Designing the Physical World with AI

Erin Price-Wright speaks with Alex Modon, cofounder and CEO at Unlimited Industries, and Davide Asnaghi, CEO at Diode Computers, about how AI is moving from software into the physical world. They discuss automating construction and electronics design, using code and simulation to model real-world systems, and how incentives and manufacturing constraints shape adoption. They also examine what it takes to scale infrastructure, reduce build times, and unlock more abundant industrial capacity in the United States.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  I want to be able to spin up a hardware company the same way that my friends spin up B2B SaaS. Like, you should be able to say, "I want to do something that's considered very hard, and just go and do it." We basically built a compiler that gives the model enough hints that it feels like it's writing a Python program instead of designing a circuit board.

• 0:19Speaker 01

  It's basically this combination of a very model-led approach that allows you to use these agents to write code, which is what they know how to do: put on rails.

• 0:28Erin Price-Wright?

  Everything is code.

• 0:29Speaker 00

  The last frontier standing is we don't have enough data. The data is the thing that we need to generate as a society if we want circuit boards to be automated by AI.

• 0:38Speaker 01

  Making sure that you design a system to actually be fully autonomous and to not be human in the loop. I think, for us at least, it feels like it's driven a very different architecture.

• 0:48Erik Torenberg?

  What happens when intelligence gets cheap, but the physical world stays slow? In the 20th century, industrial power came from the ability to design and build at scale. From assembly lines to semiconductor fabs, progress meant compressing time between idea and output. Software accelerated that loop to near zero. But in construction and manufacturing, timelines still stretch into years, shaped by fragmented workflows, fixed incentives, and systems that resist change. Now that's starting to shift. AI can write code, run simulations, and generate designs across thousands of permutations. The question is whether that translates into faster builds or just better plans. I want to understand what it takes

• 1:33Erik Torenberg?

  to actually move atoms, not just bits. A16Z general partner Aaron Price Wright speaks with Alex Modon, co-founder and CEO at Unlimited Industries, and Davide Asnagi, CEO at Diode Computers.

• 1:51Erin Price-Wright?

  We're thrilled to be here today with Davide Asnagi and Alex Modon. Davide is the CEO of Diode Computers, and they're using AI to design and manufacture custom circuit boards faster and better than before and faster than ever possible in the United States. Alex is the CEO of Unlimited Industries, an AI-native firm that vertically integrates design, engineering, procurement, and construction for big infrastructure projects. So we're here today to talk about physical world AI. And when I say that, I think a lot of people probably think about things like humanoids and robotics foundation models. But while I think robotic... Housekeepers folding your laundry is still a few years away, or maybe if you're really optimistic, a few months.

• 2:33Erin Price-Wright?

  AI is already starting to cross this chasm with use cases that move atoms. So these companies are working on physical world AI at two very different scales, from the micro to the macro. And I'm excited to get your perspectives about where we are and what's ahead. So Alex Stavide, welcome to the show.

• 2:51Speaker 01

  Yeah, excited to be here.

• 2:52Speaker 00

  Thank you.

• 2:53Erin Price-Wright?

  Maybe, Alex, to kick off, you've said that in 10 years, all construction will be fully automated, which feels like a pretty bold claim, very ambitious. But what does that actually mean? What does it take to get there?

• 3:05Speaker 01

  Yeah, I think it's probably helpful to level set on what a construction project looks like. And it starts with a developer who's got an empty lot of land and they want to make some sort of big project there. And then there is, depending on how big the project is, if you're going to build a power plant or a hospital or some large facility, you're going to spend almost a year, sometimes a year and a half, just doing design for that. And there's hundreds of engineers that touch this. There's lots of different project managers that touch this. And it's this orchestration of mechanical and process engineers and electrical engineers and civil and structural folks, all kind of working together to do the pre-construction package, which is effectively like a giant set of instructions that you can then hand to a general contractor or some builder who will order the things on there and actually construct the facility.

• 3:47Speaker 01

  That first part, that's like line of sight today of how we automate end to end. Typically, we call that final output an IFC package and issued for construction package, where you will literally feed in a site, a bunch of different requirements about what you're trying to build and anything you want to stipulate about how it gets built. And AI is going to explore tens of thousands of different permutations about how optimally done designed that facility, a button click. And then what you get back from that is a globally optimized IFC package and issued for construction.

• 4:16Erin Price-Wright?

  Optimized for what?

• 4:18Speaker 01

  It depends what the optimization function is. And so the easiest way to think about that might be like CapEx, like how much does this thing cost? But a much, much better way to think about that is like a total cost of ownership of the project. So if you really care about operation and maintenance of the facility, how constructible is the facility itself? And I think a lot of things we kind of generally see in industry today is it's there's so many different segments and slices of people that all optimize for very different things. So being able to kind of approach designing these giant things the same way you do from a software perspective, which is a very like a parametric, ultimately a super, super flexible approach you can take to optimize on any sort of main metric for that. So I guess that's, I guess, first half of it is how do you automate that end to end, which is what we're working on now and have a product to do. The second part does look like a bunch of robotics.

• 5:00Speaker 01

  And it's everything from like autonomous earth movers, which feels more tractable in the short term, to how is a site full of tons of humanoids and drones and all sorts of like autonomous robotics. So yeah, we very much think that that future is destined to happen over the next like a decade. And a big piece of that is if the incentives are properly aligned, which I'm happy to talk more about later.

• 5:18Erin Price-Wright?

