为什么物理 AI 是下一个前沿｜a16z 对话 Applied Intuition

Our mission is to put intelligence on a billion machines. We think that can have a profound impact on society. Applied Intuition is a physical AI company. We put intelligence on machines: cars, trucks, tanks, drones—it’s a physical moving thing. We make it intelligent.

Digital AI, of course, is building software, optimizing ads, and creating videos. That’s all interesting and good, but when we talk about the global economy, that’s physical AI. In this intelligence revolution, the companies that impact the physical world might actually be bigger than the companies that impact the digital world.

How many things are there where the idea of physical AI—physical intelligence—is going to matter?

There’s no reason autonomy should be this obscure, difficult technology. Our vision is that a high school kid who can make iPhone apps should be able to make autonomous systems. That platform for designing and developing is what we’re launching. It’s called Dana. Everything that we’ve built and developed over the past nearly a decade is available in Dana.

Which will we get first: a perfectly simulated real-world environment for training autonomous devices or Grand Theft Auto 6?

Qasar, Peter, welcome to the a16z podcast.

Well, thanks for having us. Your name is?

Yeah, which is one of many.

I think we’ve all known each other for too long.

Long time.

We’re lucky to both be the first investor, or among the first investors, in the first round. Of course, different check sizes, but—

And I was an investor in you even before then.

Exactly. So, let’s do that as a segue. We have a lot to talk about today. We have the biggest launch in company history to talk about today. But first, why don’t we just give an update or a status? What does Applied Intuition do for those who don’t know?

For the people who don’t know, Applied Intuition is a physical AI company. We put intelligence on machines. That’s the simple way of describing it, and all types of machines: cars, trucks, tanks, drones—you name it. It’s a physical moving thing. We make it intelligent.

The history of the company is that we originally started by making the tools that would make the intel, and then we got into the actual intelligence itself. In some ways, we’re a very boring AI company, in the sense that 83% of the company is engineering. We win by making really great products. It’s not like we’re a sales-led company or something like that. I don’t think we’re good enough for a sales-enabled company.

We have over 1,000 engineers, based in Silicon Valley, but we have offices globally—18 offices. Our mission is to put intelligence on a billion machines, and we think that can have a profound impact on society, both in the obvious things everyone talks about, like safety, and in productivity. If you really talk to somebody who’s been in a car accident, a mining accident, or a farming accident, those are real gnarly situations.

Beyond just fixing that, if you can unlock productivity, I think we’ve seen the unlock in the digital world. Everyone’s super excited about it, and you have trillion-dollar companies emerging. I’m a pretty strong believer that when we look back 25 years from now at the internet, the original internet companies that are serving and doing analytics will be interesting, but the big monolithic companies are Amazon, which delivers you stuff, and Apple. These are the true companies that came of age.

I think when we look back 25 years from now, in this intelligence revolution, the companies that impact the physical world might actually be bigger than the companies that impact the digital world.

I would love for you to talk about the following. When you first started the company, the knock on the company was, “Oh, it’s making cars autonomous, right? Self-driving cars.” But there’s Tesla and Waymo building their own self-driving cars, and then there are six or eight other car companies that matter. The company just could never get that big because there just aren’t that many customers.

Even if you take that view of us, automotive is like 30% of our business. 70% is already non-automotive. If you fast-forward another 10 or 20 years, even the manufacturers themselves, as a customer base, will be a small amount.

Our mission is to keep thinking about a billion machines becoming intelligent. Think about all the types of machines that exist. Automotive is just an easy one. It sticks in people’s heads because we all drive cars, and it’s a big market, but I think it’ll be a minority of the business. That doesn’t necessarily mean it’ll be small. Automotive is still huge. As a part of the globe’s GDP, automotive is something like 3% of all GDP.

The way we always think about it is that, as you try to get to your mission, initially the manufacturers were the distribution mechanism for that intelligence to consumers. But then you start working in defense, construction, mining, and agriculture, and suddenly the manufacturers are important, but maybe the mining operators are actually really important, or the Department of War is really important. Suddenly they become customers, and all of those are customers of ours as well.

I think if you split AI into digital AI and physical AI, digital AI, of course, is building software, optimizing ads, and creating videos, that sort of thing. That’s all interesting and good, but when you talk about the global economy, that’s physical AI. Then we’re talking about manufacturing, mining, logistics, and transportation—all of these things that supply chains touch.

Supply chains, exactly.

Let me build on that for a second. Things that move today, or historically, are things that have human beings at the wheel or at the controls in some form. Airplanes have had to get designed around a human in the cockpit. Boats have had to get designed around a human steering things. In a world of autonomy, do we already know what the things are that move, or are we going to discover that there are a lot of new things that are going to get built when you don’t need a human in the driver’s seat?

I think both. The thing that you have to remember is that you take a haulage system that’s in a port, or a Caterpillar or Komatsu dirt mover in a mine. Those are made for 20 or 25 years.

The buyers of those products might not have gotten their full-cycle ROI on them, so they’re not immediately going to buy something new, no matter how much better it is. One part of our strategy is that you have to make those things intelligent because they’re not going anywhere.

The second is what you’re talking about: that depends on a human in a cab. If you don’t have a human in a cab, the machine can be smaller. It can be shaped in very different ways. Think about mining underground. The constraint is actually the human, because the human needs to breathe and it’s very dangerous. You can build a very, very different machine. We’re doing both of those things.

