. So that's one. And you know, clearly growth investors do this, clearly, you know, but like more and more you see like early stage Series A people do this, right, which to me is a little bit discordant. Then there's what I call productive investors, which you believe you're investing in a productive asset, and the productive asset is what's creating value. And if you believe in that, as it creates value, it will create market, and as it creates value, you're going to have to find ways for the business to catch up to that value. And so the type of diligence you do if you're a productive investor is more like what is the technical approach, what are the adjacent markets, you know, how fast is this likely to evolve, how close is it to capturing value, like that sort of stuff. Is this the right team to build it? Then there's what I call frontier investing, which is I think this is the romantic idealized version of VC from like the 60s and 70s, which is kind of yolo, man. It's like founders know better than VCs and they've seen something that I don't see and they're going to go out there and do it, and like who am I to second guess them? And it's about you — they almost look like talent networks, right? Like I would say Chris Dixon, remember when he was investing in nootropics, he invested in Soylent, he invested in Oculus, he invested in Coinbase and in crypto. I mean, it's like every science fiction trope from the 70s and 80s you kind of invest in, and you just don't overthink any of these things. So it's almost like the strong form of the productive. And then the last one, and it is always so surprising how prevalent it is, is speculative investors, which they don't even care about the asset. They just want to time a market, right? And so like you'll see this like in crypto and in secondary markets where they don't know about the underlying asset. They probably don't even know very much about the actual like business metrics, but they're like, 'Well, this looks cheap. It's clearly going to go up in value. Oh, this is the top of the market, you know, like I'm going to get out.' And so they're trying to like time markets. They're actually just doing speculative. They're not really interested in like will the business grow. And again, the problem is if you have all of these archetypes, but you can justify any deal or not justify any deal. So if you don't know what your beliefs are, you'll always basically just kind of justify whatever personal bias you have.
If you think about yourself as a productive investor focused on disruption, I love this idea that disruptive technologies — you should not look at some of these traditional things that so many investors talk about. Talk about what you do look at, like when you're looking at, you know, you look at so many companies, what is your method by which to assess whether or not someone might be creating a truly disruptive piece of technology, a really valuable asset?
So I only do computer science investing and I only do infrastructure investing, and I think infrastructure as an industry has a certain shape, right? And so it's not like, you know, it's got certain buying behaviors, the technology looks a certain way, and the industry acts a certain way. And so what you're trying to do is you're trying to take a technical approach and you're trying to map it into that technical framework, right? Like this is kind of what we try and do. And the way you do that is, you know, there's a number of approaches. One of them, you're like, what are the leading companies doing internally? And this will kind of give you a glimpse on like maybe this is how platforms evolve, or perhaps the team that is doing this had did it in another context and now they're actually bringing it to this context. I think it's one of these things where like how much can you diligence without actually just using the numbers for a proxy? As far as you know, you can talk to customers, will they use this? You can see if there's like adjacent budgets that you know are growing that you could kind of argue will fragment to become this, or maybe you have early signs of willingness to pay but they actually haven't figured out the business model. Like I always think like every tech wave — I've been in the industry for 30 years — every tech wave, people like hand-wringing about margins. Everyone, like I remember for the cloud, I remember for mobile, I remember for the internet. You know, clearly now people worry about margins. So maybe you're like, there's a willingness to pay and there's a lot of interest, but like we haven't figured out the COGS yet in the actual viable business model. So like you're using partial indicators, not full indicators, in order to do this. And so my team — I mean, my team has a lot of, you know, technical people, product people, many have PhDs — and like, listen, we don't second guess the tech. Like the founder will know the tech better than us. But we can ask the question: how does this tech fit into the broader platforms? How does the tech fit into spending? How does this tech fit in with our understanding of stickiness and defensibility? And then we look at early market numbers and we'll say, 'Hey, listen, like no, like this is like a financial investor would puke at this, but like it actually shows that the customers are really interested to figure something out, and will they be able to figure something out?' So there's a lot of diligence you can do. Where with frontier investing, I don't think you do any of that. You're like, it's people.
Is this like, it's people. Is this the best team to do this crazy thing, and then you just yolo? You were saying before we hit start that nobody understands network effects. Can you explain what you meant?