  Yeah, I'm excited to talk about that. Davide, what is the kind of equivalent timeline for automation and hardware? And in particular, maybe electronics, because that's where you focus on both design and manufacturing. So maybe paint us a picture on what this looks like for you.

• 5:33Speaker 00

  So I have to be careful because my timelines are getting shorter and shorter. And I think I need to stay on the reasonable side. I will say it's really interesting to hear Alex describe AI applied to construction, because I can draw very immediate parallels to hardware. And for this discussion, I would like to stick to circuit board specifically, which is both the design and the assembly of a circuit card. I am reasonably confident, like we do a lot of work with Anthropic, for example. And the jump that we see in design capabilities between each model like tier, like publicly available model is wild. We thought it would be five years. I think that I can probably say two.

• 6:16Speaker 00

  There, I think that the caveat here is that there are very, very different types of electronics design. And like making a blanket statement about all of them, I think is not appropriate. But I do think that there are, there is a subset that I really care about. And I'm gonna go into details as to why, that I think will be fully automated in two years, like in terms of design. The other reason why it's interesting, it's like, for us, it's not just design, it's also manufacturing. And Alex said, okay, like we have this very like high optimization function on the effectively what the plan looks like. And then there will be a bunch of robotics on the manufacturing. Like, we already have the robotics. Like, the electronic industry has had robots for like years. The biggest problem is that there is an 80/20 robotic automation versus like manual labor.

• 7:01Speaker 00

  And so right now, like, nobody has really bridged that gap in the United States. So what I'm, like, what we are working on at Diet.

• 7:08Erin Price-Wright?

  What do you mean by that? Can you double click on that 80/20?

• 7:11Speaker 00

  So normally, like, there's a process called surface mount technology. So this means that there is a robot that will basically place every single component on top of a circuit board. And then you bake it in an oven and you're done. But the 20% that is like very, very hard to automate is there are some components that will not feel that like fit that very nice mold. Maybe it's like a very big transformer that needs to be soldered with a different process. Maybe it's something that hugs the PCB. Even the assembly of the like a circuit board itself into an enclosure, that's usually not fully automated as well. So there's been companies like Foxconn, for example, for Apple or like Pagatron that have solved that problem with labor. And that makes complete sense for like certain like segments and in certain geographies.

• 7:55Speaker 00

  But you're not gonna like double the production capacity for data centers in the US by just like relying on labor alone. And also, you're not gonna be able to reduce the time cycle that it takes to bring up a data center from four years to two years. If you're not able to redesign the boards, redesign them for manufacturing, manufacture them at scale in a constrained type line. So what we really are bullish on is not really AI completely automating away like design work. It's more AI being able to automate away the type of design that produces these very manufacturable outputs. And then we have a lot of moonshot ideas about how you can improve the robotics on the line. But the core goal is if the design is like constrained, you can manufacture that 100% automation today.

• 8:40Speaker 00

  Like, I don't even need to wait for robotics to get better. Like, the robots are already here.

• 8:43Erin Price-Wright?

  It is funny to think about like the design of a data center with all the complex sort of optimizations you need to do and the specs and everything you need to fit into a particular site footprint actually being kind of similar, surprisingly similar probably to the design of a very small circuit board that needs to fit lots of different components and meet lots of different specs. I don't think I actually fully appreciated that till right now. And Davide, you've said your goal is to transform every software engineer into an electrical engineer. Why?

• 9:15Erin Price-Wright?

  And how?

• 9:16Speaker 00

  I want to requalify that.

• 9:19Erin Price-Wright?

  Okay.

• 9:20Speaker 00

  Like, I think that software has incredibly good properties. And I think that empirically, we are immediately observing this with tools like Claude Code. Like, basically, agents have been able to leverage like the structured nature of code to do things that are not code-like at all. Like, you can basically use Claude Code to do things that are completely orthogonal. So the initial comment that I made was basically, yes, like, we need to take the current, like, set of software engineers that we have already in the United States and allow them to do more, like, being able to be also electrical engineer. But now, like, the total set of software engineers or people that are able to produce software is exploding. And it includes, like, agents.

• 10:03Speaker 00

  So really, what you want to do is you want to give anything that has the ability to generate code the same ability to generate hardware. That's really what you want to do. And this will include, like, greatly skilled S W Es that are building beautiful cathedrals on a PCB, like a lot of our employees, which I'm very fond of. But it will also include, like, a lot of, like, smaller designs that are completely automated away by allowing an agent to use code to build the board itself. The complexity of those designs today is limited, but you can extrapolate the derivative and, like, get to where we want to be in, like, a year or two.

• 10:37Erin Price-Wright?

  I think this gets to another parallel in both of the industries that you're working in, which is that you're working in these industries with, like, a very entrenched sort of sets of expertise, ways of working, tools that people are used to using. And I'm curious to hear, I mean, which is very different than, like, software engineering, for example, where even before AI, I think software engineers are just naturally technology curious and used to adopting completely new frameworks for doing their work, like, every couple of years. So can you talk a little bit about that? Like, how do you, how do you essentially change an entire industry's way of working?

• 11:17Erin Price-Wright?

  How much do you have to vertically integrate and own yourself versus how can you bring people along for the journey?

• 11:25Speaker 01

  I mean, on our side, there is...

• 11:28Speaker 01

  There is a lot of like, you have to earn it. And, and, and super traditional industry. This is a lot, again, with the incentives, but it's, um.

• 11:36Erin Price-Wright?

  Yeah, I mean, it's hard to think of a more traditional industry than construction.

