The Exchanges

Q the NoSQL, uh, world, so there was this, um, term, uh, that was, uh, polyglot persistence, which was this concept of, hey, it's great if you have multiple use cases, and you should use different databases for each one of those use cases, um, And I mean, it seems that people are starting to evolve away from this. Is that, is that how you think about the world as well?

A Uh, it is. I think it's less of an evolution away from it. I think if you look back 10 years, people were sort of struggling with lots of different problems, and it was very easy and sort of very pragmatic to find a product that solved a problem. Like, I need to scale this service. I need to scale this application. I need to do something. And I don't want to, like, choose a platform for the next 10 years, because frankly none of them are ready for that. So just find, you know, get me a product that solves a problem, gets me out of my current situation. And so that's what people did. So they had their big relational database, that was sort of their standard thing, and they tried some new database in order to sort of solve a problem. As all of these things have matured, as MongoDB has matured, it can solve more and more problems, and no one I ever talk to wants to have 15 different databases in production. It's just not good for anyone. It's not efficient, it's not easy to manage, it's very frustrating for people internally. They don't want to have one, but they want to have, they want to have a handful of databases. And it really comes down to is how you store the data, and what kinds of things you can do with it. And so in MongoDB's case, the thing that, you know, one of the things that I love about the document model, is the document model really, in many ways, can be a supers…

AI assessment note: “Uh, it is. I think it's less of an evolution away from it.”

Q from a product standpoint, so, so that you guys, ah, actually you were the first one in the market to, ah, partner with Databricks on, on, ah, on Spark and all the things. What, what does Hadoop look like in, you know, three, four years? Um, because we've evolved from batch to real time. What are, what are the product needs and what technology address, will address those market needs?

A The system that we commercialized in nine, ah, in 2008, Looked very little like the original software that Doug Cutting and Mike Caffarella created and that Yahoo developed in 2005, 2006. It had already evolved a long way. These days when people talk about Hadoop, what they mean is HDFS and MapReduce, yeah, yarn for resource management. You need some ingest tools, so you need Scoop and Flume, and you might even be looking at Kafka right now. Um, everybody is super hot on Spark. The technology is way easier to program. You know, it's got some rough spots. It's not well integrated with security framework yet, but it will come along. We like Impala, but you go around the industry and you'll hear, you know, 30 different MySQL is better than your SQL stories. So what we've seen is a proliferation of processing and analytic engines with a whole bunch of supporting plumbing, right? Data ingest, oh, by the way, security and data governance and data lineage and so on. And a steady improvement in the capabilities of HDFS, right? This is a really different platform than we brought to market in 2000 eight. The only thing that it really has in common, in my view, is that it's still called Hadoop, right? HDFS MapReduce is so widely deployed that it will always be used. MapReduce is so widely used that it's always going to be there. But I think we're going to see most new workloads embrace so…

AI assessment note: “percentage of cycles spent on MapReduce in Hadoop clusters generally is going to asymptotically approach zero”

Q So you guys created that concept of behavior specs. So walk us through what that is. Uh, very practically, is that a marked end file? What, what, what does it look like?

A Yeah. So the, the original idea, actually, actually my co-founder Matt came up with the idea, uh, literally about two years ago, we were talking about auto AGI. Back when we were, Um, uh, even back when we were, had agents that weren't fully, I guess, agentic as you think of them today, and they, you know, had restricted choices, even back then you still wanted to think about, okay, what kind of choice do you want it to make at this, at this fork in the road? And so Matt, we actually used to call it internally meta behaviors, uh, because the idea was that it was a, you're defining, uh, a behavior, but it's at a meta level because it's all the, across all the behaviors agent will have, you know, in all the different trajectories. And, uh, so the idea was that instead of trying to write the prompt, you have to first align on what the meta behavior is. Uh, and so that was actually the first purpose of this concept. I swear to God, literally two years ago. Um, and over time that kind of evolved, um, and we ended up calling it behaviors just cause it's, it's a bit simpler. Uh, and the idea is that you have a, a markdown file in which you write down, how do you want an agent to behave? Simple as that. It could be, it could be at varying degrees of granularity. So let's say you have something that's like very specific, like, um, you need to go to like, look at the primary sources. May…

AI assessment note: “you have a, a markdown file in which you write down, how do you want an agent to behave?”

Q This is like walk us through, like how it all came about. You started the first company as a student or right after your PhD?

