那块可能解锁 AGI 的芯片

I think AI is the next evolution of humanity. I think it takes us to a new level and allows us to collaborate and understand the world in much deeper ways.

Naveen Rao is here, an expert in AI. Naveen Rao is probably one of the smartest guys in this domain. He sees things well before anybody else sees them. You had a lot of success doing Nervana, MosaicML, and Databricks. Why start a new chip company now?

First off, it’s not a chip company per se. Most of what we’re doing, at the beginning, is really looking at first principles of how learning works in a physical system.

NVIDIA, TSMC, Google: are these potential allies for Unconventional AI, or are these competitors?

I think TSMC is absolutely going to be a partner. Google has everything internally. NVIDIA, of course, built the platform that everyone programs on today. So, are we going to be at odds with NVIDIA going forward? I don’t know. We’ll see what the world looks like, but there could be a world where we collaborate.

Has anyone called you crazy yet for doing this?

Oh, yeah. Plenty of people.

Our guest today is Naveen Rao, co-founder and CEO of Unconventional AI, which is an AI chip startup. Prior to that, Naveen was at Databricks as head of AI and co-founder of 2 successful companies: MosaicML, in the cloud-computing world, and Nervana, doing AI chip accelerators before it was cool. We’re here reporting from NeurIPS. Great to have you on the podcast, Naveen. Welcome.

Thanks. Thanks for having me.

So, you were kind of at the vanguard of thinking about what the proper hardware is for running AI workloads.

Absolutely. When you have a hammer, everything’s a nail, I suppose. The early part of my career was really about how to take certain algorithms and capabilities, shrink them, make them faster, and put them into form factors that make those use cases proliferate, like wireless technology or video compression.

You couldn’t do video compression in real time on a laptop back then. There just wasn’t enough computing power, so you actually needed to build hardware to do those kinds of things. The early part of my career was all about that. Then I went back to academia and did a PhD in neuroscience, so you still look at it like, “Hey, can I make something better that’s more efficient?”

And so, you sold Nervana to Intel.

Yeah.

And then founded MosaicML, which is a cloud company. It’s interesting to sort of cross domains like that. I would argue MosaicML was really a software company. How did you make that decision, and why do you think you have these diverse interests?

I think I was—I guess you would call it an OG full-stack engineer. “Full-stack engineer” means something different now than it did back then. I think back then it meant someone who understood potentially devices, like silicon; how to do logic design; computer architecture; low-level software, maybe OS-level software; and then applications. That was a full-stack engineer, and I had actually touched all those topics.

To me, it was very natural to think across these boundaries. Software and hardware aren’t really natural boundaries; they’re just where we decide to draw the line and say, “Okay, this is something I configure, or I don’t.” It’s about where the world is going to consume something, where the problem is, and then right-sizing and figuring out the solution to go and hit it.

Now, “full-stack” means I know JavaScript and Python.

That’s right.

You’ve had a lot of success doing both of those things, and at Databricks. Why start a new chip company now?

It is kind of crazy. It’s one of these things. First off, it’s not a chip company per se. Most of what we’re doing, at the beginning, is theory and really looking at first principles of how learning works in a physical system.

The reason to go back and do this is purely out of passion. I think we can change how a computer is built. We’ve been building largely the same kind of computer for 80 years. We went digital back in the 1940s, and in undergrad in the 1990s, when I learned about the thermodynamics of the brain—the brain’s 20 watts of energy and the kind of computations that can happen inside the brain and neural systems—I was just blown away. I’m still blown away by it.

I think we haven’t really scratched the surface of how we can get close to that. Biology is exquisitely efficient. It’s very fast, and it right-sizes itself to the application at hand. When you’re chilling out, you don’t use much energy, but you’re still aware of other threats and things like this. Then, once a threat happens, everything turns on. It’s very dynamic, and we really haven’t built systems like this.

I’ve been in the industry long enough to know that we have to have an incentive to build things. You can’t just say, “Hey, I want to build this cool thing,” and therefore go build it. Maybe in academia you can do that, but in the real world, I can’t. Now it’s exciting because those concepts are super relevant. We’re at a point in time where computing is bound by energy at the global level, which just was never true in all of humanity.

For those of us who aren’t experts, can you describe the difference between digital and analog computing systems? Why do you think the architecture has evolved the way it has, becoming more digitally focused over the decades, as you said?

Very simply, digital computers implement numerics with some sort of estimation. In a digital computer, a number is represented by a fixed number of bits, and that has some precision error and things like this. It’s just a way we implement the system. If you make it enough bits—64 bits, for example—you can largely say that maybe the error is small and you don’t have to think about it.

