Designing Printed Circuit Boards With AI

Sergiy Nesterenko is the cofounder and CEO of Quilter, an AI platform to automate circuit board layout. He was previously at SpaceX for 5 years and he has a degree from UC Berkeley. They just announced their $10M Series A round led by Benchmark, one of Silicon Valley's most storied VC firms.

(00:07) What is a PCB and its role in electronic devices
(01:30) Examples of PCBs in everyday life
(04:47) Manual components of the PCB design process
(06:26) Introduction to Quilter and its automation of PCB layout
(07:46) What can be automated in PCB layout design
(08:44) Challenges in PCB layout design
(10:12) The role of AI in automating PCB layout design
(12:25) Criteria for PCB design and material selection
(14:38) Thermal management in PCB design
(15:05) Ensuring signal integrity in PCB design
(19:44) The use of AI in PCB design
(20:08) The future of AI in PCB design
(22:07) The importance of material selection in PCB design
(23:32) The importance of thermal management in PCB design
(24:11) Ensuring signal integrity in densely packed PCBs
(26:59) The potential impact of AI on the PCB design process
(28:47) The importance of aesthetics in PCB design
(29:14) The potential obsolescence of human designers in PCB design
(30:26) Exciting innovations in PCB design
(34:42) The need for systematic access to PCB design information
(36:30) Rapid Fire Round

Sergiy's favorite book: The Lord of the Rings trilogy (Author: J. R. R. Tolkien)

--------
Where to find Prateek Joshi:

Newsletter: https://prateekjoshi.substack.com 
Website: https://prateekj.com 
LinkedIn: https://www.linkedin.com/in/prateek-joshi-91047b19 
Twitter: https://twitter.com/prateekvjoshi 

Transcript

Prateek Joshi (00:01.903)
Sergiy, thank you so much for joining me today.

Sergiy Nesterenko (00:05.076)
Yeah, thank you so much, Prateek.

Prateek Joshi (00:07.341)
All right, a printed circuit board, right? PCB, it has become an integral part of our modern electronic universe. It's everywhere and it's insanely prevalent everywhere. So in layman's terms, can you explain what a PCB is and also what it does in like our day -to -day electronic devices?

Sergiy Nesterenko (00:34.42)
Gotcha, yeah, certainly. So I like to give a little bit of context here and maybe also make some analogies to software, because maybe more people are familiar with that space. So a PCB and also designing a PCB kind of consists of a couple of different pieces. So, I mean, first, let's just define what the PCB is. The PCB, strictly speaking, is basically a piece of fiberglass and copper that connects chips.

So we have all of these different chips that do these functions. Maybe you have a processor, an ML accelerator, you have a memory, a power component. These are the individual building blocks of making electronics. So maybe in software, you might think of this as like different packages you might pip install. So what a PCB does, like the raw board itself, is it glues them together. It gives all the chips something to sit on and it connects them all in such a way that they perform some bigger function.

Prateek Joshi (01:30.223)
Amazing. And also, can you give a few examples of where an average person might see PCB in real world?

Sergiy Nesterenko (01:42.1)
Yeah, definitely. So I think maybe, you know, for me at least, the first place I saw PCB was building or taking apart a computer. So, you know, the first time I put together a motherboard and a processor and, you know, put in some memory and a video card, that's probably the first time I saw it. Other people, I don't know how many people do this, but I've taken apart all of my old phones that I no longer need. That's the somewhere you'll definitely see them. You know, and other places might be if you ever work on your car.

You might have seen your electronics control unit in there. I mean, nowadays PCBs are present in like light bulbs and switches. So if you installed a new light bulb, you might have actually seen a PCB on which some LEDs are sitting. Yeah, they really are kind of everywhere.

Prateek Joshi (02:25.839)
Right, amazing. And for all listeners out there who do break your phones and look what's inside, you'll find them. So we'll see how many people do that. All right. Now let's go into the design of a PCB. Can you explain what goes into designing a normal PCB? Just the average, the normal one that goes into our?

phones or laptops.