And then the thing that you were not talking about is that we’re all talking about intelligence almost as if it exists within a system, but the system-level intelligence is where the unlock is. We’re already doing work like that where you say, “Hey, let’s take an entire port. Let’s take an entire mine. Let’s take an entire quarry.” This heterogeneous mix of machines can all talk to each other. They can optimize and be efficient. When one machine goes down or one machine has an issue, the rest of the mine doesn’t have to stop.

When it’s human-driven, we don’t even know the machine is going to go down because there’s no announcement. The human is not plugged into the core systems of the machine.

Right.

A simple thing like knowing when a brake system is going to break is actually huge, because you can start preparing for it in advance. You’re like, “Oh, this wear and tear is higher than in other mines.” Here’s an example.

The other macro point is that if you look at agriculture, the average American farmer is 58 years old. Less than 10% of farmers are under 35. So what’s going to happen? The need for food production continues to grow. The need for rare-earth materials is growing, so these demands are only increasing, but the humans who are the bottleneck are decreasing.

Trucking is the same way. You can really unlock a lot more efficiency. One way to think about this is to imagine if the cost of food decreases because it’s way, way more efficient. What’s the downstream impact? Then imagine the same thing for goods being transported.

Let's say, instead of a few dollars a mile, it's 20 cents a mile. Suddenly, I think the unlock is very, very big, and I think that doesn't necessarily mean all the machines need to be redesigned from the ground up.

Right. Right. Got it. Makes sense. And then maybe just one more question: Give us a sense of the scope and scale of the company today.

Yeah, north of 1,000 engineers. Those engineers are, obviously, the classic software and AI engineering teams, but we also have engineers who really know safety systems. We also have engineers who really know hardware.

The important thing that we're kind of just tiptoeing around is that all this stuff is hard because it ultimately has to meet the real world. The real world has way more complexity and a lot more issues. We have engineering teams that can do that. We've deployed our models onto 50-some platforms.

Even that sounds trivial because, mostly, when you think about models, you think about deploying them through a browser or on a phone, and everything's abstracted away because you have iOS, Android, Windows, Linux, and all these systems that have already taken care of it. In the real world, you don't have that.

We also have engineering teams that can do that. Our claim to fame is that we've raised over $1 billion in the company's history. All that is sitting in the bank, and I always say that with an asterisk because it doesn't mean we're not going to spend it next month.

Good news, bad news.

Yeah, good news, bad news. But I think we're at that phase where these giant markets are around us, and we can make the decision: How aggressively do we want to pursue those? That's after a decade of, frankly, execution and deployment into production.

I think the hallmark of our engineering team is putting products into production. That really is big. I don't know—how do you think about scale?

Yeah, I think that's roughly it. The mission of bringing intelligence to 1 billion machines—that is how we think about it. Then thinking about the types of machines that will have the most impact and focusing on those areas first. But we'll get there.

Let's go deeper into the differences between digital and physical AI, and more specifically, where are we today? What progress has been made? What are some of the major bottlenecks in physical AI? Can you unpack some of that?

Yeah, I think a lot of times people think the progress in physical AI is limited to 2 use cases because they're obvious and interesting: robotaxis and humanoids. They're very visceral, they excite you, and they're kind of sci-fi. Those are very interesting, and there is real work being done by us and other people in those domains.

I think all the other domains are going to be just as important. Think about what happens at a port. There's a huge unlock there, and that's the area we're really focused on—all the other nooks and crannies.

If you look at the rise of Cisco and how networking went from individual machines, to companies getting networked, to entire countries getting networked, there's a similar thing happening with AI. AI is getting to that level. Sovereign AI is now a discussion.

Sovereign AI really is about physical AI, because that's where you're talking about AI in defense and AI in the physical machines that are moving around. If you look at the example of Waymo from America and Pony from China, they're trying to deploy in, let's say, the other country—not America, not Europe, not China. In every one of those spaces, they're much more hesitant to say, “Yeah, thumbs up, your robotaxis can run unfettered in our country.”

If you look back at this arc of the internet, when the first internet companies came, nobody was really thinking about sovereignty at all. The browser went everywhere, the internet went everywhere—that was almost the power of it. Then, when social media emerged, there was a bit more of, “Hey, actually, not every social media platform can go everywhere.” China didn't allow Facebook to come in.

Then you get into the next level of the online-offline stuff. There's more resistance to Uber and DoorDash. Suddenly, there are local players who are being favored very aggressively.

When we get to physical AI, I think there's going to be a huge geopolitical theme of more fracturing than globalization. You're going to have a demand for this AI to somehow be localized. I think that has to play into our strategy as well. We're a technology provider, so we can provide that technology across the globe. I think that's something that's understated in this conversation.

Yeah, a few other things on digital versus physical AI. In digital AI, the state of the art is that you can train models effectively on the entirety of the internet and then maybe augment that with additional data that's been collected and refined with some hired experts. This is sort of a hot field right now, but generally, you're talking about a foundation model that's built on internet data.

In physical AI, internet data is useful, too. However, to actually build a foundation model in physical AI, there's also a lot of private data collection. When we're talking about mines, logistics, or any of these other fields, the data that's useful for training models there isn't necessarily available. We have to do a lot of work ourselves, actually going out and collecting that data.

The other key factor is safety. If you're talking about building a smartphone app, you don't necessarily care whether it's a safety-critical application. But when you're talking about moving a machine that weighs many tons—or think of a humanoid that could fall over on your children—you care a lot about safety, the evaluation of that safety, and really getting to the state of the art of physical AI by proving out the safety case around some of these state-of-the-art models.