The digital computer is capable of simulating anything that you can express as numbers and arithmetic, so it became a very general machine. I can literally simulate any physical process. All of physics—we try to do computational physics, right? I have an equation, and I can then write numeric solvers that deal with those imprecisions in the number of bits.

This became computer science, the entire field now. We went in that direction very early on because we couldn’t scale up computation. It’s actually an interesting comparison if you look at that time. If you look at the papers and things, they actually looked very similar to today in terms of scaling up GPUs.

Analog computers were actually some of the first computers, and they worked really well. They were very efficient, but they couldn’t be scaled up because of manufacturing variability. Someone said, “Okay, you know what? I can make a vacuum tube behave as a high or low very reliably. I can’t characterize the in-between very well, but I can say it’s high or low.” That was where we went to digital abstraction, and then we could scale up.

ENIAC, which was built in 1945, had 18,000 vacuum tubes.

Wow.

So, 18,000 is kind of similar to how many GPUs people use now for large-scale training, right? And so, in digital computers we have transistors. Just to make it concrete, what kind of substrates are you talking about for analog computers?

Analog computers can do lots of different things. Wind tunnels are a great example of an analog computer, in a sense. I have a race car on a track or an airplane, and I want to understand how the wind moves around it. In theory, you can solve those things computationally. The problem is you’re always going to be off. It’s very hard to know what the real system is going to look like, and doing things with computational fluid dynamics accurately is pretty hard.

So, people still build wind tunnels. That’s actually modeling; that’s an analog computer. I think we still have lots of reasons to build these analog-type computers.

In the situation we’re talking about, we can actually build circuits in silicon to recapitulate the behaviors of neural networks. What we’re doing today is more specified than what we were doing 80 years ago, in a sense. Back then, we were trying to automate generic calculations, which were used to calculate artillery trajectories, finances, and maybe some physics problems, like going into space. Those require determinism and specificity around the numbers and computations.

Intelligence is a different beast. You can build it out of numbers, but is it naturally built out of numbers? I don’t know. A neural network is actually a stochastic machine.

And so why are we using a substrate that is highly precise and deterministic for something that's actually stochastic and distributed in nature? We believe we can find the right isomorphism in electrical circuits that can subserve intelligence.

That's a pretty wild idea, isn't it? Maybe unpack it one level deeper, because I totally agree with you. Computers for decades have been the complement to human intelligence, right? My brain isn't really great at computing an orbital trajectory.

That's right.

And I don't want to burn up on reentry. A computer can help us with this incredible degree of precision. We're now going in the opposite direction, right? We're actually trying to encode more fuzziness into computer systems. Go a little bit deeper on this idea of analog, and why intelligence is a good fit for analog systems.

The best examples we have of intelligent systems in nature are brains. It's often been said that human brains run on 20 watts of energy. That is true, but if you look at an animal brain, they're generally extremely efficient. A squirrel or a cat is using something like a tenth of a watt, so there's something there that we're still missing.

Not to say that we understand all of it, but part of what I think we're missing is that we have lots of abstractions in a computer that are quite lossy. In a brain, the neural network dynamics are implemented physically.

So there is no abstraction. Intelligence is the physics. They're one and the same. There's no operating system, API, or anything like that. A visual stimulus, for instance, directly activates an actual neural network and produces some semantic response.

Exactly. Those things are mediated by chemical diffusion and the physical properties of the neuron—the physics itself. So I think it's absolutely possible to build something that's much more efficient by using physics in an analogous way. That is 100% true. Whether we can do it and build a product out of it is really the question we're asking here at Unconventional.

Is part of the idea that now is the right time because AI is both a huge and a unique workload?

Yeah, absolutely. It's interesting. Just to give you some statistics, the US has about 50% of the world's data center capacity, and today we put about 4% of the US energy grid into those data centers. In 2025, we started to see news articles about brownouts in the Southwest during the summer. Imagine what happens when this goes to 8% or 10% of the energy grid. It's not going to be a good place to be.

Can we build more power? Absolutely—we should. Building power generation is very hard and expensive, and it's infrastructure. It takes time. You can only bring online so many kilowatts or gigawatts per year—something on the order of 4 gigawatts per year. By some estimates, we need 400 gigawatts of additional capacity over the next 10 years to power the demand for AI.

Wow.

