$4.5B Brain-Inspired AI Chip Company

[00:00] Why can we not build a computer that acts like biology? You founded Nirvana, you founded Mosaic ML, you got acquired by Databricks, became chief AI officer there. What compelled you to start Unconventional AI? Everything in the company we got to think unconventionally. So we have an un-CFO. I saw that. I was asking for a financial update and I heard from her un-CFO. I said, wait a second. I do want these numbers to be... [00:25] - Correct, not odd. - Well, make it correct, but I think that the key is that [00:29] Everything about this company is about breaking the old paradigm and coming up with something new. People are stochastic in how they think, and the trouble with analog computers historically was that stochastic nature. I think this will be a very special discussion because I got to look at the guest list ahead of time, and it seems like you all are super smart. So we're going to go very deep technically. We are entering a period of incredibly dynamic change in every aspect of the way we do computation. What companies do you think are going to fail? [00:59] you [01:01] Good evening. Good evening, ladies and gentlemen. [01:05] Welcome to Playground. [01:09] delighted to have you here. [01:11] I have the distinct pleasure of giving you a minute or two of context on to what this building is and what we do here. Playground is a somewhat unusual [01:22] Deep tech venture firm. Deep tech means investing in things that almost certainly won't work. [01:28] but will be consequential if they do. And we're focused on the future of computation, on semiconductors, on energy, on

[01:38] applications of advanced computation to tackling our most difficult diseases. [01:43] And my partner, Pat Gelsinger, has a storied history in semiconductors that he feels like he has unfinished business in. We have companies that are building million qubit quantum computers. We have companies that are building... [02:00] exotic forms of matter that do computations in novel and new ways. We have companies that are changing the way lithography works to restart Moore's Law. [02:13] against what is today our most sophisticated technologies, that of EV lithography machines, which have become emblematic of where technology has got to, but are starved for photons and deserve a better light source to continue their roadmaps, that they're [02:35] That machine is actually emblematic of a place that I feel like some people think we've got to, which is diminishing returns. That these exotic ASML devices are producing less and less for more and more money. That our fabs, despite the fact we spend tens of billions of dollars on building those fabs, are returning less and less for that incremental effort. [03:05] LLMs to produce less and less interesting results. And there's a sense that maybe we are at the end of something, right? That we're grinding into this process.

[03:14] this period of the end of Moore's Law, diminishing returns, and these wild expenditures of capital and resources, [03:22] are getting less and less and less for what we put in. [03:26] But I would argue that that impression may be true temporarily, but there are extraordinary degrees of freedom that we are yet to explore. [03:38] We're not at the end of Moore's law. It's not dead, it's merely resting. There are ways of doing computation hitherto unimagined, right, different architectures. We've been white-knuckling von Neumann architecture for 75 years. There are other ways of doing it. There are other things to make computation out of besides transistors and semiconductors. [03:57] that things like superconducting logic give us the opportunity to increase productivity of computation by... [04:08] factor of a hundred, factor of a thousand. Transformers are not the end state of AI, right? AI is [04:15] embarrassingly stupid [04:17] compared to many of the examples we are carrying around in our heads, both in terms of what it can achieve, but also how much power it uses to do it. [04:26] And that's not an end state either. There's nothing magical about nature. There's nothing magical about being able to think at 20 watts. But it's a bit embarrassing where at megawatts, not, you know, handfuls of watts. [04:39] So I think we are entering a period of incredibly dynamic change in every aspect of the way we do computation, from the light sources to the materials to the architectures to the kinds of models we run to the way we train them to the way we think about software. All of that is a jump ball right now. It's all up now.

[05:02] for [05:03] this incredible [05:06] rapid period of change that we all get to participate in, which is a great honor. And Playground's been lucky enough to see much of that play out within these walls. We were lucky enough to host Naveen back in the day when it was Nirvana Systems. [05:21] where we first invested because those guys were doing, before they had hardware, they were doing NVIDIA microcode better than NVIDIA was. I was like, maybe there's some guys who can beat Jensen and Jensen's game. After they sold Intel, it was like, beating Jensen's really hard, really, really hard. But it was delighted to have Naveen back with Mosaic ML, which... [05:47] demonstrated a degree of freedom right at the birth of these interesting foundation models. [05:53] And we're absolutely delighted that Unconventional is back here inside [05:59] playground because my experience of interacting with Naveen is that he is a singular focus in the world of gathering AI talent to figure out what happens next. And I think that also within these buildings, we have various forms of computation and semiconductors and logic that may inform the journey of unconventional. But it's incredibly exciting to realize that we are not at the end of [06:29] couple of years is going to be an extraordinary journey driven by companies like Unconventional. And we're super excited, super excited to have Naveen back. And with that, I will introduce the people who lean in the conversation. Of course, Naveen Rao, Constantine Bueller, and Molly O'Shea.

[06:53] So I think this will be a very special discussion because I got to look at the guest list ahead [06:59] and it seems like you all are super smart. So we're gonna go very deep technically. And as much as you wanna share, feel free to lay it out. I think it would be very interesting for the whole group to just hear all of your knowledge. [07:15] - Do my best, we'll see. - He did say in the introduction that AI is embarrassingly unintelligent compared to most brains. I think I heard most brains, so it's including all in this room. [07:28] So to start things out, I'd love to maybe start the macro perspective. We may be in the early innings of these things and the super cycle, but exactly where are we in the super cycle? What is the thick of it? What do you see going on? [07:43] I mean, it's impossible to say for sure, because you have to see the whole super cycle to know where we are, I guess. But I mean, as Peter pointed out, right, I mean. [07:53] Even 10 years ago, we were still kind of working our way out of... [07:58] traditional compute workloads, that kind of a thing. It's sort of like, that was the primordial mist of the next level of intelligent evolution. And I think we're at this point where [08:09] okay we figured out some stuff you know like uh we can use numbers and you know algorithms on existing hardware to build something that's actually demonstrating intelligence and learning adaptation but it's really bad like i mean just just think about for a moment like the errors that are made like they're they're

[08:29] There are errors that are so obvious to any biological system, even like a child or a rat or something like can reason through those. So we still have not really cracked the code on exactly how to build that. My personal take is that we need innovation on the hardware substrate. [08:44] to actually build true intelligence. You're not going to get there with this sort of rudimentary [08:50] thing that we have. So, I mean, a lot of the things that Peter was talking about in terms of investment, I mean, some of those things will pan out, some won't, but like, that's what we have to start doing. We have to start thinking very deep and very fundamentally. And biology always kind of broke across these different [09:07] abstraction boundaries. Abstractions are actually a human [09:11] construct, right? We put these abstractions in place of like, okay, well, I have software and I have arithmetic and, you know, biology doesn't work like that. And neural networks just kind of act. The physics of them act. And so how do we get to that point? It's going to take us a while to build engineered systems like that. But, you know, I think we're, [09:28] you know, not to sound too big, but this is like... [09:32] building intelligence systems at the scale that we can do as a species is something that [09:37] The world has never seen before. It's going to lead us into unprecedented ability to understand and build new things, which will continue to accelerate. That is the super cycle I think is interesting. It's like the next evolution of humanity. [09:50] And then, Konstantin, where have the dollars been following all of that? And where are we going to go next? There's been a lot of dollars following a lot of things. You guys have probably seen the news. I'll step back and talk about the investment cycle, but where are we in that AI cycle?

