← The a16z Show
The a16z Show
Designing the Physical World with AI
00:50:05
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.
Loading summary
Transcript40 lines
- [25:56]
A
A weird problem that you have to solve for that has multiple variables that are all interdependent.
- [26:01]
B
Yeah.
- [26:01]
A
So you go on into a simulation environment and explore as 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, you kind of train an AI how to use that specific tool, how to run a bunch of optimization.
- [26:17]
B
AI like goes and uses some, some all these simulation.
- [26:21]
C
Simulation software. Yeah, yeah, yeah.
- [26:22]
A
And you could do that across simulating how electrons move, how fluids move, how structures move, when the earth shakes. All 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. 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's an extremely calculable answer in almost all scenarios versus something that's a little bit more like we have to predict how it will provide an environment where we actually have no data against it.
- [27:04]
C
So I have two answers. We have a ton of simulation already because our world is made for that. We've had tools in electrical engineering like Spice at the schematics level, and then electromagnetic simulation kernels like openems on the open source or ANSYS on the 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 some reinforcement learning in the loop, you basically can build a circuit and it's very easy to determine whether or not that circuit is correct without relying 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. The thing that I am very hopeful for is that, that electrical engineers today do not, most of the times, rely on pure simulation. Anytime that they build a board, they have a really good internal intuition for why the design is done in a certain way.
- [28:26]
B
They almost use the simulation as a way to verify that as the last step versus as the design.
- [28:32]
C
That's exactly right. So what I. My eventual goal is that Simulation becomes a train time tool that you use for the model to become better at developing that taste, having that intuition. Because at inference time you don't want to. Simulation has some properties. It's not super fast. You can speed it up with 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. 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. The best engineers I know will spend three weeks simulating something 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 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.
- [29:49]
B
I mean, you said earlier, Davide, like 90%, you could 90% do the design and there's still like 10% left. Where you, a human, I mean, you have a team of amazing electrical engineers that basically goes and checks every design and finishes them. Do you think to close that last gap, do we need some sort of a fundamental breakthrough in ML or in AI research to like to understand the physical world to make that leap? Or is it more just further developing or scaling up existing systems? Or do 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.
- [30:32]
C
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 philosophical disagreement on this. 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 on circuit boards specifically. At the very least, there's a lot of things that you can do and it will work. But the data is 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 The Apple's Meta SpaceX of the world. And they will not obviously fork it over, but individually none of this is enough. You really need to pull it. So we either great collaboration, everybody open sources their data, which I think is unlikely, or we find a way to basically produce enough data that the models get 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.
- [31:44]
B
Lenny's take?
- [31:45]
C
Yeah, that's right. Which a lot of these problems are really well structured for Monte Carlo tree search, reinforcement learning style, where you can basically formulate a problem, two players playing against each other and they get better just by nature of improving recursively. I don't know. 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 the things are already on the table and we just need to kind of 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. Let's say at the current pace of improvement in both architecture and data generation capabilities, 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.
- [32:43]
B
Your area seems like, for lack of a better word, almost more permeable in that, 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 design, you know, large scale industrial projects. And Aaron, you're totally wrong. What do you think?
- [33:24]
C
I think.
- [33:24]
B
And does it take a new breakthrough to get there?
- [33:27]
A
No, I think our problem is like we are, it's even more sparse. So it's like hard to, you know, we're not an order of magnitude or 2. Yeah, yeah. So it's. 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 this space is there is an incredible amount of standards that govern how something should be built. So there will always be a optimization benefit from having more, 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 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 the model does get better and if it doesn't, then we're maybe back to the drawing board on aspects or we have something that fundamentally a gap we'll 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.
- [34:45]
B
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 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 American industrial growth. But there would probably be people that take the other side of that argument, I guess. So switching gears to that then maybe let's start with humanoids. That's a big contentious topic Davide, in the context of manufacturing on your end and for you in the context of actually running around on a construction site and moving material. Overhyped, underhyped, does it matter? Curious to get your, your hot takes on on humanoids?
- [35:40]
A
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 definitely be a core component of us. And I think humanoid is like for
- [35:53]
B
the whole like and humanoids in particular versus like specific fit for purpose robots that are like oh I'm really good at moving lumber or I'm like a concrete pore robot.
- [36:01]
A
I mean and they'll, they'll totally be like all the above on stuff. Like there's going to be just broadly more at large. But yeah, there's a reason that centralizing around a design and mass manufacturing that design, ultimately the efficiency that you get out of manufacturing a thing in that learning rate and driving costs down incredibly low outweighs the nuanced custom efficiency. So, yeah, I think that that form factor will be very, very relevant for an incredible scope. And there will be specialized robotics too, in the same way that like. Like when you engineer giant facilities, there's always these kind of volume to surface area scaling laws that say you should customize that giant vessel. And it's not always standardized.
- [36:46]
B
What do you think, Davide? Are you going to have humanoids on your PCB manufacturing line?
