Full Transcript

Transcript

Nathan Labenz: (0:00)

Hello and welcome to the Cognitive Revolution, where we interview visionary researchers, entrepreneurs, and builders working on the frontier of artificial intelligence. Each week, we'll explore their revolutionary ideas and together we'll build a picture of how AI technology will transform work, life, and society in the coming years. I'm Nathan Labenz, joined by my cohost, Erik Torenberg.

Hello, and welcome back to the Cognitive Revolution. Today, I'm pleased to introduce Autopilot, a new podcast hosted by Will Summerlin, which explores the past, present, and future of AI and automation with expert guests. Will is the founder and managing partner at Autopilot VC, and before that, he covered software and artificial intelligence at ARK Invest. He was also previously the founder and CEO of Pin AI, a venture backed startup that developed an AI powered security platform.

In this episode, Will interviews Sergiy Nesterenko, the founder and CEO of Quilter, a startup that uses reinforcement learning to automate the complex process of designing printed circuit boards. As an electronics engineer at SpaceX, Sergiy experienced firsthand the challenges of circuit board design. A single board can take weeks or months to design and validate, issues and errors are common, and iteration involves the production of physical prototypes, which can be frustratingly slow. With Quilter, however, Sergiy aims to make circuit board design as fast and seamless as compiling code.

In this conversation, Will and Sergiy dive deep into circuit board design as it exists today, tell the founding story and technology behind Quilter, and consider both how to earn trust with AI created designs and how to navigate the labor dynamics of an industry poised for disruption. While Sergiy is quite upfront about the fact that Quilter can only handle relatively basic designs today, even this suggests major productivity gains and changes to the nature of jobs. And more importantly, if Quilter achieves its vision of superhuman circuit board design performance, which given the success of other narrow reinforcement learning systems seems very plausible to me, we'll see a much bigger impact on employment in years to come.

For more fascinating looks at critical but often neglected domains and the impact that AI automation is and will continue to make on them, subscribe to Autopilot with Will Summerlin wherever you get your podcasts. And, of course, we here at the Cognitive Revolution always welcome your feedback. Right now on our website, cognitiverevolution.ai, you can chat with my personal AI clone, which is powered by Delphi AI and trained on all Cognitive Revolution episodes. Feel free to ask it any questions you have, give it a guest suggestion, or just tell us something that we got wrong. I've had a couple good experiences with it, but I'm really curious to see how it works for others. And finally for now, remember that I am helping a few different companies hire AI engineers. So if you're looking for a new AI engineering role, you can go ahead and DM the real me on the social media network of your choice.

With that, here's Sergiy Nesterenko of Quilter on the AI automation of circuit board design from the new podcast, Autopilot with Will Summerlin.

Will Summerlin: (3:05)

Welcome to Autopilot, a podcast where we discuss the past, present, and future of AI and automation.

Nathan Labenz: (3:15)

From Atlanta, Fulton Passage.

Nathan Labenz: (3:19)

That's one small step for man.

Will Summerlin: (3:22)

We explore how AI is creating and disrupting large industries in conversations with top founders, investors, and historians. I'm your host, Will Summerlin. Let's dive in.

Will Summerlin: (3:39)

Sergiy, thank you for joining us. You are the founder and CEO of Quilter. You've just raised a $10 million round led by Benchmark, so congratulations on doing that. In a little bit, we'll get into PCB design and how it's broken and how you guys are fixing it. But for context, let's start with you and the founding story. What were you doing before, and why did you found Quilter?

Sergiy Nesterenko: (4:00)

Yeah, Will, thanks for having me. It's great to be here. My first job out of college was at SpaceX. I was responsible for Falcon 9 and Falcon Heavy second stage avionics in high radiation environments. Basically, making sure that when the rocket goes to space, it doesn't die. That was my job, which involved a whole lot of electronics. The founding story of Quilter was largely inspired from that, just suffering the pain of developing hardware and all the supporting hardware to develop hardware at SpaceX.

Will Summerlin: (4:26)

If we go back to SpaceX for a second, we'll get into what you're working on now. What lessons did you learn? SpaceX is obviously an incredible company. It solves really hard problems. It seems to have a relatively intense culture. I'm curious what you took away from that culture that you're bringing over to Quilter.

Sergiy Nesterenko: (4:42)

Yeah, there's a lot. I think on a daily basis, if you ask the awesome team that I work with, I'll probably cite some learning from SpaceX almost every single day. Off the top of my head, I think maybe the biggest one is just how important it is to have something exciting to work on. The fact that at SpaceX you'd come in the door, and it's like, alright, we're actually trying to go to Mars. That's not a joke. We have to actually make decisions that eventually get us there. That's exciting. It makes recruiting really easy. It makes showing up at work really fun. It makes you work hard at your current project to get another cool project in the future. I've seen some of that here. The team at Quilter is awesome, and everybody shows up with a ton of energy every day. We've been lucky to recruit amazing talent. There's just something wonderful about working on something hard, something important, something cool. That's the biggest one. There's probably dozens of others.

Will Summerlin: (5:36)

It's like the classic missionary versus mercenary argument. Right? The missionaries will be intrinsically motivated.

Sergiy Nesterenko: (5:45)

Yeah. I've never worked in a context where that wasn't the driving case, but from my experience, that certainly resonates.

Will Summerlin: (5:53)

And when you started to think about founding Quilter, what was the founding thesis? What problem were you trying to attack?