Sergiy Nesterenko (02:57.076)
Yeah, so at the very high level, I think of PCB design in three major steps. So before you actually start thinking about designing the real circuit board, like the physical thing you might imagine, you start with something called a circuit schematic. So circuit schematic, for those who haven't seen one, looks a lot like maybe a flow chart or a block diagram. Really, its purpose is to describe in a human kind of readable representation,

what are the inputs, outputs, and internal functionality of the PCB. And by analogy to software, that's a lot like writing code, right? So you're choosing what you're gonna provide to your board, you're choosing what it's going to actually output, and you're choosing like the chips you're going to use and how you're gonna glue them together to perform that function. So the second step, that's done kind of like in a...

in a schematic editor, you can think of it as an integrated development environment where you're doing this work. You then take that design and then you carry that over into a kind of separate editor where now you're switching into actually drawing a blueprint. So the second part of designing PCB is called layout. So your task at that point is to take the logical representation that you've already built of what the PCB is going to do and make it real. So then you're finally choosing, you know, what is the size of my board?

Where do the connectors go? Where does the chip go in the middle? And then finally, how do I actually draw like copper, like physical metal that is going to connect all of the individual pieces of all these different chips to perform that function. And once you have that blueprint, more or less, you send it off to a manufacturer who could then create it. You get the board back on your desk. Usually find out you have a bug and you go pursue it and you start the whole process all over again.

Prateek Joshi (04:46.671)
Right. And historically, can you explain how people have been doing this? Meaning, is it just like humans sitting in front of a laptop and a piece of software and stitching it together with the hopes that there won't be any bugs? But obviously, of course, there's always going to be a bug or two. So can you explain all the manual components of this process from start to end?

Sergiy Nesterenko (05:13.492)
Yeah, so it is unfortunately still almost completely manual, much unlike chip design, much unlike software. So, you know, when you start out with your actual schematic, of course, you're expressing what you want your board to be. And so that's a manual process of drawing the individual components and connecting them up. One thing that you do quite a bit of is you read the data sheets. So much like you might open up the API specs of some package you're using, the equivalent of that in electronics is downloading a data sheet from a manufacturer that tells you how to basically work.

that chip. When you get into the layout piece, which is where my company is focused, that's also entirely a manual process. So in large part, unfortunately, people still have to make basically 2D, 3D models of what each of these components look like and how they're going to be soldered to the board. You're then going through these hundreds or thousands of components and finding a good location for them on your board.

trying to place common functions close to each other and predict kind of thermal and electromagnetic issues. And then finally, you are literally drawing these little things called traces, these little wire connections on the board, one by one, one after the other.

Prateek Joshi (06:26.095)
That's a lot of room for error. Meaning, if you have to do it, there's so many little connections and components and there's so much to do here. So maybe it's a good stopping point to talk about Quilter. So for listeners who don't know, can you explain what Quilter does?

Sergiy Nesterenko (06:47.636)
Yeah, so at Coulter, we've set out to completely automate the layout portion of a PCB design. So our goal is to make it such that electrical engineers can focus on creating schematics and identifying some of the top level constraints for their boards, such as what the outline of the board has to be, where they want certain connectors and screens and buttons, but then taking care of everything else, right? Everything else that isn't actually important to the product, but...

is just required to make the board work in the first place. So we, by analogy to software, we think of ourselves as trying to build a compiler for circuit boards.

Prateek Joshi (07:25.679)
Interesting, that's an interesting analogy. And if you look at all the work that goes into the layout part of this process, what, so within the layout work, what's easy to automate and what's more difficult?

Sergiy Nesterenko (07:46.196)
Yeah, I wouldn't necessarily describe any of it as easy per se, but there are gradations. So I think in order of kind of difficulty, I think coming up with a valid location for all of the components that, where the components don't collide and don't intersect and are reasonably placed around is probably one of the easier problems. Now, of course,

how you place the components has a huge impact on how they're going to be routed, all the electromagnetics and thermodynamics that are going to follow. So in that sense, it's maybe not so easy. But for boards that are not that complicated, where signal speeds are slow, currents are low, typically you have a decent amount of room on your board to make sure that things don't collide. So that's kind of like an easy version of the bin packing problem, but with some kind of optimization that you want to do on top of it.