[10:07] I'm an engineer by training, but I love history. [10:10] And I often think about this revolution as akin to the Industrial Revolution. And here's the analogy that I'd use. [10:19] Before the Industrial Revolution, say late 1600s, the vast majority of physical work [10:26] Work equals force times distance. [10:28] was done by some sort of biological means. [10:31] 90 plus percent, the vast majority of work. That's either humans or oxen or on occasion windmills. [10:39] But the vast majority of work for humans was done by biological means. Flash forward to today. [10:46] 99.999, some number of nines, of work equals force times distance is done by mechanical machines. [10:55] Just think about everything that made this possible. Of course, we see our biological work in the day to day. But think about the jumbo jets flying in and out of SFO and the amount of work that they're doing, the car that you came here and the lights, the cameras, everything around us represent the factories that are producing all this electricity. [11:15] represents [11:16] I haven't done the proper estimate, but hundreds, thousands, tens of thousands, hundreds of thousands more work than humans are doing day to day. [11:25] Now, that was the... [11:27] biological work revolution, the physical work revolution. And it turns out we didn't become disembodied. We're all still in our physical bodies. We all still go through life living in our physical bodies and physicality still matters. We're on the precipice of a similar revolution, but for the mind.

[11:44] And I think that you look 20 years out and the vast majority of [11:49] cognitive work. [11:51] is done by machines. [11:53] just like the vast majority of physical work is done by machines. So to answer your question of where are we in the cycle, I think we're actually very early on in the very macro cycle of AI. And we'll look back and say maybe a percentage or two percentage of cognitive labor is currently being automated globally. And at some point, we'll look back and say 99.9 of cognitive work is being done by machines. [12:19] Thank you. [12:21] And then the money? The money follows the work. So if that's a lot of nines, we're going to be investing in a lot of companies that are solving those important problems. [12:31] To break this down a little bit deeper, could you just explain the differentiation between all the layers of the AI stack? [12:39] It's like your life, Vivian. I guess so, yeah. All the different wares. Well, I think some of them are emergent wares at this point. But, you know, we built up on a foundation of computing that took the last... [12:56] 50 to 80 years, depending on how you want to look at it, to evolve. So that was first, we had vacuum tubes, but we kind of moved this digital abstraction. [13:04] we could build things that could process numbers and do arithmetic very fast. So that was the first layer. And that took a little while. I think ENIAC was one of the first [13:13] high-scale computing devices. That was 1945. So we didn't get to integrated circuits until the, you know, like late 60s, something like that. And that's kind of when Moore's Law started. And I would say the modern computer was really built. And so that kept evolving. Software at that point was still pretty rudimentary. I mean, I think we were kind of, you know,

[13:33] calculation and just automating calculation was really what software was about. And the application wasn't really born until probably [13:39] right around that time, probably the late 60s, when you started to think about an application. I mean, we had OSs and stuff like that. So the stack kept on moving up into actual real applications. So I'd say the 70s were kind of where applications started. That's when we started to have spreadsheets and [13:54] word processing, this sort of thing. So that was the next layer. So first you have physical layer, then you have a computing layer, then you have some sort of a useful piece of software. [14:05] That went on for a long time. Then I would say, you know, mobile and cloud are probably the next two big ones. Maybe you consider internet. I think internet was not really a layer in the same way, but... [14:14] I'm sure people will argue with me about that. [14:17] But [14:19] If I look at that next layer that happened, [14:25] putting applications together to build infrastructure and the cloud with SaaS. [14:30] that gave rise to SaaS. And that kind of took us to where we are. So AI, I think, is something that consumes all of those and works on top of it. And so now that is starting to wear, right? We have the [14:42] Neo clouds, that kind of a thing. Then we have software on top of that that orchestrates these at large scale. And now we're getting into vertical application stacks and vertical intelligence stacks. What's interesting about these verticals is that it's like go solve legal problems or go solve whatever kind of problems. It actually uses a bunch of the application stack beneath it.

[15:03] So that's good. That's great. But much of this is built on a foundation of physics that we didn't really go back and change that much. I mean, yes, we built smaller devices, faster devices, but the principles that it's built on haven't changed in a long time. [15:17] And so... [15:19] You founded Nirvana. [15:21] You founded Mosaic ML. You got acquired by Databricks, became chief AI officer there. What compelled you to start [15:29] unconventional AI. [15:31] Yeah, it's a good question. You know, I think you got to look at the motivations of different folks. I mean, my motivation in this space is [15:38] you know, it's been about this obsession of [15:41] Why can we not build a computer that acts like biology? [15:44] The single thing hit me when I was an undergrad, like 30 years ago. And that's when I learned about 20 watts of energy in your brain and, you know, just how, how, how biology kind of. [15:53] does computation. It doesn't really do computation, by the way. It's sort of [15:58] It's like dynamical systems, but we can get into that later. [16:03] My whole career has been about that, actually. When I came out of undergrad, I was like, okay, I want to build computers. How do I build those? So I actually became a computer architect. I worked on UltraSpark 3 and a bunch of processors and did a bunch of ASICs. Then went back to school to do a PhD in neuroscience. I'm like, okay, now I know how to build a machine. I've done it several times sort of from scratch. [16:24] I'm still no closer to the answer of why computers aren't as good as brains. Right. At that time, we didn't really understand what brains were doing. I think now we kind of did. But at that time, we almost thought of brains as sort of magical. Like this is the time of Deep Blue, if anyone remembers that. That was the chess computer that IBM had. I think it was 1997 and it was pure brute force. Right. It's just like now we've hit an inflection point of processing power where I can just go and like

[16:50] forward project all these different moves on a chessboard and I could be a human. Okay, that was kind of cool. It's very clear that a human doesn't do that. [16:57] I played chess. In fact, that was like a big thing I did when I was a kid. And that's not how our brains work. I did not go through and enumerate every possibility and assess it. So there's something better that our brains were doing. And at that time, it was weird. Like if you talk to people who are sort of in the artificial intelligence domain, [17:14] We kind of thought about it as magic. We had no idea. [17:16] um we also knew it wasn't brute force [17:19] So there's some kind of heuristic algorithm that we didn't understand. And I think what's changed between now and then is that we've learned that the heuristic is actually [17:28] some kind of learned estimator. [17:31] I think that was a huge breakthrough over the last 15, 20 years, is that actually brains aren't magic. They just weren't estimations of the world and actually don't do calculation. [17:42] To me, that was actually a big jump. And so now it's like, okay, now we've got a clue as to how to build a machine that is intelligent. [17:49] So, [17:50] Nirvana was actually when that inflection point happened. So 2012 was the ImageNet moment. Everyone knows about that. [17:59] And so we said, ah, neural networks are actually a new computing paradigm. [18:03] I think we need to build a new architecture around this. And it was still like we have to leverage what was possible at that time. I mean, you couldn't go and invest a billion dollars into computing for alternative platforms at that point for AI because no one even knew what AI or machine learning was. In fact, in 2014, when I started pitching the company outside of these guys and several other crazy investors, everyone told me no. I mean, I went up and down Sand Hill for a whole week and got way more no's than yes's because people didn't want to build hardware.

