The Exchanges

Q uh, of AI, at least for now? It's, uh, I guess we, we talk about Bell Labs. It's, so we go back to the Bell Labs model, but, um, I guess, how, how do you see that play out? The large companies are going to be the core, um, centers for, uh, research, and then if so, how do you contribute back, I guess, to the rest of the world?

A Right, ok, so I used to work at Bell Labs and NEC and, you know, a couple other, and I had friends at IBM and Microsoft Research and things like this, so I know a bit about industry or, uh, industry research. Um, uh, it's, Um, it's, it's a, it's a complicated, uh, issue. So it's not like everything is going to, you know, every interesting research is going to take place in industry and that, you know, academia is just going to be kind of watching us, uh, making progress. It's not like that at all. Uh, there are contributions that are very complementary from academia and industry. Uh, so of course industry has, you know, more computational resources, more data, uh, can get started quickly on, on, on projects because, you know, you're, you're next to your colleague who is one of the best specialists in the world for a particular technique. You just talk to your colleague and get started. You don't have to talk to anybody. Whereas in academia, you know, you, uh, have to get a grant. You hire students. You teach the students to, uh, uh, You know, get familiar with the techniques. Sometimes the students work out. Sometimes they don't. So you have to take another one. So every project takes longer to start up, ah, in academia. But, ah, but the type of motivation is different, and so the type of ideas that are generated is different. And I, I don't think at all that, ah, industry will…

AI assessment note: “We publish most of the stuff we do. We release most of our code”

Q And then that seemed to have evolved towards a greater emphasis on networks. Um, and network effects. Uh, how do, how do you think about, about those, uh, you know, what are some examples of proprietary data assets, and why do they matter?

A So, we came to realize that The only defensibility that you can get with data is defensibility around your own data. If you're getting data from somebody else, it's not defensible, because they can give it to somebody else as well. And even if you have an exclusive contract with them, that exclusive contract's gonna come up at some point, and you're gonna have to, if you've built a valuable business, you're gonna have to pay through the nose for it. So, um, we, we sort of realized that unless you Own your data. Um, you don't really have a defensible data asset. And then we started to look around and, and figure out where the most valuable data sets came from, and we saw these large networks, um, that have gotten built up on top of the internet that, um, create tremendous amount of data. And, uh, and so that, uh, to us started to become You call it an evolution. And I think that's probably the right way to put it. An evolution of our theme. We still very much believe in, in data. Um, you know, let's take Twitter for example. The, the stuff that Twitter's doing around discovery and, um, and who to follow and, and all those search, all those things leverage the data assets that they have. And, um, And they are making the service better and better and better because of that, but that data exists because they have this massive network of people who are contributing to their system. …

AI assessment note: “let's take Twitter for example. The, the stuff that Twitter's doing around discovery”

Q Do you need to work with a subject matter expert? You don't need to be one.

A I can, I can give a, I mean, I, I can give a very specific example. I didn't actually get to talk about my past. It's not very interesting, but one of the things, the first thing that I did, I, I worked in the intelligence community for five years. I was a computational social scientist, which I think is what they call data science back in the mid 2000. Um, and the last job I had before I went to graduate school was actually working Um, for an organization called the Joint IED Defeat Organization, so we were creating models that we would actually use to predict the placement of IEDs in Iraq and Afghanistan, and we would send down range, you know, predictions wrapped up as tactical moves to soldiers to say, here's where we think you should avoid today, here's where you think you should avoid tomorrow, and the subject matter expert there was the soldier on patrol. So like, this is a very sort of, Tactile example of subject matter expertise mattering is when you say to a soldier, I don't think you should go down that corner and they say to you that corner has been blown up for two weeks. It doesn't exist anymore. You know, that's when you really get to understand the reason why having really sort of in the loop data and subject matter experts is really impactful. So it's like, you know, sort of everything else that you do, everything else that I've done that since then sort of inf…

AI assessment note: “having really sort of in the loop data and subject matter experts is really impactful”

Q Um, couple more questions because I want to open it to, to, um, people here. Um, cloud versus on-premise, is that a relevant, is everything moving to the cloud? You mentioned transfer of data, Matt.

A I was, I was gonna raise that, that, that point earlier. It's, it's, it's a huge issue. Especially when you're talking about the enterprise, especially when you're talking about, The three letter agencies, or financial firms, or even healthcare, where the data is so incredibly valuable. There are huge privacy and security issues around it. There are a lot of places that unless you can give them an appliance, or you can load it on their servers inside their firewall, they're not going to buy it. They are not going to put their data in the cloud. I would say there, there is growing acceptance of that and comfort with security around that. But there is still a huge business for on-premise solutions, and depending on what you're, you're designing, you may need that kind of fork in your development path.