A That's correct. Um, actually during our PhD. So my co-founder Donna and I, we met at MIT as PhD students, uh, over 20 years ago. I, I now cap it because we're just old. Um, but it was, it was a fun sort of research project that we worked on, which was, um, this was around the time that Wi-Fi was emerging as a new technology. We built a, uh, research project called RoofNet. So we covered essentially the city of Cambridge, the area between MIT and Harvard with free Wi-Fi in the early 2000. So that was really exciting. It's like a hands-on kind of very practical research project. We did a bunch of academic research on routing protocols and, you know, how to build the network. Um, but the first company Meraki came out of that project, which was We thought it was tremendously cool, this idea that Wi-Fi could connect so many people, just incredibly useful. We wanted to help other people build big networks. And so we essentially took that research, um, and, and now I would use the word distilled, like we condensed it down to, uh, you know, run in a box that other people could build networks out of. And then we started essentially making that product available. So that was Meraki. Um, to be honest, we kind of thought of it as a project. Like we weren't even thinking of it as a company. Uh, we kind of bootstrapped the business in Boston. We ended up moving to California. Um, and it was …

AI assessment note: “That's correct. Um, actually during our PhD. So my co-founder Donna and I”

Q And then what, what has, um, uh, generative AI fundamentally changed for, for you guys? Um, is that video reasoning? What, what, what do you use for, for what?

A We use generative in a few different ways. I would say if we think about the overall class of models, yes, you can now reason about video. So, uh, what would have required a human in the loop reviewer, um, or, you know, and that's the kind of work that might've been done overseas in lower cost geos or something like that. You can now do at much more volume, um, in the cloud using these models. So, uh, for example, if someone slams on the brakes, right, while they're driving their truck, um, the, the naive thing to assume is like, hey, the driver was distracted and they, they, they kind of woke up. The more nuanced thing is that driver might have been avoiding a deer or a dog or, you know, some kind of, uh, defensive event. If you can watch that as a video clip, you can now say, hey, we're actually going to give the driver some positive feedback because they did a really good thing. The VLMs are able to effectively do what I just said, right? Um, similarly, like, if you want to understand, did someone run a red light, right? These things happen. You need to have a pretty sophisticated model that understands the geometry of the road and all the conditions and so on. So that would be like a JEPA-style model, for example. So we're able to use a few different model families. On the generative side of, like, actually being able to create Video. That's also very interesting because fr…

AI assessment note: “We use generative in a few different ways... you can now reason about video.”

Q so I want to do a bit of a deep dive on that on itself in a, in a second. Um, uh, but to close on this, so three bottlenecks, um, you also mentioned CPUs a couple of times in this conversation, and there seems to be a theme around the emergence of like a CPU shortage as well. Is that, so what is that true? Two, what causes it?

A So agentic AI Is a world in which AI doesn't just provide answers. It initiates action. So that action might be go to a website. It might be learn, gather some data from a website, bring it back, take another action. Those actions are done by CPEs. And so as AI gets better and better at doing things, at making instructions, calls for things to get done, we're using more and more CPUs, right? And that is driving up the consumption of CPUs and therefore the demand for CPUs. And so this sort of huge push for more CPUs is being driven by AI On machines like ours and GPUs doing agentic work and asking the, the, the CPUs to take an action, to go to a website, to order a burrito, to find a piece of information, to pull it from storage to all that work is being done by the CPUs.

AI assessment note: “huge push for more CPUs is being driven by AI”

Q Great. Let's talk about that over in that deal since it's such like a major historical milestone record making. Um, so it's, it's providing up to with seven and 50 megawatts, which is, which is interesting by the way, as a, as a metric because we're in a chip provider, but this is power. So is that shorthand for?

A It's a shorthand. I mean, it turns out right now, and we didn't talk about this because there's Sort of in the adjacent supply chain. We, we went through the shortage of, of memories. We went through the shortage of a process called COOS, three nanometer capacity. The, the other limitation in our industry right now is data center availability. And, I mean, Uh, that is a limiting factor for everybody. And that's why Anthropic did a huge and sort of very expensive deal with Elon for data center capacity. Um, uh, our deal with them, with, uh, with OpenAI was because data center capacity is a limiting constraint, measured the way data centers are measured in, in megawatts. The deal is 760 megawatts, 250 megawatts in 26, uh, on a multi-year lease. An additional 250 megawatts in 27, on a multi-year lease, and an additional in 28, a multi-year lease.

AI assessment note: “It's a shorthand. I mean, it turns out right now”

Q addition to being the application business, consumer and enterprise, and then being in the model AI research, uh, business, and then the computing data center business, since that OpenAI is in the, Uh, chip business, if that's fair, sort of completely full stack, uh, but I'm, I'm curious, and we'll, we'll go into some details about, you know, later, later, but into, like, overall strategy, where, where does that fit?