[18:33] computing is done. Like the hardware is done. We don't need to think about that anymore. You only build software. [18:38] Completely wrong, right? Completely wrong. I knew it was wrong. Because neural networks were clearly a way to define a new kind of computer. [18:46] And so that was the start of it. And so eventually we did it, it became a thing. [18:50] investors in 2015, I think, started to learn what ML was. I mean, you obviously knew what it was, but what would you say was that inflection point when investors started to know? I think the vast majority of investors didn't really look into AI ML until 2022. [19:04] The vast majority. - You think so? - Yeah, I think that there was a cycle in ML in 2015 to 2017, really around computer vision, and the breakthroughs of computer vision. But I still think that was early majority. [19:15] Yeah, you're right. And the late majority only happened in 2022. You were very, very early. Yeah. So, but I think, again, there's just this idea of like, no, this is a new way to define a computer. And so, you know, we, [19:26] We sold too early, sold to Intel too early. I didn't know what the macro was going to look like. It seemed like a lot of money. [19:32] I was able to put my kids through college, that kind of thing. How much was it? [19:40] Colleges are expensive, guys. I mean, well, think about this, though. I went back to grad school in the middle of my career, and my kids were born right at that time. And so it was like I burned through all my savings. So there's a human element to these things, too, which [19:55] maybe make a suboptimal [19:57] decision. Anyway, so fast forward then, after I was done with Intel in 2020, we started to say, like, okay, well, I know the computing paradigm is not done, but NVIDIA is kind of moving along. They're moving to this tensor architecture. Can we start to extract more out of this? So we shifted our focus to the software layer, and that's what Mosaic was. And actually, we knew that neural networks are just going to get bigger. That was very, very clear. And so that was,

[20:27] but I don't think anyone in the investor community knew what a transformer was in 2020. [20:32] So we kind of took that and we said, okay, this is another inflection point that's going to happen. ChatGPT came out in 2022. We were ready to go. [20:39] Now, what's interesting about, and this is a long answer to your question, but the reason I'm going there is this context matters is, [20:47] I want to go back and actually reinvent the machine. And what we can do here is... [20:51] actually start to [20:53] Thank you. [20:53] To use the paradigm of non-linear dynamics, every physical thing in the world has dynamics, has time associated with it. The physics have time. [21:02] And that's something that we have not utilized in computing. We've actually tried to engineer that away. [21:08] We have this very strict abstraction of hardware of being a 1 or 0 with 0 time between the 1 and 0. If I had a perfect version of a computer, I would say, do this arithmetic operation in infinite precision and do it instantly. [21:24] Neither of those things are physically possible. [21:27] But somehow the brain used what's physically possible with all its noise, imperfections, and properties, and I was able to do computation with it. So that's what I want to come back and do. And that's what this company is about. [21:39] This is it. This is... [21:41] culmination of 30 years of this and evolution of not just my thinking but everybody's thinking coming together and i think we have the opportunity to completely [21:51] open up a new paradigm that could exist for another 80 years. [21:54] Thank you. [21:55] What was step one to get the company started?

[21:58] Step one. [22:02] I don't know. I've been talking about this for a little while with certain investors. [22:07] I was an instigator. I was a troublemaker. This guy was probably step zero. [22:13] Yeah, we've been talking about it for a while. [22:15] And it was, [22:17] I mean, it was such a... [22:20] I'll get into how the partnership came together, perhaps. When we think about investing in a company at formation stage, which Naveen and I got to partner at the formation stage, even had conversation before, we think about two things. [22:35] We think about a unique and compelling insight in a big market. [22:39] And we think about an outlier founder. [22:41] And I'll start with the insight, because that's what Naveen was just talking through. And then I'll get to the founder as well. And so the insight is just such a powerful one. [22:52] As soon as Naveen and I started talking about this, we've been friends for a couple of years, but we started talking about this pretty seriously. It was an instant mind melt. Because I was an AI engineer by training. I was doing grad school when Naveen was building Nirvana. And it has always been in the forefront, the power limitations and the compute limitations [23:22] which is the reality that digital computers were adopted because of their reliability and the need for precision. It makes so much sense as soon as you hear that. Whereas AI models are stochastic. You've all seen hallucinations, but that's just one of those symptoms. The reality is AI operates in a stochastic world and analog computers do. By the way, so does our brain.

[23:47] Now, if I asked someone in the audience to remember the number 37, [23:51] In a week, I asked about that number 37. They might remember 37. They might remember 73. [23:57] They could remember 38, 36. And they could be totally off and remember 429. That would be a problem. But overall, people are stochastic in how they think. And the trouble with analog computers historically was that stochastic nature. [24:12] But Naveen pointed out this is the moment to dramatically reduce power, which everybody knows is a, if not the major constraint of AI, and have a tradeoff that's much more reasonable in the AI world. So that was number one mind meld, just awesome set of conversations. I love this idea. And then the second piece is I like thinking about things with historical analogy. [24:42] And I talked about Industrial Revolution a little while ago. And with the internal combustion engine, we started the first steam engine. [24:49] I say we as if I was part of it. Humanity. I had the first team. Part of humanity. So yeah. Yeah. Right. Thank you. It's in 1702. And 1702 is when we really. [25:01] humanity had that first steam engine and it wasn't until 1913 that you have the model t [25:07] And that first combustion engine that's driving a normal car. And of course, there's many innovations in the engine in between there. But the reason for that was electricity was not very reliable. So the entire industry moved towards combustion.

[25:23] And in college, I took some chemical engineering coursework, and I learned about the Carnot limit, which is the theoretical limit for how efficient a engine could be, an internal combustion engine. And as Naveen and I were mind-melding on this, I said, this just feels like an ice, an internal combustion engine, to an electric vehicle. And then he told me about the Landauer limit, which I had not known about, which is this theoretical limit to destroying or really destroying a bit in information theory. [25:53] sense that just like we had a flip from combustion engines to EV, for some, not all of the world, we can have a flip to analog. And so that was insanely unique, insanely compelling, just immediate mind meld, the kind of thing that you look for in this job. And then the other piece is Outlier Founder. [26:14] And Naveen is kind of definitionally outlier founder. He's humble about his exceptional accomplishments, of course, at Nirvana and Mosaic ML. He's also humble about being an amazing race car driver, which is pretty darn cool, and another point of bonding. But that combination basically said, I got to be part of this formation stage, and we got to work on it together from there. [26:38] Yeah, so I think that was... [26:40] part of it, like, okay, [26:42] just kind of need someone to sort of push you over the top. And, you know, I was at Databricks. I did leave Databricks early. [26:48] from my, you know, [26:50] total earn out as part of the acquisition deal. But, you know, I had had the conversation of with with Ali and team about like, hey, look,