The next difficulty comes in just connecting everything. And I like to give people an intuition for why this is difficult because it's a little bit deceptive. The basic problem you have is you have all of these connections you have to make. So you have a starting point and an ending point and you have to find some path between them. And, you know, creating your first path is actually really, really easy, right? You have a starting point A, you have some obstacles, ending point B. That's just a maze solver, right? CS1A kind of problem.

The hard part comes in the fact that once you've solved that problem, the solution to it becomes an obstacle for the next problem. So your second trace is a little bit harder. Your third trace is harder still. Your fourth trace is harder still. By the time you get to trace 10 ,000 or trace 100 ,000, making sure that you still have room for it is exceptionally difficult. And this is something that humans have an kind of a underrated ability to do. Like I think our...

just evolution has made us rather good at this kind of spatial planning and prediction that is actually rather difficult for computers. So, and of course, if you don't complete that job, there's no point in manufacturing your board, it's not going to work, you must reach a hundred percent. That's the second difficulty. And of course, the final and the really big difficulty is even once you've solved all those problems, you then have to check all of the physics. And there's no guarantee necessarily that you got all the physics right. And then...

Sergiy Nesterenko (10:08.02)
going back and completing that whole loop is the holy grail.

Prateek Joshi (10:12.239)
Yeah, that seems fairly complicated. And when you think about the numbers here, yeah, the first pad seems okay, it's easy. The second is a little harder. And by the time you get to, I don't know, 10 ,000, yeah, it's gonna be insane. So how are humans doing it today? Because doing something like 100 ,000 is a big number. So is there like, I don't know, physics -based simple like,

physics -based software or basic automation software. Like how are people doing this without something like an AI infused tool?

Sergiy Nesterenko (10:51.7)
Yeah, I mean, it's just a painfully manual process, right? So the process of just actually connecting all the dots on a big board, it's very slow. You look at like a SpaceX flight computer or Apple motherboard, for something of that complexity, you're easily talking about two or three months of work, right? This isn't something that is done quickly. It's also not something that can really be paralyzed that well. So for example,

at some of these big companies, like the best that they can do to get this job done quickly is to work during the day, send it to somebody halfway across the world, let them work during the night, send it back and keep going. And even if you have all the money in the world, that's the only thing you can do to get this done faster. You know, the way that they kind of end up solving these problems of like, how do you make sure there's enough space for the last trace is you, you often don't, and you have to go back and redo things like you'll.

you'll be working on a board, you find that, oh no, I'm gonna painted myself in a corner, you go and undo and delete some part of the design you've already worked on, or you shove it around a little bit, make space, and then try kind of a new approach and solution. And it's just very, very iterative like that. And then as far as the physics side of things, there are some kind of industry standards, heuristics, rules of the trade and rules of thumb that people have passed down that are basically conservative.

rules that you can use that kind of guarantee or close to guarantee that the board's going to work and that's mostly what people follow rather than like full wave simulations or anything like that.

Prateek Joshi (12:24.719)
Right, and now that you're automating this process, you're using AI to make it faster, better, hopefully cheaper. How does AI help here? Or rather, how are you using AI to do this?

Sergiy Nesterenko (12:42.612)
Yeah, yeah. So the analogy I like to give, especially to folks in the ML community who are familiar with this, is we take a similar approach to what Tesla did with autopilot, especially in the early days, right? This is like maybe in the 20 year or a hundred year future, one day this will be like a single model that takes in a schematic and just directly predicts copper in one image or something like that. That is nothing like what our stock looks like today.

as exciting as that will be one day, that's just not possible for what I can tell. So what we end up doing is we have a combination of, you know, classic models and classic algorithms where they're appropriate, and then AI at a kind of high level strategy level where appropriate. So to make it even more explicit, we cast this problem as a problem of reinforcement learning, rather than just traditional supervised learning.