A As we begin to, as a pretty big fraction of the world's population, AI usage is exploding, uh, inference is obviously becoming a big fraction of our work. It's consuming a lot of compute. And one of the other realizations is because we know what is the work toward exactly, what is the model we want to run, uh, we can co-design the hardware. To be super efficient in delivering those models, right? And so the, the strategic thesis we have in here is how do we take advantage of knowing what the end workload is, what the model itself is, and design chips, um, that are very efficient in serving those models. So it really allows us to drive efficiency advantage, drive more tokens per watt. So the key metric that Jalapeno is optimizing is maximizing the number of tokens you can produce per watt. And because the world is constrained by power today, so the more tokens you can produce for the same amount of watt water, it's better for everyone. So we look at it as a very critical ingredient in scaling how we deliver intelligence to the world.

AI assessment note: “look at it as a very critical ingredient in scaling how we deliver intelligence”

Q You touch upon some, some important developments that happened since we last chatted in terms of like overall maturation of the, uh, industry. You mentioned the agenda commerce protocol, which I think came out Last fall. What is that? Do you think that's something that you guys built in partnership with OpenAI?

A Yes, we built it in partnership with OpenAI. The agentic commerce protocol is just a, uh, standardized way for businesses to work with agents. And there's a couple different components of it. Um, and this is sort of wrapped in our broader agentic commerce suite. Uh, one is how do businesses expose their Product catalog, their inventory, their prices to agents. Um, and you know, you could argue, oh, the agents could go out and, uh, you know, search or infer, but sort of inventory is a thing that you want, like, deterministically known. Um, and we don't want businesses to need to kind of register their product catalog or register their inventory with every single new agent that comes online, because in the same way you and I Like to work with a lot of different model providers, uh, or a lot of different models within those model providers. In many cases, both. Uh, we, we similarly are seeing businesses not want to place bets on just one agentic surface. Um, they may be selling B to C and B to B. They may be wanting to reach a wide swath of consumers across many different, um, surfaces. And so agentic commerce protocol lets them, um, expose their product catalog once, um, and then opt in to, um, all of the agents Uh, who, who, uh, work with, with that protocol. Um, it also includes the, the shared payment token, and so this is about making sure that, um, in the moment of transacti…

AI assessment note: “Yes, we built it in partnership with OpenAI. The agentic commerce protocol is”

Q It's been dead for years? Like, why, why is that?

A Well, you just look at the, the progress, uh, in semiconductor manufacturing, you know, the, the original statement of Moore's law was economic, right? It was about, we can afford to put twice as many transistors on the same chip in every, whatever, 24 months, whatever the, the time period is. And, um, these days that is, Absolutely not the case. It hasn't been for probably five or 10 years, right? Now, we are still scaling our systems, right, um, through a number of ways. One is just applying a lot more silicon to it, right? Uh, we are also getting, transistors are continuing to get smaller and, and more efficient, although at a slower pace, but they're also getting quite a bit more expensive at the same time. Um, uh, so the, uh, you know, in an era where, where Moore's law was alive, the best way to make the system of the future was to take the system of the present and then just shrink it and, and maybe double it at the same time, right? But in an era where, where we've been living for a while now, where you don't get economic benefits from taking your existing design and shrinking it, uh, you really have to be more clever about how you use every part of the system. Uh, that, that's, uh, you know, an era where accelerated computing is, is much more valuable than ever because the, the work of thinking through the prop problem from first principles and co-designing absolutely …

AI assessment note: “you don't get economic benefits from taking your existing design and shrinking it”

Q To double click on this, uh, at a high level, NemoTron family is focused on urgent reasoning with a particular focus on making it efficient. Is that, is that the right headline?

A That's right. Yeah. Um, NemoTron has always been, um, uh, speed first approach to building models because NVIDIA is an accelerated computing company. As I was saying, we're trying to think through what is the problem here computationally from first principles. And, um, you know, NemoTron, ah, three family has a lot of things in it that are, ah, we're really proud of. For example, ah, NemoTron Ultra and Super, ah, were pre-trained using four-bit arithmetic. We pre-trained those in MVFP four, um, which, you know, ah, is, ah, a not trivial thing to do, to invent the algorithm so that your model can converge to an excellent result using such coarse arithmetic, ah, required a lot of invention. Really proud of that.

AI assessment note: “That's right. Yeah. Um, NemoTron has always been, um, uh, speed first approach”

Q Great. Can you talk about the multi-token prediction, uh, which is also very interesting?

A If you're running at a low batch size, um, which is when you are trying to get the most interactivity if you're in a data center, so you want, you want the model to respond as quickly as possible, and it's okay for it to be more expensive. Your token, your cost per token might be higher, but you want the result as quickly as possible. Or if you're running, um, locally, um, so you might be running At batch size one just because you're the only person using it. It turns out that the GPU has extra execution capabilities that are just lying there unused. The bulk of the work when you're running in these scenarios is actually fetching the weights from memory, and then you push the token past those weights, and then you fetch more, more, uh, weights from memory. But it turns out if you, if you push two tokens or even five tokens, Through those same, um, weights, it would cost basically the same amount of time because the, the expensive thing is not doing the math to push the token through the weights. The expensive thing is just reading all of those weights from memory, all those parameters they have to come in. And so the idea with multi-token prediction is to take advantage of this by having the model predict multiple tokens at once. Let's say that the model predicts five tokens. We know the first token is correct. The next four tokens may or may not be correct. So then what we do …

AI assessment note: “having the model predict multiple tokens at once.”