I've realized that like even the most common terms we use, there's no consensus on. And so, for example, network effects is the favorite one that I pick on. Like I asked you before...
No, you didn't mess it up. You said what like a third of people say. So I said, so what is a network effect? What did you say?
I said a product that gets better as it adds a marginal user.
So to me this is almost a tautological definition because most products will get better as you add a user, right? Like they will find a bug, they will provide feedback, they will tell other people, you know, and so that doesn't say anything about defensibility. But other people, when they use the term network effects, they actually mean the original meaning of the term network effect from the 70s, which is a telephone network where when users join, they interact with each other, that grows super linearly, and that's why you get like natural monopolies. And other people use the term network effect more to mean like an actual flywheel. So like a self-reinforcing flywheel. So not just like things get better, but like it self-reinforces in some sort of a way. This is where you hear the term like data network effects. But even in that sense, like you don't know if like that self-reinforcement is super linear or not. So you don't know if it makes it defensible or not. So you have this very common word that's on 70% of pitch decks where a founder is sitting in front doing a pitch and nobody has consensus on what it means.
So how would you define it?
So I think the root of the confusion is not a definitional problem. I think the root of the confusion is conceptually we conflate the mechanism with the shape of the moat. And let me explain. So there's many things that could create a moat, right? Like a brand, scale, a two-sided marketplace, a network effect. But that moat — to reason about that is only really useful if it has some properties that make the company defensible against somebody else, right? So if that moat grows superlinearly with scale, like a classic Metcalfe's law network effect would, that's really defensible. If it grows sublinearly with scale, it may not be useful at all. And so I think, so you and I have a conversation of defensibility, I think the first thing we need to do is not conflate the shape of the moat with the mechanism that caused it. So first I think we should say the moat that we have is superlinear, it's linear, or it's sublinear with scale. And then the second thing — a lot of things actually contribute to a moat, right? To that, like for example, like we mentioned network effects, or scale effect. What's a scale effect? The scale effect to me is as the companies scale, it can reduce unit economics normally because it has pricing power, for example, or maybe CAC goes down because of brand. What's a two-sided marketplace? That's very different to me than a network effect. That's where basically you've got, you know, self-reinforcing supply and demand. So I think that there are many mechanisms, and then there's the shape, and then to have a discussion I think you want to make sure that the shape is something that actually provides long-term defensibility.
Then the last thing I'll say is it's almost impossible to have these conversations in isolation because we're actually talking about competitive equilibria. Because if you have a network effect, then maybe the competitor has a network effect too, and maybe if they have more capital than you and grow faster, that probably impacts the shape of your moat, right? Like you may go from, you know, like you may not grow as fast marginally because they're stealing your customers or something like that. And so I just think that we need like very, very basic concepts. We need to not conflate the notion of the actual shape with the actual mechanism. And then I think we should actually like enumerate the mechanisms because something like the definition that you had just doesn't really help anybody.
Can you give examples of superlinear, linear, sublinear just to bring it to life?
Yeah, for sure. A common view of where network effects came from was from the 70s. I forgot the name of the economist that did it, but he worked at Bell Labs and he noticed for like train networks and communications networks, the interactions of users grows with the square of the number of people that join, right? And so as a user joins, they can communicate to many other people. So the actual number of connections grows more than the number of users. And then, you know, therefore the value grows more than just linearly with embedding. The same thing with like destinations on a highway or destinations on a train. So that would mean that if you have two competing companies and one takes, you know, $5 million of funding and the other takes $5 million of funding, but this one is marginally better, marginally larger — this one's a little bit larger — for the same amount of investment, it will grow faster than a challenger, right? So like the growth actually increases over time, and that's why you get a monopoly. Whereas sublinear means if I'm a leader and I invest the same amount as a challenger, I grow marginally slower than the challenger. That would be sublinear. And clearly this is very important when it comes to moats. Because if you have two competing companies and there's a sublinear moat, the challenger will always catch up to the leader because the leader has to spend ever more money to grow. And, you know, all things being equal...
