In the future, my kids are going to have access to the workforce equivalent of millions of people worth of labor just like at their fingertips, like on their phone. If you have incredibly high agency and curiosity to solve a problem that speaks to some other human, you're going to be able to just will that into existence.
The three or four who took to OpenClaw first, it lasts 20 days. They became literally five times more valuable than the people who didn't.
Economic consequences of this technology are going to be large. I think all of us can agree. And if I were somebody in any of those industries, I'd be thinking, how can I leverage it? Let me become the expert so that I'm at the bleeding edge of this technology.
All right, everybody. Welcome back to This Week in AI. This is the new podcast from your bestie who started All-In Podcast and This Week in Startups. I am absolutely addicted to talking to AI founders and building with these tools. I think it's going to be the greatest change in the world. So, I said, you know, it would be interesting if I could go all-in on AI and bring together three of the top leaders each week in AI to talk about the week's news. And if you're new to the program, go to thisweekinai.ai.
And at the top, you'll see a link to Spotify, to YouTube, to Apple Podcast, and to our daily email newsletter. So, you know, you can go deep with me today on the program. Chase Lochmiller is here from Crusoe. You know them because they are one of the, if not the leading infrastructure companies building out this insane data center buildout and we're going to talk about it today. Maybe just tell everybody a little bit about what you do at Crusoe.
Thanks for having me, Jason. You know, I'm the co-founder and CEO at Crusoe. Crusoe is a vertically integrated AI infrastructure provider, which means we do everything from the hardware angle of building out this massive infrastructure boom to support the infrastructure of intelligence to the software aspects of orchestrating and running reliable large-scale computing workloads on large clusters of GPUs and delivering, you know, managed services like managed Kubernetes as well as further abstractions like our managed inference product that, you know, sells folks tokens. So, you know, we're sort of in the business of either renting data center space, renting clusters of GPUs, or selling tokens.
And your biggest partnerships are?
Our largest, most well-known partnership is in Abilene, Texas where we're building a 1.2 gigawatt campus for Oracle and OpenAI. And, you know, that's, I think, the largest single cluster, you know, being, you know, operational and in development right now.
And that would be called Stargate, I guess, is the code name.
That's what kids on the street are calling it. Yeah.
Yeah. Okay. We'll get an update on how that's going, but it's incredible to have you here. Last time we saw each other, we were both in Davos. So, Kumbaya Davos, had a good time running around there. Also with us, Naveen Rao, who is with Unconventional AI and they're rethinking the computer from the ground up from first principles to be more power efficient obviously for AI. So Naveen, explain what that means and what you're actually building.
We're kind of thinking about the computer that we all know and love. It's something that's an 80-year-old paradigm. You know, this was invented in the 1940s to actually compute trajectory of artillery back then. And so we've gotten a lot of mileage out of this architecture, but now we're at this point where our application is intelligence. And how can we kind of rethink what the computing primitives really are for this application? If you think about what a neural network is, it actually just runs on the physics of our neurons. There's no numeric representation or floating points or anything like that in our brains. So can we start to build circuits that actually have similar kind of properties as those physics of neurons and actually build something that's not just 20% more power efficient but orders of magnitude, like three orders of magnitude is what we're shooting for in the next five years.
How will that manifest and when will we see it? I guess.
Yeah, product-wise, I mean, this is a research project for now. And it's products we're definitely heading toward products in four to five years. It'll manifest as just greatly reducing cost. Basically, the figure of merit we care about is joules per token.
So you can think about it, and I'm sure Chase can back me up on this, but when you build a data center, it's about getting the power contract and then it's about how do I monetize every one of those watts? And so we provide the path to greater monetization for each watt.
And you think you can do this a hundred, a thousand, perhaps even more effectively than the current solutions.
So there's actually some theoretical results from the 60s which suggest that we are maybe somewhere between seven to 10 orders of magnitude away from actual physics of computing. So we think three orders of magnitude is something we can accomplish in the next few years. After that we're going to have to rethink the actual fundamental substrate, like can we do this with electronics and silicon? I'm not sure. But what we're doing will open the door to new substrates, new kind of physical manifestations of computing.
It's unbelievable. I mean Elon's always talking about like the amount of energy it takes to run our brains and how efficient they are.
That's kind of, is that the north star, the human brain when you build this new product?
Yeah, we actually want to break past that to be honest. So if you think about it, human brains are 20 watts. If I sum up 8 billion people times 20 watts is 160 gigawatts. The US has alone about a thousand gigawatts of capacity and the world is about 9,000. So we have a lot of energy in the world. We just have to use it more effectively. And just think about what we can do if we can harness that for intelligence in a synthetic world.