So what does that mean? That means that we have the freedom to create our own game that the AI will play. And, you know, unlike chess or Go where the game is well -defined and you simply have to go solve it, we have the power to choose what that game is. We have the power to choose what the actions are. Of course, we invest in making that game as fast as possible to evaluate, to give the AI the best chance to have the most data to learn from. So we invest a lot in how do we give the game the...

easiest actions to use, right? Because the easier the game is to win, the more likely the AI is to find a good solution. And then at a separate level, once we have this kind of game with really fast collision detection and kind of more sophisticated actions that take into account some things that the AI shouldn't need to learn, then we let it do the high level planning of like, okay, where roughly do I think these components should go? Where roughly do I think these traces should go?

And then the game engine will kind of snap them into the precise positions that prevent collisions and short circuits and things like that.

Prateek Joshi (14:38.223)
In your literature, you've talked about criteria like number of pins, number of components, density, signal frequency, current. So can you explain these criteria just for in layman's terms and also give us a sense of what it means to have 100 pins versus 1000 pins or 10 components versus 100 components?

Sergiy Nesterenko (15:05.204)
Yeah, so I think of, so all of these speak to the difficulty of designing a board. So what we're really talking about is, is it an easy board to design or hardboard design? How long is it going to take? I roughly think of these criteria as breaking into two major categories. One is kind of speaking to the raw combinatorial difficulty of actually putting together the board. And the other is speaking to the physics, right? Like the kind of...

sensitivity of the physical imperfections that you will have on that board and how they impact your circuit. So in the componentorics perspective, you know, you can roughly think of this problem as kind of like a bin packing problem with a routing problem. So of course, if you have lots of very large components and they take up, you know, the majority of the area of your board, that's a much harder bin packing problem than if you have.

you know, a 1 % density or 10 % density. So that's one kind of metric that we think about is how dense are the component outlines on the board? You know, the second one is the number of pins. And that basically means the number of connections and number of traces that you have to actually run. It kind of also speaks to density, right? Like if you have a board that has a ton of space and only a few traces, it's a really easy problem. But if you look at like a, you know, a motherboard, you'll see that every last square inch is full of traces back to back to back.

And that's a harder problem. So that's kind of on the combinatorial side. And on the physics side, roughly speaking, it breaks down into, for us at the moment, electromagnetics and thermodynamics. So with thermodynamics, the question is, how much current are we carrying? And therefore, how hot is the board likely to get? And what do we have to do to prevent that from being a problem? And on the electromagnetics perspective, the faster your signals, whether digital or analog,

the more wave -like properties they have and the more non -trivial imperfections you're going to see in your circuit scheme.

Prateek Joshi (17:02.511)
Right. I want to talk about design rules. It's a very critical concept when it comes to circuit board design. So can you explain design rules, like what it means? And also, how do you evaluate the designs before you get to work?

Sergiy Nesterenko (17:24.372)
Yeah, good question. So design rules are intended to help you make sure that your design, your board is going to be good on the first try, right? That's what they try to do. And they attempt to capture all of the different, you know, real things that your board has to go through. So the very first and most obvious set of design rules that's very important to consider is to do with manufacturability. So one of the simplest rules might just be,

you know, what is the width of the thinnest trace that you're allowed to make? Or what is the width of the smallest separation between two traces that you're allowed to make? And what that speaks to is that different manufacturers have different tolerances on their etching process. So if you violate that rule, there's some chance that the copper will bleed over from one trace to the other, and that ruins your board completely. So there's basically a large set of design rules that are...

to do with that sort of consideration, right? How do we make sure that the traces are good, fully connected, they don't collide, that they're not gonna get etched away too much, that everything is solderable and so on and so on and so forth. The second set of design rules is drawn more at the attempt of kind of like considering these physics considerations.