AI assessment note: “there is still a huge business for on-premise solutions”

Q So would coding be the brain and then everything else becomes the, uh, legs, uh, And the hands that are connected to the brain, or is that the wrong analogy?

A Yes, it's kind of like, you can kind of think of the model is the brain, and then what people call a scaffolding, agent scaffolding. That's sort of the affordances, right, the things that you can actually do. All these affordances for software, at least, you know, for digital intelligence, are going to be primarily through code. That is, the other option is teaching a model how to drag a mouse around, and it's called computer control. Some of that will happen, but I just don't think that that's going to be the majority in which, um, you know, A language model interacts with software. So if you solve coding, you've just solved how does a language model, how should it interact with software?

AI assessment note: “you can kind of think of the model is the brain”

Q That's complicated, right? Because you can have a user that's not the same person as the buyer, and if you make your product fully available to everyone, you may have use cases that you don't really want. How do you think of that?

A So it's, it's, as far as we can tell, it's way more complicated than building a consumer company, um, in that, in a consumer company, you know, to first order, if you have users, you can monetize. Right? And so, like, basically, all you wanted to do was get users, then the assumption was is that you could use advertising or something to monetize it. And even that's changing now, by the way, um, as GAFA takes control. But that's kind of the way you thought about it. In the enterprise, that's not the case at all. I mean, you can have an organic motion that you simply can't build a sales motion on. So I'll give you a few examples. One of them is you can have an organic motion that sell, that is adopted by one person in an organization, but the buyer is someone who is entirely different, and the person that organically adopted doesn't have budget. Another common failure mode is, let's say you have an organic engine, so you're, like, a bunch of people are using this product, ah, but you made it too good, and so now you've cannibalized any ability to monetize later. Um, another problem that I'm seeing in a lot of companies is, let's say you build a good organic motion, but not a great one. So you're doing an open source project, and it's pretty good, but not great. The problem is, is now if you build sales, as soon as your sales surpasses your organic motion, you're SOL. It's really …

AI assessment note: “adopted by one person in an organization, but the buyer is someone who is entirely different”

Q And just to verbalize, uh, the question early in this conversation, uh, what is so bad about super intelligence? Obviously, uh, there's a lot of talk about, uh, scientific progress and curing cancer, um, and your, uh, documents, uh, effectively recommends pausing the rates towards the super intelligence. So why, why is it so bad?

A Yeah, I wouldn't say super intelligence is bad. I would say it's dangerous. Like it's a very dangerous Thing to create. So the most straightforward story for why, um, it's dangerous is that it seems pretty likely that on sort of a trajectory similar to the trajectory we seem to be finding ourselves on, um, you end up with AI takeover as a result of building super intelligence because the AIs are in a position where they can take over due to being highly capable, widely deployed, and, um, you know, uh, building basically, uh, the huge amount of industrial capacity potentially. Um, and we could talk more about what a takeover would look like. And then if they're in the position where they could take over, then there's a question of like, would they want to, or how would the motives shake out? And it looks like we don't really have that much control over the motivations of AIs. And it seems like that problem gets harder as they're, you know, much more capable and built via a process where AIs are automating AIR and D and we maybe are losing our understanding of how that process works. There's sort of this, like, we don't necessarily control the technology, misland AI takeover. Another concern is that Historically, like, you know, um, at least in recent times, uh, the distribution of power among humans has been reasonably distributed, though not necessarily, like, super, super dist…

AI assessment note: “I wouldn't say super intelligence is bad. I would say it's dangerous.”

Q Yeah, reading Redwood research stuff over the years, it seems to, that there has been an evolution from being focused largely on interpretability to much more AI control. Is that, is that fair?

A Yeah, that's fair. So I would say that like our arc as an organization was we, when I joined the organization, I just finished up a project on adversarial training and was interested in getting into like doing interpretability and what we would call like model internals work where it's like, can we take advantage of the fact that we have white box access to these models to do something, you know, better than just the naive methods of sort of prompting and training when like trying to like align these models, understand their motives, like know what's going on. And we explored that area for a while and then for a mix of reasons decided it was like quite a bit less promising than we'd initially hoped and decided to move on to other things. And one of the things we moved on to shortly after that was, uh, AI control, um, which is the idea that maybe it would be a good idea to prevent AIs from being capable of accomplishing problematic things, or basically make it so the AIs aren't able to cause huge problems, even if the AIs wanted to. With, you know, there's a bunch of different stories for why this is a good idea, but basically the idea is, like, there may be some intermediate period, an intermediate period that I would say we're currently in, where the AIs are maybe capable enough to cause at least moderate problems, and then I think increasingly able to cause quite large proble…

AI assessment note: “Yeah, that's fair. So I would say that like our arc as an organization”

Q And as an aside, by the way, there's, uh, one of the parts of the write-up that I find the most fascinating is, uh, uh, your model says, uh, world GDP could grow roughly 200 X, uh, during the 20 thirties under the restrain plan. Can you, can you talk to that? Like the, the number is staggering. I mean, we're in a world of, uh, three percent GDP growth.