Ramp is the only platform built to make your finance team leaner, faster, and better, saving businesses 5% annually on average so you can stay focused on growth. Ramp just opened for business in the United Kingdom. So if you're running a business in the UK, you can now use Ramp's AI-powered finance platform to manage cards, expenses, bills, approvals, and accounting all in one place. I run my business on Ramp, and so should you. Learn more at ramp.com/invest.
Rogo is the AI platform purpose-built for financial institutions, serving tens of thousands of bankers and investors at hundreds of leading firms worldwide. Rogo's AI agents autonomously execute a large chunk of your firm's workflows. They can screen deals, draft sims, run buyer outreach, and due diligence on data rooms, all with full security and compliance handled. Every action feeds Rogo Intelligence, the firm's context layer, turning years of deal experience into a secure, governed institutional memory that makes each new transaction stronger than the last. To learn more, visit rogo.com/invest.
OpenAI, Cursor, Anthropic, Perplexity, and Vercel all have something in common. They all use Work OS. And here's why. To achieve enterprise adoption at scale, you have to deliver on core capabilities like SSO, SCIM, RBAC, and audit logs. That's where Work OS comes in. Instead of spending months building these mission-critical capabilities yourself, you can just use Work OS APIs to gain all of them on day zero. That's why so many of the top AI teams you hear about already run on Work OS. Work OS is the fastest way to become enterprise-ready and stay focused on what matters most: your product. Visit workos.com to get started.
Do you have like favorite company, recent company examples of these different types?
So I would — so actually I would — so how many actual network effects in the Metcalfe's law of superlinear do you think there have actually been here?
I feel like this almost never happens. Like I think we can all agree the telcos.
That's why we had to break them up.
We could probably all agree LinkedIn, like, you know, like they haven't had to like change very much and they seem to, you know, basically have a permanent status. Social networking. And but all of these examples, it's like the interconnections of people. I would say GitHub was another one that had a...
Which was effectively interconnection.
It's interconnections of users, right? So I think there are not very many examples, but when you have them, these are these very durable, high-value companies, right? So that to me would be like a classic network effect. I would say other companies probably are more scale effects, right? Like for example, people will say Uber. I'm not sure that that would fall under the definition of a network effect. Maybe a two-sided marketplace because you have the drivers and you have the passengers. Maybe you could use that, but even then you don't have a lock on either of them. Like all drivers drive for everybody and all passengers could use anybody. So that seemed like that was more like a funding scale effect just because they could raise more money and therefore the unit economics went down. And so, you know, my meta observation — so first off, I actually think the network effect thing was more of like a personal annoyance and me being a pedant — but I actually think defensibility most is kind of, it's not really even worth talking about most of the time. Let me explain why. So here's my experience for what it's worth. If a market you're investing in is growing, then companies are not going to want to compete. There's like a lot of new white space that gets created. And so what they'll try and do is they'll try and just move into other white space. And so talking about defensibility is kind of a second or third order thing if the market is growing.
And if the market is slowing down or shrinking, normally it consolidates anyways independent of moats or not. And so I do think we spend a lot of time as investors hand-wringing about defensibility when in practice it doesn't really matter, especially in growing markets which we invest in. And we've seen this with AI. Like I remember early discussions on models. I mean, I can't tell you how many investors are like, 'I will never invest in a model. It'll never accrue value. It's not defensible.' That's fine, but as long as the market's growing, in a large market it's going to accrue a lot of value. I remember this about the Neoclouds. 'They're never dependable. The quality of revenue is bad,' you know. And so I think that we kind of trick ourselves as investors by thinking in terms of defensibility when if there's a growing market, it's kind of not even worth considering.
Do you think that's a unique privilege of your stage of investor? And then, you know, once you get later stage, public — like someone that's buying the Anthropic S-1 needs to probably think about defensibility.
I mean, it depends on your thesis on the size of the market. I mean, if you believe we're 0.1% penetrated, like, you know, this has got a long way to go and value will accrue. And if you're liquid, you can go in and out of the market if value accrues. I mean, I think that would — you know, I think the nature of venture has really changed actually in general. You know, if you read back kind of, you know, the time of like, you know, Arthur Rock, it was really frontier-like investing and people didn't think about quality of revenue, you know, they didn't think about defensibility so much. It was people and it was tech and it was new markets. And you know, over the last 30 years, like the tech market has grown so large and the amount of money that's available to investors has grown so much, and then the size of the private companies has grown — like a private company today rivals public companies of just, you know, 10 years ago. And so I say in some level VCs have actually become public investors, right? And then a lot of the people that have entered VC have come from like more business school and traditional finance background. So I think a lot of the mindset comes from that world, but I think it gets applied too early.