So we have a huge shortfall, and we really need to rethink this. The 15-year-old sci-fi nerd in me says, “Wow, we're mobilizing species-scale resources to invent the future.” We are. Then there's the practical side: even if we add 400 gigawatts of production capacity, our 1970s-era transmission grid is probably going to melt under the load. There are very serious infrastructure hurdles to this.

It's hard to get a lot of humans to act together. That's just the reality, and that's what has to happen to solve these problems. What trade-offs do you think this entails—the path you're pursuing versus the mainstream digital path now?

I don't actually see it as digital or analog. It doesn't work like that. I think there are certain types of workloads that are amenable to these analog approaches, especially workloads that can be expressed as dynamical systems.

Dynamics means time. They have time associated with them. In the real world, every physical process has time. In the computing world—in the numerical computing world—we actually don't have that concept. You simulate time with numbers.

Actually, simulating time is very useful for certain problems. I think we should still build those things, and we should still have those capabilities for the problems we need to solve that way. But for problems where, as you said, things are a bit fuzzier—trying to retrieve and summarize across multiple inputs—that's what brains do really well. They can take in tons of data and formulate a model of how those things interact. Sometimes those models can be extremely accurate. Look at an athlete.

Alex Honnold, who climbed El Capitan. Just think about the precision that's required. It still scares me every time I

It's insane, right? If he slips—if he's off by a millimeter in some places—he dies. That's true for every top-level athlete and anyone who's at the Olympics.

Steph Curry—the story is that he set up a special tracking system so he could make sure the ball was hitting the middle of the rim, not just going through.

The level of precision these guys achieve with a noisy neural network is actually quite high. Neural systems can do a lot of precision under certain circumstances. What's interesting about these situations is that Steph Curry, when he shoots a ball, is never going to shoot it under ideal circumstances in a game. It's always a unique input, and there are a lot of different input variables coming at you: where the other players are, precisely where you're standing, whether your shoes are different, whether the surface is a little different, whether the ball is tackier, or whether your hands are sweaty.

There are so many inputs, and we put them all together and integrate them while still producing very accurate behavior. Brains are exceptionally good at this. That's a set of problems that is very useful to solve, and now we're approaching those problems. But it doesn't mean we don't still use computational substrates to do actual computation. This is an intelligent substrate.

What types of AI models or data modalities do you expect your hardware will be well suited for?

We're obviously starting with the state of the art today: transformers and diffusion models. They work and do really good stuff, so we shouldn't throw that out. Diffusion models, flow models, and energy-based models are actually pretty interesting because they inherently have dynamics as part of them. They're literally written as ordinary differential equations.

That makes it possible to ask: Can I map those dynamics onto the dynamics of a physical system in some way that's either fixed or has some principled way of evolving? Then can I use that physical system to implement the model and do it very efficiently with physics? That's the nature of what we're doing. We will be releasing some open-source work and other things around this to let people play around.

Transformers are a big innovation because they made the constructs of a GPU work extremely well. That doesn't mean it's wrong, but I don't think there's anything natural—there's no natural law—about the parameters of a transformer. A transformer's parameters are a function of the nonlinearities and the way the whole thing is set up with attention. There will be some kind of mapping between transformer parameter spaces and these other parameter spaces. Transformers, I think, have used lots of parameters to accomplish what they do.

I have to ask: Since you mentioned energy-based models, and Yann LeCun has been writing quite a lot about this, do you think pursuing these sorts of paths gets us closer to AGI—whatever AGI means?

Honestly, I do. The reason I feel that way—and again, this is hand-wavy; I'm going to be really honest—I don't—

That's why I'm putting quotes around it. I think the discussion is necessarily hand-wavy.

It's got to be, because we just don't know. My intuition says that anything where the basis is dynamic, with time and causality as part of it, will be a better basis than something that isn't.

We've largely tried to remove that. A lot of times you can write math down that's reversible in time and things like that, but the physical world tends not to be, at least the way we perceive it.

So can we build out of elements of the physical world that do have time evolution? I think that's the right basis to build something that understands causation. I do think we'll have something that's better and will give us something closer to what we really think is intelligence.

Yes, we have intelligence in these machines. I don't think they're anywhere close to AGI because they still make stupid errors. They're very useful tools, but they're not like working with a person, right? I think most people at that—

That's actually really interesting.

So the sort of thing that's missing in AI behavior—which I think a lot of us see, that there's something missing but can't quite put a name to it—it sounds like you're arguing that part of that is a real sense of causality. And that training, and a more dynamic sort of regime, may impart this kind of apparent understanding of causality better than what we have now.