Q Yeah. The, the concept of network, uh, is, is fascinating and seems very central to the story. And it's a network of users, but it's also physical network, right? Like what, what you guys did initially was to create this edge network. Like, how does one even go about building that as a young startup?

A I mean, in the very beginning, um, you know, we, when, when we, when we launched, um, we had, um, servers in five cities around the world. Uh, the, the Chicago, Ashburn, Virginia, uh, San Jose, California, um, Amsterdam and Tokyo, except Tokyo sort of, if we had been totally honest, deserved an asterix because we couldn't totally figure out the routing. And if we got it wrong, a whole bunch of traffic from the U S ended up hitting, hitting Tokyo. And so we would literally turn the Tokyo, um, servers on and off throughout the day, depending on, uh, you know, where, where, who is, who is awake. Uh, around the world. Um, today, fast forward, and you know, we're in over 350 cities worldwide. We're in over a thousand data centers, uh, worldwide. And, and we will go from, you know, the minimum deployment is sort of like a rack of servers. In some places, we'll have hundreds of racks, uh, of servers, just depending on what's, what's needed there. Um, but I think that that was the, that was, you know, a huge focus, um, behind the scenes, was how do you figure out How to get into all of these places around, around the world. And it was, um, I mean, it's, it's a, a series of, of kind of stories of how did we create enough value to, you know, the ISPs, uh, around the world, the, the, the internet providers for consumers that they would invite us to be into their, into their buildings. And…

AI assessment note: “how did we create enough value to, you know, the ISPs”

Q And when we say bots, uh, do we mean agents? Do we mean crawling bots? What do we mean?

A I mean, I think that Agent, bot, crawler are all synonyms. It all means the same thing. It depends on whether you want to describe it in a sort of normatively positive way, in which case you call it an agent. If you want to describe it kind of the ickiest way, you call it a crawler or a scraper, but it's the same thing behind the scenes. It's, it's a, it's a machine that's accessing resources as opposed to kind of a human eyeball driving a browser, accessing those, those resources. And, um, and again, in this particular case, Um, you, you still have a whole bunch of this sort of scraper, crawler, hacker traffic. That hasn't changed that much. You still have some of the sort of bots like Google bot that are, that are building search indexes. That hasn't changed that much. If anything, it's gone a little bit down. Um, but the thing that's really driving this is what I think we would all call agents, which is, you know, if, if I go to chat GPT or I go to Claude or I go to, you know, any of these systems and, and I'm shopping for a digital camera. You, you're gonna see, you know, thousands of times more site visits for that same activity, that same job to be done of, of shopping for a digital camera than if I did it myself. I might visit five sites, uh, if I'm, you know, personally trying to figure out what digital camera to, to buy, whereas my agent might visit 5000. And that's, t…

AI assessment note: “Agent, bot, crawler are all synonyms. It all means the same thing.”

Q In terms of breaking, what, what does an agentic internet mean in terms of security?

A Yeah, I mean, it's, it's, that, that security's gonna be, um, I, I think for the next two years, uh, we're going to see just a series of really scary, um, things happen, happen on, online. Um, you know, the, the, probably the worst kind of bug that was out there for a long time was this thing called log for J. Where you could send a very simple command and compromise, um, uh, almost any, any server running anywhere. Um, you know, I think for, for the next, um, uh, you know, a 104 days, a 104 weeks. So two years, um, you're going to see a log for J like vulnerability every single week. Um, because like, and again, we were, we were part of, Glasswing and, and got to use, um, you know, mythos to, to look at software. We, we get early releases of, of, um, you know, various, um, uh, open AI models. We've built our own security models. These models are incredible at finding vulnerabilities and they're going to find them like crazy. And I think security is going to be, it's going to seem really scary for the next little bit. There's going to be a bunch of VCs that invest a bunch of money because you're going to see just, You know, an explosion in cybersecurity companies. They're stopping this. But what's interesting is I think it's going to be a little bit of a head fake. Um, because I, I think that two years from now, um, what's going to happen is software is going to get just massiv…

AI assessment note: “for the next two years, uh, we're going to see just a series of really scary”

Q How real is the movement against data centers from the global community and how do you think about, uh, how to respond to it?