Sergiy Nesterenko: (6:00)

Yeah. So when we founded Quilter, it actually wasn't focused at all on circuit boards, believe it or not. This was 2019. It was a slightly different capital environment than it is today. I got the opportunity to meet an investor. That was Michael Dearing at Harrison Metal at the time, and just talk about what it might be like to start a company, what we might do. We first actually headed in the direction of making low cost sensors for manufacturing. That was the first idea and thesis. Spent a few months thinking about that direction. Wasn't a good idea, at least not for us. We were too late to it, I think. Shifted to doing analysis on those sensors. That turned out not to be a great idea. Then shifted to reliability analysis on circuit boards. And then I started thinking, why do analysis if I can just design it correctly in the first place? So it was almost a whole year of this wilderness period of what am I actually going to do with this company before we landed on creating circuit boards.

Will Summerlin: (6:58)

So when we think about circuit boards, maybe we can start at a fifth grade level, and then we can dig a little bit deeper and get more advanced. If you're explaining this to me like I'm a fifth grader, what are circuit boards and if I'm ripping open appliances in my house, where would I find them?

Sergiy Nesterenko: (7:13)

Yeah. So first of all, circuit boards are just an important part of electronics, which is typically the thing that connects all the different pieces together and makes them come to life. A circuit board is kind of this rigid piece of fiberglass and copper that connects a bunch of chips. So these chips on a piece of electronics are the things that perform these different functions, like do some processing, store some memory, have some power, and the circuit board glues them together. It's a passive element, and it just is meant to connect them logically. So when you're taking something apart, you'll find them just about everywhere. Obviously, something like your computer or your TV is probably primarily a circuit board. You'll see it in your light bulb. You'll see it in the control center of your stove or whatever else. Anything that has any amount of sensing and actuation will probably have electronics and therefore circuit boards.

Will Summerlin: (8:11)

Does every electronic device have its own unique circuit board design? I would assume that my laptop has a different circuit board design than my iPhone.

Sergiy Nesterenko: (8:19)

Yeah. That's a really good question. In general, yes, very much so. The thing that is fixed and hard to create and therefore is reused between a lot of products are the chips. In general, creating a new chip is a multimillion dollar proposition. So it's very difficult to go create a unique chip for a unique product. The circuit board is kind of the opposite. You can take all of these chips that already exist, glue them together in some new way, and make a new product. So, yeah, definitely, the board in your iPhone is different than the one in your Android, different than the one in your computer, so on and so forth with relatively few exceptions.

Will Summerlin: (8:55)

Is the complexity of the circuit board correlated with the sophistication of the device? So, like, an iPhone would have a much more sophisticated circuit board than an alarm clock?

Sergiy Nesterenko: (9:06)

Oh, absolutely. Yeah. The sophistication of a circuit board or the difficulty scales with roughly two things. One is just the combinatorial complexity. So how many things are there on that board? How densely packed are they? And then the other is kind of physics related. Does the board have a lot of very high speed signals which are difficult to work with, or does it have a lot of very high power signals which create heat in the chip to manage? And yes, of course, the difference between an NVIDIA video card and an alarm clock is night and day.

Will Summerlin: (9:39)

That makes sense. Maybe we can shift into the design phase of this. How do you, without Quilter, say we're using the traditional methods of human labor, human knowledge work labor, how does one go about designing a circuit board? What would I be optimizing for if I were to design one? And what tools would I be using?

Sergiy Nesterenko: (10:01)

Yeah. So I like to think of the process of designing a circuit board in three major steps. And I often like to make the analogy to software, because more people, I think, have written some code than have designed a circuit board, so I find that helpful. So the first step to designing a circuit board is to create what's called a circuit schematic. If you've never seen one before, it looks a lot like a block diagram or a flowchart. So you're in a relatively human readable language trying to explain what are the inputs, outputs, and function in the middle that the board will do. So you might say, I have a connector. I have a processor. I have some memory. I have a radio antenna, and then I drive that antenna or something like that. You can think of that a lot like writing code. So writing Python, this kind of human readable representation is a lot like creating a schematic. That's the first step.

The second step is to convert that schematic into a set of blueprints that a manufacturer can actually use to turn into a real physical design. Once you have the schematic, you have to explain to the manufacturer exactly how to arrange copper, fiberglass, solder, and components together to make that schematic real. And so I think of that kind of like compiling code. That's the second step.

And then, of course, the last step is to take that blueprint. It's called a layout. Send it to a manufacturer. The manufacturer creates it. You get the board on your desk, and usually you figure out that you messed something up and you try all over again.

Will Summerlin: (11:25)

How long does it normally take to go from step one to having a manufacturer actually produce something? Is that typically a week long period or a year long period?

Sergiy Nesterenko: (11:36)

Yeah. It depends very much on the board you're designing. Once you finish the schematic doing this process called layout, it can be a couple of hours for something really, really simple with a few components, or it can be many months. If you're thinking of a Starlink antenna or an iPhone motherboard, you're easily looking at two, three months of manual labor to to make that design happen. Ironically, the process of doing layout is much faster in a lot of cases than manufacturing itself. In many cases, especially if you're willing to pay for it, once you have a design, you can have a prototype on your desk within 24 hours or even just a few days if you're willing to pay for fast shipping from China. But the process of doing layout varies between typically a few days, but it can be a few months.