A Yeah, yeah, yeah. So, so a key part of our Perspective is like, even, you know, even non super intelligent AI systems at the level of capability we discussed would be radically transformative across the, you know, across the world and like, and just like for all kinds of different things. And we are imagining sort of slowing down AI development some, or like going at a more cautious pace for some period, and then eventually hitting a level of capability where the AI can basically automate basically everything that humans can do. And staying at that level of capability for a while, where we work on safety and security and at that level of capability, those AIs would be capable enough to be doing huge amounts of autonomous R and D. And in addition, It would be totally possible to have robots that are basically like, you know, more capable than humans at manufacturing and industrialization and so on. And in particular, you could have robots build robots. And so you can end up in a situation where you have huge amounts of robotic industrial capacity that is itself building more robotic industrial capacity that can then produce downstream goods. And that total capacity can basically grow very fast. I think we propose like Limiting that growth somewhat for various reasons with like various types of taxes, but like, uh, we can, we're imagining sort of the robot population or like, you…

AI assessment note: “robot population... doubling or quadrupling every year, which... means the economy would double”

Q were a couple of parts to the discussion that I found particularly interesting. In particular, uh, there was this argument that, uh, RSI may be very good at automating the process of creating the next generation of AI, but it may or may not have the, uh, the kind of intuition, uh, that one needs For a scientific breakthrough and truly novel ideas. So what's the rebuttal to that argument?

A Yeah. So, um, the way I put this argument is AIs might be very good at sort of the more mechanistic or nitty gritty parts of AI development, like writing code, running experiments, but not as good at sort of the broader conceptual leaps. So first I would say that like AIs seem Significantly better at engineering and grungy stuff and sort of just keeping trying than they seem to be at conceptual breakthroughs, but their ability to do sort of these. Breakthroughs, especially in easy to verify domains are improving. And like, you know, one example, um, is like their ability in math, but even in like, you know, ML, their taste has been improving. I think it continues to improve. Um, and so it's not, it's not so clear to me that this will lag super far behind. Another thing is that, uh, you can measure how good these AIs are at intuition or research taste or having breakthroughs, especially in domains that are relatively easier to verify. And if you can measure it, then you can take your grungy AI Um, labor or even just your human laborers and try to optimize that. So, you know, if it can be measured, it can be hill climbed on very roughly speaking, at least. And I think this is a case where you could just hill climb on how good the AIs are at making these sorts of breakthroughs in a wide variety of different settings. And then I expect that would transfer to making the actual break…

AI assessment note: “if it can be measured, it can be hill climbed on very roughly speaking”

Q All right. So, uh, going back to AI, 2040, give us a quick version of, uh, what it is, who wrote it and what is the main thesis? And then we'll go into some details.

A Um, there's sort of two components. AI-TV is a scenario, um, focused on, like, what the authors, including me, think is, like, a plausible good route for things to go, or, like, a reasonable plan, at least in some circumstances. Um, and it's written by Thomas Larson, Daniel Cucutello, me, uh, Eli Lifland, um, Brendan and Romeo. And I would say, like, The, the, the basic story is like, how would you do a deal with China to make AI development both be safer and also so that we can sort of like hang around at a point that's short of super intelligence, but where the AIs are still really, really capable for a long time so that we can study those systems and have a longer time to sort of integrate them into the economy, understand how things will go and so on. Where I think a concern that we have is like on the default trajectory, you maybe go straight from like AI systems that are like competitive with humans to AI systems that are wildly superhuman in a very short period of time. And that seems like. Uh, quite scary in a variety of ways. Um, in addition to that, we worry about like AI development being insufficiently transparent for sort of third parties to provide a reasonable check to AI companies and whether their plans will work. And we sort of have a unified proposal that solves a bunch of these different problems and makes it so that, for example, you can pay a huge amount o…

AI assessment note: “it's written by Thomas Larson, Daniel Cucutello, me... the basic story is”

Q And you know, as you were describing this, uh, we were talking about continual learning earlier. So whether that's continual learning or something else, if there was a technique that appeared that, um, made the compute a lot more efficient and, uh, you were able to, uh, just vastly reduce the, the compute effort, especially for those pre-training runs, how would that impact the plan?