• 11:40Speaker 01

  Yeah, yeah, yeah, yeah. It feels, yeah, even manufacturing feels like incredibly like pro-technology and like light-years ahead of where the construction industry really feels. And a lot of that is, again, it's, it's, the incentive structure is like helpful to understand because really the, the source of capital sets the incentives all the way down through the project. And like, um, we certainly didn't used to always do it this way, but how it happens now is a, basically a project that you want to go build goes through this pretty like stage gated process of removing risk so that you can fund it. And then eventually you get this kind of project that feels fundable to an investor. And this is a very different profile than venture, right? This is people who like, all they want is a.

• 12:17Erin Price-Wright?

  Part of finance.

• 12:17Speaker 01

  Yeah, it's a stable IRR.

• 12:18Erin Price-Wright?

  They want a, a set rate of return.

• 12:20Speaker 01

  Yes, I like, I have no ability to take that going to zero and I don't even care about capturing the upside. So all those kind of incentives force the adoption of new technology down. Like when you get to someone who might have some new piece of technology, it's just super unincentivized because there's no upside. Like no one actually wins from that environment. And it looks actually really similar to defense, you know, in a pre-Andrill Day, if you will. Like it is a, someone's got budget, I'm going to pay you to.

• 12:44Erin Price-Wright?

  Like four companies sued the government.

• 12:45Speaker 01

  Yeah, yeah, well, I mean, yeah. And that's, that's the process of saying, hey, this is like, it's, it's just, it's just wrong. It's just like broken. And, um, and that change just hasn't been able to kind of permeate through to this industry. And then, and we're kind of left with like, you'd walk into one of these firms and you'd look at people's computers and you'd feel like you're stuck in like the late 90s. You're like, what is going on here? So for us, you know, vertical integration is like, you know, we have to own enough of it that we can actually do a clean interface to the industry rather than, um, you know, trying to pick off one small part and then force people to, to kind of change and adopt that technology. So yeah, it is, it is a process of us kind of owning a big enough portion where we can reasonably do it as a startup and, and then earn the right to continue to, to chunk more and more of that from a vertical integration perspective.

• 13:28Erin Price-Wright?

  So do you have to go out and hire and look for the like? Forward leaning, I mean, for example, like you, you, you need a, a registered civil engineer, I presume, to sign off on a drawing before it goes to construction. Like, how, how do you think about constructing the team, no offense intended, of people who are like actually, you know, willing to think differently, given the right incentive structure?

• 13:52Speaker 01

  Yeah, I mean, um, most of our team is actually not software engineers. Like, we're mostly multidisciplinary, mechanical, electrical, civil,

• 14:00Speaker 01

  lots of simulation controls. We definitely have people who have like AI and software background, but it, in our experience, has been much easier to teach a multidisciplinary person the latest and greatest AI tools, you know, than, than, uh, the other way around. So, uh, yeah, the, the team construction thing is, is, uh, as long as you can kind of think about AI as a core primitive to how we build it into the product, it's much easier to take the domain expert to it. And the good thing is that like, you know, um, there are electrical engineers who are like incredibly bright and who love sitting at the cutting edge of a product. And our team doesn't, you know, we have some folks that come from large capital projects in oil and gas and energy world, but the majority of people are like, they work on, you know, cutting edge vehicles, rockets, um, that type of background.

• 14:43Erin Price-Wright?

  Davide, like, maybe similarly, similarly in some ways, I don't know if the same sort of capital incentive structure is aligned, but like, electrical engineers are sort of famous for, no, no offense, because you yourself are an electrical engineer, for, um, what's the right word, like being fixed in their way?

• 15:01Speaker 00

  Opinionated and tasteful. That's, I think, what you're looking for, but.

• 15:04Erin Price-Wright?

  Tasteful, yes.

• 15:07Erin Price-Wright?

  Very Italian, very Italian of them. How do you, how do you get people to change the way they work?

• 15:13Speaker 00

  So, this is something that we had to think about, like, very hard at the, like, beginning of DIAID. Like, when the idea was in its, like, very infancy, it was more about, there is a pain point that I am feeling myself when I, like, go through the process of designing hardware. And there is a new technology that very clearly holds the keys to solving a lot of this pain point. How like if I had to like give this technology to the world, how would I give it? And so one somewhat obvious like answer that you could have thought about in like 2020 would have been, let me build like, you know, CAD software. And that that's fine. And like CAD software has its own like idiosyncrasies as an industry.

• 15:57Speaker 00

  It's a very high like switch cost. There, there's a lot of like reasons why building a like CAD software business is complicated. Our goal was if you truly make the leap of faith that the models will do it themselves eventually. What you want to build is you want to build the infrastructure for the models to be able to like go and do these things, which are not currently able to do. And you don't want to convince people to buy your software. You want to convince people to buy the end product. And this is a much harder company to build. It does require a much larger width of like the market. And so today, the, the way that we like explain ourselves to our enterprise customers, like Fortune 100 companies, is not, hey, we're gonna sell you like a software product.

• 16:42Speaker 00

  It's more you are used to like effectively working with companies that deliver the exact same product, which is I give you my specifications, I get back a physical product. We fit into that mold. And then how we do it is an implementation detail to the company. But we do it faster and we do it cheaper, which is what matters at the end of the day. Also, we have some major advantages, like we offer an inspection into how the process happens. You see in real time, like your board being like designed and delivered. Something that like engineers are very used to, like software engineers, like you can think about it as like pull request on GitHub, like you see the code like streaming in real time. This was not really the case.

• 17:25Erin Price-Wright?

  The checks passing.