The financial mindset gets applied too early.
Like a traditional investing financial mindset gets applied too early. And I think it's kind of confused. I mean, it's very often, for example, I talk to like a junior investor and they're like, 'Oh, the quality of revenue is bad.' I'm like, 'Homie, we didn't even know how to monetize the internet.' Like business models always lag disruptive technologies.
Right. And so like you can't — if you truly believe you're investing in a disruptive technology and that's what you truly believe in, A, you should believe it creates TAM. So why would you worry about market size? Because it's creating market size. And B, it's disruptive. So we don't even know what the pricing model is. Like remember like how long it took for us to go from like perpetual to like recurring? That was like years of like whatever and hand-wringing, etc. And so, you know, I mean, I think my caution is like maybe most investors should know what kind of investor they are. Like maybe you're a financial investor and like that's how you view the world, but then you're not actually evaluating tech. You're not evaluating the disruptive power of tech. You're literally looking at the quality of revenue, which by the way is going to naturally move you towards non-disruptive technology because if it's disruptive, you probably don't know what the business model is. You probably don't know what the TAM is.
On the other hand, if you really believe you're in — like if you believe I invest in tech and tech creates value and that's the type of investor you are and you do that early, you should not consider these other things. And so I think it's less about is there a right way to invest? No, these are all legitimate, coherent, right ways to invest. It's more about don't be confused in your basic...
Kind of approach, because otherwise you become incoherent.
And so then like you're like, 'Oh, I'm a disruptive tech investor, but I'm looking for like whatever good NRR' — like it doesn't work that way.
How would you describe the state of the union today in your world in infrastructure? Like obviously it's changing really, really fast. There's new kinds of companies, but there do seem to be like some — we could like name some layers or some categories. How would you describe like what it looks like today?
So for me, so the definition of infrastructure is like the differentiation is actually technical, where I don't view SaaS or apps as that way. Like do you view like if you invest in like a SaaS or apps company as like a technical investment?
Right. Because it's like a business now.
Yeah. Especially now. I mean it's like a business model innovation or something like that.
And so because for us as a technical innovation, it has to sit kind of within the traditional technical stack. So the traditional technical stack for infrastructure is compute, network, storage, and databases, right? And each one of those has subcomponents and they all have buyers and they all — you know, now we're seeing that being kind of reformed with AI a bit. And so all of them are being disrupted, but also now you have models, which I would consider infrastructure. They look to me more like ASICs than they look like, you know, traditional software. I mean like the capital needs, the, you know, the competitive pressures on them. And then I think they're disrupting the entire application stack above them. Like we don't even know if they will layer — like we don't even — like I mean maybe you could argue with Jevons they do kind of, you know, you can kind of put them in software, but traditionally they're almost like, you know, their own separate entity that you can't marry to software very well. And so I would say we're in kind of a point of maximum disruption, but all of the infrastructure supporting software, whether it's vibe-coded or not, or whether it's to build models or not — I mean that doesn't change the major components. And so I think we do have a bit of a traditional mindset of compute, network, storage, databases, and then models. Yeah. And then but we also throw up our hands. We're like, this is all changing so fast because for example when you worried about dev tooling in the past, you worried about ergonomics for a human. No, you don't really worry about that. And then actually a lot of the design decisions that we made as technologists were like, well, a human's going to be using this so it can't be too complicated. And then you would trade off things like safety or you trade off things like reliability. We don't really have to do that anymore because the models are doing it. So I think this is kind of the biggest refresh of computer science infrastructure we've seen maybe since the internet, maybe forever.
Would the shorthand for this be that each of those five portions of infrastructure are being refactored for agents? Is that like overly simplifying?