Will Summerlin: (12:22)

Is there some level of validating the design before you go to manufacturing, not only to make sure that the system works, but also to make sure that it's manufacturable?

Sergiy Nesterenko: (12:32)

Yeah, that's a good question. So there is some amount of this. Yes. In general, basic questions about manufacturing are pretty well handled today. So one of the first things that you're doing when you're creating a board is you're deciding where exactly do all of the chips go? How do I make sure that none of them are colliding? And then how do I connect all of them with copper on the board? So making the basic checks around, hey, nothing's colliding. All the chips are connected the way that your schematic said, and all of the connections are far enough apart that the copper doesn't bleed into itself because of etching imperfections, stuff like that. That's called the design rule check or DRC, and it runs very quickly in any of these tools that you use to design boards. That's not so bad.

The hard part starts to show up when you start to think about more complex interactions. I'll give you an example. One of the difficult things to understand in circuit board design is this notion called crosstalk. So if you have two connections between chips that are not tied to each other, not actually physically touching, they can still couple a lot of energy between each other. So you'll have a signal on this wire A. That signal, because it's a wave in electromagnetics, can jump over to the wire B. And in the most extreme cases, you can actually couple more energy to the other wire than the one that you intended. So this can very quickly ruin your board from working at all.

And so you have, as a designer, potentially hundreds or thousands of these possible imperfections to consider. And in general, it's very difficult to make sure you've gotten them all right. Designers typically use rules of thumb to design boards, quote, unquote, correctly, as they make an attempt at this. There are some simulations that you might run on cases where you know there will be issues, but there is not a one click button that definitely gives you a sure thumbs up that you are completely confident it's going to work. And practically speaking, it almost never does work. You almost always find something that you have to go back and redesign for.

Nathan Labenz: (14:34)

Hey, we'll continue our interview in a moment after a word from our sponsors.

Will Summerlin: (14:36)

So if I'm going to replay this a little bit, the design process today without Quilter is almost based on tribal knowledge that sits within a designer's head, and it's a relatively manual process to solve the puzzle. What is the designer's life like with Quilter? How much of that do you guys automate today, and how do you think that's going to evolve over the next five years?

Sergiy Nesterenko: (15:00)

Yeah. So the principle of what we're trying to do is to completely eliminate layout altogether. We want to get to the point where doing a PCB layout is like compiling code, where in practice, pretty much in all cases, people just click a button and it runs. Now, of course, that's much easier said than done. And so that's an easy thing to say, but very, very difficult thing to do.

So with Quilter, for the boards that are in our scope today, which I could talk about in a second, that is very much the process. The designer will finish their schematic. They will open up the beginnings of their board file in order to enter some constraints. So for example, you might have very physical requirements on how big the board can be, where connectors go, where buttons go, where a screen might go. And then everything else that you need to do, you simply leave off the board in the design file, upload it to Quilter, and then Quilter drives it all the way home straight to something you could submit to a manufacturer.

But, of course, I should make it clear that it's limited what kinds of boards we can do. If you give us a computer motherboard or an NVIDIA video card today, we can't do it. And so you're just not going to get a good result. You probably won't even get a completed result. But for boards that are sort of within our parameters, we've done this many times and and it works just fine.

Will Summerlin: (16:18)

Is the thesis that over time you'll sort of move up the stack of complexity and get to a point where you can design a board for a laptop? Is the market big enough at the bottom?

Sergiy Nesterenko: (16:29)

Yeah. I mean, I think that there is a significant market at the bottom of the complexity chain for sure. Even quote, unquote easy boards where the physics isn't that complicated, they still can take a week to do or two weeks to do by hand. So there's still a very, very significant value to eliminating that work and, even more importantly, reducing that latency of your product iteration. But we, of course, want to be able to do all of them because we can eventually unlock the potential to design better boards than humans can, and that opens up a whole new set of considerations.

Will Summerlin: (17:02)

How much are humans currently paid? Do you have any rough idea of the current labor spend on board design?

Sergiy Nesterenko: (17:10)

Yeah. Totally. So on average, this is actually a profession. So there are people who specialize in this specific task of doing PCB layout. I believe the median salary in this profession is about $100,000 to $110,000 per year. And it's not just those people who do this task. Almost anybody who designs electronics eventually has to do their own layouts, and really, professional PCB layout folks are typically reserved for designs that are really complicated and really mission critical.

So, again, thinking back to SpaceX, right, we had a team of a few dozen people who would do PCB layout, but they could only support flight computers and the intense radios and everything like that. Even for someone in a position like myself doing radiation testing, we had to do our own designs, because that team just simply couldn't support us. And it could easily take up half of your time designing these kinds of boards. So in total, we estimate that there are you know, it's hard to get a precise number, but $20, $30, $40, $50 billion a year of labor being spent on this task.

Will Summerlin: (18:18)

What are the labor dynamics like? Are there labor shortages? Is there an oversupply of board designers?

Sergiy Nesterenko: (18:25)

Yeah. I think there's actually a drastic undersupply of board designers and electrical engineers in general. It's definitely a big problem. I think what you're seeing is that and I don't have numbers for this to make it really concrete, but I think what you're seeing is that a lot of the young engineering talent nowadays is choosing to go into software and AI and robotics, and fewer are going into electrical engineering and specifically these skills of intense PCB design. I mean, just from my own experience at SpaceX, I wish we had more support for doing layout with that central layout team because then you don't have these folks in all these adjacent roles doing that same task. I think there's a significant shortage of this kind of skill out there.