A Well, it's certainly, it's like very popular in the news right now. I'd say that, um, it's definitely very real. I mean, I think that rightfully communities that host any type of large capital project, whether it's a power plant or a, uh, solar farm or a data center or a distribution center, right? Those communities want to have a seat at the table. I'd say in general though, I spend a lot of time reading through a lot of the comments from communities, and people want jobs. They want tax revenue. Any major capital development Is going to bring a lot of tax revenue and it's going to bring a lot of jobs and it's going to bring investment into their community. And what they really are voicing, I think is one is having a seat at the table while, while this stuff is, you know, being developed. I think that that's an important thing is just to have their voices heard and that, that the developers coming in and actually understanding the community. The other thing to kind of, I think, keep in mind is That there's a lot of misinformation out there. So for example, Every single modern deployment of, let's say, a Blackwell class or a Rubin class GPU, you know, the VR, GBN VR GPUs. Um, these are oftentimes in a closed direct-to-chip liquid cooling system that's connected to a dry cooler, which means that there's almost zero evaporation. It's not using evaporative cooling. It's using a dry…

AI assessment note: “I'd say that, um, it's definitely very real.”

Q Super helpful. If two companies have the same chip fundamentally, how do they extract more value from it? What, what needs to happen to maximize the Usefulness of that chip.

A If you look at the cost structure of let's say one GPU hour of time, you know, we were talking about H 100. The largest part of that cost structure is the depreciation that is associated with that GPU hour. And, um, basically you can think of a utilization metric as being like kind of a multiplicative factor on that. So one over the utilization. So if you, If you use your capital asset, 50% of the time, you will have on a per hour basis, twice one over 0.5, the amount of per hour depreciation expense associated with that. And so I think that the number one way that companies are, you know, sort of gaining a unique advantage is, well, how can I build a cloud product that is beloved by people that is going to drive a high utilization? And, um, you know, in addition to that, the market, as we mentioned earlier, for on-demand compute basically The retail pricing is obviously much higher than the wholesale pricing. So the retail is like on demand, spin up a GPU, spin down a GPU, normal cloud service. The wholesale is sort of buying 10,000 GPUs for five years, for example. And so one of the things that we do at Lambda is really try to figure out, hey, how can we sort of get the most dollar utilization and percentage utilization out of The capital deployments that we do. And that's, that's by making great cloud software that makes it easy for somebody to spin it up and down. So for ex…

AI assessment note: “number one way that companies are, you know, sort of gaining a unique advantage”

Q And, uh, quickly, what was your path to OpenAI? So how did you go from studying physics to being where you are today?

A I did a PhD in theoretical physics, uh, from, from MIT, thinking about the intersection of quantum gravity and quantum information. Thought a lot about black holes and quantum chaos kind of thing of what if you throw something into a black hole? What happens to the information? Does it, does it come out? How, if we think about black holes as computers, how fast are they? I was very interested in this fundamental question in theoretical physics, which is how do you find a quantum theory of gravity? I also got very interested in this interplay between computation and the laws of physics. You know, any computer exists in the universe in, in, you know, behaves according to physical law. So the sort of computations you can do are bounded by the laws of physics, and there's some sort of interesting relationship there. Black holes are pretty interesting because they sort of saturate some conjectured bounds around processing of, of information. And from there, uh, I did a postdoc at the Institute for Advanced Study. And around that time, I'm pretty old now for, at least for this field, so that was about 20 16, uh, was when the DQN Atari paper from DeepMind happened in 2015, and then AlphaGo was in 2016, and I got very excited about these, um, about the possibility of, uh, machine learning, and then, and then deep learning was statistical Science that lived in a similar framework to the…

AI assessment note: “I did a PhD in theoretical physics, uh, from, from MIT”

Q And, uh, going back to, um, generalization as well, are there examples where, um, actually getting better at one domain makes the model worse at, uh, the rest? A little bit, uh, to what you were saying about, like, some people are very good at math. Some people are very good at English. Pretty often they're not the same people.

A In domains, usually not. What will happen though is, um, You will make decisions based on which domain we optimize for. And if you optimize for one domain, you will be able to optimize less for another one. So it's not necessarily that optimizing for one thing will make the other one worse. It's just that as a result, you can optimize less for the other one because your compute constraint, your data constraint, you have like, like your, your, uh, human bottleneck also in terms of that work. What does happen is, uh, you can have negative kind of generalization, like bad generalization or negative transfer. More for these horizontal aspects of the model. So I'll give you a very concrete example. Explicit instruction following versus implicit instruction following. If I, if I have a model, and this is, we often hear, for example, from OpenAI models, that they tend to be really good if you tell them exactly what you want. Um, but as a result, sometimes we hear also that they're, like, less good if you are not as, as specific about what you wanted. For example, if I make, if I make a typo, and I say, like, change this file, and I make a typo in this file, um, an extremely good model at, like, explicit instruction following will change the wrong file, the one that has a typo. But, like, humans would probably realize that you made a typo. Um, and, and like, as a result, there are case…

AI assessment note: “In domains, usually not.”