[26:59] things are working like things are working great actually at Databricks. I mean, I think when I left, like when we came when I came there, we had about a 20 million dollar business and it was, you know, seven or 800 million when I left. Something like that. I think it's crossed well over a billion. So. [27:14] Somebody can do that. Other people can do that. It's sort of a less interesting problem for me to keep just scaling that up and making it bigger. I mean, it's great. There are people who are excellent at that. [27:23] Am I the right person to do that? I can do it if I need to, but... [27:26] Other people can do that. I can... [27:28] Let's like help me go back and. [27:31] innovate and ally was on board with that and so actually data breaks is on the cap table as well so um we came to an agreement then we were off to the races [27:40] Not to deviate a little bit too much, but you were one of the first... [27:45] chief AI officers. Like, this role did not exist before. So what was that like [27:51] Well, I mean, oddly enough, I did something similar at Intel as well. There was no such role there. I don't even know if it was officially ever called that, but it's effectively what I was. But I mean, you know, when you have these new roles that come up, like, I don't know, chief product officer is one of those things that came up, what, like 10, 15 years ago. [28:11] you sort of define it yourself to some degree. And I think there is some latitude in these roles. I mean, you know, I think some of it is obviously defining a strategy around AI. That's important. Defining products, defining who the important customers are. It's kind of like being a CEO, a piece of it in a sense. But then outside of that, you know, what are your responsibilities? I don't know. I mean, I think there's enough latitude and it's nebulous enough that you can make it work. And I think it just has to make it fit into what makes the company

[28:39] successful. And so, you know, Ali and I had just sort of figured that out. I would end up doing a lot of the, you know, public speaking and press and stuff like that around AI. And that was a very hot topic or is a very hot topic for Databricks. So I think it was good to get that, to get Databricks established in that space, you know? [28:56] What's your official title at Unconventional? Is it Un-CEO? Yeah, I feel like Naveen can't take a normal title. I think it has to be created as a new one as part of that. And you're a trailblazer. [29:08] - Who you are. - Everything in the company, we gotta think unconventionally. So we have an un-CFO. - I saw that. I was asking for a financial update, and I heard from her un-CFO. I said, "Wait a second. I do want these numbers to be [29:21] correct not odd well make it correct but but i think that the key is that [29:27] Everything about this company is about [29:30] breaking and [29:31] breaking the old paradigm and coming up with something new and so it's actually really hard like for me too not just you know for ever for any human it's very hard to break out of what you've been taught and what you've you know no works so like you know as uh like we talk about analog engineers all the time like analog engineering has largely been about making circuits act linearly linearly predictably what's really interesting is if something is linear and predictable [29:57] it actually is very easy to simulate. If you can simulate in a digital system, it's less interesting. [30:02] It turns out like silicon has all these beautiful physics that are associated with it. We just try to like minimize. We say like, where can I make it look as linear as possible? That's the least interesting part of it. And that's the part that everyone's been trained to go after.

[30:17] so [30:18] Now we're saying, okay, open it, open the aperture, look at the places where it's ugly, quote unquote, right? And nonlinear. That's where the interesting things happen. And incidentally, those things are extremely hard to simulate on a digital machine. When it's hard to simulate a digital machine is doing something interesting. [30:33] Now it's about harnessing that and making it work for us, right? And that's engineering. That's what we have to do. But that's where the interesting parts come from. If you're just doing something that's [30:43] That's easy to simulate a digital computer. Just use a digital computer. There's nothing new here. So anyway, that's why everything is un. Like you have to be outside that normal way of thinking, break some of the assumptions that you've grown up with, [30:56] maybe for 30 years of your career, and think differently. - So how did you get the team together, and how big is the team? - I think the team will be like maybe 24 or so by the end of this month. [31:09] Yeah, that's a hard one. [31:13] We've got to find people who are deep experts in their field. I mean, everybody in this company is smarter than I am on each of these elements. I am not... [31:21] a deep expert on all these components by any means. So, [31:26] at the same time you've got to have that expertise which means you probably did go through these kind of conventional routes and then have the right attitude to think outside of it i think i usually do sort of the attitude check about it it's like well yes you've been doing it this way but [31:40] Are you open to trying other things? What do you do outside of work? Do you do things that require risk tolerance?

[31:49] So I think these are the kind of like, [31:51] you know, personality attributes you need to build a team. And our hiring bar is very high. It takes a long time. It takes a bunch of interviews for us to hire someone. So it is what it is. We're human limited at the moment, for sure. But we are we're building building a team as quickly as we can. You came out of stealth in December, and you disclosed a very large seed round. How much have you raised a date? And how much do you think you're going to need to raise to get this company to escape velocity? [32:17] Yeah, I mean, it's a good question. So we raised 475. There's actually interest to do more. So we'll probably probably raise some more in a short time frame. [32:25] - I think to get to actual first product, it's probably a billion and a half is my guess. [32:32] It sounds like a huge number, but I think the opportunity is enormous. So this initial money will get us through much of the exploration and... [32:40] sort of [32:41] not productization, but like what it takes to build a product. So there's a lot of engineering required to get [32:47] from the science to this is something I can productize, then the actual productization will probably be the last billion. [32:54] or so and it's just because we have to build a bunch of infrastructure we have to we're going to do a whole stack like we're not just going to sell a chip you can't do that because we're changing the whole paradigm [33:02] So we do need a fair bit of money to do that, but that's [33:05] a few years down the line. With capitally intensive businesses, Konstantin, I'm really curious, in the current state of the AI cycle, there's huge rounds, there's smaller rounds, and there's midsize rounds. But I'm curious, from your perspective, do you think these very massive rounds are reasonable, unreasonable? What becomes the breaking point?

[33:28] It's all about the company. [33:30] So around is meaningless without the context of the company. And when you're set up to go after a massive problem, [33:42] that you're uniquely positioned to go after. [33:45] That societally and individually and as an investor warrants a significant investment. In the case of unconventional, [33:54] It is [33:55] singular, I'd say, in their ability to aggregate this type of talent. [34:00] If you want to be in analog computers, if this sounds interesting, if this is an area that you want to spend your life in and you think can be transformational, first of all, this is the company for you. And secondly, Novena is the person to be leading that. [34:11] for years and decades to come. And when you can aggregate that kind of talent in one space, [34:18] and form the best talent in the world, and you believe that a space is sufficiently important. I think most of us would agree that massive reduction of power in order to do digital cognitive processes [34:31] is sufficiently important, you can build something very large and impressive. To go even further, how do you feel about valuations? [34:40] I'd always like them lower, Amali. Of course. I was going to say, you're asking an investor. Yeah. How do you feel about it? He wants them higher. I want them higher. Actually, I don't. I think you want them at the right... [34:53] - Yes. - In the right range, 'cause valuations to me, as an entrepreneur, are expectations, right?