• 17:26Speaker 00

  Yeah, exactly. That's exactly right. And so, so basically, this is how we pitched it, to the point where like our core compiler tool chain is open source. Like we actually, if any electrical engineer wants to take it and run with it, please be my guest. Because what we did is like, once you generate an artifact, it will work in our ecosystem. We will be able to take it and send it for manufacturing. This is like, we want to own the infrastructure, not the like core design primitives. And we are very bullish that those design primitives will be very helpful to both humans and like agents. So thi- this is the bet that we made. It's a, the infrastructure is open source. You can go look at it on GitHub, diodin slash PCB. We have a lot of people that actually are trying it in like a closed beta,

• 18:11Speaker 00

  where they give us feedback and they use our internal tools that we haven't released yet. And we will release more and more tools. But the way that we are harnessing that power is by doing everything end to end. And like the, you know, Alex put it beautifully, like you need to do enough. If you try to do too much at like too early, you dilute yourself and your focus becomes really complicated. The other benefit of like this strategy is a lot of the times the AI is not a hundred percent of the way there. And like, you want to be able to place company like defining bets, even if the product itself may not quite be a hundred percent there. And so like, for example, we are 90% more efficient at building boards than we were in a world without our tools. But if we were to give these tools, like there's still 10% of work required,

• 18:56Speaker 00

  which is why people like pay our current services right now. And like the eventual goal is as the models get better, that percentage goes down and down and down, until it's like fully self-service and you can reap the rewards and basically manufacture right away.

• 19:09Erin Price-Wright?

  Yeah, I mean, I think you both touched on this sort of like time, money. It's like sell the thing that the customer knows how to buy.

• 19:16Speaker 01

  Yeah, totally.

• 19:17Erin Price-Wright?

  I mean, a project finance, like truly time is money. Like if, if it's a, a project that gets one year, it takes one year to build versus a project that takes five years to build at the same cost is a very different IRR.

• 19:28Speaker 01

  Totally. I mean, the value prop for us is not that it's cheaper from an engineering perspective. Engineering is such a small percentage of the cost anyways, that no one wants to compress that. It is that if you can bring in a schedule by three months or six months, that actually materially impacts.

• 19:40Erin Price-Wright?

  The finance ability, yeah. So maybe moving to, you know, building physical world AI, how, how have you guys? Worked to get AI to understand kind of your physics and real, physical and real world constraints that that matter like there's no common crawl for you know, construction design data or maybe there's a little bit more data availability on the on the board side, but probably not much that's that's digitized and that's that's easy to access. So how do you get enough data to even work with to to to train the models that you guys are using?

• 20:16Speaker 00

  Maybe I'll start because I have the easiest path like in general, by the way, my company is much easier than like Alex. I admire Alex and I respect him a lot because I think that his domain is like a very beautiful and complicated domain. I do simple things. I build electronics. So we cheat like instead of requiring a ton of data, we reframe the program with something that there is a ton of training data on, which is code like traditional electronics is not designed in code at all. And in fact, very tasteful electrical engineers despise the idea of writing code. And that's fine. Like, we don't want anybody who doesn't want to write code to write it. Let the model write the code. The model is amazing at writing

• 21:00Speaker 00

  code. It understands the concepts. We basically built a compiler that gives the model enough hints that it feels like it's writing a Python program instead of designing a circuit boards. So this is how we got around the schematics problem. There's a lot of like actual physical, like, you know, the, you know, physics problem.

• 21:20Erin Price-Wright?

  Yeah, like you're, you're literally putting something on a, on a square that has to fit into some sort of form factor. So you do have physical constraints beyond just the code.

• 21:29Speaker 00

  That's exactly right. And there, there is no data. Like, you're right that there is much more data than in construction, but like the data is like two orders or three orders of magnitude smaller than what you would need to train a foundation model to actually do the the proper job. But because this is a physics based problem, that you you see companies in the space trying to tackling the like, you know, the actual hey, how do I represent it like physically? And we actually see like agent models being able to Slowly and kind of painfully, like, starting to be able to do it the same way that a human would, for example. We also have like some moonshot projects in the works that maybe like we'll chat about it next time. Like, I'm basically, I will just say I'm very bullish on diffusion as an architecture.

• 22:13Speaker 00

  That's what I will say, like, for this specific problem. And the, the, the thing is that you need to bootstrap it somehow. And for us, the bootstrapping comes from code first, build a library of validated blocks. Those blocks become training data for the next round of models. And so you, you basically build a compounding interest. Also, if you become the rails that everybody can design for free on, that is data that comes your way. Like, and like, as long as you give enough incentives and say, hey, I'm going to manufacture in the US at a cost competitive with Asia, which is unheard of effectively, and you can do it profitably, which is very important. You will be able to actually like generate the amount of data that makes the model just like skyrocket in accuracy.

• 22:57Speaker 00

  So this is like where we want to eventually go.

• 23:00Erin Price-Wright?

  How about you guys?

• 23:01Speaker 01

  A super, a super cool idea. Yeah, same philosophy as like, everything is code.

• 23:04Erin Price-Wright?

  Yeah.

• 23:05Speaker 01

  And if you can play in a language that the models already understand, your life is like a bajillion times easier. And so for us, it's like, the industry really works much more maybe deliverables based or output focused. We're always thinking about it. It's like, everything's a snowflake. It's like always N of one. Which is the core fundamental problem of why it takes so long and we don't get better. Like, we're not improving our rate of building these projects. And so our world looks much more like model led. It's like, how do you actually embody all the relationships into some really robust model? That.

• 23:38Erin Price-Wright?