I split it in two pieces. For me, there's like people are vibe-coding a lot of stuff. There's more software being written than ever before and that stuff has to run somewhere and that has nothing to do with — like that's not agents. That's just normal software. Maybe it's an agent that created it. Maybe it's a human being created it, but like it needs databases and it needs compute, right? So that's one thing. And so I just view like it's like the product of agents. And then the other one is like we're building these new systems that are kind of — that have AI at the heart of them and then how does that change it? Those are the kind of two. So just assume way more software being built and has to be run and deployed somewhere, and then think that now we also have systems that like actually create agents.
It seems like each new wave of technology, like there's a new kind of database that's needed. Are you seeing that across all the areas that you care about that there's sort of because of AI whole new entrants?
I think a lot of infrastructure is getting a bit sleepy and now because agents are very sophisticated, that's no longer the case. Like so for example, let's take OLTP. What is OLTP? OLTP is a transaction database, like think the Oracles of the world. That's a pretty mature industry. You didn't see a lot of movement. Maybe you saw like movement to the cloud or something like that. But now we're seeing really interesting companies that are like, you know what? The hard thing about using a database from a programmer standpoint is databases always had their own languages and they've had their own languages for technical reasons and it's really, really hard to have them work with normal programming languages. But now because of AI — and one of the reasons is like if you change that language, like anybody that knows databases wouldn't know how to use it and you'd have to like train a whole version of programmers to use this new language. But now we're actually seeing databases come out which are entirely new ways of thinking about databases which have way better properties. They don't really have to train anybody because the agents already know how to do it. And so, you know, I think these things are path dependent. Like some sectors, a new entrant is going to come and change it. Like databases, OLTP may look entirely different in five years. And some, the incumbent may figure it out and be like, hey, listen, we're just going to add a few things. Networking, of course, which is — listen, I did my PhD in networking so it's an area of like special interest — like that's been so sleepy and boring for the last 15 years. And now because these clusters that we're building are so large and they've got like such rigid performance constraints, we're seeing, you know, net new networking companies come out for the first time since Arista. You know, we're actually having to change — you're seeing like actually, you know, changes in the materials used to build them. We're seeing optical come up. And so there's a lot of interesting stuff in networking. Compute, I mean, you know, the whole notion of like what are the abstractions to run an agent on? Is it like a container? Is it a VM? Is it a durable workflow like something like Temporal? Like that is changing. But I do think we're at a point of maximum disruption where many points of the infrastructure stack can be rebuilt.
Can we spend a few minutes on models? Anthropic's S-1 came out this week, which, you know, obviously a beautiful view into the business itself. But I'm actually more interested because the infrastructure stuff — in what you think about just the state of these things today. I'm interested in all the debates: closed versus open, the public frontier versus the private frontier that the model — the companies keep for themselves, you know, movement of talent, opportunities, things that don't exist that should, like whatever — whatever's on your mind about these things and how you see them evolving. I'm fascinated to hear.
So I think the biggest question is what happens if and when the market slows down, because I think we've got three years of data of what happens in the existing very big, fast-growing market. And like what happens there is the frontier — being on the frontier means you've got tremendous pricing power. Even if you're just an epsilon better than the next guy, you got tremendous pricing power. There's no demand problem. You've got basically infinite demand. And so like these kind of frontier models that are the best tend to get 80% plus of the market dollar-weighted, and that's happened for three years, right? But you could argue the reason that's happening is because we've not saturated the market at all and people are going to want the smartest thing. And so then under that view, there's nothing structurally going to change as long as the market's growing. And you could argue this is going to go on for quite a while because I mean, what is it, penetration? 1% or something like that. They say 3% of the enterprise, but actually from budget-wise, it's like 1% or less, right? And so there's just a lot to grow given how useful the technology is. The more interesting question to me is, okay, like what happens if the market starts to saturate in pockets, right? So if that happens, you know, will then things shift to open source? Will you lose pricing power on the frontier? You know, one thing that we don't really do a good job of accounting for when we model out the numbers for these companies is, you know, what happens if you invest in a model and then like you don't get the next model out really quickly, right? Let's assume that like Anthropic stumbles and they don't release a model for six months. What is the lifespan, do you think, of the previous model?
A very short period of time.