Yeah, and I think humanoid data collection has been its own little area of interest. But when you talk about collecting data in places like South Korea, where you have North Korea, they don't allow mapping companies, let alone allowing an American company to come in and collect data.

Over the years, whether it's in the Middle East or Latin America, we've figured out how to get into these countries, work with the governments, and get the thumbs-up to collect proprietary data. In that way, it's similar to other digital AI systems: Your proprietary data sets, scaling laws, and all that stuff are the same. It's just applied in a very, very different way.

The way to think about it is that the diffusion of these models is very different because not everyone can just access them through a phone. That, ironically, plays in our favor. Once we have a massive proprietary data set—which we've been building—we already have hundreds of petabytes of data.

Then we have our own tools, like synthetic data tools and NeuralSim. We can use our own tools and our own proprietary data, and that allows us to build some of the best systems in the business.

Qasar, is there a chicken-and-egg thing where, in order to build an autonomous physical thing, you need a lot of data, but to gather that data, you need a lot of physical autonomous things running around collecting it? Once you have a giant network of physical things running around, you have the data that makes them all work. Is there a flywheel aspect to that? What's the level of difficulty involved in booting up that flywheel?

It's difficult, but it's also not difficult. We have one of the largest data-collection fleets on the planet, frankly speaking. That's how you bootstrap your way into it. That's just money, resources, and technical knowledge.

But it's not like that knowledge is extremely obscure. There are probably more than 5 companies that have it. I think what's more difficult is how you actually have a model that's going to work on lots of different hardware and is tested appropriately.

You saw it with Cruise. Cruise was a company that did amazing self-driving work, and then there was 1 accident. General Motors owns them, got super scared, and pulled back. Just getting these things into production is actually more difficult than it seems.

I think we believed synthetic data was going to be important, so we started our synthetic data team more than 5 years ago.

More than that, yeah. More than that at this point.

We were strong believers that synthetic data can accelerate autonomy development. We’ve just seen that. And then there are lots of other secondary and tertiary technical innovations. Obviously, the transformer revolution hitting self-driving was massive. Basically, everything done in self-driving pre-2021–2022 is relevant, but you’re almost like, “That’s kind of the starting point.”

But it’s also different from today being the starting point. Those 4 or 5 years are actually—there has been a lot of work done. You can see it most clearly with Tesla, but there are other folks. In that process, the actual techniques historically—and let me just simplify here—imitation learning was the name of the game: you collected a bunch of data, and then the models would basically imitate what human drivers do.

The real state of the art right now is end-to-end reinforcement learning in a closed loop, in your tools. So it’s a little simplified to say the system learns itself. It identifies where the issues in the self-driving system are, and essentially you then find data like that or synthetically create data like that. Then you close that loop and see, “Are you performing the same scenarios better and better?”

I think if you fast-forward some years, that will be a completely closed loop with no humans intervening. Right now, you still have the fog error that we saw. We still see errors in the real world that impact self-driving.

Oh, yeah. So it’s like, what are the bottlenecks? And the bottlenecks—there are plenty of them—but whenever you’re dealing with physical systems, you inevitably hit a lot of gnarly hardware problems. It could be anything from overheating to a sensor being slightly miscalibrated. A funny issue I saw yesterday was basically a fogging sensor, fog impacting a sensor. These are the things that you actually have to solve for this stuff to work very reliably in the real world.

Yeah. So I’m going to ask you a three-quarter question, and we can decide whether you guys want to engage on it or not. It might be an opportunity, or it might hit a question, which is: were you surprised? Cruise was a super high-flying Silicon Valley autonomy startup that was kind of running neck-and-neck with Tesla early on and so forth, with a very top-end team. Then they famously got bought by General Motors, and they—

One of my first distributions personally, so I had enjoyed it.

There we go. Y Combinator company. And, you know, top-end team, and they were, by all accounts, making excellent progress. They got bought by General Motors. They became the GM autonomy program. GM got a lot of praise, at least in tech circles, for being the legacy automaker with the biggest investment—

I called Peter before it was announced, and I said, “Hey, Cruise just got bought.” You know, he’s also GM family. We’re both GM families. Peter guessed it. He said, “Nvidia?” I said, “No.” He said, “Apple?” I said, “No.” I said, “General Motors.”

So that’s surprising to people who are from GM.

That they were willing to buy the thing.

Yeah, that they did it.

Okay, that they did it. And then, by all accounts, they were—I mean, as far as I ever heard, they were making excellent progress.

Yeah.

And then they had this—there was an accident. Was there an injury or fatality?

It wasn’t a fatality, but it was a serious injury. Somebody was dragged 20 feet.

Yeah, serious injury, bad press, and then the GM CEO and board put a bullet in the Cruise project. I know at least some of the senior Cruise people were extremely upset by the aftermath of that. Was it surprising that they reacted the way that they did?

Full disclosure: General Motors is a customer, and I went to the General Motors Institute, so we have a lot of love for the company. Coincidentally, I should say, I’m reading this very famous book, which I had actually never read before, called On a Clear Day You Can See General Motors. And DeLorean’s book. Have you read it?

Have you read that book?

I have, years ago. It’s one of the great all-time book titles, and we should just pause to say John DeLorean was the super genius of the car industry.

Yeah, he was going to be the next president of General Motors.