  When you say model, what do you mean by model? Like an AI model or more of like an ontological model?

• 23:43Speaker 01

  Ontological model, like a, a, a repossess, a, a simple representation of a complex system.

• 23:49Erin Price-Wright?

  Sorry, my, my, my, my palate.

• 23:50Speaker 01

  Yeah, I did a kick in for a minute. And you, you basically have a combination of like, provide an, an environment that agents and.

• 24:03Speaker 01

  LLMs can generate a bunch of code that fits within this framework that we've provided it, and it has the ability to use all these deterministic tools, almost the same way a regular engineer would. And if you can do that in this like model-led way, you basically have this like parametric relationship to everything. And what's really important is that, well, every time you tweak and adjust, it just gets better and it improves its resilience. So, yeah, probably like the biggest way that that shows up for our customers is that, you know, if you spend six months designing something, and say you have another like three months before you kind of lock everything and you fund the project and you build it, if you want to change something six months in.

• 24:42Erin Price-Wright?

  Oh, God.

• 24:43Speaker 01

  Yeah, you'd like start over.

• 24:44Erin Price-Wright?

  Cascade.

• 24:45Speaker 01

  Yeah, and it's a total nightmare. In this version, it's like, it's, everything's just a, it's just like an updated variable. So it's highly both optimized, but also you can iterate on these things. So, yeah, it's basically this combination of a very model-led approach that allows you to use these agents to write code, which is what they know how to do. Put on rails.

• 25:11Erin Price-Wright?

  Everything is code.

• 25:13Speaker 01

  Everything is.

• 25:14Erin Price-Wright?

  Everything is computer.

• 25:15Speaker 01

  Yeah.

• 25:16Erin Price-Wright?

  How much do you guys use, it sounds like both of you are doing some of this, but use simulation, either for data generation or other purposes right now? Is that something that you're, you see using more of in the future, using less of, or maybe even other techniques other than simulation?

• 25:37Speaker 01

  We certainly use simulation to like compute basically the values that are required for the design, but we don't have.

• 25:44Erin Price-Wright?

  Give an example.

• 25:45Speaker 01

  Like, I want to understand what a mass flow rate is of a fluid that's going through a pipe with a certain consistency at a certain temperature and pressure. And that's like a weird problem that you have to solve for that has multiple variables that are all interdependent.

• 26:00Erin Price-Wright?

  Yeah.

• 26:01Speaker 01

  So you go into a simulation environment and explore is a pretty wide space. And then ideally, you're goal-seeking, you're optimizing towards something. So that's just a calculator. We've done that simulation software for like decades. And those are tools for us. It's like you kind of train an AI how to use that specific tool, how to run a bunch of optimizations.

• 26:16Erin Price-Wright?

  So your AI like goes and uses some optimization.

• 26:20Speaker 01

  Yeah, yeah, yeah. And you could do that across simulating how electrons move, how fluids move, how structures move when the earth shakes, all of that through simulation. And the good thing about at least our world, the easier thing that we have in our world is when you think about designing any of these big projects, it is just, I mean, it's just Legos on hard mode. Like you are empirically saying, the only way that you validly fit these things together is if it's either been done before or it perfectly matches up from an inputs and outputs perspective. So it is, it's an extremely like calculable answer in almost all scenarios versus something that's a little bit more like, we have to predict how it will provide in an environment where we actually have no data against it.

• 27:03Speaker 00

  So I have two answers. Like, we have a ton of simulation already because our world is made for that. Like, we've had tools in like electrical engineering, like Spice at the schematics level, and then like, you know, electromagnetic simulation kernels, like Open EMS on the open source, or like ANSYS on the like actual board level. Both of these things exist and are used in the industry and are very important. I think that the way that we currently use simulation is to give a grounding to the model if there is like some reinforcement learning in the loop. Like, you basically can build a circuit, and it's very easy to determine whether or not that circuit is correct without relying

• 27:48Speaker 00

  on a golden reference, like, I blessed this specific configuration. You can build it in many different ways. As long as you achieve your output, you're good. So simulation in that case, very important. We have it. We will continue to use it.

• 28:03Speaker 00

  The thing that I am very hopeful for is that

• 28:10Speaker 00

  electrical engineers today do not, most of the times, rely on pure simulation anytime that they build a board. Like, they have a really good internal intuition for why the design is done in a certain way.

• 28:25Erin Price-Wright?

  They almost use the simulation as a way to sort of verify that as the last step versus as the design.

• 28:31Speaker 00

  That's exactly right. So what I, like, my eventual goal is that simulation becomes a train time tool that you use for the model to become better at developing that taste.

• 28:43Erin Price-Wright?

  Having that intuition.

• 28:44Speaker 00

  Because at inference time, you don't want to, like, the simulation has some properties, like it's not super fast. Like, you can speed it up with, like, parallel kernels, and you can do a lot of things. But simulation is fundamentally something that we hope to use at training time rather than at inference time. And we built all the hooks to do it because it's important to ground physicality. Like, you basically need to provide ground truth somehow. But we are really seeing emerging properties where if you train the model with enough data, it will develop that taste. And in 95% of the cases, that's what you want. You get the product very fast. And it's very, very hard to beat actual manufacturer hardware. Like, the best engineers I know will spend like three weeks simulating something,

• 29:29Speaker 00

  and then they will be like, I don't trust it until I actually built it and put it in an enclosure where we'll change completely my electromagnetic properties, for example. So it's just like, simulation needs to be a training tool, and then you kind of need to get physics to tell you you're right or you're wrong. So reducing the training time is probably the most important part.