It's very, very short, which is entirely different than we've ever had in software. In software, we've always assumed some long tail. But it turns out like what happens if like, you know, like you put out the SOTA model and everybody loves it. Then something internal happens. I don't know if it's regulatory. I don't know if it's technical. I don't know if it's a breach. Whatever. And then you slip three months, right? It could be the case that you lose the entire market. I mean, my experience, and I see a lot of these model graphs from apps companies and from, you know, pitches, I would say that the lifespan of a model is three to six months. So that means you have to release the next model. So again, for this conversation, I think it's a very interesting conversation. We should tease it apart: like what happens if the market continues to grow and doesn't saturate? I think we already know because you have...
You have three years of data and it just means the frontier wins and they do a great job and they create great models. Then what happens is if either the use cases saturate, meaning you don't need more intelligence, or the markets start slowing down for another reason — nobody knows the answer to that. I think there's a very real world where the value comes crashing back down into open source and, you know, we'll see kind of like a shift that we've never seen because software has always had a long tail of retention.
On this trip I've been asking a lot of people what percent of software coding tasks are already saturated, and the typical answer is a very, very high percentage. That's been the biggest market, and yet it still seems like people are spending like crazy on frontier tokens to do coding, which it seems a little at odds — like if I could do it with a six-month-old model or something that's vastly cheaper.
Well, these are two different statements. I think we need to be very — this is a great thing to dig in on, but we need to be very careful when we're talking. Please tease it apart. Yeah. This is great. Great. So there's two very different statements. So statement number one is I am not using more tokens, right? Because I, you know, I have enough tokens. I've automated 95% already, so I can't automate anymore. That's — and I think that we're getting pretty close to that. Comment number two is therefore I can use a cheaper model — does not follow from comment number one. Right? So if you and I are like — you and I are competing startups. We got Martin's startup and we got Patrick's startup. Right? We've got competing startups and because we're in competitive equilibria, are you going to choose a dumber model to compete with me?
In competitive equilibria, you wouldn't. You would always pay 10 times more for the smartest model.
Yeah. Always, always, always. And so, you know, there's this very open question of, um, will you saturate tasks so that you don't need a smarter model? And I think it's very hard to do that in a Darwinistic economic equilibria, like, you know, and when you're actually competing. And so some of them may, but I think at the end of the day it's about can you get software out faster, can you make it better, and can you compete in the marketplace, and that will necessarily push you to the more intelligent thing. Now that said, clearly things will slow down because you're actually not spending more tokens. And so either you have to increase prices or you're going to have to go broad. But in my experience, the pricing is maintained for sure, and because of the competitive equilibria, like you always want the smartest thing. And then we're just seeing like a broadening use of these outside of code, right? So general knowledge work — like I'm a huge Grok user. It's like my favorite thing ever. And so I think we're starting to see it take over more and more of these.
Say more about your personal experience with that. What you're learning, what you're doing with it, what you're learning from that experience.
My new Grok. This is so funny.
I know you have a home automation thing going on.
I know. It's so good. Like I just feel like I've waited my entire life for this technology to show up, right? Like I want Star Trek. I want to go home and I want my house to greet me and I want to talk to it and I want to have the face show up on the monitor and I want the whole thing. And like it's never — we've never been able to do it and now it's kind of trivial. So I use Grok for a lot of stuff. I use it for personal finances. My big hobby is modular synth — like I like synthesizers, like making music, but with like the 1970s like patching, tweaking type stuff. It helps me with a lot of that stuff. So I use it for a lot of like standard things, but my new kind of passion project is I want Grok to embody itself. So like what I tell it, I'm like I say, I say, you know, I want you to be the living AI in my house. That's what I want. And I want it to be just like Star Trek. Help me make this happen. And it's literally...
Buy stuff. So I was like, 'Oh, let's try this.' So like yesterday, I bought a few hundred dollars worth of gear that's coming to the house and like I'm going to have it embody itself. So...
So what gear is it buying?
I don't — I don't even know. It's like these weird open-source projects and like I mean it's like this really obscure — like I know some home automation stuff, like I know like Zigbee, LIFX, I know some home — but it's like really obscure stuff that it's decided to buy. But like I just have in my head this vision of like, you know, like that kind of vintage view of what the future looked like, like the '70s view, like you watch the house.
