A (1:56)

Absolutely. I mean, you know, it's like when you have a hammer, everything's a nail, I suppose. But the early part of my career was really about how do I take certain algorithms and capabilities and shrink them, make them faster, put them into form, factors that make those use cases proliferate, like wireless technology or video Compression, you couldn't do video compression real time on a laptop back then. There wasn't enough computing power so you actually needed to build hardware to do those kind of things. So early part of my career was all about that. And then I went back to academia, did a PhD in neuroscience. And so you still kind of look at it like, hey, can I make something better that's more efficient?

B (2:31)

And so you sold Nirvana to Intel and then founded Mosaic, which is a cloud company. It's interesting to sort of cross domains like that, I think to be able to look at hardware and software. I would sort of argue Mosaic was really a software company. How'd you make that decision and why do you think you have these diverse interests?

A (2:47)

Well, I think I was, I don't know, I guess you call it an OG kind of full stack. Now Full Stack Engineer means something different that it meant back then. I think back then I meant someone who understands potentially devices like Silicon, how to do logic design, computer architecture, low level software, maybe OS level software and then application. That was a full stack engineer and I actually had touched all those topics. So to me it was very natural to kind of think across these boundaries. To me, like software and hardware is not really natural boundary, it's just where we decide to draw the line and say, okay, this is something I make configurable or I don't. And it's like where is the world going to consume something? Where is the problem? And then right size and figure out the solution to go and hit it.

B (3:29)

Now full Stack means, I know JavaScript and Python.

A (3:32)

That's right.

B (3:32)

So you've had a lot of success doing both of those things and at databricks. Why start a new chip company now?

A (3:37)

It is kind of crazy. It's one of these things like actually I was first off say it's not a chip company per se. Most of what we're doing is at the beginning is theory and really kind of looking at first principles of how learning works in a physical system. And the reason to go back and do this is just 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, you know, the dynamics of the brain, like the brain's 20 watts of energy and the kind of computations that can happen inside of brain and neural systems, I was just blown away then. And I'm still blown away by it. And 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. 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. And once a threat happens, like, everything turns on. It's very dynamic. And we really haven't built systems like this. And 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 I go build it. Maybe in academia you can do that, but in sort of the real world, I can't. And 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.

B (5:06)

And so for those of us who aren't experts, can you describe the difference between digital and analog computing systems? And, like, why do you think the architecture has evolved the way it has sort of more digital focus over decades, as you said?

A (5:19)

Yeah, I mean, very simply, digital computers implement numerics and numerics with some sort of estimation. Right. I mean, 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, like 64 bits, you can largely say that maybe the error is small. You don't have to think about it. And so 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, I can then write numeric solvers that sort of deal with those imprecisions in the number of bits. And so this became obviously computer science. It's an entire field now. And we went that direction actually very early on because we couldn't scale up computation. It's actually kind of an interesting conversation if you look from back then, not that I was there, of course, but if you look at the papers and things, they actually look very similar to today in terms of scaling up GPUs, analog computer directly from the first computers. And they worked really well, they were very efficient, but they couldn't be scaled up because of manufacturing variability. So someone said, oh, okay, you know what I can actually say 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. And so that was kind of 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. Right. For large scale training. So scaling things up is always a hard problem. And once you figure out how to do it, it makes a paradigm happen. And I think that's why we went to digital. But analog still is inherently more efficient because it's actually analogous. Computing is the way to think about it, like, can I build a physical system that is similar to the quantity I'm trying to express or compute over? You're effectively using the physics of the underlying medium to do the computation.

B (7:16)

And so in digital computers, we have transistors just to make it sort of concrete. What kind of substrates are you talking about for analog computers?

A (7:22)

Yeah, I mean, analog computers can be lots of different things. There's. Wind tunnels are a great example of an analog computer in the sense where I have a race car on a track or an airplane and I want to understand how the wind moves around it. And you can, in theory, solve those things computationally. The problem is you're always going to be off. It's very hard to know what the real system's 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. That's an analog computer. I think we still have lots of reasons to build these analogous type computers. Now in the situation we're talking about, we can actually build circuits in silicon to recapitulate behaviors of neural networks. So what we're doing today is more specified than what we were doing 80 years ago, in a sense, is that then we were trying to automate generic calculation, which was used to calculate artillery trajectories. It was used to calculate finances. Maybe some physics problems like going into space, things like that, those require determinism and specificity around these numbers and these 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 the substrate that is highly precise and deterministic for something that's actually stochastic and distributed in nature? So we believe we can find the right Isomorphism in electrical circuits that can subserve intelligence.

B (8:52)

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 sort of the compliment to human intelligence, right? It's like, hey, my brain isn't really great at computing an orbital trajectory.