A Yeah. So if, let me give a hypothetical and then talk about what I think the realistic case is. So if hypothetically it was the case that tomorrow a recipe for training super, super intelligence on like, you know, 64 H 100 or like some small amount of compute just dropped. I think we'd have super intelligence very fast or like that would be my sense. And it would not be possible to do this sort of deal, but I think there's, there's, it's not, um, but, but the hope with restricting compute Isn't just that like, you know, AI development is currently very compute hungry. It's that also that R and D is very compute hungry. And so even in a regime where you were doing continual learning, before you had the version of continual learning where you could do everything on a single, you know, on like some tiny amount of compute, you're going to have a shittier version that can do it on a moderate amount of compute. In order to develop the version that can do it on a tiny amount of compute, based on the history of error progress you need, you would need a lot of compute to do that research. Or I shouldn't say need, but in practice that, that would come about through a lot of compute. Now, if it was the case that there was some alternative research direction, Which in practice was a lot less compute hungry, and which got quickly very developed, quickly developed during this period, and whi…

AI assessment note: “it would not be possible to do this sort of deal”

Q about a minute ago, there was a different letter that came out with, uh, 1100 people. And, uh, and, and this time both anthropic and open-ended sign that asked government to help, uh, deliberately pace the frontier of automated AI research. Uh, so basically the industry kind of asking, um, for a slowdown. Are you, are you, In that camp, or you think that's just not the way it works?

A I did sign this letter. Uh, I agree. I agree that I think we should, uh, we should go there. I'm also, and you probably have the same feeling as, as an investor. I have the feeling that even if we stopped right now, we would still have quite a good companies we could build on top of we have what we have right now. I feel like there's a lot of things we, we can already do with these models. And they're already extremely interesting. I feel like there's a lot of things we need to understand and we should do in terms of open science and sharing how they work. So I'm not, I'm not in the camp of we need to rush really quickly right now. Uh, the main question is, uh, if we want to slow down a little bit, also that would be great because maybe then we don't have four announcements per day that we need to mix in one podcast. Uh, maybe I can take one day of the holiday in the summer. Um, but no, I, I think the main question is, uh, can we do it right? That's the main question here. I think a lot of people would be fine with AI going a little bit, a little bit slower, being a little bit more open, being a little bit more caring, a bit more like, uh, reflexive and, and, and trying to understand better, you know, how to, how to do that really well. Um, but the main question, how can we negotiate, uh, how can we organize, uh, a slow down there, uh, without having, Bad incentives where, wher…

AI assessment note: “I did sign this letter. Uh, I agree.”

Q And did you pick accounting because of how interesting that was from a agent building perspective or the other way around?

A It's probably the other way around, but I do think it is actually quite interesting from an agent perspective. Accounting is interesting for a lot of reasons. It is, one, one of, if not the largest knowledge work profession in the country. There are over three million, you know, combined kind of accountants in the country. And what I think is so cool about accounting, actually, is that most people don't really think about accounting. They don't think like, well, yeah, why is that even there? You know, probably most listeners have never thought, like, why does it even exist? And, um, I know we're gonna talk about agents, maybe quickly, 30 seconds, just to convince everyone how cool accounting is. Uh, if you think about the real world, uh, so much stuff happens. Economic activity, right? Like, you know, I was just drinking a water bottle there, like, you know, someone, um, uh, that bottler had to choose to, like, go, uh, uh, buy from that factory or that supplier, um, or decide to open, you know, some additional, uh, store or hire a salesperson. And these are all economic decisions that stem from understanding The real world. What's in the real world? You know, money moves hands, someone signs a contract, someone delivers the inventory. It's like all these events that occur. And so much of modern capitalism relies on the ability of all these actors to make decisions on these even…

AI assessment note: “It's probably the other way around, but I do think it is actually quite interesting”

Q So do you think of, um, doing your own RL as some kind of moat against being swallowed up by model performance? I mean, as, as you think about applied AI companies of the future, will, will they all be RL labs of some sort?

A So generally, and if anyone's currently trying to found a company, I recommend thinking this way. Technical moats are not real moats. Like, I, I, there's no portion of basis's long-term terminal value that stems from some, you know, secret RL trick we found that nobody else found. So that doesn't matter. Um, uh, what matters is that right now we are obviously very good at building long horizon agents, um, that, you know, can be reliable and deployed in production, and we'll continue to be the best at that. And that allows us to win market share and get deeply embedded. And that's why we move really fast, because the work we're trying to do is to, um, uh, go and proliferate maybe before, like, you get to AGI, whatever you want to call that. So I think that Most of the moats that will exist will be business moats. That is true in the AGI era. I would argue that's also been true in the pre-AGI era. Like, I don't think that, you know, Salesforce can write a better SQL query than I can. Like, the mode that Salesforce has is not related to their technology, right? It's related to their business position. You know, it's the, it's the powers. It's the, it's the, it's the, the workflows that they own. It's, it's all the, it's so many of these different things that, uh, come with being, you know, embedded in, and that's what matters, uh, not the technology. The technology is a, Is a, uh,…

AI assessment note: “Technical moats are not real moats. Like, I, I, there's no portion of basis's”

Q And, um, maybe walk us through how AI Changes that whole discussion. So, uh, you know, there was industrial automation, obviously that's been going on for decades and perhaps centuries. Then there was a whole wave of IOT and, uh, famously your, your ticker as a public company is, is IOT. And now there's AI, um, how, how different is the current moment?