And then later on he started his own car company, which was in Back to the Future, and that whole thing collapsed for a variety of reasons. He was a legend. He was one of the principal drivers of innovation in the car industry.

Exactly. Lee Iacocca bought lots of these categories.

Yeah. You’ve got to remember this is in this era.

Sorry, repeat the title of the book.

On a Clear Day You Can See General Motors.

Why was that the title of the book?

It’s a very large complex. It’s like a nation-state. Really, these companies are extensions of the state. Hyundai is an extension of the state. Kia is an extension of the state. Volkswagen—Volkswagen board members are members of the government. So these are extensions of the state in almost every way.

There used to be an old saying: “What’s good for General Motors is good for America.”

Right.

You cannot understate how important General Motors is to the history of the American corporation. Sloan’s My Years with General Motors and Adventures of a White-Collar Man—if you run a large engineering organization, you should read those. The modern corporation that we talk about didn’t just emerge. Sloan and Kettering—with Kettering as head of engineering—created this architecture, with levels and vice presidents, and how you do functional and matrix organizations. It really is the source code.

Then comes John DeLorean, and he writes that he’s going to be president and is so fed up with the company. But what was controversial was that GM was doing really well at the time. GM was number 1 in the Fortune 100. When we say GM was number 1 in the Fortune 100, it was number 1, 2, and 3. It was everything, and it was seen as the best company in America. So somebody openly criticizing the company was controversial.

He writes this book as he quits, out of how annoyed he was at General Motors and how it was being led. After he sobers up, he’s like, “I don’t want that book published.” He fights for years for his co-author not to publish the book, but the co-author still publishes it. So it’s a real insight into a large corporation.

Incidentally, I’m reading it now, even though I worked at GM 20 years ago and knew a lot about the company. What’s shocking is that it’s not only about GM. Most of the major manufacturers still operate that way on the inside. The point for everyone to take away isn’t that the people who run these companies are stupid. They’re not stupid. It’s kind of like when you’re selling to the Department of War and people say, “Why are you doing that?” It’s like, “Well, the distribution defines the business.”

This might be out of date, but when I worked in safety systems 20 years ago, I remember GM used to pound into your head that, of the top 5 consumer lawsuits in American history, 3 are automotive. We got the majority, right? So you have to be extremely careful.

We had these weird things inside the company. It wasn’t red, yellow, green; it was purple. You’d always have those decoders, because when they go to court, they’re like, “You let a safety system that was marked red go to production.” And I was like, “No, it was marked magenta.” Can you imagine how infuriating that is? Every time you’re like, “What does orange mean? Does this mean I have to—?”

Well, Ford’s slogan for a very long time was “Quality Is Job One,” right?

Yeah, which fits with safety as well.

Yeah, exactly.

And that’s the one-two punch of automotive: quality and safety. Quality and safety. Quality really became important because the Japanese really reset that stage. That’s a whole separate automotive history. We could talk about automotive history for an hour, but the punchline is, you have the Silicon Valley company meeting this immovable object.

There is a parallel universe where Cruise is out there right now, even as part of General Motors. I think you always have to take it into the context of where the company is and where union negotiations are happening literally that year. If you’re the union, you’re like, “You can't make a million dollars for us, but you're funding this thing.”

That’s killing people, and it’s sloppy. I’m not saying precisely that’s what happened, to be very clear, but it’s a multivariate problem. My other hot take is: I worked at both companies, right? Google and General Motors. Those companies are way more similar than they are different—way more similar than they are. Literally, people don’t need to know this, but the Google leveling system is the same as General Motors’ [laughter] leveling system.

I used to say this inside Google meetings. It was like, “Hey, actually, some of the engineers I knew at General Motors are better than the engineers here.” And people would look at me like I was saying there’s no God in church. They were like, “How dare you, you metal-butt-bending monkey from Detroit?” [laughter] I was like, “No, actually, making a modern combustion engine is extremely complex. It’s not just—it’s not simple stuff.”

The macro point, I think, is that it’s a bunch of things. I think safety is always at the top of their list. We’ve hired lots of Cruise people. I think the way they dealt with that specific issue with the government—you have to dance a particular way when that happens—and they just didn’t dance exactly right. That gives government bureaucrats more ammunition to go after you.

You’re a big target. Like General Motors, you have to—it reminded me, did you guys ever see that movie, Goodfellas? One of the last scenes, “House of the Rising Sun,” you know, all the old bosses go in the back of the courtroom and they’re like—and that’s what happened. The board was like, “What do we do about Cruise?” They’re like, “What can we do?” [laughter] It’s like, “Kyle’s a good guy, but—” [laughter] And then it’s like, cue “House of the Rising Sun.” People are running through San Francisco—just kidding. [laughter] Don’t make that an AI video. It’s going to get a mean text from Kyle. So, I think there is a universe where it would survive, but it’s tough.

So, then a lot of what Applied Intuition does is, as you said, that dance. It’s how to be a great partner to these companies.

Exactly.

Bearing in mind their own very real issues and constraints.

I think GM also had the issue of business model, right? Cruise was going after the robotaxi concept, but GM makes its profits from personal car ownership. Those things can be a bit odd. So, I think that was also a bit of the—

Oh, right.

Yeah, and I think it wasn’t clear. By the way, you, of all people, spoke at YC in 2013. I was in the audience; I was a partner at the time. You said something which I think is very recursive here. We’re feeding each other your own advice.