A (9:07)

That's right.

B (9:07)

And I don't want to burn up on reentry. So, like, a computer can help us with this incredible degree of precision. We're now kind of going the opposite direction. Right. We're actually trying to encode more sort of fuzziness into computer systems. So go maybe just a little bit deeper on this idea of an analog and why intelligence is a good fit for analog systems.

A (9:25)

Well, I mean, the best examples we have of intelligent systems in nature are brains. And it's often been said, you know, human brains run on 20 watts of energy. That is true. But if you look at mammalian brains, generally, actually extremely efficient, like a squirrel or a cat, it's like a tenth of a watt. And so there's something there that we're still missing. And not to say that we understand all of it, but part of what I think we're missing is 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, you know, OS and, you know, some sort of API and this and that. It's like.

B (10:04)

So there's some visual stimulus, for instance, that directly activates a actual neural network and produces some somatic response. That sort of thing.

A (10:11)

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

B (10:33)

And is part of the idea that now is the right time because AI is a. Both a huge and a unique workload?

A (10:40)

Yeah, absolutely. You know, it's. It's interesting. So just maybe some stats here. Like, the US is about 50% of the world's data center capacity, and today we put about 4% of the energy grid of the US energy grid into those data centers. And this, this past year, 2025, was the first time we started to see news about brownouts in the Southwest during the summer. And, you know, just Imagine what happens when this goes to 8%, 10% of the energy grid. It's not going to be a good place that we're in. So can we build more power? Absolutely, we should. Building power generation is very hard, expensive and it's infrastructure like it takes time. You can only bring online so many kilowatts or gigawatts per year, so something on the order of four per year. By some estimates. We need 400 gigawatts additional capacity over the next 10 years to power the demand for AI.

B (11:33)

Wow.

A (11:33)

So it's a huge shortfall and so we really just need to rethink this.

B (11:37)

The 15 year old SCI fi nerd in me says like, wow, we're mobilizing species scale resources to invent the future.

A (11:45)

We are.

B (11:46)

And then there's the practical. It's like even if we add 400 gigawatts of production capacity, our 1970s era transmission grid is probably going to melt under the load. So yeah. So there's very serious sort of infrastructure hurdles to this.

A (12:00)

I think it's hard to get a lot of humans to act together. Right. It's just a reality. That's what has to happen to solve these problems.

B (12:08)

What trade offs do you think this entails? You know, sort of the path you're pursuing versus the mainstream digital path? No.

A (12:13)

Yeah. I actually don't see it as, you know, it's 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 the ones that are, that can be expressed as dynamical system dynamics, meaning time. They have time associated with them. In the real world, every physical process has time. And in the computing world, like a numeric computing world, we actually don't have that concept. You simulate time with numbers. Actually simulating time is very useful in certain, certain problems. So I think we should still build those things and we should still have those capabilities for the problems that we need to solve that way. But for these problems where, you know, like you said, it's a bit fuzzier. I'm trying to retrieve and summarize across multiple inputs. That's actually what brains do really well, Right. They can, they can take in tons of data and sort of formulate a model of how those things interact. And sometimes those models can be actually extremely accurate. Like look at an athlete, you know, you know, Alex Honnel climbed El Capitan, right. Like, just think about the precision that's required.

B (13:20)

It still scares me every time I see.

A (13:22)

Right. And if he Slips like just he's off by a millimeter. In some places, he dies. Right. And that's true for like every top level athlete. They're someone who's, you know, at the Olympics.

B (13:34)

Yes, Steph Curry, you know, the story is he set up a special tracking system so he can make sure the ball was hitting the middle of the rim, not not just going through.

A (13:41)

So the level of precision these guys hit with a neural network that's noisy is actually quite high. So neural systems can actually do a lot of precision under certain circumstances. But what's interesting about these situations is Steph Curry, when he shoots a ball, is never going to shoot it under ideal circumstances in a game.

B (14:00)

Always.

A (14:00)

It's a unique input and there's a lot of different input variables coming at you. Like where the other players are precisely where you're standing. Maybe your shoes are different, maybe the surface is a little different. Like maybe the ball is tackier or your hands are sweaty. Like there's so many inputs and we kind of put them all together and integrate them and still have very accurate behavior. So brains are exceptionally good at this. And, you know, that's a set of problems that is actually 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 kind of intelligent substrate.