Q Does reinforcement learning, uh, create new capabilities, uh, or does it make the model better at existing capabilities?

A It's really hard to say because pre-training, when it's trained on all of the internet, arguably already has all capabilities in it. Um, so it's, it would be even hard to answer this question scientifically, um, because arguably everything is, is already there. What I would say is that if you look, uh, at models that we were training or that we were post-training, like, two years ago in the open source world, uh, for example, I, I worked on one of them, Alpaca, where we used 50,000 examples for SFT, and, like, now when you look at reinforcement learning from From models like Kimi or, or from DeepSeq models, it seems that they are closer to one million data points. So definitely people scaled up a lot the reinforcement learning stage. Um, and from this, it seems that they've learned like new capability, like this reasoning aspect, this fact that you can check your answer and, and, uh, and try to improve it. Um, so you can, you can really think for longer to get, to get a more correct answer. So all this to say that Arguably everything is already in pre-training, but we were definitely able in the last one year and a half, even in the open source world, um, to have more capabilities after reinforcement than we used to before.

AI assessment note: “it seems that they've learned like new capability, like this reasoning aspect”

Q And is there a fundamental technical difference between ephemeral and long running? So if you have an agent that runs for 24 hours, how does that translate in terms of sandbox requirements?

A The reason most sandbox environments do not run forever or even is because it's a technical problem. If you think about servers underneath, these servers also have to be managed, right, and maintained. And so if your sandbox can run forever, that means that you can never reset. You can never reboot the underlying server. You can't update it. You can't patch it. You can't do all these things without turning off all the sandboxes. And so the way you solve that, the easiest way to solve that is having sandboxes that have a termination time. It's like, they will only last an hour, 24. It doesn't matter. Pick your time. And then if you decide that you have to do something with the underlying machine, you just, you just flag that machine as non-schedulable. And so at some point in time, there's nothing else on that machine. You can do what, you can fix it. You can like reboot it. You can do whatever you want. Easy peasy done. Right. And so because historically most of the workloads were femoral, you didn't have to try to solve that problem because you didn't care. Like most workloads, like a Lambda function usually runs what? Five minutes, 10 minutes, like whatever. It's not a problem, right? You never had that restraint or constraint. Now that you do, to have something that can run forever, you have to be able to live migrate the sandboxes itself between the machines so that you can…

AI assessment note: “you have to be able to live migrate the sandboxes itself between the machines”

Q And is there a fundamental technical difference between ephemeral and long running? So if you have an agent that runs for 24 hours, how does that translate in terms of sandbox requirements?

A The reason most sandbox environments do not run forever or even is because it's a technical problem. If you think about servers underneath, these servers also have to be managed, right, and maintained. And so if your sandbox can run forever, that means that you can never reset. You can never reboot the underlying server. You can't update it. You can't patch it. You can't do all these things without turning off all the sandboxes. And so the way you solve that, the easiest way to solve that is having sandboxes that have a termination time. It's like, they will only last an hour, 24. It doesn't matter. Pick your time. And then if you decide that you have to do something with the underlying machine, you just, you just flag that machine as non-schedulable. And so at some point in time, there's nothing else on that machine. You can do what, you can fix it. You can like reboot it. You can do whatever you want. Easy peasy done. Right. And so because historically most of the workloads were femoral, you didn't have to try to solve that problem because you didn't care. Like most workloads, like a Lambda function usually runs what? Five minutes, 10 minutes, like whatever. It's not a problem, right? You never had that restraint or constraint. Now that you do, to have something that can run forever, you have to be able to live migrate the sandboxes itself between the machines so that you can…

AI assessment note: “you have to be able to live migrate the sandboxes itself between the machines”

Q Where do skills fit into the picture of co-work?

A So skills are essentially just markdown files that explain to the model how to do things. And I'm always surprised at how well this works. If you treat model, the model Claude in this, in this case, like a coworker, you get very, very far. My recommendation to like everyone I always talk to is just treat Claude the way you would treat a coworker. So a skill is fundamentally just a text file. And in the text file, you explain how to do a certain thing. My default example is always say booking a flight. At Anthropic, we have a specific particular vendor that helps us with our travel booking, so you can't just go to Google Flights, you need to go to this, like, particular vendor portal, and then we have various travel policies, and the same way I would explain this to a co-worker, I can explain it to the model. I'll just make a file that is like, here's how you book flights. You go to this website, and on this website, please consider the following things, and then maybe you also sprinkle in, like, a few personal things, right? Like, in my case, avoid red-eye flights, but also, I do actually enjoy my weekend quite a bit, so, like, Try to book a flight. If I have to fly to New York from San Francisco, try to like take the four p.m. flight. That's my favorite flight. And you put all of those things in the text file and the model then is extremely capable of understanding the instruc…

AI assessment note: “skills are essentially just markdown files that explain to the model how to do things”

Q So, uh, detailed system prompt, planning tools, sub-agents, file system, is that the list of, like, core components of the modern agent architecture?