[35:00] When you set them higher, your expectations are higher. That's what a good founder thinks. That's very that is truly very good because it is the life. I mean, you're a repeat founder. You've actually seen it. I find first time founders are 0% calibrated to that type of thinking. [35:12] And a few years in, it is very difficult for them. And then repeat founders, of whom I work with several, always have a very similar logic because they've seen what happens, even if it's not to them, to their friends. [35:25] Once they've set those expectations, and even if they surpass it, it doesn't feel as good for the entire company, or if they fall short, it feels much worse. So that is a very experienced way of looking at it. Do you feel pressure? [35:40] No, not really. Because this is... [35:44] I don't know. I just feel like this is the path I was meant to be on, right? Because I feel like this is the thing that we need to do as a society. I mean, [35:52] just kind of put it in context [35:54] I'm sure someone in this room knows how much money do we spend on energy generate electricity generation in the world. [36:00] Does anyone have an idea? I don't know, actually. [36:03] I know it's a really big number. I should know this number. A lot, Naveen. [36:07] uh 20 trillion 10 10 trillion 20 trillion something in this order right [36:12] Half a billion dollars is a drop in the fucking bucket. Like if you can solve one tenth of a percent of that problem, you've made an enormous company. So I think that's what it comes down to here is that like the scale of the problem is so big. [36:27] No, I don't feel pressure because

[36:30] are we going to hit a thousand x in the first go maybe maybe not i don't know there's a some percentage chance that we we we will hit it but i think we're going to hit [36:39] 50x with a very high chance. If we do that and we have a path to getting to 1,000 and getting to 10,000 and 100,000, now we've created something really big. So to me, it's about like, let's keep pushing this and be as unconventional as we can, get to a new paradigm and then keep on pushing that. That's another 20 or 30 years of innovation. [36:58] so [36:59] All I can think about is excitement. Honestly, I don't think about pressure at all. Of course, I have investors and I have to make sure that... [37:07] I take care of them. But at the same time, like I think they're on, they're on, [37:10] They're on the journey with me. [37:12] We were talking about this a little bit beforehand, but clearly capital is not a bottleneck. Power is a bottleneck. And you have a hot take on this. Do you want to share what your take was? [37:22] Yeah, I mean, you know, everybody here I'm sure has heard all the hype, right? Everything about, oh, we're going to do all this build out in data centers. We're going to have this giant advantage. I'm not going to name company names, but... [37:32] You probably know who I'm talking about. [37:34] None of this is physically relizable, right? That's what's really kind of annoying to me. I'm like, okay, you're talking about, you know, a trillion dollars of GPUs. [37:41] where the fuck are you going to get the power for that? I'm sorry, right? Like you're not going to be able to build these things physically. And the things that have been realized, I mean like the XAI clusters and stuff like that, [37:52] you know valiant efforts no doubt about it and it's amazing to push through and get it done but it broke a lot of rules and is doing a lot of stuff very inefficiently so it's not really something you can do over and over again to me you really start changing the world when you make something that's repeatable and you can sustain it when i sustain say sustain i don't necessarily necessarily mean from any ideological standpoint i mean just economically like can you do this thing over and over again does it get cheaper you know can you keep scaling it

[38:17] And so that's important. You have to have this. So doing something as a one-off is not super exciting to me. [38:22] Doing something where I can now [38:24] you know, go from that one-off to something that's scalable, repeatable, will continue to get cheaper, will continue to scale, and will kind of hit these Jeevans paradox sort of dynamics, that's interesting to me. [38:37] What are your principles as a founder and a leader? [38:41] Well, it's interesting because I think [38:44] as a founder you change some of these things depending on the company to some degree of course you know [38:49] truth-seeking we don't want to have these ideas of [38:53] you know i'm just going to defend my idea because it's mine and and i don't care if it's it's true i mean you have to have this have to have people at the beginning of a company who who set the stage and are and are really truly um [39:06] about solving the problem. Actually, I've seen this impedance mismatch happen when those people go into a big company because in big companies, the incentives are very different. Many times it's not about solving the problem. It's about getting promoted or something like that. And so oftentimes you get these very mission driven people in those scenarios that it causes problems. So I think being mission driven. [39:24] uh going after the problem and caring about it that's that's an important thing i think that's job one um [39:30] I think, um, [39:32] trying to find the [39:35] the quickest way to get information is sort of a cultural, cultural, [39:42] I don't know, attribute you need to have, like, [39:45] Try to think efficiently, even if you've raised a bunch of money. It doesn't matter. It has to be about...

[39:50] do I need to do all these experiments? What's the cheapest way to get there? That, I still think, even at any scale, will only help you. [39:57] So I think that's an important attribute. [40:01] I think going after... [40:04] the original source of information. You know, like we want to keep our organization as flat as possible. [40:09] um efficiency drops really precipitously as you go over like 80 to 100 people that's what i've seen over and over again um what's that uh number called 150 people uh it's particular now not that cal no some people that they're the cortex [40:26] - Yeah, but there's a word for the number. It's the something number. I forget the name now. But anyway, over 150 people basically [40:34] Dunbar's, that's it, yeah. There you go. Thank you, chat, JBT. [40:39] I had a phone in a friend there. Yeah, so Dunbar's number, right? Like when you break that, you start to see these massive inefficiencies because coordination starts to take over. So I think keeping things flat, going to the source of information is very important. [40:53] And, you know, [40:54] always just [40:56] being i want to be humbled by the people that i hire i want everybody to feel that way and we will just keep getting better at the after the 20th hire i probably would never have been hired at the company you know what i mean that's kind of my goal so i think having these sort of attributes in everybody's mind [41:11] just makes the company better. And it's hard to keep that going. It's just [41:14] You see it over and over again, but we're in a research mode for a while, so I think we can adhere to these principles for a number of years.

[41:21] So I think those are kind of the high level ideas that I like to keep. [41:25] Konstantin, what are your principles? You work with many amazing founders, including Naveen. What have you learned from them? Interesting. I'll say for investing or for company building? [41:37] Investing. Okay, so for investing, [41:40] The first principle is honesty and integrity. [41:42] without question, working with people who are [41:46] honest, say things the way they are, have clear communication with me and genuine trust. That's non-negotiable and high integrity on all these things. It has to be a true relationship. I'd always rather know bad news, honestly, than the opposite. And frankly, that's the kind of relationship that I have with my founders and I'm really grateful for. [42:08] And partly that's because when there is tough news, which by the way, if you do something important, there's going to be a lot of tough news. Always. If there's no tough news, you're probably doing something unimportant. [42:20] So when there is tough news, the response is level-headed. Or you're lying. Yeah, or you're lying. Exactly. And when there is tough news, you know, the way that these relationships actually grow are by responding with ability. [42:33] you know, taking responsibility and responding with ability and saying, okay, this is what we can do. This is how we can help you. This is how we can [42:38] put this on track or just listening and solving it together. So the first two are always honesty and integrity. And then the next is I'm looking for a spike in the person, something that they're amazing at. I'm not looking for someone that checks every box. I'm looking for someone who's unbelievable in some particular area. And the other areas we can help, you know, that's the job. I'm not here to say, okay, well, you're really great at this, but how about we teach you these things

[43:08] Instead, it's, "You're great at this. Be the best in the world," is the saying. "Run incredibly fast, and we will help complement you in the areas that you don't know." [43:17] And the third is [43:18] Make no small plans. [43:20] There's a Daniel Burnham quote. He's an architect of the World's Fair. And he said, make no small plans because they have no way of making men's blood boil and probably will never be realized. So if you're going to do it. [43:37] do something big. What companies do you think are going to fail? Um, I would rather not answer that. Um, [43:50] And the reality is we can get into that, but I also don't focus on that. My job is to focus on the things that can be successful. Let me think on that, if there's any category of issues. [44:20] insurance companies and the like, where you can buy revenue for less. But it's even true in software businesses in an AI era where you're acquiring customers and the long-term uniconomics aren't attractive. And so anytime you see a business growing really, really quickly, I like to see that the founder's thinking about the long-term.