Will Summerlin: (19:07)

SpaceX is somewhat unique in terms of vertical integration. Do most companies have designers in house, or are they third party services that people typically rely on or companies typically rely on?

Sergiy Nesterenko: (19:20)

Well, so I think most companies that develop hardware have their own internal electrical engineering talent. That doesn't mean that they necessarily have their own layout folks. That can depend case by case. But even a company that is five or ten people developing IoT products will probably have a few engineers doing the PCB design, the electronics design. And either they will choose to have someone do the layouts, or they will pay someone externally to do the layouts for them.

Will Summerlin: (19:51)

Got it. That makes sense. Maybe we can talk about the incentive structure a little bit. If I'm a hardware company and I need to design a circuit board, what do I care about? Do I care about the time it takes to design it, the cost to design it, the efficiency of the system? What are the incentives that I really care about?

Sergiy Nesterenko: (20:11)

Yeah. I think that really depends. That really depends on who you are within the company and the application. So we can maybe talk about a few different cases. If you are designing an early prototype for a board, the most important thing is probably the speed. What's the fastest you can get it on your desk and start playing with it? And just to give some perspective to this, for every one flight computer, you're probably going to design tens or hundreds of early prototypes, validating little pieces of the design, validating memory chips and power chips, and also creating support equipment to fully test it.

On the exact opposite extreme, if you're Apple and you're designing a motherboard for the iPhone, you probably care most about size, performance, and maybe to some extent cost. I mean, obviously, to an extent cost, but maybe that's not the thing you push down the most. So it really depends whether you're building a one off prototype, in which case I think speed is very important, and, of course, it has to work, or if you're building a mass production unit, in which case reliability, cost, form factor, maybe even aesthetic are the most important things.

Will Summerlin: (21:21)

Maybe we can double down on performance. What does a well designed high performance circuit board look like versus a poorly designed low performance one?

Sergiy Nesterenko: (21:31)

Yeah. So the simplest way to look at this is when you just look at the raw PCB itself. So when you've got a schematic and it does some function, like, maybe it's an NVIDIA video card, some fast processor with fast memory, the job of the circuit board is just to faithfully recreate that schematic. So a poorly designed board is just going to prevent your chip and your memory from working at the speed that they could work at. What you'll find is that if you run them really, really slowly, everything is okay. But then as you turn up the speed of the communication between the memory and the chip, eventually, it starts to fail. And, of course, that's terrible. You don't want to be bottlenecked by the performance of your board.

So there's a threshold you have to meet on how good your design is to not ruin the schematic that you designed. The other thing is that there's other limitations. So for example, to make a product in United States, you have to pass FCC testing. You have to make sure that your board doesn't emit radio waves that jam everything else in your house. So, of course, if you've made a poor design, that's going to happen, and you're not going to be allowed to sell the product. But if you've made a good design, that won't happen. There might be some cases where that's really the ultimate goal, is making sure that the board itself is not the bottleneck for your schematic, that it's the raw chips that you're working with that are the bottleneck for speed or power performance of your design.

Will Summerlin: (22:58)

And if you think about the arc of using some level of AI and automation to design circuit boards, if we start as trying to get to a point where you're as good as most human designers, the average human designer, and then the end of the arc is you're better than the best human designer. Where are we on that arc? And do you think we'll get to a point within the next five years or so where your system is better than the best human designers?

Sergiy Nesterenko: (23:22)

Yeah. I mean, frankly, I don't think we are as good as the average designer on anything physics and density related. This problem is very, very difficult, and I don't think we've reached that point yet. I think the only thing that we are significantly better than humans at is just time. So on a board where these physics constraints are not as significant, we can do in an hour what would take someone a week to do. But, of course, the interesting and difficult parts of this come up when we are as good as or better than most humans at all of the physics things. And I absolutely think that's achievable within five years.

Will Summerlin: (23:59)

You mentioned earlier, perhaps we'll be able to do things in the future with better designed circuit boards than we have today. What are those things? What is the horizon of interesting problems we can solve with really good circuit board designs that we just don't have today?

Sergiy Nesterenko: (24:15)

Yeah, I think a really obvious one is cost and size reduction. So one of the things that a designer has to consider is if they sit down to do a design that's going to take two or three months to do, their worst fear is being a few weeks away from completion and realizing that the design's not going to close. They just simply can't get everything to fit and be set up in a functional way. That would require them to go back and resize the board, for example, or add layers to the board, which could easily add more weeks or even a month or more to the cycle, which is unacceptable.

So what you find people doing is making a lot of conservative assumptions. You make the boards a little bit bigger than they have to be. You give them more layers than they have to have. You use thicker traces or more copper pours than you might need to use. And so that ultimately results in larger boards that have lower yields to manufacture and are therefore just more expensive. I think that's probably the most obvious thing that could be done to improve boards.

Will Summerlin: (25:14)

And if we dig under the hood a little bit, as you guys think about building your technology stack, obviously, you're not building on top of GPT-4, this is a different problem set. Maybe you can talk through the type of machine learning approaches you guys are taking and specifically whether or not you need a lot of proprietary data and how you go about acquiring that data.