The key thing in the new-technology business is that everyone figures out the technology. Though, that’s still hard; it’s still hard sometimes to build really complex things. It’s when and how you deploy them into the market. The when becomes really important. You’re 2 years early, and you’re doomed. You’re 2 years late, and there are too many competitors. You have to hit it at the right spot.

A controversial thing to say is that I actually think Cruise was certainly moving at a much faster pace than Waymo. They started way behind, and you’re talking about neck and neck when, ultimately, the plug was pulled. So, who knows what happens in the long term?

Our hypothesis in that same equation is actually the distribution: you let the manufacturers do that. We run self-driving trucks right now in Japan. They carry commercial loads. There are safety drivers there, but they’re autonomously running. You won’t know that because the brand is Isuzu; that’s the customer.

Why it’s so good for us to partner with Isuzu in that case is that the company’s been around almost 100 years, right? If I’m not mistaken, it’s a pre–World War II company. They know the government. They have test tracks. They know safety. They know their own trucks very well. So, when we provide them with the intelligence and the integration into their physical machinery, that’s a fantastic one-two punch.

I think today the world is ready to consume AI in the real world, and that’s largely because of ChatGPT, Anthropic, and everything that’s happened. People are no longer like, “What’s a self-driving car?” This is because of Waymo and Tesla. The market is ready to consume, and I think you just have to meet the market in the best way possible.

In our view, that has always been: you go through some of the people who run the economy right now. Whether it’s a mining operator, whether it’s the Department of War, whether it’s the manufacturers, we work within each vertical with the right partner. That’s a fundamentally different view than a Tesla or a Waymo, which are going to be vertical, whereas we’re really playing the horizontal.

The way we think about our company is that we’re kind of like a chip maker. We actually look, talk, and walk a lot like a silicon company, except we obviously don’t make chips. But we have design wins and really large, long-term relationships. Once we’re in, we’re in. It’s really hard to take us out. So, you need deep trust. Our partners have a lot of deep trust in us, and we know their markets really well.

I think one of the things Jensen knows is his customers. That’s why NVIDIA does well, beyond the fact that they obviously make very complex technology.

So, how are these legacy car companies preparing for the future? Are they making more acquisitions or going to build and partner with you? How are they going to compete with tech-native companies?

It’s like saying, “How are governments dealing with AI?” It’s such a broad topic, and each manufacturer—even if you take Honda, Nissan, and Toyota, 3 Japanese manufacturers with long legacies—they all approach it very differently. They’re roughly in a spectrum from “We’re going to build” to “We’re going to buy.” More than ever, “We’re going to buy” is the common answer because they’ve been trying.

For the folks that are going to build, we provide them tools, and we talk a little bit about our new product that we’re announcing here. For the ones that just want to buy, we sell them the actual intelligence that goes on the machines. So, we meet the customer wherever they’re ready in their journey.

The more nuanced version of that is that every product is a different product. The amount of silicon and dollars you can put toward it, toward sensors, and what the customer’s willing to pay all depend on what actually gets in. In the long horizon, all these things will be fully autonomous, but the intermittent steps are very much what we saw in the PC.

You have a slow step up to one day when nobody really looks at laptop specs, and maybe, frankly, phone specs. But that wasn’t the case from basically 1985 to 2005, when people finally stopped speccing at all and were really moving to laptops. There’s a similar kind of 20-year horizon there.

And broadly, when you talk about machines becoming intelligent, fundamentally, a machine is a collection of these different components that are integrated, right? Whoever does that final integration is often the company that puts its badge on it—the brand name—but many, many companies are building technology that goes into those machines. So, we now have a bunch of technology components and platforms that can go into these machines. We also sell the core technology that can be used to develop them as well.

If you look, by the way, under the hood of a dirt mover, combine, or diesel truck, they’ll have Cummins engines in them. But nobody says, “Well, because all these guys buy Cummins, this means that they’re”—whatever. Caterpillar is a good company. It’s like, no, that’s just a component that they buy. They have a different role.

When you look in any of these verticals, it’s just a complex web of folks. That’s why I always say the chip analogy actually works quite effectively: none of those companies make chips, but they all buy chips. I think that’s a good way to think about it.

So, self-driving cars—we’ve all been talking about self-driving cars for, I think, the whole thing started around 2012, 2005, or something, with the DARPA Grand Challenge originally. Then Google engaged on the program shortly after that.

Yeah, late 2000s, yeah.

Late 2000s. So, almost 20—basically around 20, less than 20 years, maybe. There have been lots of predictions over the last 20 years that self-driving cars are imminent, at any moment. I guess the bad news is we’re sitting here today and most cars are not self-driving. The good news is there are now self-driving cars.

Yeah.

The Waymo cars are driving all over the places where they’re deployed.

I think people in San Francisco now treat it as routine that they get into a car.

I think you can call Tesla. It's kind of like the AGI thing: if we're talking about 20 years ago, everything we're seeing right now is mind-blowingly AGI. The goalpost keeps moving. The Tesla stuff is amazing.

You can look at a bunch of manufacturers. BlueCruise, Super Cruise, BMW, and Volvo's Pilot Assist are all quite impressive systems. They're not full self-driving.

Right.

But, yeah.

Well, it's full self-driving—except for whatever remote—

[snorts]

Monitoring is happening.

The Tesla—we have a home in Los Angeles, and if you guys may recall, there was a large fire in Los Angeles last year.