• 29:49Erin Price-Wright?

  I mean, you said earlier, Davide, like, 90%, you could 90% do the design, and there's still like 10% left where you have a human. I mean, you have a team of amazing electrical engineers that basically goes and checks every design and finishes them. Like, do you think to close that last gap, is there, do we need some sort of a fundamental breakthrough in ML or in AI research to like?

• 30:16Erin Price-Wright?

  To understand the physical world to make that leap? Or is it more just further developing or scaling up existing systems? Or you think that's the wrong question to even ask? Like automating that last, whether it's 10 or 5 or 1%, doesn't really actually matter that much?

• 30:32Speaker 00

  So my co-founder, his name is Lenny, he is the smarter one. So I will say this is a very important premise. Him and I have a philosophical disagreement on this. Like, I think that basically we already have all the building blocks. My personal take is that the largest, like the last frontier standing is we don't have enough data, like on like circuit boards specifically, at the very least. Like, there's a lot of things that you can do and it will work, but the data is like the thing that we need to generate as a society if we want circuit boards to be automated by AI. The data exists, it's usually siloed into like, you know,

• 31:14Speaker 00

  the Apples, Meta, SpaceX of the world, and they will not like obviously fork it over. But individually, none of this is enough. Like, you really need to pull it. So we either like, you know, great collaboration, everybody open sources their data, which I think is unlikely. Or we find a way to basically like produce enough data that the models get like inherently better. And if a new architecture comes out, amazing, you can be more efficient and you can have less data. That's my current take. But there is another competing take, which is Lenny's side of the argument.

• 31:43Erin Price-Wright?

  Lenny's take, yeah.

• 31:45Speaker 00

  That's right. Which, a lot of these problems are really well structured for like Monte Carlo tree search reinforcement learning style. Like the, you know, where you can basically like formulate a problem where like two players like playing against each other and they get better just by nature of like improving like recursively. I don't know. Like, I usually defer to his opinion. So I will say that maybe there is a world where we don't need any more data and like the things are already on the table and we just need to kind of like tickle them the right way. In the meantime, because this is an open question, we will focus on building the thing end to end and we will bridge the 10%, however we need to. But I'm very hopeful. It's like, at the current pace of improvement

• 32:30Speaker 00

  in both architecture and data like generation capabilities, like breakthroughs will happen and you need to be able to harness them. And you want to be in a position where you benefit from them and you're not on their path.

• 32:42Erin Price-Wright?

  Your area seems like, for lack of a better word, almost more permeable in that it's less of a controlled system. So there's a lot more variables, like the wind, or I don't know, whatever. There's stuff that sort of blows in and might affect you that is hard to be represented by a model. So from like outside in, my read would be, there will probably have to be this more of a human in the loop forever. But maybe you disagree with me and you're like, we are going to have fully end to end designed, you know, large scale industrial projects. And Erin, you're totally wrong. What do you think?

• 33:23Speaker 01

  I think you can-

• 33:24Erin Price-Wright?

  And does it take a new breakthrough to get there?

• 33:25Speaker 01

  Yeah. No, I think our problem is like, it's even more sparse. So it's like hard to, you know, we're not in order magnitude or two. Yeah. So we don't even really have that as an option on the table. I do think that most all those problems can be bounded and that the kind of benefit of the space is there is an incredible amount of standards that govern how something should be built. So there will always be an optimization benefit from having more data points that you can feed in or more nuanced information. But to kind of beat status quo is just the bar is so unbelievably low. So yeah, I would

• 34:07Speaker 01

  say take the under on it. And I think we will be at end to end. I think it's actually an important design paradigm is like for us specifically is like making sure that you design the system to actually be fully autonomous and to not be human in the loop. I think for us at least, it feels like it's driven a very different architecture. Now, we similarly have put the bet that like the model does get better. And if it doesn't, then, you know, we're maybe back to the drawing board on aspects or like we have something that fundamentally a gap will have to close. But yeah, the system itself needs to be kind of designed with the requirement that says this is designed to be end to end automated.

• 34:44Erin Price-Wright?

  Maybe switching gears to talk a little bit more. I mean, I think- Physical world automation, physical AI, whether it's robots running around or something else, feels very like from a sort of society level perspective. I think, you know, people are scared of it. People are thinking about it. It's very much in the zeitgeist. I think it's a truly exciting boon to, you know, American industrial growth, but there would probably be people that take the other side of that argument. I guess like, so switching gears to that then, like, maybe let's start with humanoids. That's a big contentious topic. You know, Davide, in the context of manufacturing on your end

• 35:27Erin Price-Wright?

  and for you in the context of like actually running around and construct on a construction site and moving material, you know, overhyped, underhyped, does it matter? Curious to get your, your hot takes on, on humanoids.

• 35:40Speaker 01

  Well, I'm like super excited for the future to like, feel like the future. Yeah, I think that that's like incredibly inspiring. Like, I want to live in that world. It'll, it'll definitely be a core component of us. And I think humanoid is like for the whole like.

• 35:53Erin Price-Wright?

  And humanoids in particular versus like specific for purpose, robots that are like, oh, I'm really good at moving lumber or I'm like a concrete pourer robot.

• 36:01Speaker 01

  And they'll, they'll totally be like all the above on stuff. Like, there's going to be just broadly more automation at large. But yeah, there's, there's a reason that like, centralizing around a design and mass manufacturing that design, ultimately, you know, the efficiency that you get out of manufacturing a thing and that learning rate and driving costs down incredibly low outweighs the nuanced custom efficiency. And so, yeah, I think that that form factor will be very, very relevant for like an incredible scope. And there will be specialized robotics too. In the same way that like, when you engineer giant facilities, there's always these like kind of volume to surface sales, surface areas, scaling laws that say you should customize that giant vessel and it's not, you know, always standardized. But yeah.