B (14:36)

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

A (14:44)

Yeah, so we're obviously starting with the state of the art today. Like transformers, diffusion models, they work, they do really good stuff. So we shouldn't throw that out. And diffusion models and flow models are actually energy based. Models are actually pretty interesting because they inherently have dynamics as part of them. They're literally written as an ordinary differential equation. So that makes it such that, hey, 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? And then can I basically use that physical system to implement that thing and do it very efficiently with physics? So that's kind of the nature of what we're doing. And we will be releasing some open source and things around this to let people play around. But transformers are really, they're a big innovation because they made the constructs of a GPU work extremely well. And it doesn't mean it's wrong, but I don't think there's nothing natural. There's no natural law about the parameter of a transformer. Transformer's parameter is a function of the non linearities and the way that whole thing is set up with attention there's going to be some kind of mapping between transformer parameter spaces and these other parameter spaces and transformers are I think have kind of used lots of parameters to accomplish what they do.

B (16:06)

I have to ask, just since you mentioned energy based models and Yann Lecun has been, you know, writing quite a lot about this, do you think pursuing these sorts of paths that you're talking about is, is gets us closer on the path to AGI? Whatever. Whatever AGI means.

A (16:24)

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

B (16:29)

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

A (16:33)

It's got to be because we just don't know. But my intuition says that anything where the basis is dynamic, which has time and causality as part of it, will be a better basis than something that's not. So we've largely tried to remove that and a lot of times you can write math down, it's reversible in time and things like that. But the physical world tends not to be at least the way we perceive it. And so can we build out of elements of the physical world that are.

A (17:01)

Do have time evolution? I think that's the right basis to build something that understands causation. So I do think we'll, we'll have something that is better and will give us something closer to what we really think is intelligence. Because yes, we have intelligence in these machines. I don't think they're anywhere close to AGI because I mean they still make stupid errors. They're very useful tools, but they're not. What it's not like working with a person. Right. I think most people at that.

B (17:26)

That's actually really interesting. So the, the sort of thing that's missing in. 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 part of that is sort of a real sense of causality and that training in more dynamic sort of regime may impart this kind of like apparent understanding of causality better than what we have now.

A (17:50)

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

B (18:23)

Put unconventional in the context of the broader industry for me, like Nvidia, tsmc, Google, are these potential allies for unconventional. Are these competitors? How do you think about it?

A (18:36)

Yeah, I mean a couple of things that we set out to do when we built, we were starting this company was see if we can find a paradigm that's analogous to intelligence within five years and then at the five year mark we should be able to build something that's scalable from a manufacturing standpoint. So you know, 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. So we need to have something to say, okay, go build 10 million of these things. Right? So I think TSMC is absolutely going to be a partner forward. 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.

A (19:19)

Google, Nvidia, Microsoft, all these guys are, you know, at the forefront of where the application space is. Obviously Google kind of has everything internally and I think they're working on sort of lower risk but you know, continual improvements for their hardware.

B (19:36)

With TPUs, you mean?

A (19:37)

With TPUs? Yeah. From what I can see, you know, just publicly is it makes total sense, right? They have a business to run and they're trying to make their margins better. And you know, how can I do that with all the tools I have at, you know, in front of me? Nvidia, of course, you know, they, they, they've built the platform that everyone programs on today. So is it, are we going to be at odds with Nvidia going forward? I don't know, we'll see what the world looks like. But I mean we are trying to, to build a better substrate than Matrix Multiply. There could be a world where we collaborate on such solutions and you know, we're open to all of these things.

B (20:15)

Where do you personally get the motivation to get up in the, in the morning and build this company? I mean, you've had a lot of success in your career this year. In startup. What, you know, what's exciting about this to you?

A (20:25)

I don't know, it's a weird thing like if you haven't worked in hardware, it's hard. I've had been fortunate to work in hardware and software and you know, I love writing a bunch of software and then hitting compile and seeing it work. That's a good dopamine hit. But man, when you work on a piece of hardware and you turn that thing on, that's a big dopamine hit. That's like, this is like celebration jumping, you know, jumping up in the air, high fiving, it's a different thing. And I don't know, you sort of live for these moments, you know, like when I was at intel, like 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, see what happens. Sometimes you turn it on, it's like.

B (21:04)

You see the little puff of spoke come out.

A (21:06)

You're like, uh, oh, that's not good. But you want to be there, you want to be part of the moment. But I think that's part of it. I think 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. So every technology has negatives, but the positives can be so far outweigh it. And the only way we're going to get to ubiquity is 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.

B (21:50)

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, understand the natural world better. I don't think it's at all what some of the doomers think of replacing sort of human experience.

A (22:04)

That's a short term thing. I mean there will be bumps along the way, right? Technology does that.

B (22:10)

That's what happens when you've seen too many sci fi movies.

A (22:12)

That's right. But look at Star Trek.

B (22:14)

Yeah, yeah, totally, totally.

A (22:16)

It's great.