A Yeah. Well, it's interesting. You mentioned centuries because that is like the right timescale to think about physical infrastructure. All the way back to, like, the Roman era, like, there's pretty significant infrastructure out there. So many of these processes of, like, how do you maintain a roadway have been in place for a very, very long time. A lot of that process was manual, right? Like, let's go inspect the condition of the road. Let's understand when it was last worked on. You know, let, let's kind of dig it up and see what we find. If you think about the ability to digitize that and then use sensor data, it's a huge unlock. So the question becomes, how do you get the data? Then how do you process it? And then how do you come up with a meaningful insight or really an action? Like what should we do about it? Uh, and that's, that's I think now possible. The last call it two decades was around reporting. Like how do we ingest the data and give you a really cool like table so you can look at it and reason about it, figure it out. Now what's awesome really in the last two, three years is the AIs are able to reason about this kind of information. They can look for other context clues and then give you the insight. And now we're actually seeing agentic AI, of course, which is It can take an action for you can maybe schedule the work to be done or, or start performing some of t…

AI assessment note: “Now what's awesome really in the last two, three years is the AIs are able to reason”

Q As a thought, uh, so selling to a bunch of different personas, especially in like more traditional industries, uh, especially as you've added this AI layer recently, how do you go about it? Like, how do you convince people in You know, typically non technology industry to choose to buy.

A Well, you know, I think the, the great part about this technology is very tangible. And, um, it's the kind of thing that when you see it, you get it very quickly. So what we do is we tend to go on site, we will demo the technology and do trials. So you can easily, these are plug and play. So you can easily try it out in your environment, in your industry. And, um, like I said, there's so many challenges in physical operations. It's tends to never just be one thing. So Yes, we want to reduce the number of accidents we get into, but I think we're also, like, leaving our trucks idling a lot because it's just a bad habit, or we're leaving tools behind at the job site, and we'd like to get those back because we spend millions of dollars replacing them. So we will often find multiple challenges like that, and then we demonstrate to them at small scale, like maybe a team or, you know, a region or something like that, that this works, and when they see it, they get it immediately. These are people who are experts in their industry, so they would say, I immediately see the value or the ROI, but they have to see it in that kind of tangible way. They're not just buying it because it's AI or big data or something like that. They're like, no, if this solves problems for us in our construction business, great, let's do it.

AI assessment note: “we tend to go on site, we will demo the technology and do trials.”

Q Um, is there a role for, uh, Local AI and local chips. NVIDIA had some announcement around just building chips for Windows computers. Is it for inference? Is that something that you guys, uh, think should be part of the multi-cylical ecosystem?

A Yeah, I think, um, if you look at the way the ecosystem for apps and cloud emerged, uh, everything you can do, you should do on your phone or your laptop. But the, the, the ability to get real processing power to a phone or to a laptop is constrained because they're generally working off a battery. And so you, you want to do as much as you can, as, as close to the data as you can. But the truth is, is in most situations, for real compute, you have to go to the cloud, to the data center. And that's exactly the way it's going to be with AI. We're going to do a lot of work Uh, uh, on the cell phones, uh, on a laptop, but for the big work, you're going to go to, to the data center. And that's where our focus is. Our focus is, is data center compute free.

AI assessment note: “for the big work, you're going to go to, to the data center”

Q Yeah. And to the general chip versus pistolized chip, uh, it was actually interesting that you guys in your city celebrated when Grok was, uh, acquired. Was that, uh, what was that? Was it a recognition by NVIDIA that your vision was right all along?

A Yeah. I, I think, uh, One of those ideas sort of most durable modes was the perception that the GPU could do everything, and it was all you needed for AI. And the acquisition of Grok for twenty billion dollars, and the structure, and the speed with which they chose to do it, made clear to everyone that that wasn't true. That, uh, the GP architecture couldn't do, could not do fast inference, and that this market was large and growing quickly. And we were the fastest at it, and the largest, and, you know, our sales were more than 10 times the Grox, and they paid twenty billion dollars for the number two collector. So that was a good day. That was a great day.