• 36:45Erin Price-Wright?

  What do you think, Davide? Are you going to have humanoids on your, on your PCV manufacturing line?

• 36:51Speaker 00

  So I have the luxury of loving all robots equally. I don't care if they're humanoid. I don't care if they're special. They all have PCBs. I love them. They have PCBs inside them. I love you. I think that for our specific process.

• 37:08Speaker 00

  I think that we, like electronics specifically, and like circuit boards particularly, are already so automated that your goal is to bridge that gap. And so like it can be a humanoid. I'm very, very bullish on VLA's. Like I suspect that there will be a lot of improvements to that like remaining 20% of work that needs to be done, that you can do with, you know, a robotic arm, which already exists in terms of hardware, with like smarter ability to discriminate components and like have the ability to do computer vision on the fly. I do think that there will be a little bit of that. And in fact, like I think that this is probably like a great optimization that you can do. Like I'll give you like a very specific example to our use case.

• 37:52Speaker 00

  There like if you want to solder very chunky components, you have two options. You have either like you do it by hand. So you take a human and you do it. Or you have a machine called like a wave reflow oven, which is basically a huge molten pool of tin. That like I went like to Europe. Like I was visiting some manufacturers and they told me it's literally so expensive to heat up the tin in terms of energy that we don't turn it on because the volume that we do is like not worth it. So we'll just like do it by hand. So that kind of thing feels like it's very achievable to automated robotics. But this is like a marginal thing in our assembly line. What I don't think is a marginal thing is like, if you look at the implication of what like

• 38:37Speaker 00

  automating knowledge work looks like today, like at some point, there will like there is a vastly larger like chunk of the economy that needs like you to do something. Like if you need to mine ore, like somebody needs to mine that ore. And it can be a machine or it can be a human. Hopefully not a human, like not the greatest job that like you want to like, you don't want to put humans in harm's way. Like you want to be able to actually have the robots that do this kind of stuff. And so what we want to do is like, you want to be able to be a part of this. Like you want to be able to facilitate this. Like one example is actuators. Like we have some IP in like motor controlling. I have a good friend, David Hansen,

• 39:21Speaker 00

  who builds beautiful motors with like Western magnetics materials.

• 39:27Speaker 00

  And we want to build them. We want to build this kind of stuff. We want to be part of this. And I don't know if I would like 100% bet on a specific form factor. Like, we invest and we really like robots that are specific, like pick and place machines are basically robots. But I think that there will be more and more and more. And so this is like part of the bet that we're making.

• 39:51Erin Price-Wright?

  You kind of touched on something that I think is important, which is this sort of tacit knowledge that I think Dan Wang calls it the process knowledge, which China has in spades, which, you know, the US has to some degree, but it's we have an aging workforce and some of these kind of skilled labor, whether it's manufacturing engineers or highly skilled, you know, construction engineers or civil engineers or electrical engineers. These people that, you know, work in these physical fields and have developed an intuition and a taste and understanding of what works and what doesn't. How important is that to capture? Is there a way to start encoding that in these models? Or should we be really thinking, you know, as a society about how to train up

• 40:36Erin Price-Wright?

  the next generation of this type of worker that historically has been really valuable and is retiring quickly?

• 40:44Speaker 01

  I do think there is a lot of, there's a lot of tacit knowledge in the industry that helps be more like a shortcut or a rule of thumb to the right answer, that you can just first principles derive. So there is a scenario where you, you know, you could do a heck of a lot more work. When your marginal cost goes to zero, it's totally fine. We solve the problem. But like where it really is, is in like the trades, like the electrician and how they work. And there is an incredible amount of tacit knowledge there. Which, yeah, I think is both a, yeah, like a challenge. And yes, we need so many more of them. I don't know if this is, this could totally be the wrong number, but I would guess that the average. Salary of an electrician in Texas right now is higher than like a Silicon Valley software engineer. Like it's incredible, like super demanded.

• 41:26Erin Price-Wright?

  Yeah, well, I was talking to the, I think this was last year, I was talking to the CTO at Microsoft and he was telling me that one time, Microsoft employed a third of the electricians in the state of Georgia.

• 41:39Speaker 01

  Oh, my goodness.

• 41:39Erin Price-Wright?

  When they were building a big data center there, which is just wild.

• 41:42Speaker 01

  People are turning to manufacturing for a lot of the data center scope, just because there's not enough trades to, there's not enough people in the trades to build these projects. And so the only alternative is you, you mass manufacture these things where you can concentrate labor in a, in a modular scenario, even at a premium from a cost perspective. So, yeah, there is, there is, again, it's a little bit of an all of the above strategy where you would need to say, yeah, we should totally be training more people on very practical skills that, that are going to be needed for a while. And, and then hopefully start to codify a lot of that. So that when we really want to go into scale mode in a, in a real world of abundance, when like intelligence falls to zero, then yeah, it'd be great to embody that into robotics, too.

• 42:22Erin Price-Wright?

  What do you think, Davide? Do we need more PCB manufacturer technicians?