All right. And also joining us from Arena, Anastasios Angelopoulos.
Anastasios Angelopoulos
6:01 ↗
Angelopoulos. Yeah.
Angelopoulos. Obviously Angelopoulos. My Greek brother.
Anastasios Angelopoulos
6:08 ↗
My Greek brother.
Yes. I mean but Calacanis is hard enough. Angelopoulos. Actually, Angelopoulos is not so bad. And you're obviously running Arena, and this is when people say in the arena ranking the AI models, that's actually your company. So, this is incredible to have you on the program. And it makes sense to me that the arbiter of all language models and their performance would be Greek. That makes total sense. Since we've created science, philosophy, theater, democracy, all these things for society, we might as well be the judge of how AI is doing. Is what is your business though? We know you have the arena. People fight in the arena to move up the rankings, but where did this come from and what has it turned into in terms of a business?
Anastasios Angelopoulos
7:00 ↗
The project started as it was a research project at Berkeley while we were doing our PhDs. Wei and I and many other students at Berkeley along with Yan, and we never really approached it with any commercial sort of ambitions at all. We just wanted to, it was at the time that ChatGPT started like 3 years ago and we were all sitting around a table and thinking hey how do we evaluate this system? It's doing well on the benchmarks but it's not doing well in real life. You try to chat with it and it's dumb. Not ChatGPT but other models. And so that's how this arena paradigm emerged of, hey, let's put it in front of users, have them chat with it, two responses come up instead of one, and then they can vote for the one they like better.
And the platform has evolved quite a bit since then. You know, starting from the early days and then suddenly we grew and grew and grew in terms of users until we were at Berkeley trying to run basically this industry-scale project out of academic grants and support from labs and from VCs that were kind enough to donate money as gifts to try to keep this project running. And it had become this central force in the ecosystem for neutral independent evaluation. And at some point we decided how are we going to have the impact that we want in this world? How are we going to help measure intelligence and ensure that it's reliably and responsibly deployed? Can we do this through an academic setting? Can we do it through nonprofit? And we decided no because it would starve a bit of resources. So we decided to start a company.
The customers are the language model companies. Since then the company's grown quite a bit.
Anastasios Angelopoulos
8:43 ↗
Independent AI executives. What's the product? And traverse the tradeoffs between performance, latency, cost, and so on.
All right, I actually have a question. How many users were on LMSYS? We were big users of it at Databricks when we built our models. As you know, just kind of curious how many people were actually using it providing feedback sort of on a daily basis.
If you want to be a data-driven founder, and trust me you do, you're going to spend time in spreadsheets but a lot of these spreadsheets they're just stuck in the 1990s. Thankfully now there's Quadratic finally bringing the productivity boost of AI into your spreadsheets. But this is not the same little chatbot in the corner who answers random questions. No, it's an AI native platform that handles all the number crunching and organization for you. You just describe what you want to do with your data and Quadratic makes it happen right there in the spreadsheet. Now you can get instant insights about your business without fighting formulas and trying to remember different ways to where to put the comma and the parentheses. It just works and it instantly shares the results with your collaborators. No setup or payments are required upfront. So you can start using Quadratic right now and experience the magic. Just visit quadratic.ai/twist to sign up and use the code twist to get 1 month free of their pro tier subscription. That's quadratic.ai/twist.
Well, these days it's tens of millions. We have tens of millions of MAU. And we have like on the order of half a billion or more conversations. And so it's funny at the time that we were sort of chatting with you guys at Databricks in the early days of Arena, in the very early days of Arena, we were actually going around Berkeley and giving people Amazon gift cards to vote. The platform was that small. It really and it was really supposed to just be an academic paper. And then it sort of eventually spun out. I think probably with basically we just got lucky that the space got so competitive and people were looking for a battleground and then we became that default battleground and we took that and sort of ran with it and said okay how can we continue to retain these users and how can we, you know, retain the users that are giving us the valuable feedback that's going to drive our understanding of real world performance and so that's been sort of our north star that's why our user base is heavily skewed towards the programmers. It's about 28% software engineers.
Well, and I noticed I mean you guys would actually provide that way of testing models out before they were released. So you sort of get some sense of where that model will be and it was like a secret model, right? We kind of did that with you guys as well.