So for example, like one design rule that's really common is you know that you have, I don't know, let's say a five volt power powering your board, and it might need to carry two or three amps of current. So you'll set a design rule that basically says that the width of that trace has to be some limited amount. And so if you do a really, really good job enumerating as many of those design rules as possible, you have a better chance of making sure that all of that physics works. So unfortunately, especially on the physics side,

design rules are almost always incomplete. Like you can't actually really as a human enumerate every single possible imperfection and set the design rules sufficient to really make sure it'll never happen other than on the really simplest boards. So that's where electrical engineers and layout persons use their judgment and experience to basically keep the design rules in their head and keep the kind of simulation in their head as they're designing the board. So of course our goal,

Sergiy Nesterenko (19:38.036)
It requires us to enumerate that set completely rather than what is done today.

Prateek Joshi (19:44.399)
And again, this is maybe a big open -ended question. With sufficient data, can an AI model learn the best design rules? Or is that still something that we have to explicitly specify for it to work? How do you think this is going to pan out?

Sergiy Nesterenko (20:07.924)
Yeah, that's a good question. So I actually don't think that learning what the design rules should be is necessarily the right application. So if you think about the source of the design rules, what they're really coming from are primarily like your intention as a designer for how you want to use the board and data sheets. So I'll give an example. Let's say we have a power converter that's stepping down from five volts to two volts or something like that.

What you can often see in the design is that, okay, you're using some sort of maybe linear regulator and that regulator has some input, some maximum current on the input, some maximum current on the output. That information is in a data sheet, like a PDF that the manufacturer of that regulator provided that specified those maximum inputs and outputs. So the way we think of this, to lift that burden off the user, to like have to go and exhaustively specify every possible constraint and design rule,

we are reading these data sheets to then say, hey, this part, the maximum input it can do is, you know, whatever, two amps, the maximum output it can do is two amps. So let's just make sure that the board isn't the limiting factor here. And let's make sure that the board can support those currents and the input and output. Of course, there's still some designer input. It's possible that a designer, you're only ever intending to use it for a half amp maximum or a hundred milliamps maximum. And you can override us and tell us that. But really the...

Most of the design rules can be found in the data sheets for these parts.

Prateek Joshi (21:39.119)
Another critical element here is the material. So material selection, it can impact how good or bad your PCB is. So how crucial is material selection in PCB design? And also what constitutes a good material? Does it depend on a given design or is it universally, hey, here's some material that's good for most PCBs?

Sergiy Nesterenko (22:06.95)
Yeah, so the material selection is primarily to do with kind of more advanced features. So if you're doing a really basic board, like a, I don't know, like maybe the simplest board you might think about is something like an Arduino Uno or something like that. It really almost doesn't matter what choices you make. At the end of the day, the signals are low enough and the kind of currents are low enough that it's gonna work. So where you get into things like Rogers materials and kind of fancy materials is where,

you might get into like high speed signals or RF traces or something like that. That's where you start to consider that the signal you're trying to work with is really a wave, right? It's not just like a single constant voltage on the line, but it actually like waves up and down. And so all of the properties of waves start to come out. And so, you know, you might think of like looking through a piece of glass.

and how transparent that piece of glass is and how much light reflects off that piece of glass. Well, if that light that's going through is actually your signal, you want close to zero reflection and you want entirely good transmission. Well, there's an exact equivalent of that very consideration inside of a PCB because you can control the speed of light effectively of that signal in the PCB and that's what controls the reflections. And a good material can make it easier basically to accomplish that requirement.

Prateek Joshi (23:32.207)
Let's talk about thermal management. And when it comes to hardware, it's an extremely critical part of the design because you don't want your hardware to just melt under stress. And I've seen it happen. If you go back far enough, I've been alive for long enough to see like melting laptop, which is not a thing. Like maybe people may not see it that much, but it is a thing. So thermal management. How do you predict?

and mitigate heat issues in densely packed PCBs. How often does it happen? Is it that important? And also, how do you plan around this?