Yeah, yeah, yeah.

And the California power grid was buckling even before that. It turns out that among the things Cybertrucks are good at is being very good batteries for powering your house.

Yeah.

So, literally, we have a Cybertruck as our backup battery for the house. As of last year—whatever the FSD release was, I forget the exact one—there was one that, at least, a lot of people thought really turned the corner.

FSD 14, yeah.

And that thing drives people. I talked to somebody yesterday who has a Model Y and let the thing do the full route all the way up Highway 1 through Big Sur.

Yeah, I think the mean time and the miles per disengagement are really high. I think the miles are in the thousands, which is very impressive.

Yeah, I know. For people who have driven Highway 1 through Big Sur, that's a stressful drive. He said it was great the whole way. Anyway, I wouldn't have been talking to him had it not been—

[laughter]

Would have gone right off—

Because he unbolted the steering wheel, so—

Right off, right off the—

Cliff.

And then Tesla's rolling out its robotaxi. It's starting to show up in the wild. So, on the one hand, those exist; on the other hand, 99.999999% of cars are still not self-driving.

Maybe one other thing would be the self-driving trucks. There's been this recurring panic in the press that if trucks become self-driving, all these truck drivers will be out of a job. Sitting here today, I don't know: are there any trucks on the road that are self-driving and don't have at least a safety driver in the truck? I think the answer is probably still—

Yeah. So, let's split the multiple points that were brought up here. One is personally owned vehicles and why they're not more ubiquitous. Part of that is the manufacturers are not good at deploying technology. Part of that is they want to be safety-conscious, but most of it is cost, cost, cost.

What you're seeing in China is a different EV ecosystem, mainly because they don't care about profits. When you're talking about a business that doesn't care about profits, it changes the entire calculus of the industry. What you're seeing is L2++ systems.

We can simplify the entire self-driving conversation to: Is there a driver behind the steering wheel? The driver behind the steering wheel is still there, but generally, Tesla drives everywhere. They're sub-$1,000. There's an aggressive cost curve—that's the chip, sensors, the package, the software, everything. We anticipate a very aggressive decline. Once you get to around $500, the automotive OEMs will actually subsidize it for free.

This happened with navigation systems. If you remember navigation systems, it used to be a big thing where you'd pay $4,000 or $3,500 to get one, and then suddenly it became free and just became the default. There's a weird thing where getting into a subset of your cars costs X dollars, and getting into all the cars costs X plus a small incremental amount.

There's a fixed cost, and then you have the number of vehicles, the assembly line, homologation, all these testing regimes—all this stuff. So, I think you'll have a wait, wait, wait, and then a lot.

Yeah.

Every single OEM, without exception—even the lowest-dollar OEMs—is working on an FSD competitor. So, it'll come.

A good analogy for thinking about self-driving in the personally owned ecosystem is mobile phones. We had satellite phones, then we had the Qualcomm brick phones, then we had the Motorola Razrs. From the late 1990s to the late 2000s, there was a huge question of, "When's mobile going to come?"

And then it comes. By 2007, from the iPhone launch, in about 4 years you get Uber, Instagram, WhatsApp, and Snapchat. Those are the killer applications. I think there's a very, very similar wait, wait, wait, and then it's basically ubiquitous in every vehicle.

If you had to ask me what that number is: 2028 SOP, 2029 start of production, 2029–2030, and then by the early 2030s it'll start becoming very cheap to free.

I think routinely, by the early 2030s, you would just buy a car and assume it's self-driving.

Exactly. Or it has the driver-in-seat L2++ system, to be very specific.

Like Cybertruck or Tesla—the Tesla equivalent of what Tesla owners have today.

Today, exactly. Yeah, it'll be the default.

The question then—the other side of this—is why don't we have a bunch of Waymos everywhere? Specifically, Waymo has a different technology. Without getting into the nuances here, Tesla, many of the Chinese companies, and Applied Intuition were very much in this end-to-end model architecture. This is a new way of doing self-driving.

Waymo, for lack of a better word, is not that. That doesn't mean they're not learned; it just isn't one end-to-end system. It's not one monolithic model. One of the proclivities of that approach is that it depends on HD maps. Therefore, there's a geofencing concept.

I think Waymo's trying hard to remove that bottleneck so it can expand geographically faster, but the reality of today isn't there. The other thing is, when you have researchers—which Waymo really was, coming out of an Alphabet research organization—they didn't put commercial constraints on it.

The sensors are bespoke and expensive. The cars and the compute in them are just not economically feasible. They've tried a lot to get that down, but it's a lot easier to go from something that's really cheap and make it more featureful than to take something that's overbuilt and try to trim it down and make it really, really cheap.

And that's the big debate: who's going to get there first—Tesla with full self-driving, or Waymo with cost and geographic ubiquity? But you know what we're not debating about? Is it going to happen?

Right.

You know what we're not debating about? Is there a big technical breakthrough that needs to happen? None of those things. So, now we're clearly in the engineering side of self-driving, which is just this grind down to dollar-per-mile efficiency.

The moment that it's cheap, all the OEMs are smart. They'll just adopt it. It's not that OEMs are resistant because they don't think consumers want it or they don't understand the technology. It's because they want a price envelope that allows them to keep their razor-thin margins, at a scale that's deployed across 100-plus countries in V1.

If you're just doing a small deployment, it's very different. The last thing I would say is that the buyer of a Subaru or a Suzuki has very different brand expectations than the buyer of a Tesla, including the age of the consumer and what they think will happen or won't happen.