Sergiy Nesterenko: (25:38)

That's a great question. The best analogy for how we approach this problem is to look at what DeepMind did with AlphaGo. We use a branch of machine learning called reinforcement learning, where you create some sort of game, create an agent that can play that game, and you have a score or reward that tells the agent how well it did. Obviously, with DeepMind, they did this with the game of Go and famously defeated the world champion. What we're doing is creating, quote unquote, the game of circuit board design. We're creating agents that play that game, and then the score is physics and manufacturability. How cheap did you make this board? Is it actually going to work? That's what tells you whether you won or lost.

The nice thing about reinforcement learning is that you don't actually need any external data. This is something that's been tried multiple times in this space. A lot of people have this first principles idea of taking a bunch of human-designed boards, showing them as examples to a neural network, and trying to recreate them. That's not how we approach this problem, and that's not what we do. In part because it's actually just difficult to get really good designs. There's not a GitHub of millions and millions of designs out there. Furthermore, the best designs are locked away at NVIDIA, at Apple, at SpaceX, and not in the open source.

But second of all, I actually think long term that's the wrong approach. If you focus on learning from human examples, you're likely to peak at roughly human performance, which is actually what happened to DeepMind when they first started playing Go. They looked at a whole bunch of human examples of moves, trained on them, queried the network to make the best human-like move, and they got to a grandmaster level, but not a world champion level. The best strategies came from the network trying new things, seeing how they work, and remembering them. I think that's what will eventually lead to superhuman level board designs for Quilter.

Nathan Labenz: (27:41)

Yeah. There's all this unexplored solution space outside of where humans have been playing. And if you can explore that, you might find much more efficient designs than what humans know today.

Sergiy Nesterenko: (27:52)

Hey, we'll continue our interview in a moment after a word from our sponsors.

Nathan Labenz: (27:58)

As you think about the advantages of scale in this type of business or product, presumably the more boards you design, the better you get at designing boards, which creates more customers, which creates more boards. There's this sort of compounding moat. I've always kind of thought of compounding moats in AI as a data advantage, right? If you need proprietary data, the more customers you have, the more data you generate, et cetera. How does that play out for you guys? Does that compounding moat exist? And as a result, do you think you'll get harder to compete with as you scale?

Sergiy Nesterenko: (28:32)

Yeah, I definitely think so. One of the things that's possible for us to do, and this isn't something we currently do, but I've personally thought about a bit: when you look at any one customer, a lot of their boards and designs start to look really similar. I'll give you one example. At SpaceX, we were forced to use a relatively small set of components in our library of designs, and it's for good reason. You want to make sure that each part you're going to put on a vehicle is well screened, understood, and has a good supply chain. What that means is when you go to design a new board, you're going to reach into that library for a component you've probably previously used that has already been screened, rather than the ten other alternatives that exist in the market.

Even without that, even for an individual designer, you're likely to work with the tool that you already know how to use. If you're writing code, you're going to reach for a package you've used before. If you're designing a board, you're going to reach for a chip that you've used before. So there's potential to make little mini networks for each customer that have gotten good at their designs and only tackle their designs to keep the data safe to them, but then get special precision on the kinds of stuff they're asking for.

The other way that Quilter can grow with scale is by integrating with manufacturers. One of the other pains that exists today in design is you actually have to go and figure out who's going to manufacture your board first and then design to their specifications. Whereas if you think about it with code, it's exactly the opposite. You write your code, and then you choose at the end which processor to compile it for. As Quilter integrates with more manufacturers, gets to know manufacturers, and closes the loop, that's another place where we can benefit from scale of serving manufacturers a lot more designs at once, helping them with their margins, building those relationships, bringing costs down, and eliminating pain for the customer to submit those designs to the manufacturers.

Nathan Labenz: (30:34)

If you think about the manufacturing side, is most of the manufacturing done in China today, or is it being distributed into other regions?

Sergiy Nesterenko: (30:45)

Yeah, I think it's probably true. By volume, there's no question about it. By number of unique designs, it's probably still true that China manufactures the most. But there is a significant amount of manufacturers here in the States that do local designs. There's two reasons that folks will choose to manufacture boards in the United States. One might be you're required to do so. If you're SpaceX, you can't send your designs to China because of ITAR. The other is latency. When you have somebody fab your board onshore, they can build it and drive it over the same day, and you can be working on that board. Whereas if you have to fly it over from China, you're limited by shipping. I don't have an exact number for how many designs are manufactured uniquely in the States versus China, but the biggest fab in China does about 8,000 unique designs per day. So it's a lot.

Nathan Labenz: (31:43)

That's insane. Is the manufacturing process and the manufacturing layer relatively commoditized? Or do different manufacturers have some proprietary process that is a lot cheaper or a lot more efficient?

Sergiy Nesterenko: (31:59)

The answer is kind of both, depending on what part of the difficulty scale you're looking at. If you're thinking of a really simple board with high margins and thick traces and everything like that, sure, every designer can do it, and the margins aren't that high for anyone. But the different fabs certainly distinguish themselves once you get into special features, once you get into high-speed signals with special materials or tolerances or yield, or whether they focus on quick turn or high volume. That's where you start to see some differentiation.

Nathan Labenz: (32:33)

When you think about Quilter and your ability to automate the design phase, I would assume the process today is you're still good at manufacturing, they ship you a physical board that you can run tests on. Will that be the case in five years, or will you guys get good enough where you can run essentially a simulation and all the tests, and by the time you get a board back from a manufacturer, it works exactly as expected?