Sergiy Nesterenko (24:11.636)
Yeah, yeah. It's of course very important. I mean, it depends on the kind of board you're designing. So if you're, a lot of the boards we see today don't really have these problems. So if you're thinking about like a IOT board or a temperature sensor in your house or a thermometer or something like that, you're purposefully trying to keep it low power so that you can constrain the battery to run as long as possible. And that means that by definition, you're just not generating that much heat. So no big deal.

But if you're talking about something that's powering a motor, okay, that's definitely gonna be a concern. And it's an important concern. So at the PCB level, right, the impact of the actual PCB design, there's a few different considerations that come to mind. The first is that you have to make sure that the copper you've put on the PCB can actually sustain that current without basically overheating too much. So copper is not a perfect conductor. It has some resistivity.

And so what you have to do is basically make sure that there's enough of it in parallel to carry that current. So the simplest consideration there is how wide is your trace, or maybe a trace is not sufficient and you do an entire, what's called a pour, so kind of like a bigger polygon to carry the current. The other choice you have is how tall to make it. So we have these choices about how many ounces of copper a board has, which roughly correlates to the kind of the vertical thickness of the copper.

And that can also help you care, spread some of that current and spread some of the heat load. The other considerations that come up are components that are sensitive to heat. So for example, if you have a voltage regulator, you are gonna have it heat up. Like no matter what you do with your PCB, it's going to get hot. It's gonna reach 30, 40, 50 C, 100 C, depending on what your specs are. And with that being the case, you then have to make sure that the board can sustain that.

Right, so the raw PCB, there's kind of no problem, but you might have neighboring components that are sensitive to that. So maybe you have a crystal that if it gets hot, changes its clock cycle, and that could be really bad. And so you want to keep those things apart. You also might consider putting in a heat sink and good airflow into your mechanical enclosure and fans and stuff like that. And that's a little bit outside of our kind of domain and control.

Prateek Joshi (26:27.663)
Now let's say that you build an amazing PCB, it's very dense, it's high speed, it looks good, everything is great. Now, within all of that, the signals need to actually go from point A to point B without getting lost. So how do you ensure signal integrity in these situations? Like what are all the tests you do? Do you do simulations beforehand to make sure you don't have to redo the work?

So how does this happen?

Sergiy Nesterenko (26:59.316)
Yeah, so there's the difference between how it happens today and how I think it ought to happen. So what happens today, practically speaking, is there's not a whole lot that you end up doing until you've basically finished the design, right? Maybe you can do the sensitive parts first and run a quick simulation on that. But realistically speaking, you might move those things around later on and have to redo that test. And so practically, you end up finishing most of your design.

and then maybe you can run a couple of simulations. The simulations are, you know, for that kind of consideration are basically some version of the Maxwell equations. And it ranges anything from just like a heuristic where you might say, well, this is a high speed trace, let's avoid vias. Let's avoid having this trace come close to noisy sources. Let's make it as short as we can, right? Those are things we can do without a simulation. The next level might be to do kind of like,

quasi -static approximations to the Maxwell equations. So you'll basically compute like the capacitance of the trace or the mutual inductance of that trace with some other trace. And in the quasi -static approximation that gives you some idea of the interference you're going to have. And then there's just different ways of solving the full blown Maxwell equations, whether it's FEM or FTTD or a method of moments or any of these other kind of simulation techniques where you can fully, fully characterize like.

what's happening to your signal as it goes through the trace and what's happening to signals that are near it that might couple. Realistically, the problem with designers is that you have to know which questions to ask. So you have to kind of say, I only have time to run 10 simulations. So I'm going to ask these things that I'm worried about and call it a day, build the board and see what happens.

Prateek Joshi (28:47.599)
Right. And if quilter works, like in an ideal case, everything works and you automated. Now, will that make like the human doing the layout, will they become obsolete? Obviously, there's always room for people, but just like in the extreme version, are you replacing that human in this process?