• 42:32Speaker 00

  My take on this is the biggest cultural disconnect, which by the way, like this is something that when I was like visiting and like living in Hong Kong, you see in spades, is that a lot of the, our, the result of being able to just like send your designs, like kind of like ivory tower, like you are designing in the US and then sending to manufacturers somewhere else. You're obstructing the manufacturing somewhere else. You kind of don't feel that pain. Like the, the pain is really like disconnected. And that's why like the design for manufacturing muscle kind of atrophies. I think that more so than the know-how on the line,

• 43:17Speaker 00

  it's this idea that the person that designs the board will be the same that manufactures it, or very close, like they're friends, for example. If you look at a lot of electrical engineering design done in China, it's designed, even if it doesn't matter, so that it's easy to make. It's visceral. I had this friend who would... I was like, "Why do you make your boards so cramped? Why just do it on double side?" He looks at me and he's like, "But then you have to do two passes on the SMT line." And I'm like, "But it's not you." And he's like, "No, but I know the person who's gonna do it, and my board is gonna arrive later, and it's gonna be more expensive. I'm just gonna spend a little bit more time designing it." And this is very cultural. And I think that what is missing, at least in the circuit board,

• 44:02Speaker 00

  is this very visceral connection. And I don't know that you can just artificially manufacture it or hire your way into it or pull people out of retirement. I think that the only way is through at this point. And so you basically need to find a very cheap way to generate these DFM-ready designs. And Claude doesn't care if you bash it and say, "Yes, this is good, but make it more manufacturable." Or you say, "Hey, these are 150 checks. Go through all of them and painstakingly change the artifact until it's easy for me to make it." And then, of course, there are very smart people already. All the contract manufacturers for PCBs in the US are very talented.

• 44:47Speaker 00

  Actually, the capabilities of contract manufacturers in the US are super high because they only bid on military contracts, which require the highest possible capabilities. But what you want to build is this new set of mass production capabilities, which has been evaporated by the industry because the economics didn't make sense. And we are betting on doing it by teaching the designer, which in this case is not a human necessarily, to do it for you.

• 45:14Erin Price-Wright?

  I think that's a really, really important point. It brings me to my closing question for both of you, which is... Both of you are talking about how do we use AI in the physical world to like do more of something? Whether it's build more, design and build more PCBs in the US, be able to, you know, pay for and construct and design and construct more kind of large scale industrial projects. How do you guys think about the second order effects of that? Like, maybe this gets to the mission statement of why you guys are working on what you're working on. But maybe that's a good place to kind of leak to close the conversation.

• 45:55Speaker 00

  For me, like, I feel this pain personally. Like, I want to be able to spin up a hardware company the same way that my friends spin up B2B SaaS. Like, you should be able to say, I want to do something that's considered very hard and just go and do it. And I think that the second order effects are, today, we have what is effectively like the equivalent of curing cancer, like happening in artificial intelligence. It's like this marvelous thing that if you, three years ago, you gave somebody like Claude code, they would have like thought that you were, I don't know, like a sorcerer or something. Like, it is that good, but it's also so bad at like actually delivering physical products. And it's like marvelous because it can do it in some capacity.

• 46:40Speaker 00

  But we really need to basically have the same like stepwise improvement that we had for software. We need to have the same thing for physical design in order for American engineers, American like teenagers to be passionate about building physical things and say, I want to build a CubeSat and like put it in orbit and do cool things with it. And it's easy because I can just spin up and like have my boards manufacture next day. I think that those are the second order effects that I am interested in. Like, and the only way that we get that is if we teach models to actually do like things in the real world, which is my, like, I enjoy this a lot. Like, this is why we started the company.

• 47:21Speaker 01

  Yeah, I mean, for us, it's like. I don't know, if you if you look at like basically any in the US at least, any construction metric, so like labor productivity or adjusted capex numbers over the past like 50 years, we're getting worse. And I, you know, my past life, I come from the world of software, where it's just like de facto, everything gets better. There's always progress. And that is like clearly not true in this space. And you extrapolate that line out. And yeah, we just like lose how to lose the muscle of knowing how to build large ambitious projects. And yeah, that's like the graph I see when I like close my eyes the other night. And so for us, it's like, you know, are we able to do an aspect of the of kind of what this lifecycle of building these projects looks like in order of magnitude better so that we're in the right to redo the whole thing?

• 48:05Speaker 01

  And and solve it from, you know, kind of incentives down is how do you build and just an absolute like order of magnitude or orders of magnitude more from a project's perspective? And that's everything from like the energy that we need to actually like win in AI and build all these data centers to all of the like advanced manufacturing companies that we're doing to reindustrialize and just to build just a massive amount of stuff that we need, all the way through critical minerals. So yeah, there is just like the core bones of how basically all this stuff that you see when you look around works, like we're getting worse at. And that's like a very, very concerning thing.

• 48:43Erin Price-Wright?

  Cool. Well, this is really fun. I'm glad both of you are working on your respective problems. I'm glad you're in our portfolio because it's, it's, you know, I'm leaving this conversation optimistic. But yeah, thanks so much.

• 48:57Speaker 00

  Thank you.

• 49:01Erik Torenberg?

  Thanks for listening to this episode of the A16Z Podcast. If you liked this episode, be sure to like, comment, subscribe, leave us a rating or review, and share it with your friends and family. For more episodes, go to YouTube, Apple Podcasts, and Spotify. Follow us on X at A16Z and subscribe to our Substack at A16Z.substack.com. Thanks again for listening, and I'll see you in the next episode.

• 49:26Erik Torenberg?

  As a reminder, the content here is for informational purposes only, should not be taken as legal, business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any A16Z fund. Please note that A16Z and its affiliates may also maintain investments in the companies discussed in this podcast. For more details, including a link to our investments, please see A16Z.com/disclosures.