- [36:52]
C
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, I think that we like electronics specifically, and circuit boards particularly, are already so automated that your goal is to bridge that gap and so it can be a humanoid. I'm very, very bullish on VLAs. I suspect that there will be a lot of improvements to that remaining 20% of work that needs to be done that you can do with a robotic arm, which already exists in terms of hardware with smarter ability to discriminate components and 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, I think that this is probably a great optimization that you can do. I'll give you a very specific example to our use case. If you want to solder very chunky components, you have two options. You have either you do it by hand, so you take a human and you do it it, or you have a machine called a wave reflow oven, which is basically a huge molten pool of tin, that I went to Europe, 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 not worth it. So we'll just do it by hand. So that kind of thing feels like it's very achievable to automate in robotics. But this is a marginal thing in our assembly line. What I don't think is a marginal thing is if you look at the implication of what automating knowledge work looks like today, at some point, there is a vastly larger chunk of the economy that needs you to do something. If you need to mine ore, somebody needs to mine that ore. And it can be a machine or it can be a human. Hopefully not a human. Not the greatest job that you want. 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 you want to be able to be part of this. Like, you want to be able to facilitate this. Like, one example is actuators. We have some IP in motor controlling. I have a good friend, David Hanson, who builds beautiful motors with Western magnetics materials. 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 100% bet on a specific form factor. 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 part of the bet that we're making.
- [39:51]
B
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 the next generation of this type of worker that historically has been really valuable and is retiring quickly?
- [40:44]
A
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 solved 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, A challenge. And yes, we need so many more of them.
- [41:15]
C
I'm.
- [41:15]
A
I don't know if this is. This could totally be the wrong number, but I, 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.
- [41:27]
B
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.
- [41:39]
A
Oh my gosh.
- [41:40]
B
When they were building a big data center there, which is just wild.
- [41:42]
A
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, to build these projects. And so the only alternative is you, you mass manufacture these things where you concentrate labor in a, in a modular scenario, even at a premium from a cost perspective. So yeah, there is, there is. We 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, modify in a real world of abundance, when intelligence falls to zero, then yeah, it'd be great to embody that into robotics too.
- [42:23]
B
What do you think, Davide? Do we need more PCB manufacturer technicians?
- [42:33]
C
My take on this is the biggest cultural disconnect, which by the way, this is something that when I was like visiting and living in Hong Kong, you see in spades. Is that a lot of the result of being able to just send your designs kind of like ivory tower. You are designing in the US and then sending to manufacturers somewhere else. You're abstracting the manufacturing somewhere else. You kind of don't feel that pain. The pain is really disconnected. And that's why the, the design for manufacturing muscle atrophies. I think that more so than the know how on the like line, it's this idea that like the person that designs the board will be the same that manufactures it or very close, like their friends. For example, if you look at a lot of electrical engineering design done in China, it's designed like, even if it doesn't matter, so that it's easy to make. Like, it's very like, like it's visceral. Like 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 going to do it. And it's like my board is going to arrive later and it's going to be more expensive. I'm just going to 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 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, make it more manufacturable. Or you say, hey, these are like 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. 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 kind of 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.
- [45:15]
B
I think that's a really, really important point. It kind of brings me to my closing question for both of you, which is both of you are talking about how do we use, use AI in the physical world to do more of something, whether it's design and build more PCBs in the US be able to 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? 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, good place to kind of leak, to close the conversation.
- [45:55]
C
For me, I feel this pain personally. I want to be able to spin up a hardware company the same way that my friends spin up B2B SaaS. 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 the equivalent of curing cancer happening in artificial intelligence. It's like this marvelous thing that if you three years ago you gave somebody Claude code, they would have thought that you were, I don't Know, like a sorcerer or something. It is that good. But it's also so bad at actually delivering physical products. And it's marvelous because it can do it in some capacity. But we really need to basically have the same stepwise improvement that we had for software. We need to have the same thing for physical design in order for American engineers, American teenagers to be passionate about building physical things and say, I want to build a cubesat and put it in orbit and do cool things with it. And it's easy because I can just spin up and have my boards manufactured 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.
- [47:22]
A
Yeah, I mean for us it's like, I don't know 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. Like, and I, 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 in it. 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 at the end of the night. And so for us it's like, you know, are we able to do an aspect of the, of, of kind of what this life cycle of building these projects looks like in order of magnitude better so that we earn the right to redo the whole thing and, and solve it from you know, kind of incentives down is how do you build an, 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 the like advanced manufacturing companies that we're doing to re industrialize and just to, to build just an, 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.
- [48:43]
B
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.
- [48:57]
C
Thank you.
- [49:01]
D
Thanks for listening to this episode of the A16Z podcast. If you like this episode, be sure to like, comment, subscribe, leave us a rating, or review and share it with your friends and family. For more episodes, go to YouTube, Apple Podcasts, and Spotify. Follow us on X16Z and subscribe to our substack@a16z.substack.com thanks again for listening and I'll see you in the next episode. As a reminder, the content here is for informational purposes only, should not be taken as legal, business, tax or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any A16Z fund. Please note that A16Z and its affiliates may also maintain investments in the companies discussed in this podcast. For more details, including a link to our investments, please see a16z.com disclosures.
- [49:58]
C
It.