AI assessment note: “made clear to everyone that that wasn't true”

Q comes from the big labs, which, uh, you know, themselves are financed by, uh, venture capital, private equity, edge funds, where you call it, you know, sovereign investors. Um, is, is there any concern that, you know, demand for chips comes from the labs, which Are maybe artificially financed and that the, you know, if you had a true circle of deals here and there, then you sense any infrigility?

A Um, I, I, that's not exactly our experience. I mean, obviously we have, uh, enormous demand from pull open in. But we have huge, you know, dozens of other customers who, who are trying to place very big orders. And historically bubbles were when supply got out ahead of demand, right? When, uh, in the nineties, we built out a telco infrastructure, right? We built out fiber years before it was Going to be used and it took six or eight years and it all got used, but it was a sort of, if you build it, they will come mentality. Whereas what's different about AI right now is we're all trying to catch up. Um, we're trying to build data centers faster. We're trying to increase our, our demand, our, our, our supply chains for demand that's already here today. And only a very small portion of the world are using AI anywhere close to its potential. And we're already sort of overwhelmed with compute. We're, um, overwhelmed with, uh, the demand for, for memory, which was a real weakness in the GPUs. It's not a problem we face. And the ecosystem, there are constraints left and right, and that, that doesn't feel like a bubble.

AI assessment note: “that's not exactly our experience”

Q Including chips. You also released a few weeks ago, uh, MRC. Which is a networking protocol. Walk us through that. What is it and why is it a big deal?

A It's a new, uh, Networking, uh, protocol routing technology, if you will, uh, uh, uh, to scale these really large cluster fabrics. So imagine you have, uh, a 100,000 GPUs, uh, they need to be connected together, and when you're doing large training runs, uh, they're constantly communicating with each other because these models are so large that The processing of the models is happening over the entire 100,000 GPU plus, for example. And you can imagine, imagine the number of links and switches and NIC cards that need to be there to connect all of these chips to them. At this scale, stadiums are fun, right? It happens all the time. Uh, you can't really even enumerate all the ways things could fail. So the strategy beyond MRC is how do you design algorithms and protocols that can gracefully And mask all these failures, and make sure that the training workload does not get impacted, right? It just, the network is an abstracted system that the training job does not have to worry. It's always going to be there. It's always going to find a path, even if a link fails. So it is all about reliability. It is all about availability. So how do we design protocols that contain the complexity of such a, such a big cluster? And make sure that we don't get stopped because of failures, which are very common in the system. So MRC is like a multi-part spraying protocol, where you can spray packets…

AI assessment note: “It's a new, uh, Networking, uh, protocol routing technology, if you will”

Q uh, it's a bit of a new world, right? So obviously not quite a pivot because obviously the all AI research is, uh, going, uh, you know, full speed ahead, but like building a whole new business within the, the, the company, is that, is that fair? Is that how people think about it? Because at least building models is one thing, building data centers, it's a whole different world.

A Yeah, I mean, I think OpenAI has always had a fundamental belief that computers are the foundation of everything, right? Computers are the foundation for intelligence, and the way we keep continuing to scale intelligence and distribute intelligence is by having compute, and so that has never been different, and it's never always been the belief. I think what's becoming clear is to build the kind of compute we need and at this scale, Uh, we have to not just rely on getting compute from our partners. We increasingly have to take a much more active role in building and getting that compute that we need. So it does, uh, absolutely feel like a new muscle that we're building in the company.

AI assessment note: “So it does, uh, absolutely feel like a new muscle that we're building”

Q It's a closed loop. And by the way, you mentioned Texas and rural areas. Uh, since we talked about data centers at the beginning of this conversation, uh, why, uh, do, uh, OpenAI and other companies pick rural areas? Like, how do you select a site for a data center?

A Uh, so many factors. So one is, of course, land, like, plentiful land. Uh, number two would be, uh, permitting. Like, can we build these things? And we want to build these things Such that they are, uh, not affecting any neighbor roads, right? So land that is somewhat removed as their ideal candidate. Of course, access to power, right? So a strong grid, strong gas, uh, availability, all of those are important factors. And then four is neighbor, right? So how quickly can you build these things? So availability of labor, construction labor, qualified electricians, bloggers, All of this can play with. So all of those factors go into every single site selection decision. Um, and I know we obviously Texas isn't popular because it fits a lot of these criteria, but it's not the only state. I mean, we have data centers all around the market in LA plus.

AI assessment note: “so many factors. So one is, of course, land, like, plentiful land.”

Q uh, all of the sense in the world, uh, given the scarcity. Um, and then, uh, it seems that Target has evolved, so from, uh, the, what was going to be a joint venture with Oracle and, uh, SoftBank, Um, to what now seems like, it seems like it's more like an umbrella term for the complete strategy these days is a, is a fair way to describe it or?