So, if you're Suzuki, you're like, "My buyer doesn't want this stuff, so I'm not going to jam it into the car." It's not because they're not technically competent. This is a different area.

When do you think it'll be routine in the 200 biggest American cities? When will it be routine to walk outside and just take it for granted that a robotaxi can come pick you up?

It's 2026 now. I mean, certainly by 2030.

Sorry. Okay.

Yeah, certainly by 2030. The big variable there really is because what Waymo will say is that the dollars and cents per city already work. It's like, "Well, a company that has basically unlimited capital—why are they not already in 200 cities?"

But then you see their launch schedule is pretty aggressive. You're like, "That can get there." So, maybe it has been aggressive. I'd say 2028.

Yeah, okay.

Like 2 years.

I would say available in 2030 but routine in maybe 2032 or 2033.

Because there's a scale-up, there's volume. And then also, if you're living in LA, 5 years ago I'd go to LA and people would be like, "What's Applied Intuition? I don't know what self-driving cars are." In the last couple of years, now they all know self-driving, and some of them even know Applied Intuition because they know it from the other manufacturers.

I think you fast-forward another 2 to 4 years, and everybody knows it. Now, does that mean everyone's taking Waymos exclusively? The answer is no, actually. If you look at the numbers, if you're Uber, you have to be scared. I mean, they're just eating into ridesharing.

Yeah, but to get 100% autonomy, I mean, that's another—yes, to be extremely cheap. And what about long-haul trucking?

Long-haul trucking—that's the passenger side. Long-haul trucking has completely different economics and a completely different business model. There are many companies right now—I would say probably north of 5—that are running long-haul trucks with drivers, carrying loads in America and China. I think China is probably getting into double digits.

It's there, but the reason you don't know about it and the reason it's not top of mind is that it's not a consumer product. Unlike on the Waymo and Tesla side, where investors are willing to essentially give you a market-cap adjustment for the potential of the business, they say the trucking business is—what's the word?

You buy a car with your heartstrings. You buy a truck with a calculator.

Yeah, it's a calculator business. It's pure dollars and cents. As the provider of self-driving trucks, if you're doing the whole thing, like some of the companies are—which we're not—you have to show every mile: “I'm going to save you this many dollars.” And it's, “For sure, for sure, for sure,” because the buyer's unsophisticated. They're just like, “Well, I already have staff that can drive,” and they're not inclined.

Where we're playing in Japan, it's not random that we're doing trucking there. There's a massive labor shortage today and an imploding demographic situation. There's demand from almost every sector, and that's why we've picked that market to really grow.

You can take even more obscure examples, like quarries, where you're literally moving cement and dirt.

Yeah, quarries.

Those are rock, stone, cement.

Rock, stone, cement. When are those?

I can tell you the people who own and run those things wanted it today.

The macro point that people don't talk about, though, is that in legislation and in the economics—the political economy—of this conversation, you see this big pushback in digital AI, because people are like, “I don't know what's going to happen to my job,” and VCs, I'm sure all of your associates, are very scared. But in our universe—

They're debating whether they need us.

Yeah, in our universe, it's the other way around. I'll meet these operators, and they're like, “We'll give you everything. If you can do this, we'll give you everything.” So then it's just up to us to get there as aggressively and physically as possible.

Well, the fear for a long time has been that trucking, for some reason, triggers the press's imagination on apocalyptic levels of job loss.

But that's so wrong.

Go ahead.

There's not enough truck drivers, and guess what? Nobody wants to be a truck driver.

Why is that? Let me explain.

Because it's a terrible job.

I grew up in a town whose main feature was a truck stop, so—

It's like—

But, yeah, why is truck driving not attractive to you?

This is like talking to my kid who's like, “Why can't I put my hand on the stove?” It's like, “Because it's going to burn your hand.” After the third “Why?” it's like, “Come on, buddy, let's do this.”

Make sure everybody knows I did not do that with my son.

Yes, yes. So what's hard? Why is being a truck driver a difficult job, or why would kids not want to do it when they grow up?

This is what it is. Let me use a parallel analogy, which is very clear. People will say, “Nobody wants to work anymore,” and they'll say, “McDonald's has all these job openings.” No, actually, what it is is that the people who used to work at McDonald's now do DoorDash and Uber.

Right.

Because it's better for them. They can start and end their hours when they want, they don't have a boss, they don't have to stand on their feet, and they can surf their phone in between orders. That's the reason. It's not random. The market is efficient.

In the truck-driving example, why does somebody not want to be away from their family for 4 to 8 days in a row doing long-haul trucking? The sharper example is in Australia: Why don't people want to get on a plane, go to a mine, and work there? Or go work on offshore oil rigs? Those jobs exist. If you want a job that pays 6 figures, they exist.

Even with such lucrative pay packages, it's not enough, because people are like, “You know what? I like being around my family, and I'm willing to take an incremental decrease in how much money I make.” Today, more than ever, things like back pain, being exposed to the sun, cancer, and people caring about those things are now part of the—

This is the thing, so tell me if I have this right, but I believe commercial long-haul truck drivers have a life expectancy 10 years less than their peers. I think it's a consequence of several things. One is some combination of nutrition and sleep. It's very difficult to eat well and exercise.

What's your sleep score if you're a long-haul trucker? Let me guess: They don't even sleep on that schedule.