Sergiy Nesterenko: (32:57)

Yeah, our goal is definitely to get the board right on the first try. I think that's absolutely achievable. The answer to "will this board work or not?" is in principle computable. I don't think it's easy. I think it's going to take time, but I think it's possible. We're really talking about capturing the process of manufacturing, Maxwell's equations, and laws of thermodynamics. There's a lot to do. It's a lot of compute. It's a lot of things to think of, but it is possible.

We should do it because the problem with waiting for a board to come back, getting it on your desk, finding an issue, fixing it, and resubmitting—that's the thing that takes the most time away from a designer. With code, it's relatively easy to write a bunch of print statements, debug, and keep rerunning it. With a board, it's actually very hard to figure out what you did wrong. That can take weeks or months, or some boards will ship with an issue never figured out. So I think there's a lot of value in getting it right the first time if you actually can.

Nathan Labenz: (34:00)

You mentioned compute. Are scaling laws here the same as they are in LLMs, where the bigger the model, the more compute you use to train it, generally the better the performance?

Sergiy Nesterenko: (34:12)

I mean, increasing model sizes, increasing compute leads to better performance, but it's not the only bottleneck. In reinforcement learning, you have these three totally different pieces, right? What is the game that you're playing? What is the agent? And what is the score? You can't just invest in a huge agent but then have a very difficult to play game that's super slow to evaluate and a score that doesn't actually reflect real life. So we invest not just in more compute, more model. We invest across all of those pieces.

Nathan Labenz: (34:46)

Have you guys been constrained in any way by GPU access, or has that not really been an issue?

Sergiy Nesterenko: (34:53)

Not at this point. We try to focus on problems where it's not requiring some massive million GPU for three months kind of run or problem. I mean, that's very expensive, but also very slow to iterate on. It's a surprising amount that you can do in reinforcement learning with a single GPU in a few minutes. One of my favorite demonstrations to show people is that you can train a robot to walk in a simulation in five minutes on a single GPU.

Nathan Labenz: (35:25)

Yeah, it's really... that's five minutes of training compute?

Sergiy Nesterenko: (35:28)

Yeah. Yep. This is worth talking about. There's a cool library called JAX, I think coming out of Google, that basically gives you a whole bunch of CUDA access for training models, building models, and so on. What they've done is they've written a physics engine in that language, so to speak, that also evaluates on GPU. So you can simulate a humanoid or a hexapod or a cheetah or whatever sort of robot system that needs to walk on a GPU, and you can evaluate about 50 million of those episodes—taking a thousand steps, roughly speaking—in a matter of a few minutes.

Given that that's possible, we try to look for things like that as much as possible. Because then we're not just hoping for the best with a huge model. We're really getting at: are we using the right model? Do we have a game that's that fast to evaluate? Are we looking at the right things? We can iterate faster.

Nathan Labenz: (36:28)

Outside of circuit board design, what other problems or domains of human knowledge work do you think could be solved or partially automated with reinforcement learning?

Sergiy Nesterenko: (36:40)

Oh, that's a good question. Most explicitly, the stuff that I've been thinking about is tangential to circuit board design. There are two other similar problems that look almost exactly like circuit board design, and that's chiplet design and then potentially chip design. Those are really interesting. But we can classify that as electronics design. But I'm excited for it in a lot of different cases. I mean, I'd love to see more solutions doing mechanical approaches like this. When you consider something simpler like designing a fan or a heat sink—what's the right way to arrange fins on a heat sink? That's something that could be interesting to do with reinforcement learning. All the way through to: can we design a plane? Can we design a wing that is more efficient and has less turbulence and whatever else? These are not something I know anything about, but it would be very interesting to explore and study.

Nathan Labenz: (37:31)

Yeah. It seems like any sort of design optimization where you have well-defined constraints, like the laws of physics, you sort of create these outer bounds of constraints and explore a wide range of solution space within that.

Sergiy Nesterenko: (37:45)

A hundred percent.

Nathan Labenz: (37:47)

That makes sense. As you guys think about go-to-market and monetizing the product, who are you typically selling to? Is there a head of electrical engineering inside of a company, or is it somebody else?

Sergiy Nesterenko: (38:01)

Yeah, that's a really good question. Our current go-to-market is primarily driven by our technical capabilities. In the five-year future where we can solve any board and do a computer motherboard, sure, there's going to be a head of electrical engineering or a head of PCB design that we can directly approach who's going to see millions of dollars of value for their company almost immediately. But we just can't do that yet. So we're in this trickier phase of: where can we actually solve problems today while we build to that future?

The most direct interesting segment has actually been what I think of as electrical engineering adjacent roles. I'm thinking of folks who require a circuit board to do their job, but that is not their primary function. Myself from SpaceX was exactly this kind of example. My primary function was to test electronics and design and predict radiation effects in flight and make sure that it was all okay, which required me to design a lot of boards. But that wasn't the main thing I was trying to do. So for me, being able to automate that piece and just have it work was really attractive, and I was not designing the kinds of boards that were so complicated that it couldn't be done.

We're seeing interest from folks that are firmware engineers, robotics engineers, automation engineers, test engineers—lots of these tangential roles that are not designing the main Apple motherboard, but are designing the things that are immediately adjacent to it.

Nathan Labenz: (39:31)

That makes sense. And as you think about pricing and the long-term pricing model, if you're able to do designs as good or better than humans, do you think about a pricing model that reflects the human labor costs that you're replacing, or do you think about it in a totally different way?