A Yeah, yeah. I mean, we look at Stargate as our compute strategy, and it, uh, is varying degrees of, uh, us designing or building the compute for ourselves. Uh, for example, with, uh, with Oracle Close Partnership, uh, we help them, uh, design, we help them with how to operate AI compute, which is effectively a new kind of compute. Uh, for us, uh, we work with SoftBank Energy, which is public. Uh, we basically have co-designed the one shell with them, and they're executing on that one shells, and, uh, we will be kind of figuring out how to operate our chips, uh, in these data centers ourselves, using the new chips. So Stargate to us is that umbrella strategy for across all of these different things, and think of it as an evolution That will continuously be on, because it's never going to be tomorrow we wake up and do only one kind of way of building compute.

AI assessment note: “Stargate to us is that umbrella strategy for across all of these different things”

Q Okay. So therefore, uh, financing is outsourced to, uh, apartments. Okay. We talked about jalapeno, like, let's, let's go into a bit more detail there. Um, because in particular, it seems that, uh, you guys went incredibly quickly in designing it. Like I read somewhere in nine months from design to tap out. So maybe Maybe walk us through that, and what was the reason it went so quick?

A Yes, it was incredibly quick. Uh, nine months is very, very fast. Probably the fastest I've seen in my career. Uh, I think several reasons. One is, uh, it, it's a team. It's a strong team. Uh, they have, many, many of the team have designed TPU chips at Google in the past. So very well-experienced team. We have a great partner in Broadcom that has a very strong track record of delivering XPOs, ASICs. So I think a strong partnership with Broadcom and, and making this happen. Three, I think, ah, perhaps an OpenAI unique point, ah, in most chip companies, or when you design chips, you don't know what you're designing it for, because you're a vendor, the customer who eventually runs the workload is someone different. Ah, there's a unique advantage here of us knowing what the future models might look like, and therefore, being able to short circuit A lot of the decisions you need to make, design decisions you need to make on the chip side, so that, that's super helpful. And finally, increasingly, AI itself, helping design and optimize the chip, that is usually the one that takes the longest time, because human, you're basically limited by how many human, human, much human time is there to process all this data and run the experiments, and we can do a lot of those iterations much faster.

AI assessment note: “I think several reasons. One is, uh, it, it's a team.”

Q are the biggest roadblocks right now? Uh, in, in particular, do you think that the, uh, issue ultimately is more just like technical capabilities? And we'll talk in a second about some of the stuff that you guys have built, or is that a human question of like trust and just accepting to have the machine do things for you? Especially when your money, your personal money is at stake.

A Yeah, I think the, the two primary, um, blockers that, you know, we'll need to move through to really scale this up are, um, one trust, and actually we've, we've done a lot on the trust side. We talked about the shared payment token. Agent doesn't have any access to the credentials. Every shared payment token includes radar scores, right? Both is this a legitimate buyer and is this an agent acting in a legitimate way on behalf of the buyer? Um, Maybe we'll talk a little bit about link as the wallet for agents, but we've done a lot so that consumers can set guardrails around what the agent can spend, right? So it's, it's a little different than like a one-time use virtual card, which are like, uh, pretty maniacally scoped credentials, but in the case of a link wallet, you, you very much have, have the guardrails to set, but even with the sort of trust layer from a, from a technology or infrastructure perspective, like, I just think it takes time for any market to Build trust, especially when you're talking about, um, making decisions for, you know, what I buy and spending my money. Um, I think it's very natural for humans to kind of want to build their way up to that, and so, um, that's a, that's a big reason why on the consumer side, what we're mostly seeing is people are discovering things inside AI apps, but they're still choosing the exact thing, and they're still disproport…

AI assessment note: “two primary, um, blockers that, you know, we'll need to move through... are, um, one trust”

Q And technically, is that a one-time usage credit card, or is that completely different?

A Good question. Okay, so the very, like, when we were talking 12 months ago, the very, very first version of Agentic Commerce on Stripe, and I think in the world, was these one-time use cards, and, uh, actually the genesis of that was one-time use cards that we used, um, for, like, platforms and marketplaces, right? Like, when I order, um, a salad from DoorDash, the driver has a one-time use virtual card, uh, that can be used to pay for my salad, and That feels good to me because, uh, neither the driver nor the restaurant, which I don't have any affiliation with needs to see my payment credentials. And so, uh, in the first version of agentic commerce, which for us, uh, I believe the very first, uh, meaningfully live volume was on perplexity shopping. We used, uh, these, these one-time use virtual cards. And basically the, the human consumer would say they wanted the thing, their payment credentials were Used to basically buy this, like, one-time fund, this one-time use virtual card. The agent was handed this one-time use virtual card, and then they went off on the internet and, and, and purchased with it. And that was, uh, very scoped, uh, by definition of being one time, and it was generally, um, you know, scoped to a single provider and scoped to a very fixed amount. Uh, I would think of, uh, link wallet as much more flexible. Right? So I can set a budget, uh, to be used acros…

AI assessment note: “I would think of, uh, link wallet as much more flexible.”