B (22:18)

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

A (22:28)

There's a Number of data points. Of course, like I said, the brains are existence proof. But there's also 40 plus years of, of academic research which is showing a lot of promise here. People have built different devices, albeit not in the latest technology, 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 just pure dynamical systems and math theory do start 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. And you know, that's what great engineering is all about, is like, you know, exploiting this thing that someone else built for something else, exploiting that thing. And it's. Engineers are kind of like the opposite of theorists. It's like, well, all right, that thing doesn't quite fit. Sand it down and make it fit, right? So it's like, we gotta do a little bit of that right now and then we can build something and put it all together.

B (23:28)

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

A (23:31)

Oh yeah, plenty of people.

B (23:33)

Is it like everybody?

A (23:34)

Well, it's, I'm used to this at this point. You know, my family I would call 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 that's that you need crazy people to go out and explore. I mean, if you think about humanity out of Africa, all that, the crazy.

B (23:52)

People went out, we would be lost without crazy.

A (23:54)

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

B (23:57)

And so what kind of people are you looking to bring onto the team of a very ambitious goal who should be interested in joining you?

A (24:06)

Yeah, I mean, I think some of the traditional. Traditional ish. When I say traditional, I mean over the last five years this field of AI systems has evolved. Like 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 in different ways. You know, 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 a richness of dynamical systems and relating that to neural networks. So there is a theory aspect of this. Then there's folks who are kind of at the system architecture level. It's like, all right, here's what the theory says. This is what I can really Build. How do I bridge that gap? And then there's the people actually physically building this stuff, like analog circuit people. Actually digital circuit people too. We're going to have a mixed signal here. So that's. That's the whole stack. The stack is. It's hard because these are all things that no one's really pushed to that level. Like 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. First time you do something, things don't usually work the way you think they.

B (25:17)

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

A (25:23)

Something like that. Yeah, yeah, yeah, exactly right.

B (25:26)

Put a few vacuum tubes on top for. For effect.

A (25:28)

Yeah, we can. I need blinking lights.

B (25:31)

Yeah, exactly.

A (25:31)

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

B (25:40)

This is a funny time for. For top AI people, right, where you have sort of the option if you want to start a company, there's a lot of venture capitalists who probably would fund you. If you want to get a cushy job at a big company, you can get a very cushy job and kind of do some interesting things. Y or people can join a startup like Unconventional that has a lot of the nice aspects people look for in sort of AI careers and are taking super sort of big swings. I'm just sort of 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?

A (26:17)

I think you get such a breadth of working at a startup that at the beginning of your career that will pay dividends later on. Because like I said, the reason I can think across the stack is because I did all those things very early in my career. I built hardware, built software, build applications. And in big companies, it's not aimless fault, it's just the way it is. Like 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 very good at, but slightly good at a lot of things.

B (26:56)

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

A (27:02)

Absolutely, yeah. I mean first few years it really is open ended. I don't want to close doors. Like I'm really specific about this. Like I always try to bring the conversation back because those people are like, oh that's gonna 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 and you know all the trade offs they're in but if you make those trade offs up upfront, you don't go into a good place. So yes, we're really thinking wide open but with an eye on the future. We are building a product and to.

B (27:33)

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

A (27:42)

Yeah, I mean I think what I've done really well across the companies I've built has been going after hard problems which kind of lends itself to smart people wanting to come in and try solve them. They see a challenge, it's like here's the Mount Everest of climate. But then giving them agency and I sort of look at it like what decisions can I make as a leader to increase agency of the org overall. Like me making top style decision may be a globally better for the company in the short term but I think long term we will do better if more people have agency and can and try more things out. So 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, what you really want to do this that makes sense. Go for it, you know, and then you own it. You own both the good and the bad. Right. And that's agency to me is like you got to, you guys feel like, okay, I fucked up and this wasn't, that's okay too. But give people the room to do that, you know.

B (28:37)

Anything else you want to want to say before we wrap up?

A (28:39)

I mean I think this is like an opportunity to do something that is generationally will be felt, you know. To me that's, that's what gets me up in the morning is you know, you can go work on a product and make a tweak and people will use it. That's great. But like in five years many times people forget those things. But if we are successful here, the world will not forget this. For a very long time, right? This can be written in history books and so I feel like those opportunities are rare.

C (29:10)

Thanks for listening to this episode of the A16Z podcast. If you like this episode, be sure to like, comment, subscribe, leave us a rating, or review and share it with your friends and family. For more episodes, go to YouTube, Apple Podcasts and Spotify. Follow us on X1 6Z and subscribe to our substack@A16Z substack.com thanks again for listening and I'll see you in the next episode. As a reminder, the content here is for informational purposes only, should not be taken as legal, business, tax or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any A16Z fund. Please note that A16Z and its affiliates may also maintain investments in the companies discussed in this podcast. For more details, including a link to our investments, please see a16z.com disclosures sa.