A Those are the four that when we launched Deep Agents, and so the story behind launching Deep Agents was we saw, we saw Manus, we saw Cloud Code, we saw Deep Research, they all had these four things, um, and we were like, okay, that's, that's pretty common, let's put it into a Python package and, and, and, um, and make it easy for people to build their own versions of that. So those were the four things at the time. Those are still probably the core things. Some other things that are frequently used. I mean, uh, bash and executing code is, is a big one that's not always used because, uh, sandboxes like Daytona are still new. And so people are still discovering how to run them and how to manage them. And so it's often easier not to do that, but we're seeing more and more want to do that. Um, and so that's where things like sandbox has come in handy. Skills are new primitive that didn't exist when we launched deep agents, but are now very, very, very interesting.

AI assessment note: “Those are still probably the core things.”

Q Since you mentioned sandboxes a second ago, since we are the Daytona Compute conference, Zetona being a leader in sandboxes. Let's talk about the compute layer of agents for, for a minute. So starting at a high level, why do agents need a sandbox?

A Yeah, I think, uh, the main reason in my mind, and you should have Ivan on to definitely correct me, but the main reason that we see so far is to, is to write and run code. Um, so, uh, I would draw a distinction between kind of like file systems and sandbox. Access, as mentioned before, you could have a file system interface that actually does not exist in an actual file system. But if some of those files are code, you might want to run and execute those code, that, that code. Why, why, why is that interesting? Why is that valuable? One, like this code could just be like scripts that are loaded beforehand, but they, but you can parameterize them. You can call them as CLIs or something. And that lets the agent, it's a different form of tool calling, um, that can often be easier. Two, the agent can write its own code and then run it. And in particular, like this last one is, is Why you need sandboxes. Anytime you want the agent to kind of like run untrusted code or do arbitrary things, you don't want that kind of like happening on a, on a shared server on your, or even on your like, uh, local computer. I think you see this a little bit with the open claw stuff, right? Like open claw, um, you know, uh, it, it does a bunch of things under the hood, including kind of like writing and running code. That's why people are buying Mac minis as a, you know, um, primitive way of sandboxing…

AI assessment note: “the main reason that we see so far is to, is to write and run code.”

Q And, uh, voice cloning, is that a, is that a use case?

A Actually with Gradium, we have the best of, um, of the, of the industry. And the best means not only replicating like the specific characteristic of someone, but I mean, also the accent, some unusual recording condition. And so in a lot of contexts, for example, if you want to create a vintage sounding character with old radio effects, and it's something that we do pretty well. If you want to have A robot voice. We can do that pretty well as well. If you want to have any kind of accent or speaking style like posh or, uh, more, uh, you know, laid back or more urban, any kind of social aspects of, uh, and individual aspect of voice is something that we, uh, that we replicate pretty well. I think what is interesting is that, uh, voice learning in itself where I see the most potential is about Uh, creative, creating interactive experiences around, uh, licenses. I tried to pitch, for example, you know, who wants to be a millionaire? Uh, there is a video game, uh, and the questions can be generated on the fly. You would like the host to, the voice of the host to pull on them all the time. So this one makes a lot of sense with cloning a specific voice because you want to replicate the voice of a character, of a person, of an athlete, of a K-pop star, whatever, I don't know, you know, any kind of experience where people want to engage with a voice that they know. But they want to go th…

AI assessment note: “voice learning in itself where I see the most potential is about Uh, creative”

Q the year of RLVR and GRPO. So you had, uh, like a nice timeline where you said, uh, the year of RLVR and GRPO. So we'd love it if you could walk us through those techniques. So fair to say, so both of those belong to the world of post-training. Let's pick, uh, Uh, RLVR and, uh, let's start with the definition. What, what does RLVR mean versus regular RL?