Like, yeah. Like, I mean, I just want that. And so like I've basically just been kind of feeding it pictures and descriptions and I want it to create that by itself. And so I'll let you know how that goes.
What is this and other experiences using this technology, since you're probably in the 0.001% of, you know, creativity using it? What does it tell you about what's coming?
Yeah. No, this is it. I just feel people keep thinking of these things as software and it's not. And I think we just keep mapping it to things like search and it's not. And we really have created general flexible technology that can bring us like many of these future things that we've always dreamed about, you know, but like, you know, you just didn't have the technology to implement. And I just feel like we're so constrained from our presuppositions of what software can do that people aren't doing as many of these things. You know, I'll give you another example. You know, I work with Fei-Fei and World Labs, great company. And one thing I realized as part of that and what they did is they created, you know, they created a model that would generate, you know, 3D scenes. And, you know, they released a model called Atlas a few months ago and, you know, that would like generate basically any scene but also supported dynamics so it would move. And I realized, you know, we're this close, we're this close to being able to step into a movie. I mean, like imagine like, you know, Miyazaki probably wouldn't be happy with it, but imagine like having Spirited Away running and you throw on your VR glasses and you walk in, you know, and like you're there. And like, um, did you ever see Sleep No More? Like the...
Of course. Yeah, I went to it.
Like, so like why isn't there a Sleep No More in VR...
...that, right? That you're walking in, like the whole thing is going on. And like, just for those that are listening, Sleep No More is an immersive play which is three hours. It was in this kind of three-story hotel or something like that building in New York and you're like running around trying to figure out and like Hamlet is being played out or whatever.
And so I just feel like we're starting to realize how powerful this stuff is, and that a lot of like these tropes and visions of the future from like the 50s and 60s we can actually make. And I just think that like the biggest constraint is no longer the tech. It's literally our creativity. And it's just because we've been beaten down by 30 years of software being — I mean, let's be honest, software has sucked. Technology's been hard to use. Like, old people hate it. Like, it's most often broken. And now we're actually in a point where this stuff could actually start working. And so for me, it's like I'm so excited for what this brings.
It's tempting to say you could just freeze the models and their current capability and still get all of what you described. So there's like a diffusion question I'd like to explore. But before we do that, what, if anything, do you think is missing from the models? Like how do you hope or wish or dream that the models themselves continue to get better?
So I think again, I think if you step back and you look at like what we know as an industry, I feel the current tech stack, anything that's digital, I think with enough capital you can solve, right? And the reason is because anything you can simulate on a computer, you can create RL environments for, whether that's 3D using Blender, whether that's general application use of a computer, whether it's code. So, you know, and you and I have talked about this before, I think when it's all said and done, the most interesting thing about AI, the meta thing, is that like you can take a money kid and then you can point it at a problem and it will solve that problem. Like maybe it's expensive, but at least you can solve it. In the past, you know, you'd be like blocked in engineering and you'd die, but like now you can actually just direct money at problems and solve these problems. So I feel very confident saying anything in the digital world, because you can create an environment and you can RL it and you can train on it, we can solve. So I just feel very comfortable. Anything you do with a computer, AI will be able to do. So maybe it doesn't do it now, but like you can just see that path happening. What I don't know is does the same thing apply to like the physical universe, right? You can't like — sure, sure you could set up a lab to like nick speakers or whatever, you know, to do that. But like are the scaling curves the same? You know, is the distribution the same? So anything to do with like robotics and the physical world and science and medicine, I don't — I have no idea if it follows the same laws. And so I think that's still an unlock. But I feel very comfortable that the current digital side is done. I mean, it's done.
How have you processed all this interesting security stuff, breaches? I've heard it called breach apocalypse. It's obvious that we're like entering a very interesting era in cyber. What do you make of all this?
I lived through on the ground, you know, with formal studies in security. I lived through the internet and it was way worse than this. Like if you like by this time with the internet, you know, we had worms that had caused, you know, billions of dollars of economic damage. We'd taken out critical infrastructure. I remember when it was — remember Windows 95? You're probably too...