[44:38] as opposed to the short term. So categorically, businesses that chase revenue just because it's attractive to investors and it makes the employees and team feel good, I'd be very worried. And I really like businesses. I feel really excited when I meet a founder who's taken no shortcuts [44:54] and done the really hard thing up front and knows the revenue can follow. What do both of you think about all the circular deals that are going on at the top? [45:06] Well, I mean, there are some games definitely being played. But at the same time, I think the bet is... [45:14] that [45:15] Yeah, I know we talk about it looks bad or whatever, but the bet is that the demand will come. [45:19] And I think that's actually a good bet. So I see it as like, all right, fine. It looks, it doesn't look great in the short term. And I can see why people are irked by it. [45:29] because there are some sort of circular dependencies on these things, but [45:33] If. [45:34] If out of the results of this, you create AIs that start to do very interesting things, which we are doing. [45:40] and real demand happens, you will see societal, secular shifts. [45:45] We are seeing it already, right? I think software engineering is probably one of the areas that's probably is being hit first. [45:52] And I don't see hits the wrong word. I think being changed first first. And we're seeing people have a willingness to pay for those solutions. And so [46:01] Once that demand is established, all of these games don't matter anymore. It's like, okay, well, yeah, you did that deal and it pumped up your valuation, but then you grew into the valuation because the demand came.

[46:11] So I think that's what's been irking people is that the demand wasn't there and many of these things happened earlier. [46:16] So I actually don't see it as a terrible thing. I think I saw a very similar thing happen in the early 2000s. There were all kinds of [46:23] crazy shenanigans that happen. [46:25] some of the biggest businesses in the world now [46:27] came out of that. [46:29] - Before we open it up to Q&A, I think we're almost there. I wanna talk about race car driving. - Okay, happy to do that. - How did you get into this sport and how does it make you a better CEO? [46:41] Um, well, I got into this for because I just like cars. I mean, I'm an engineer, you know, and cars are [46:47] a thing you can engineer that's also beautiful and they make cool sounds and all that. There's something very visceral about them. It is sort of a form of moving art in a sense, right? An experience. [46:59] learning to do race car driving stuff. I actually just was pretty good at it when I first started. And I raced, you know, like racing go karts years ago, like in the early 2000s. I didn't do anything for a long time because I, [47:11] went back to grad school had kids and all of these things um after i sold nirvana i didn't want anything i stayed in the same house and all that kind of stuff but i wanted a ferrari i bought a ferrari i took it to the track and that [47:26] That just started again. The addiction. It was like that first hit. And yeah, so then I started to like-- [47:34] I started racing in that series and sort of kept on upping the dose. And actually, I'm right in the middle of a race at the moment. The Daytona 24-hour race is happening on Saturday. I was up.

[47:44] the track all weekend because we had our practices this past weekend. I'm literally here and I go back to Daytona tomorrow. So, um, it's something, how does it make me a better CEO? I think. [47:53] There's actually a lot of lessons to learn from, from racing. It's, [47:56] It's precise execution. So people have this out in both racing and in entrepreneurship, people have this like romanticized or or just incorrect notion that like, you know, if you're racing, it's just like, just go out. You just have like, you know, [48:10] massive cojones and you just go really fast. It couldn't be farther from the truth. Being a race car driver is about precision. It's like you hit the mark perfectly every time and it's just tiny, tiny movements, right? Same thing with being an entrepreneur. You have to be precise. You have to [48:27] in the way you hire, in the way you talk about things, in the way you focus the team. It's not something where it's just like, "Oh, I talk about some high-level bullshit and then everything just falls into place." No, there's a disciplined execution. [48:42] that happens. And so those things are actually kind of similar, just the timescales are compressed. And I think for me personally, it's just like, it's a way to just break my mind out of a local minima. You know, it's like when you're in a race car, you're in a complete zone. [48:56] like there is nothing else outside of that i mean i had uh i was at lamar um in the summer last year and you know i did a quadruple stint which is quite a long time like two hours 45 minutes in the car straight [49:09] I can't remember the time. You're literally just in there, just boom. I'm just doing the thing over and over again. Like you come out and you have this Zen feeling. It's really, there are probably much cheaper ways to accomplish that, mind you. But I love it. - Get a one week meditation retreat. - Yeah. - Just silent meditation. Safer, we need you healthy. - That's no fun though, come on. - Do you have a key man clause or anything? - No key man clause. - Don't ask about those things. Why do we have to go there?

[49:38] Um... [49:39] No, but race cars are pretty safe. [49:40] so but it's actually i don't know there's a competitive element to it also like being strategic [49:45] you know, judging risk. Like everything you do in a racetrack is about risk reward trade off. Right. I'm going to pass this guy now. I'm in a 24 hour race. I'm an hour one. Is it worth it? Maybe not. Right. So you have to make these micro decisions all the time, which again, is very similar to running a company like you're constantly making tiny decisions. [50:04] the success or failure of that company is a is a culmination of all those little decisions same thing in a race 24-hour race the decision you made over a three millisecond period might have changed the outcome [50:17] So make good decisions. And I think these are the kind of things you actually learn. I think there's a lot of parallels. And that's probably true for many sports, but... [50:25] you know. [50:25] How fast can you go? [50:28] Depends on the track. Well, that was probably the fastest track. We hit about... [50:31] 200 miles an hour on that track. [50:34] okay that's nice yeah um constantine any hobbies on your side [50:39] I work all the time. Okay, well, I think we'll open it up to the Q&A if anybody has any questions. [50:50] We got one open. [50:52] Or I guess we have mics, but we'll come back. It's OK. [50:59] So my question to Naveen is, [51:02] What do you think of combining the physical and virtual systems? So creating a thing that's

[51:11] what [51:13] create a new kind of symbolic system based on topology of the object. [51:17] I'm talking specifically about Hopfilms, but Hopfilms, it's a note of light. [51:24] And so it's possible to create a kind of new alphabet or more complex, just a language, and describe it in the Hopfion, based on the topology of the object. And my other question to Konstantin is, what in case of mind, like people, like the evolution of mind will be getting further [51:48] Question is, what is coming after Heidegger? [51:52] Heidegger is existentialist. And so, yeah, that's it. [51:59] Well, I'll attempt to take the first one. I think you're saying virtual instead of doing it in an existing numeric machine. [52:07] The knock of life is in the mode. [52:10] in the greci [52:11] Okay. Uh... [52:13] uh but combining with bullshow is me i mean the [52:18] one deal, but make the work in column. [52:21] All right. [52:22] So, okay, I mean, I would argue that all computers... Thank you for giving that one to Naveen. [52:28] So all computers are that, right? So any... [52:32] computation or memory does result in a physical change. [52:36] We cannot represent information without a physical change. In fact, that's that's fundamental to this Landauer principle that we that Constantine was talking about earlier. So it's about.