A I would say RLHF is the biggest leap in LLMs we have seen in a long time, because that was taking GPT from GPT to chat GPT, you know, like the RLHF, the reinforcement learning with human feedback. And in that sense, uh, it's almost like LRVR, which is reinforcement learning with verifiable rewards took That other leap basically from just simple chat model to a reasoning model. Both RLHF and RLVR have the RL in it. So both are based on reinforcement learning. But I mean, this reinforcement learning is a bit different from the reinforcement learning that plays Go. It's almost like a special thing and a simpler thing in the context of LLMs. But the idea is that instead of doing next token prediction, just predicting what's the next token, it's more like looking at the full Answer. And then based on that answer, you give a reward, like, um, in RLHF it's, you have multiple answers and you say, which do you prefer? Or in the case of LRVR, you look at the full answer and then let's say it's a math problem. You say the math problem is correct. The final answer of the math problem is correct or incorrect. That's like the main difference between next token prediction and pre-training and then the RL here. So RLVR was kind of like popularized by deep seek R one, which was based on deep seek version three. And that came out R one came out January, 2025. And with that, they also introduced …

AI assessment note: “LRVR, which is reinforcement learning with verifiable rewards took That other leap”

Q the, the world of LLMs, you mentioned MOE and that triggers a question, which is what I think a lot of people are wondering, which is like, are we seeing Uh, real architecture breakthroughs within the LLM world, or are we effectively at this point polishing what we already have within the LLM world? What are you seeing that's moving the needle in terms of architecture, uh, improvement or optimization?

A Improvement is not so much coming from the architecture anymore. It is basically the post training, but like coming back to the architecture, I think it's still an interesting question because there are so many different architectures and Almost no one uses the same one. Like they're all very similar, but they are not identical. I think a lot of it is coincidental, uh, where there are some tweaks. And if you look at the loss in some cases on some training data and some, uh, training pipelines, maybe like moving the normalization, uh, the RMS norm before or after makes a small difference. I mean, there are theoretical justifications, but also for example, um, Olmo, Olmo three, which is very transparent. They moved the RMS norm, uh, placement. Um, so then Gemini had a post and pre-norm. They had both on both ends. And so there is some justification where, okay, ablation studies show this stabilizes the training, but while assuming a stable training, it's not gonna, I think, make your model magically perform better. I mean, this is just like people tune their cars a little bit by, you know, putting in different air filters and something like that. So I think it's on that level where you can make small tweaks, but it's not really changing the engine itself. The one thing though, what we've seen is a lot of Large architectures now using MOE that I think that's a new, 2025 thing. Of …

AI assessment note: “Improvement is not so much coming from the architecture anymore. It is basically the post training”

Q Okay. Fascinating. All right. Dan, what is your perspective on all of this?

A I really appreciated Tim's post because I think one, one thing that, that I really appreciated is that there's some, some AGI talk that if you just kind of like trace the exponential, at some point you get, you know, the, the thing that will eat up the, the universe or, or whatever, um, which I, I always found a little bit odd to, to think that way. I appreciate the, the thing in terms of the actual physical constraints, because Uh, you know, like Tim said, these are physical systems with physical inputs, um, and, uh, and actually doing physical computation. I think my perspective was that if you look at where the systems are today and you look at the, the models that we've trained, uh, we are just so far from being, from even using the last generation of hardware as, as efficiently as, as possible. So, and not to mention all the new hardware that that's being built out. So I think on the technical side, I'd say there, there are two major points I wanted to make In my post, which was one, if you look at the models that are kind of the, the really great ones, um, the, uh, that, that we know today, and I, in my blog post, I mostly talked about open source models because they talk a little bit more about how they train the resources behind it. Um, we don't have public figures behind, you know, how much open AI and, and Enthropic are using. Um, but if you look at the deep seek mode…

AI assessment note: “I think my perspective was that if you look at where the systems are today”

Q And Dan, you mentioned a second ago in reference to, uh, self-driving cars, that moment where, where things flip, switches, is turned on. Has that happened? With agents already. You talked about software singularity. Are we, are we at that moment for agents?

A Yeah, I think that so personally in, in my life, I'd say that moment was last June ish. So June, 20, 25 was the moment that, that it really flipped for me to give some context here. So what I do in my day job at Together AI is we write a lot of these GPU kernels. I don't know how popular, but the, in the, in the general ML zeitgeist, GPU kernels are thought of as kind of like the final boss of the thing that you learn how to program. They're very hard. They're very highly parallel. Uh, you don't write them like you have to write in C++, which is this old language that the old systems people use, you know, decades ago or whatever. They're not in Python, et cetera. When you're trying to hire for people who can write kernels, it's very hard. It's a very challenging skill set. It's certainly, you know, the, the, the, the, the tip of the spear in terms of, um, the, the programming strength. And last June we had this really interesting realization where we realized that cloud code, cursor agent, these agentic coding assistants were actually very good at writing these kernels. So, um, there was one week where I think I wrote like three or four different features that usually would have taken me a week each. And I wrote all of those in a single day. Um, and I was like, oh my God, this thing is making me five times more productive as a kernel expert. Um, I got my team on it. Now my team…

AI assessment note: “personally in, in my life, I'd say that moment was last June ish.”