Q Very cool. All right. So that's Vibe deployment. Tokens as money is, uh, also a fascinating topic. It's like taking it from the, the, the top. What does it mean to monetize tokens from a Stripe perspective?

A It's a big question because what does it mean to monetize tokens from a world perspective is kind of like, okay, actually, how's the whole next generation of B to B and some B to C going to monetize? Look, like, the, the, the last kind of decade plus of SAS, uh, had, like, pretty beautiful and simple to monetize economics, right? And in particular with SAS, like, you build a product once, and then you get one more customer, and it costs you basically nothing to serve them, right? Marginal costs are near zero, and that's why SAS margins are really good, and that's why, you know, the fixed fee subscriptions or seat-based licenses work really well, um, in SAS. Um, AI, and I say AI generally because like you could literally be selling LLMs, but you could also be selling, you know, some product that's a wrapper on top of LLMs or a product that's heavily powered by LLMs and requires a lot of tokens. Um, it breaks that model because obviously every prompt and every API call and every task, um, has a, a real marginal cost all of a sudden, which it didn't have in SAS. The inference isn't free. And so you now have all these businesses where How your customers use your product directly determines, uh, whether you make or lose money, and that's a very different game, um, and, uh, you know, from our vantage point, like, part of what that boils down to is the need for usage-based billing as …

AI assessment note: “part of what that boils down to is the need for usage-based billing”

Q And I think we talked mostly so far about agents buying, at least like in the example we gave where the agent represents the consumer. Presumably there's a concept of agents selling as well. Like what's the, what does this look like in the future? Like if all technical problems are solved and adoptions happen, is that basically Two agents negotiating something. What is the ultimate vision?

A So when I, when I, when I like step back with my economist brain, I'm like, that would be really efficient, right? Agents are really good at discovery. We've already seen that. Um, agents are really good at integration. Agents are pretty good at like finding optimal pricing, matching, negotiating. They're incredibly persistent. Uh, their time is worth a lot less than human time and they can get those, get those back and forth done. Um, Much more, much more quickly, um, and then they're also actually really good with, like, integrating and actually adopting the thing, so especially if you think of, like, B to B buying, and maybe we can talk about, um, Stripe projects a little bit later too, but just, like, actually, like, not just finding the service and negotiating it for the price and contracting on it and buying it, but actually, like, getting all the way to integrating and using the product, um, I think agents are going to help with a lot, and so I'm, I'm definitely Um, imagining an economy that is much more efficient because you have agents on the buy side, and as you note, also on the sell side, and they're kind of hyper-efficient on all of those dimensions, uh, which of course, um, you know, in the, in the, in the, uh, Ronald Coase, Nobel Prize winner Ronald Coase version of the world would basically, um, just like remove frictions for firms to work with each other, would…

AI assessment note: “imagining an economy that is much more efficient because you have agents on the buy side”

Q You mentioned Tempo at some point, should we cover Tempo? Like, is, is how relevant is Tempo to the agentic commerce conversation?

A So I think there's, there's a couple, um, components of our work with Tempo that I think are, uh, really interesting. So, um, when you think about agentic commerce on the business side, we launched the machine payments protocol or MPP and we built it with Tempo and it's an open standard and the way it works is, um, Quite elegant, right? Like an agent requests access to a service, um, and it can be whatever an API or an MCP server or whatever, and then the service responds with a payment request and then the agent pays, um, and there's no kind of account creation or checkout UI or human in the loop or sort of the way you and I would traditionally, um, engage on the internet. It's just this very kind of machine readable, standardized way for agents to, to buy from businesses. Um, and that is really, uh, MPP is really the primary, um, mechanism that we're seeing businesses use, uh, for kind of agents, agents as buyer. Um, the other, uh, collaboration with, uh, with Tempo that I am, um, super bullish on is, is more related to, to fraud, uh, because, you know, agents are increasingly becoming the users of AI products, and agents can burn through tokens very, very quickly, And so we were talking a little bit about this, like, uh, dichotomy that a business faces where either you can like siphon off self-serve and be, uh, really safe, but grow slowly or open it up, including to agents,…

AI assessment note: “we launched the machine payments protocol or MPP and we built it with Tempo”