[52:47] what kind of change can you use to represent information? And so, you know, we're very good at building semiconductors that can use charge. [52:54] to represent information and we can manufacture those at scale. [52:57] That is by no means the only way to represent information. So I think what we're doing, unconventional, I would say is, [53:04] Can we change the way we think about computing instead of becoming fixed bit with digital numeric arithmetic? Can we think about it as like a dynamical system that evolves with some sort of time varying physics? If I do that. [53:22] Now I can represent information on on physical media to have their own dynamics. [53:27] First one is going to be silicon because we know how to build it. [53:31] again when someone comes up when i say i said when not if but when someone comes up with a better way of representing information it's smaller you know more energy efficient denser whatever we can use that [53:43] But we have to move away from that abstraction in a sense. So I'm open to any new ideas, but we have to be able to build it and manufacture it. That's the five-year problem. [53:52] So I don't see anything else that fits that bill right now. [53:55] Thank you. [53:56] I'll jump into your answer about the mind and what happens next in this [54:01] era where we have massive uh artificial cognitive scale um so there's the physical and then there's the metaphysical i'll leave the metaphysical because that's a matter of belief and religion and i think that that is you know its own very important set of things but on the physical um

[54:22] I think that we're going to have more and more computation doing more and more cognitive work. And I don't think that that replaces the most fundamental human interactions or the fundamental needs of the mind. [54:34] Thank you. [54:35] And I'll go back to the physical world again. Theoretically, [54:40] it makes sense for all of us to be moved by mechanical objects. Because the cost of [54:46] driving mechanical scooters everywhere. Everyone in this room being on a mechanical scooter to exert zero effort so we can move faster between destinations and doing everything over Zoom. This is a packed room, but theoretically we could do this without physical exertion at all. [55:02] would say if that technology was allowed, humans would all be sitting in front of a screen from a distance and minimizing physical exertion of effort. [55:11] If that would happen, that would suggest that when cognitive processes occur, automation of cognitive processes, that will also minimize human cognitive processes and offload everything. [55:24] to machines. So I don't think we ever do that. [55:27] I don't think we ever offload everything to machines because there's this old Greek philosopher that on the physics would say the man is the measure of all things. The person is the measure of all things. And I think that we will choose to keep our minds largely as they are, just like we've chosen to keep our bodies largely as they are even after life. [55:46] the industrial revolution and the automation of the vast majority of physical work. [55:50] Guthrie, up in the front, what was your question? If the ultimate output of AI is the completion of a cognitive task, to your point, how do you actually measure the efficiency improvement?

[56:03] across many different types of tasks. [56:05] Yeah, that's a great question. I mean, we have a convenient way of doing that now because of tokenization. [56:11] at least in the near term, the way we're measuring it is joules per token. [56:15] pretty straightforward. It's very hard to game. Right? So, you know, with ISO quality, holding the quality constants and joules per token. Yeah, beyond that, though, you could argue that, okay, is the token the right [56:30] the right atom of that task. [56:33] maybe maybe not right so that that starts opening up um some gamification so i think when you start moving away from that thing being the unit of measure it starts to open this up and that will happen [56:43] absolutely so then it'll come down to well okay how quickly do you can you do it for or how much is it cost to do it will eventually come into cost and i think at some point there will be some tasks still that humans will probably still be pretty good at or at least in guiding the system um because it's going to have different um attributes than a human brain in terms of what what it's good at what it's fast at what it's slow at what's energetically intense at what it's not energy [57:13] between different tasks at some point. [57:15] Thank you. [57:16] So these previous approaches like spiking or computing memory, they've been tried, they've [57:27] much as we would like compared to artificial [57:30] annual networks. Can you comment on that and how that is different from what you're thinking?

[57:35] Yeah, so first off, just because something is tried does not mean it's wrong. [57:39] And it is absolutely the wrong reason not to do something. I've heard that so many times throughout my life. It's absolutely not the right reason to not do something. Was it flawed? [57:49] I don't think fundamentally flawed. I think it was flawed from an execution standpoint. We have many neuromorphic approaches actually [57:57] started very early and it was just from a pure biomimicry standpoint. Also, I think because we didn't understand learning systems like 20 years ago, [58:07] it was very hard to actually put some kind of frame in what's useful. [58:11] out of these. Like the question that was just asked about how do you evaluate it, that wasn't even clear 20 years ago, which now it is. So I don't think neuromorphic stuff was wrong. I think it was just maybe a bit early. Now, [58:22] What we're doing differently is we're not being so literal about biology. [58:26] Biology is built out of very different stuff than silicon. [58:30] We are choosing silicon because we can manufacture it. The nonlinear dynamics of silicon are different than biology. I'm not saying better. I don't know if it's better or worse, but they are different. So that might necessitate different ways of thinking about it, like spiking in a neuromorphic system. [58:47] You don't need to make it spiking. If there's something that, if it solves a problem for us, like transmission information or something like that, in a particular energy envelope, sure, we'll do it. But we're pretty open about how to build these systems. So we're trying to build the systems kind of [59:00] for what their attributes are versus like trying to fit a paradigm on top of it.

[59:05] So I think that's a big difference. And also, you know, we are [59:10] We do have these... [59:13] almost not ground truth, but sort of a golden model in a sense. Like I can compare it to a digital version now, which I couldn't do before. [59:20] I remember the neuromorphic stuff back in the 90s. It was like, oh, I can... [59:24] you know, do some sort of [59:26] classification or something like that, which was very constrained and not super... [59:30] comparable to a digital system. Now I can directly compare and I can say what I'm better at. So I think just having that framing actually makes it successful. But we're going to make it work in silicon and make it manufacturable and it doesn't need to look like biology. I'm really glad you asked that because it's partly what's so exciting here. [59:49] The understanding is that, [59:53] Perhaps... [59:54] This is a field that has been tried. But the reality is the world is full of fields that have been tried, but have not been brought to their completion. I mentioned earlier electrification of vehicles, so the EV revolution. And that happened very recently. [1:00:16] even though the electric motor has been along here for a very, very long time. And there are so many benefits because when you come up with the internal combustion engine, you have to deal with all sorts of physics and transmissions and coolants and all these things that you don't have to deal with when you go with just the fundamental basics of the physics of an EV. Similarly, with analog computation, there's this