Yeah. Again, hand-wavy, but yes. Look, you have kids—little kids—and you see them. Children kind of innately understand causality in some ways: this happened, then that happened. You can say it's reinforcement learning or whatever; that's some part of it, but there's something innate that we understand about causality. In fact, that's how we move our limbs and all of that. I know that if I send a certain command to my arm, it'll do something. So I think there's something innate about the way our brains are wired, built out of primitives that do understand causation.

Put Unconventional AI in the context of the broader industry for me. Nvidia, TSMC, and Google—are these potential allies for Unconventional AI? Are they competitors? How do you think about it?

Yeah, a couple of things that we set out to do when we were starting this company were to see if we could find a paradigm that's analogous to intelligence within 5 years. At the 5-year mark, we should be able to build something that's scalable from a manufacturing standpoint. You can think about building a computer out of many different things, but if it's not scalable from a manufacturing standpoint, we can't intercept this global energy problem. We need somebody to say, “Okay, go build 10 million of these things,” right?

I think TSMC is absolutely going to be a partner going forward. I met with them recently, and we want to work closely with them to make sure we get what we need, get fast turnaround times to prototype, and all of that. Google, Nvidia, Microsoft—all these guys are at the forefront of where the application space is. Obviously, Google has everything internally, and I think they're working on lower-risk but continual improvements for their hardware with TPUs.

With TPUs, you mean?

With TPUs? Yeah. From what I can see, just publicly, it makes total sense. They have a business to run, and they're trying to make their margins better. How can I do that with all the tools I have in front of me?

Nvidia, of course, has built the platform that everyone programs on today. So, are we going to be at odds with Nvidia going forward? I don't know. We'll see what the world looks like. We're trying to build a better substrate than matrix multiply. There could be a world where we collaborate on such solutions, and we're open to all of these things.

Where do you personally get the motivation to get up in the morning and build this company? You've had a lot of success in your career, in startups. What's exciting about this to you?

I don't know. It's a weird thing. If you haven't worked in hardware, it's hard. I've been fortunate to work in hardware and software, and I love writing a bunch of software, hitting compile, and seeing it work. That's a good dopamine hit. But when you work on a piece of hardware and turn that thing on, that's a big dopamine hit. It's like celebration—jumping up in the air, high-fiving. It's a different thing, and you sort of live for these moments.

When I was at Intel, I was one of the only execs who would go to the lab when the first chip would come back. I'm like, “I want to see it turn on, see what happens.” Sometimes you turn it on and you see a little puff of smoke come—

That's not good.

But you want to be there; you want to be part of the moment. I think that's part of it. For me personally, we have this opportunity now that we can really change the world of computing and make AI ubiquitous. I'm the opposite of an AI doomer. I think AI is the next evolution of humanity. I think it takes us to a new level, allows us to collaborate, understand each other, and understand the world in much deeper ways.

Totally agree.

Every technology has negatives, but the positives to me so far outweigh them. The only way we're going to get to ubiquity is that we have to change the computer. The current paradigm, as good as it is and as far as it's taken us, is not going to take us to that level.

I think that's such a great way to say it. AI actually can help us understand each other better, help us understand ourselves better, and understand the natural world better.

I don't think it's at all what some of the doomers think of as replacing human experience.

That's a short-term thing. There will be bumps along the way, right? Technology does that. That's what happens when you've seen too many sci-fi movies.

That's right. But look at Star Trek.

Yeah, yeah, yeah. Totally. Totally. It's great.

This is a really big swing, right? This is a very ambitious company. What gives you confidence that it's going to work, or has a reasonable shot of working?

There's a number of data points. Of course, like I said, the brain's an existence proof. But there's also 40-plus years of academic research showing a lot of promise here. People have built different devices, albeit not with the latest technology or with professional engineering teams, but they have built proofs of concept that actually show some of these things work.

We've also, from a theory standpoint—both from neuroscience and from pure dynamical-systems and math theory—started to understand how these systems can work. So I think we now have pieces at different parts of the stack that show, “Hey, if I can combine these things the right way, I can build this.” That's what great engineering is all about: exploiting this thing that someone else built for something else, exploiting that thing, and then—

Engineers are kind of the opposite of theorists: “Well, all right, that thing doesn't quite fit. Sand it down and make it right.” So we've got to do a little bit of that right now, and then we can build something and put it all together.

Yeah.

That's awesome. Has anyone called you crazy yet for doing this?

Oh, yeah. Plenty of people at this point.

Is it everybody?