Topic one, this will fit right in your wheelhouse, Anastasios. We have looked at OpenClaw now becoming, this is all like in the same week. OpenClaw goes to number one in terms of stars. It's kind of like votes or follows on GitHub. Additionally, I don't know if you guys saw this Qwen 3.5 small model. This is Alibaba is open source. It was, it's running on a, you know, I think this is an iPhone 17 Pro. Now we're starting and we were wondering when this would happen. Here it is. It's happening. And this is obviously on airplane mode. So this is just running in the background. And Reflection AI, they do open weight models. They've hit reportedly a $20 billion valuation. And the world has just become enamored with OpenAI to the point at which if you run it like we are you eventually Chase get to the point where you're like I need to run my own silicon and I need to run an open source model because it's too expensive to use Claude or ChatGPT. We were trending, I think, with 15 people getting on it towards 500 bucks a day each, $7,000 a day, $2 million in tokens. And this is like in month one of the project. I think we literally would have exceeded the salaries of everybody. So, Chase, what are you seeing in terms of the open-source models and their performance and their adoption?
I think there's this, you know, kind of constant debate between open source versus closed source and, you know, it's a matter of like who's going to win. And I don't really view it that way. I think there's like sort of room for, you know, multiple solutions in the future. I think we're going to have a future of both, you know, a lot of closed source and a lot of open source. It is very exciting to see a lot of the developments in open source and, you know, we are certainly seeing a lot of adoption from a lot of our customers in Crusoe Cloud that, you know, are AI native applications that are building and scaling and running a lot of inference workloads on open source models and, you know, I think it's a very exciting and innovative ecosystem that's sort of taking hold. And, you know, I think a lot of the efficiency gains that you're seeing and, you know, how people are able to distill the knowledge of these larger models into smaller mini models that can run, you know, like you demonstrated sort of on device. I think that's like a very fascinating trend. You know, and sort of the quality, the outputs are still quite good. So, you know, I think all this goes towards this broader trend though is like, you know, models are getting smarter, they're getting more useful. You know, things like OpenClaw are actually, you know, providing, you know, these agentic platforms to go do stuff for you, right? I mean, it's like, you know, I was talking to someone yesterday that runs a very large investment manager and he was saying how he's like gone fully all-in on OpenClaw and he's like running all this different stuff. And he's like, I'm convinced I'm calling this AGI. He's like, the only way I like, you know, I...
Do you think AGI, Chase? Or does it feel like this is the year we hit AGI? I guess is another sort of...
I think there's like a, you know, it's kind of like, Anastasios is kind of the king of benchmarks over here. But, you know, it does feel like one of these like moving benchmarks that like as it gets better, it gets more impressive. We sort of like move the goalposts a bit. You know, I kind of had this, you know, personal definition like having come from more of a, you know, math background, I felt like if, to me one of the big tests is going to be when, you know, there's this set of very difficult math problems called the Millennium Problems that are hosted by the Clay Mathematics Institute. And to me, like the breakthrough is going to be, you know, there's a million-dollar bounty. There's one of them has been solved of the seven. The guy actually never even collected the bounty. But, you know, to me the definition is like when AI is able to solve one of those problems that's like...
Anastasios, this is a problem that humans have not solved.
So artificial intelligence.
So maybe that's like, maybe that's super intelligence. Maybe that's exceeding.
Sure. Sure. Sure. Sure. So ground us here at, I think you're making the point, Chase, is that we are, we're like in the red zone. We're within 20 yards of the end zone and we're like, you know what, let's push it another 10 yards. It's a 120-yard touchdown. Anastasios, how do you look at it? Define it for us here.
Yeah, I mean, so it's very funny. I had actually a similar bet with a friend of mine who's a professor at Stanford stats that we made the bet just, it was a few months ago. So, you know, I think things are moving in my favor that in 10 years, two of these problems will have been solved autonomously by AI. The reason why it's important that these problems, you know, to sort of Chase's point, are unsolved by humans is that once they've been solved and the information's out there, they're no longer good benchmarks because they overfit. And so I do think that there's a component of this that's when we create benchmarks for AGI or what have you or just performance in math and coding and instruction following and production settings. We need to make sure that the data is constantly fresh to avoid overfitting.
What do you define it as then? Do you have like a personal definition you like? And then I'll go to you because I know you have a lot of thoughts here too.
Yeah. I honestly don't think about the definition of AGI much because I think that basically all these debates reduce down to what's your definition question mark. And so to me it's much more about how do we ensure that when we deploy the system we can do so reliably and responsibly that we're helping people not hurting people and use it as a tool. And I just, you know, the AGI aspect of things is not really part of my daily life in terms of my thinking. However, what is is that I think that a large fraction of human labor is going to be commoditized by AI and that's going to have very wide-reaching economic impacts and probably require a new economic system to deal with it.
All right, we just hit on all seven possible issues on the docket. Let me get Naveen involved in here because this is kind of what happens when you have people who are in it every day. Yeah, these things are turning into a flywheel in our minds. You have OpenClaw driving Jevons paradox, you know, then you have Goodhart's law in action. It's all kind of happening at once. So, just to ground us again, AGI and then what's happening inside of enterprises and what you think of these open source models. Take it where you want to go, Naveen.