[1:00:40] whole space [1:00:42] littered with [1:00:43] just gems of information over the decades. [1:00:47] from analog computation that [1:00:50] The theoretical limits are even better. If you take it to the extreme of where the physics can go, it's even better. And it also reminds me of neural networks, which have gone through phases. So neural networks-- I'm an AI engineer. It's been the cornerstone of my career for the past 15 years. [1:01:06] During that time, neural networks have been out of fashion and in fashion, not nearly to the extreme as they were in the 80s and 70s when they were completely in a backwater. But let's not forget. [1:01:19] that [1:01:20] The folks at OpenAI were unbelievably courageous because they did something that the vast majority of academia thought was not going to work. [1:01:30] Remember, the vast majority of academia, when OpenAI was doing their research that now has changed the world, would have thought that that would fail. [1:01:38] And you can say that with confidence, because otherwise, academia probably would have been working on it. [1:01:42] So it's so exciting as an investor to find areas [1:01:47] that have just the laws of physics and science and the promise of what can happen on their side, and a great team to go after it. [1:01:56] I think we have time for one more question. [1:02:03] - I have a pretty loud voice, I can go for it. Thank you. [1:02:08] Awesome, again, congratulations, super excited. I have a question, like where do you think is the hard wall today?

[1:02:16] 2026. Is it the energy constraint that you have to fundamentally deal with upfront? Data movement, memory, materials? I know there's a lot of race with materials, foundational materials across the world, particularly with China. So what do you envision like just just going into 2026? With existing paradigms, you mean? Yeah. With existing paradigms? [1:02:39] Starting with the new paradigm that you're beginning to build, but you still have to deal with some existing remnants. [1:02:46] Well, yeah, I mean, if you look at an existing digital system, the vast majority of energy actually goes into data movement. So that doesn't go away. Like to move data, like I said before, all information has to be represented physically. [1:03:01] and when you move information something physical has to change [1:03:05] work equals force time distance. That's still true. If it's charge or if I'm moving something around, I have to physically move it. So that is still true. We cannot get away from this. I mean, arguably some of the optics people would say we can. [1:03:18] maybe at some point that'd be great or quantum people would say we can um there's still a lot of a lot of hard stuff to solve there so [1:03:27] That absolutely is going to be there. I think from a material standpoint, [1:03:32] Lithography is kind of hitting physical limits. Yeah, we've gone to EUV. [1:03:36] Now we have... [1:03:37] 2 nanometer, 3 nanometer, 4 nanometer. They're all very similar density. They kind of have just different attributes. Some are better at this and worse at that. So we're kind of in this [1:03:46] now trading off space. Just to kind of give you an example, again, I go back to biology because biology is sort of an example of a system that we could have observed in the past and had gone through changes.

[1:03:56] As a trained neuroscientist, I can look at the, what we call the cytoarchitecture, basically the cell structure of the cortex of a rat [1:04:06] And a human, I can't tell the difference. [1:04:09] So the evolutionary biology version of Moore's Law kind of ended about 150 million years ago. [1:04:17] The circuit didn't change very much at the micro level. [1:04:20] The macro level changed, the architecture of the brain changed and other things, but the circuit didn't change so much. And so I think maybe we're getting to some kind of limit there with the current lithography techniques. Now, there are people building new lithography techniques, and that's what I think is interesting. It's going to necessitate what we're doing. [1:04:38] If we don't think about how I can use new materials that have different properties and new ways, and if we just put our blinders on and say it must be digital and it must act this way, I think that's going to limit what we can do physically. So that's actually what I'm even more excited about beyond just what we're going to do in five years. It's like, what are we going to do in 20 years? [1:04:56] And I'll say, and I'll add to this, which is a perfect... [1:05:02] example of the physical world and processes where we're limiting. What I see as limiting factor is [1:05:09] is not talent, but the matching problem for talent. [1:05:14] To this day, the vast majority of problems [1:05:18] are still out there. [1:05:20] They're ahead of us. And the amount of creativity and courage necessary to solve them [1:05:26] is still way undercount.

[1:05:28] And I actually do think we have immense talent in this world, in this room, but in this whole world. And finding and developing and accessing and placing that talent in the right places is so under-optimized, so radically under-optimized. There are so many brilliant minds who can solve things in their own unique ways, and for a million different reasons, are not in their best and highest use. [1:05:50] And so that is, if I could optimize one variable to maximize the success of humanity, that's the variable that I'd push on. And frankly, it's a variable I feel like I get to work on every day. [1:06:01] Wow, what a good place to end. [1:06:08] Thank you. [1:06:08] There is one extra question. A bonus round. Okay. [1:06:12] One last thing. [1:06:17] When do the robots take over? No, actually. So you mentioned architecture between the rat and the human. And given your neuroscience background, the one element, you didn't talk about neurocomputing that strikes me. You talked about analog, stochastic, right? That's nonlinear. It's also recurrent. [1:06:33] And if you look at works of Antonia and Hannah de Mazia, for example, or Francisco Varela, a generation earlier, that is the defining feature, right? That the insula tracks, the body state, the change. How do you think about bringing that into what you're doing? And does it play any role at all at this point? Or you're just doing the neurons? No, no, no. 100% it does. So this is what I mean by non-linear dynamics. [1:06:54] So if you look at these recurrent systems, [1:06:58] We can study them in crude ways, which would

[1:07:00] dunk a bunch of electrodes into a brain and watch what happens. You actually see that like there's there's a ringing state. So you can perturb the state and it moves through these like transitions. [1:07:12] These transitions actually are computation. [1:07:15] That is how our brain does computation. It's done through this evolving state. So you perturb it from some kind of outside stimulus and you get this kind of [1:07:22] computation, those states are learnable and changeable depending upon experience. [1:07:26] So we're building systems like this. [1:07:28] exactly that. They are recurrent. And we can model some of this stuff and we have modeled it. In fact, if you look at what a transformer is really doing, [1:07:36] I got like a transformer neural network [1:07:41] if anyone knows what a KV cache is, like you get a bunch of tokens in, you do all this stuff where you perturb the system, right? You basically, you set up a new set of, [1:07:52] desired states for the system. It walks through, it puts another token out, it jams that back in, and it does it again. [1:07:59] That's recurrence. The whole thing is actually a dynamical system that is being simulated in this... [1:08:06] horrific algorithmic way that's order n squared in a digital system. But it is a dynamical system that sort of kind of converges. [1:08:15] In fact, it doesn't really converge because sometimes it gives bad answers. Convergence would be giving good answers. But that's exactly what it is. It is a recurrent system. So now can we think about how do we physically build something like this? [1:08:26] that does converge and uses the physics in our favor instead of having to simulate every step of the way

[1:08:31] Fantastic. Okay, that was a good bonus round. Well, first off, thank you so much to Playground, Peter, Hadley, and the team here. And second off, thank you so much to our wonderful guests, speakers, leaders of the frontier of AI. [1:08:50] Naveen and Konstantin, thank you. - Thanks for the great question. - Thank you, Mike. - Thank you, Naveen. Thanks for your service.