Sergiy Nesterenko: (39:50)

Yeah, a hundred percent. I think that we do think about it with the existing way to value this work, because there's already an existing budget item, because people are already familiar with it. The way that this is typically done in the space today, if you were to outsource a layout to an external fabricator, is it's done in proportion to the amount of time that it's going to take. People will take a look at the whole board, count the number of pins or the number of connections that they need to make, and charge somewhere between one to ten dollars per pin or one to ten dollars per connection. It depends on where you're doing it, which country, and so on.

So that's a budget item and a pricing structure that already exists that we can directly tap into. Maybe at first we'll start out on the cheaper side of things, and then over time as we provide more value and more simulation results, we can come up to the higher end of that.

Nathan Labenz: (40:48)

Where have you seen the most market pull so far? Is it in companies that are designing—I know you guys are right now focused at the bottom of the pyramid on simpler designs. Is that typically a consumer electronics company? Is it a defense tech company? Where are you seeing the most market pull right now?

Sergiy Nesterenko: (41:05)

Yeah. So the interesting thing is that it's actually less segmented by the vertical of the company and more by who at the company you're talking to. Whether you're SpaceX or a consumer company, you probably have both really, really difficult intense designs and really, really easy, you know, an intern could do it type of designs. That's where this electrical engineering adjacent thinking came from—we were just seeing folks who were doing these boards that didn't want to be doing them and could be better using their time somewhere else. We see that everywhere. We see that in aerospace. We see that at these chip makers. We see that at consumer companies, you name it.

Certainly, I think if we had to pick a type of company that has seen the most interest, it's those who value speed. For whom having a board on your desk tomorrow rather than next week is critical—that's who's approaching us most.

Nathan Labenz: (42:05)

That makes complete sense. As you go about building out your own team internally, obviously AI talent is hot right now and quite competitive. I've seen this at other places and heard this from other founders that it's actually easier to recruit talent if you're working on something interesting. And you brought this up at the beginning of the podcast, being missionary-driven versus building another chatbot for X, Y, Z. I'm curious if you're seeing that play out in real time.

Sergiy Nesterenko: (42:37)

A hundred percent. Yeah, without a question. The things that make it attractive to work on something like this is, well for one, reinforcement learning itself is a more rare application of machine learning. There are probably ten to a hundred times as many applications of more traditional supervised learning in the industry than there is reinforcement learning. So immediately, experts in reinforcement learning are drawn to us because we're an example of an application of their work, and we've seen some of that.

The other thing that makes it really exciting for people is having something physical that you can imagine. We're not talking about a slightly better user experience score at the end of a chat conversation. We're talking about something you can hold, something you can physically work with. For example, one of the goals that I have for the company—this might be a bit ambitious, but we'd like to do it—I want to be able to take this video call on a Quilter-designed computer by the end of this year.

Nathan Labenz: (43:57)

That's awesome.

Sergiy Nesterenko: (43:57)

Yeah. I think if not this year, we'll pull it off shortly after. But imagine writing some code and then eventually having a board on your desk that you made with algorithms that then does that. That is so satisfying to think about and to do. Certainly that draws people to tackling this problem.

Nathan Labenz: (43:57)

That'll be in the Computer History Museum one day as an exhibit. That's awesome, and that's an incredible goal. Another question on talent. It seems like the trajectory for a typical research scientist focused on AI or machine learning broadly kind of went through undergrad, went through a PhD program, did a bunch of research, and then went to one of the big tech companies and continued doing research there. A lot of that research was essentially academic research, so it was open source. Now it seems like that trajectory has shifted. Now people are going straight from undergrad into companies, whether that's OpenAI or you guys or somewhere else, and doing research that's sort of closed within that organization.

I'm curious if that is a risk, both in terms of the quality of talent, but also in terms of: reinforcement learning is great, there's a bunch of examples from DeepMind that you guys can pull on. As companies sort of close themselves off a little bit more, that becomes less accessible. I'm curious if either of those two things are existential risks for the AI industry at large.

Sergiy Nesterenko: (45:05)

I hesitate to think that that will be. I mean, just historically speaking, when companies go and require a whole bunch of talent and therefore inject money into a bunch of problems, you don't end up seeing something completely isolated. You end up seeing a whole bunch of open source examples and alternatives and a vibrant ecosystem for that. It's hard for me to imagine that wouldn't be the case here. Even if the best RL engineers are going to, I don't know, DeepMind and DeepMind starts to close source everything, maybe they'll maintain an edge for a little while. But just the raw interest, the number of people who are going to go study it, and the number of people who are going to make an open source project to try to get a job at DeepMind is only going to go up. So yeah, I guess it's not something I've considered as a potential threat.

Nathan Labenz: (45:55)

That makes sense. Yeah. And I'm curious, if you look at the AI landscape at a very high 30,000-foot level, there's this big battle between open source and closed source. I'm curious if you have an opinion on how that's going to shake out over the next five years, whether the closed source foundation models run away or open source continues to keep up and starts to create some level of commoditization.

Sergiy Nesterenko: (46:22)

I don't know that I have a good prediction on that, to be completely honest with you. I mean, I think, again, historically, open source has been incredible and underestimated. As a good example of this, take Windows versus Linux. Windows is probably still the most popular OS for home use, but for production servers, Linux runs everything, even though Microsoft spent plenty of time trying to make production-grade servers and Windows Server and whatnot.