Well, I think, okay, I think there's a few things here talking about the definition. I mean, Chase is talking about one specific one where I would say it's kind of what we think is what smart people are. It's solving these kinds of like mathematical problems. That's kind of intrinsically difficult for humans and there's only a small subset of humans that can actually do these things. What I actually think is even more interesting is that we haven't solved problems that every human and actually many animals can solve still. So this is why this definition keeps sliding around because, you know, it's a little like porn. You know it when you see it kind of a thing.
Many times we build a model and it solves some benchmark as Anastasios was talking about and maybe it even solves some really hard problem but then it falls flat on some really stupid thing that everybody can do. So then you're kind of like is that really AGI? Right? So I think we're in this place where we're creating a new kind of intelligence that solves a certain set of problems. Now every computer has existed for a purpose and, you know, there's a lot of economic value regardless of whether we call it AGI or not but I'm a motor control neuroscientist, okay? Like I studied how brains can actually solve physical problems in the world. I think physical problems in some ways are very hard, they're very high-dimensional and there's so many sources of error and noise that come into play and these models do not do well in those domains.
Notion is the AI-powered connected workspace for Teams. It brings all your notes, docs, and projects into one space that just works. And with AI built right in, you spend less time switching between tools and apps and more time creating great work. And now with Notion's custom agents, busy work that used to take hours or never got done at all runs itself. And my company Launch is run entirely on Notion. It is the single source of truth for our company. And custom agents are now here and they can specialize anything you can come up with. Unlike AI chat bots, which, you know, they just stick around, they wait for you to ask a question. You can assign your custom agent to ongoing tasks. So you can have them do regular updates across the entire system like, 'Hey, alert me to what people are working on or keep track of the docket for the next podcast episode and even checking when I got new ads to read.' We're leveling up the This Week in AI newsletter and I assigned hire a new writer to our task routing agent. Now Oliver has started this project the moment he opens his computer in the morning. And now producer Oliver just starts that project the second he logs into his computer in the morning. Can you imagine an AI to keep an eye on what everybody's doing? Not in a creepy way, but in a productive way like a coach and dropping a new task to the next available person. That's what this is all about. Try custom agents now at notion.com/twist. That's all lowercase letters. Notion.com/twist to try custom agents today. And when you use our link, you know you're supporting our show. Notion.com/twist.
Right? We're still struggling with self-driving edge cases. Yeah. You know, here we are 25 years, what is it? 20, 25 years after the DARPA challenge. Like it, two decades of work on it. Still lots of work to be done.
And then in terms of open source's role, do you, and how small these get because you're working on a whole new paradigm, right? Chase is working on scaling this, finding new energy sources. Obviously Jensen's working on next level chips and HBM and at the same time the models are getting smaller and then Apple silicon which Steve Jobs conceived of in 2008, executed on in 2018 when the first I guess M1 and some of these things started to come out. You know what an amazing pressing thing, you know, obviously Steve Jobs, he figures these things out early and now we have people saying, you know what, I'll just run some stuff local and we're going to have this crazy world where things are working on multiple platforms.
Chase, humans we sort of have this like system one, system two like thought process like you have sort of a, you know, really fast system that, you know, is very intuitive and then you have like your system two like higher dimensional reasoning system. And I think, you know, because, you know, you made the comment that like models are getting smaller, they're simultaneously getting smaller and bigger, right? We have things that are bigger, more capable models that are these higher level reasoning models and then we also have things that we're able to get better accuracy, better output, sharper intelligence with smaller distilled models. And I think what's important for that is like when you think about efficiency of, you know, how intelligence not only gets created but actually gets distributed to the world. It's really important to think about like where does the compute actually happen and, you know, that's part of, you know, the point you were making around like wow these things, you know, this new version of Qwen can run on an iPhone in airplane mode and wow that's pretty cool.
Actually I kind of, I'll sort of add a little bit here is that so I think the bigger models what we found over time, now we're still learning how intelligence works and these bigger models are very good at retrieving lots of different kinds of information. So pre-training and adding more parameters basically gave us more access to information. Now I think we're moving to this realm like you said bigger models or reasoning models. I actually think we're moving to a realm where potentially we can separate out the reasoning capability at least partially from the retrieval part of the system. And when you do that you actually might be able to build something that has the reasoning part that's small and then you give it access to data in different ways. That's kind of a paradigm I think is a little bit more like biology frankly. Like our brains don't necessarily remember tons and tons of really specific information, but we have mental models that we can reason through new data presented to us.