Sergiy Nesterenko (29:14.516)
Yeah, in the extreme version, I hope so. And again, thinking by analogy to compilers, before compilers existed, people used to compile their own code by hand. And what inspires me is to think, would we have had this software revolution over the last 40 years without the compiler?

And it seems impossible. It seems impossible that we would have had Windows or PyTorch or a web browser or any of these wonderful things if we're still literally like writing to registers in a processor manually. That's ridiculous. So I get inspired by what will be possible when we have the compiler of hardware. And that is very much our goal.

Prateek Joshi (29:57.263)
Right. I think that's a very good analogy. And also, first of all, who wants to sit and compile? It just doesn't make like that. You should human capital, human intellect and energy should be devoted to higher pursuits because compiling is something that the machine can and should do. And it's doing right now. Right. That's a fantastic way to look at it. All right. Now, what innovations in PCB design are you most excited?

about like today, meaning what's happening that's getting you really excited here.

Sergiy Nesterenko (30:30.996)
Yeah, that's a good question. You mean in the process of like building boards or yeah, yeah. Yeah, that's a really good question. So I wish there was more to be honest, I really do. But there are good things happening. So there's a lot of pain in like even just submitting the boards to manufacturers. So like this process of you've got a design done, now you have to create these like actual blueprints that turn into masks that then do lithography.

Prateek Joshi (30:34.575)
Yeah, yes, yeah.

Sergiy Nesterenko (31:01.446)
The manufacturer has to review that manually. That takes a while. Sometimes you find out days later that there was some issue that you need to correct. And that's really, really slow. And I think software and maybe even quilter will help to solve some of those pains. I think on the actual like manufacturing side of things, it's been exciting to see over the last 10 or 15 years, how cheap it has gotten. You know, it's if it takes me a week to design a board, I can have it manufactured in China and shipped to me.

for 20 or $30 and have it come within 36 hours. So like, it's ironic that compiling the board is significantly more expensive and difficult than actually building it. Like it shouldn't be that way, but it's props to the people who are building these things. I think it's also really cool to see that there are starting to appear like basically 3D printers for really, really basic circuit boards so that you don't have to send it out to a whole fab that costs a hundred million dollars to set up. You can just have a printer.

that makes a basic board for you. I think that's really exciting. And personally, I mean, this is not something I've seen, but I'd love to see more three -dimensional boards, right? If you think about it, like being stuck on a plane is not always a good thing for a product. And there's not a lot of innovation, but there's at least one company that I've seen that is starting to play with like, how do we make three -dimensional arbitrary shapes for boards where that might be a better fit for the product?

Prateek Joshi (32:28.911)
So many interesting things happening here. Also, I have one final question before we go to the rapid fire round. And it's about the use of AI. So when it comes to using AI for PCB, start to add like everything, every process that is part of it, let's consider that. What are maybe your one other area where you really wish AI can really kind of

should be doing that. Like, for example, the way you said, humans shouldn't be compiling, it's just a waste of time. So you're doing it. So what does that one other process in the entire supply chain that AI can just solve the problem?

Sergiy Nesterenko (33:19.956)
Yeah, I think the next most obvious to me and something that we're starting to work on because we have to, but I wish we didn't have to, is taking all of this information that is currently locked away in just human readable text and making it systematically available for scripting. So let me make that more concrete. Like when you look at a circuit schematic, unlike code, it is not information complete.

Like you can't look at an arbitrary schematic and nothing else and fully determine all the signals that are happening, all the thermals that are happening. You have to look at what each part is in that schematic. You have to open up the data sheet, read the whole data sheet. And then potentially there's some things that are still not there that are only in the designer's head. And I think that is a big barrier for just any software applications to this space, right? If the information is not there, all the best software in the world can't do anything about it to automate decisions from it.

So right now we're stuck with these just raw big data sheets and PDFs that have all this information and a lack of place where human designers put this information. And I think AI can, especially large language models in the giant boom we just seen, can go a long way to start to convert that kind of human readable representation into something that both humans and software can access very easily.