Well, I'm used to this at this point. My family has called me crazy. I was called crazy going back to grad school years ago, when I had a very good career in tech. So it's fine. I think you need crazy people to go out and explore. If you think about humanity coming out of Africa, the crazy people who went out—

We would be lost without crazy.

You need some crazy in there. So it's okay. I'm fine with that.

What kind of people are you looking to bring onto the team? It's a very ambitious goal. Who should be interested in joining you?

Yeah, I think some of the traditional-ish—when I say traditional, over the last 5 years this field of AI systems has evolved—people who are really good at taking algorithms and mapping them very effectively to physical substrates. Those folks who understand energy-based models, flow models, gradient descent, and different ways—this kind of thing is what we need there.

We need theorists who can think about different ways of building coupled systems, how I can characterize the richness of dynamical systems, and relating that to neural networks. So there is a theory aspect of this. Then there are folks who are kind of at the system architecture level: “All right, here's what the theory says. This is what I can really build. How do I bridge that gap?” And then there are the people actually physically building this stuff—analog circuit people, and digital circuit people, too. We're going to have a mixed-signal team here. So that's the whole stack.

The stack is hard because these are all things that no one's really pushed to that level. When we build this chip, our first prototype, it's going to be probably one of the larger, maybe the largest analog chip people have ever built, which is kind of weird. The first time you do something, things don't usually work the way you think they—

So you can get in on that Cerebras–Jensen game where they were each pulling the biggest possible wafer out of an oven.

Something like that. Yeah, exactly right.

Put a few vacuum tubes on top for effect.

Yeah, we could. I need blinking lights.

Yeah, exactly.

We're not going to have cool heat sinks. It's going to be super—it's going to be cold. You don't need big heat sinks, you know? So I hope they make something that looks interesting here.

This is a funny time for top AI people, right? You have the option, if you want to start a company, of a lot of venture capitalists who would probably fund you. If you want to get a cushy job at a big company, you can get a very cushy job and do some interesting things.

Yeah.

Or, you know, people can join a startup like Unconventional AI that has a lot of the nice aspects people look for in AI careers and is taking super-big swings.

I’m just curious. You’ve been on all sides of this. Do you have any advice for younger people starting out in their careers, or how do you think about this?

I think you get such a breadth of experience from working at a startup at the beginning of your career that it will pay dividends later on. The reason I can think across the stack is because I did all those things very early in my career. I built hardware, I built software, and I built applications.

In big companies, it’s not anyone’s fault. It’s just the way it is. You get hired to do a thing, and you do that thing over and over again. You get really good at doing that thing, and that’s fine. You need people who are really good at doing specific things. But if you want to be prepared for change in the future, being really good at one thing is probably less valuable than being slightly good at a lot of things.

Yeah, that’s interesting. Is it fair to say Unconventional is sort of a practical research lab? Is that the kind of culture you’re going for?

Absolutely. Yeah. I mean, the first few years, it really is open-ended. I don’t want to close doors. I’m really specific about this. I always try to bring the conversation back when people say, “Oh, that’s going to be hard to manufacture.” Stop. Don’t think about that. Will it work? First, come up with existence proofs. Then we go back and try to engineer it, with all the trade-offs therein.

But if you make those trade-offs up front, you don’t go into a good place. So yes, we are really thinking wide open, but with an eye on the future, who we are building a product—

And to your point, it takes not only people with diverse skill sets, but people with high agency to try new things, learn new things, and integrate across the stack.

Yeah. I think what I’ve done really well across the companies I’ve built has been going after hard problems, which lends itself to smart people wanting to come in and try to solve them. They see a challenge—it’s like climbing Mount Everest—but then giving them agency.

I look at it like, what decisions can I make as a leader to increase the agency of the organization overall? Me making a top-down-style decision may be globally better for the company in the short term, but I think long term we’ll do better if more people have agency and can try more things out.

Personally, I like to find ways to get out of the way when I see people who are very passionate about trying something. It’s like, “Okay, you really want to do this. That makes sense. Go for it.” Then you own it. You own both the good and the bad, right? Agency to me is like, you’ve got to be able to say, “Okay, I screwed up. No, this wasn’t right.” That’s okay, too. But give people the room to do that.

Anything else you want to say before we wrap up?

I think this is an opportunity to do something that will be felt for generations. To me, that’s what gets me up in the morning. You can go work on a product and make a tweak, and people will use it. That’s great, but in 5 years, many times people forget those things.

But if we are successful here, the world will not forget this for a very long time. This will be written in history books. I feel like those opportunities are rare.