The thing I think that's really different here is with a neural network, open source shines when there's a critical mass of people interested in contributing to it, and a large number of relatively small contributions results in a really fantastic product. So you think of VS Code where you can have 10,000 people write a small extension, and because of that, there's a huge ecosystem of things that help you. That makes a really great product. But it's not clear that a thousand people spending an hour a week produces a better neural network or something like that. So that maybe is an argument against open source being quite as competitive as it has been in the past. But I don't know. I don't think I have a strong bet or opinion in one direction or the other, to be honest.

Nathan Labenz: (47:40)

No, that's a really good point. You have these step function improvements like transformer, right, that sort of lead to step function increasing capabilities. And if those happen within a closed environment, perhaps that's an unfair advantage. That makes a ton of sense. We're running close on time here. Maybe if we go back—you've been building Quilter for a few years, and I'm sure you've had a lot of experiences, positive and negative. If you could take everything that you've learned in the process so far, go back to day one of founding the company, what advice would you give yourself?

Sergiy Nesterenko: (48:18)

That's a good question. I mean, I wish I could just remember everything we did wrong and then replay it very specifically. Like, for example, let's skip the whole wilderness period of a year of exploration, and let's go straight to this. But of course, that's not the intent of the question.

I think, honestly, probably the single most useful thing I wish I invested in more early on that maybe is helpful to other founders is accounting for the fact that building a startup is really difficult and really lonely and brutal at times. The things that have been most effective for me in dealing with that is, number one, being surrounded by really great people. If you have a cofounder, that's fantastic. If you have friends who are building really difficult things or building other startups, that's fantastic. When you're talking to investors or partners, you should be evaluating that probably more than financial terms and whatever else because it's going to be so important. Because at the end of the day, most founders just quit at some point, and it's obvious why. Just surviving that is so important.

I think the other is just self-care. Going through a company is a lot more stress and self-criticism than working at a big company, even like SpaceX. I mean, maybe watching a rocket launch on the day of launch is more stressful than an average day at running a startup, but on the whole, it's no comparison. So investing in self-care—having a routine, making sure that you're always carving out time with family, making sure that you're exercising and eating well and sleeping well. All these things are obvious, but so easy to forget when you get sucked into the work and you keep running.

Nathan Labenz: (50:08)

That's great advice. Yeah, I remember I had a startup in a prior life and I was so focused on what we were doing, I didn't have a couch in my apartment. We just spent all the time in the office. Finally, after like six months, my girlfriend came over and was like, "You really need to buy a couch." I was like, "Oh," and that just didn't even cross my mind. But I agree. Yeah, it's certainly important to take care of yourself. Ultimately, in many ways, your job as a CEO is to influence people and make decisions, right? And if you're not taking care of yourself, it's hard to do either of those things. That's great.

You raised money from Benchmark. Benchmark is obviously a legendary firm that many founders want to work with and many investors want to be like. I'm curious, what drew you to Benchmark and what excites you about working with them?

Sergiy Nesterenko: (50:58)

So I got the introduction to Benchmark through one of my existing investors who spoke very highly of them. I'm working directly with Eric Vishria there, who I got to meet and get to know a little bit before we made the commitment. I think there are two things more than anything that drew me to working with them. Number one was experience. I think that there's a lot to be learned from previous failures, but there's potentially even more to be learned from previous successes. Working with someone who has seen that story of seed through public company and seen it multiple times is super attractive. The hope is that that results in avoiding lots of mistakes that aren't necessary and unforced.

The second thing is, as any founder should, I did a lot of diligence on Eric, both on people he recommended I speak to and people I found in my own way. I've never heard such positive founder references. Bar none, I've never done a reference call for an employee or a coworker or a boss or a partner or anything that came up that positive. And so that made my decision for me really quickly.

Nathan Labenz: (52:09)

That's awesome. Well, congratulations on the journey so far and excited to watch it unfold. And if you have a spare laptop at the end of the year, I'll certainly take one off your hands. If people want to follow the journey, or if I'm a consumer company or a company that needs to design electronic circuit boards, or an engineer that wants to try to find a way to work with you, how can we get in touch or follow the journey?

Sergiy Nesterenko: (52:34)

Yeah. Easiest thing is our website, so quilter.ai. If you search for Quilter PCB or Quilter AI on Google, I'm sure you'll find it as well. Certainly, our careers page—we're open to people joining us. We're in open beta, so if you're an electrical engineer, you're welcome to upload designs and try it out. I'm sure we have social links there as well for updates along the way. And yeah, once—well, it probably won't be a full laptop. It'll probably be kind of a Raspberry Pi looking thing that we build—but happy to send you one once we have it.

Nathan Labenz: (53:02)

That's great. Well, thank you, Sergiy, and really appreciate you joining.

Sergiy Nesterenko: (53:06)

Alright. Thank you, Nathan.

Nathan Labenz: (53:07)

Thanks for tuning in to the Cognitive Revolution. If you're using AI to create or disrupt a large industry, I would love to learn more about your story. Drop me a line on Twitter or go to our website. If you want to support the show, the best ways are to leave a review wherever you're listening and subscribe. The Cognitive Revolution is part of Turpentine, the podcast network behind Moment of Zen, Turpentine VC, Age of Miracles, and more shows for experts by experts in tech.