Prateek Joshi (34:42.639)
Alright, with that we are at the rapid fire round. I will ask a series of questions and would love to hear your answers in 15 seconds or less. You ready? Alright, question number one. What's your favorite book?

Sergiy Nesterenko (34:51.988)
I'll do my best.

Sergiy Nesterenko (34:56.948)
Lord of the Rings.

Prateek Joshi (34:58.511)
Oh, perfect. You're my best friend already. In all these episodes, rarely people say that. So that's fantastic. All right. I love that book, all the books. All right, next. Yeah, yeah. Next question. What has been an important but overlooked AI trend in the last 12 months?

Sergiy Nesterenko (35:11.124)
Yeah, I couldn't put it down.

Sergiy Nesterenko (35:21.844)
Yeah, I think I'm starting to see applications of ML in accelerating physics simulations. And I think that that doesn't get nearly the attention that it deserves. I'll give a shout out to a friend of mine who just announced the company Navier AI that's doing CFD with ML. I would love to see more of that and more attention given to that.

Prateek Joshi (35:42.831)
I agree on this one. Actually, it is a very important field and especially as the 2024 and beyond, a lot of the call it hard tech or deep tech is coming to the forefront and simulation is just part of life when it comes to physical systems. So yeah, I agree that that's a great one. All right, next question. What's the one thing about designing circuit boards that most people don't get?

Sergiy Nesterenko (36:08.692)
Yeah, I think we touched on this a little bit, but I think most people, even designers themselves, don't actually understand how good people are at spatial reasoning. I think we take that skill in ourselves rather for granted.

Prateek Joshi (36:22.319)
That's another good one. All right, next question. What separates great AI products from merely the good ones?

Sergiy Nesterenko (36:30.356)
Yeah, I think great products are those where a real problem existed, where ML happened to be a good solution. Whereas the not so good ones are maybe those where the ML algorithm is just really in search of a problem.

Prateek Joshi (36:43.503)
Right. What have you changed your mind on recently?

Sergiy Nesterenko (36:49.108)
Yeah, the biggest thing, we made a decision early in Quilter to make boards that are purely optimized for the signal propagation properties, but not for looks. And on principle, we thought that mattered most, but a lot of people didn't like it. It just looked way too weird and alien. And so I changed my mind about being principled about that and just saying, okay, we have to accept that these have to look familiar.

Prateek Joshi (37:04.557)
haha

Prateek Joshi (37:13.583)
That's actually very interesting. You would think that aesthetics, the place it would matter the least would be like in a densely packed circuit board. I mean, who cares? Turns out people care. If it looks too weird, they're like, nope, this doesn't look right. Let's redo it. That's the way. Yeah, I don't trust it. All right. Next question. What's your wildest AI prediction for the next 12 months?

Sergiy Nesterenko (37:28.756)
They don't trust it. Yeah, they don't trust it.

Sergiy Nesterenko (37:38.836)
Yeah, I'll take two. So I think the most optimistic prediction I have for the next 12 months is that sample efficient reinforcement learning, like the stuff we're seeing from Sergey Levine's group in Berkeley Robotics Group, is going to have a chat GPT style moment and find a lot more application in the real world because of what it brings. My most pessimistic view might be that we're going to go through a trough of disillusionment with LLMs before we really see the great products invented there.

Prateek Joshi (38:07.887)
All right, final question. What's your number one advice to founders starting out today?

Sergiy Nesterenko (38:15.668)
I think the one thing I underestimated as a founder is that it's going to be a hundred times harder and a hundred times less glorious than you imagine. And so it's really critical to surround yourself with people who will actually be there when it gets like that.

Prateek Joshi (38:31.087)
Right. So founders, if you're listening, if you're okay with that 10 ,000 X markdown that Sergiy just did on your founder live, still do it. So Sergiy, this has been a brilliant episode. Love the depth and clarity of how you talk about this. This is brilliant. So thank you so much for coming onto the show and sharing your insights.

Sergiy Nesterenko (38:53.266)
Thank you Prateek, it was a lot of fun.