Exactly. Obesity, heart disease, hypertension, and so forth are all very high. The second is the vibration; it's very difficult and stressful. Then, as you mentioned, cancer—truck drivers have a much higher rate of melanoma on their left arm.

Exactly. There are photos of a truck driver who's been driving for 30 years: one half of their face looks completely different from the other half because it's exposed to the sun.

An even more stark statistic: Mining is 1% of the labor pool globally and 8% of work-related fatalities. Do you think people are rushing to work in mines when they hear stats like this? Most major mines have a fatality regularly, which means once, twice, or 3 times a year. If you ever visit a mine, you'll see that everything is based around safety, because once you experience one of your coworkers dying, you're like, “What am I doing here?”

Yeah.

I understand you're trying to enumerate this for the audience, but these are not good jobs. This is not a mining podcast. This is not a podcast about how great long-haul trucking is. They're just not attractive jobs.

Even truckers don't want their kids to become truckers. It's a very—because of that, they want their kids to be in, at the very least, a safer line of work. Notwithstanding all that, how long do you think there will be safety drivers in long-haul trucks that are self-driving? Or, let's say, somebody else in the cab to deal with what happens when they arrive?

We know multiple companies that have driver-out goals right now. They're working to get drivers out right now, without going into our own details.

To be honest with you, it's not long. We're talking a few years.

I think on the long end.

Yeah, on the long end. There's a software-technology component, which is one part of the problem, but the other part is the redundancies you need in hardware and the validation necessary for those redundancies. In many cases, that can actually be a long pole.

It's like, they're productionizing a fully redundant steering system and a fully redundant braking system. That's not in high-volume production yet. Once you get that in high-volume production, you get the quality up, then that's validated, and now you can actually do these—

You need the price down.

Exactly.

Do you guys see a world where there are a billion of these little delivery robots running around?

Yeah, I think so. The product we're announcing, which I think will probably come out around this time, is called Dana.

You can simplify everything that Applied Intuition does into 2 buckets. We've been talking mostly about the models that go on the machines. We call that onboard software, or onboard AI. Then there's offboard AI: the tools to design and develop these same systems—the models that actually go on the machines.

Our vision for that, with the delivery robot as a great example, is that a high school kid or a middle schooler can make iPhone apps, so they should be able to make autonomous systems. Why can't they? Just ask that very simple question: Why can't a 9th grader make a delivery robot at home?

Well, they don't have the actual environment in which they would first develop the scenarios. They would define the requirements: “I want this robot to go around my high school campus, around, let's say, these 4 buildings.” Once you define the requirements, then you have the scenarios made. What are all the scenarios that can be made by using, let's say, a satellite image of the high school?

Now you have to train the robot, so you need some data. Where do you get that data? There's maybe enough publicly available data to train a fairly rudimentary robot. You get that data online—maybe from YouTube videos and a couple of other places.

Now you need to deploy it onto the actual machine. You deploy it onto the machine, and then the robot runs into the wall. “Okay, what happened there?” The loop closes. That platform for designing and developing is what we're launching.

It's called Dana, which is the street that Applied Intuition is headquartered on. This comes from our tooling background. If you look at how tooling has changed in the digital AI world, and at what Claude did to all of you, you all remember—from Mixpanel to GitLab, GitHub, and all these tools—everything has moved into a very different, almost IDE-like environment, frankly speaking.

We think the same thing is going to happen in the physical world. That's what we've built and what we're launching. We already use it in-house to develop our autonomy system, and we're working on the most scaled, complex systems on the planet across all these different verticals. So we're pretty confident that it's actually quite useful, and we've seen massive productivity gains.

But we also think other companies will use this to build their own systems, because it gets to that mission of 1 billion intelligent machines.

Fundamentally, Dana is our agentic platform for physical AI. Everything that we've built and developed over the past nearly 1 decade—every tool, every technique—is available in Dana. It's actually very easy to use with the agentic interface, so workflows that used to take days or weeks to run can now run in minutes in many cases. This lowers the barrier to entry for building these systems.

We're just lowering the bar for what it means to develop an autonomous system. Autonomy, still within the scope of software, is quite exotic. It's not because of the things that we've talked about; we've just brought that down very aggressively.

The old adage for how you make a great product in software is that you increase safety, convenience, or cost. We want to try to do all 3 of those things with Dana. Our hope is, just as you said, that kids can develop robots for their own use.

That extends to humanoids. We're not just talking about land-based systems. You can do humanoids, and you can do drones. Right now, writing drone software and deploying it is quite obscure and almost hobbyist. We want to make that absolutely—not child's play, maybe, but teenager play.

So this means a world with a lot more experimentation and entrepreneurship in agriculture, with bots, and basically in every domain—construction, defense. All of a sudden, you have a much larger number of people applying creativity, coming up with ideas, and making things that move.

With Claude, it's one thing to make engineers more efficient or bring more people into engineering, but when these agents really run, you're getting into the iPhone example. You couldn't imagine Instagram before the iPhone. Imagine 2005 on laptops: “In 10 years, there's going to be this app, and you can put photos in it.” “Well, the phones don't have cameras.” “Yeah, but it's going to be social.” What the hell? Facebook. It's just hard to imagine.

So we think that by lowering that barrier, you're going to get way, way more creative autonomy products.

Okay, I want to say that you can decide whether to include this or not: my kid is building autonomous bots in Factorio. That's one of his projects, and because the toolkit isn't available yet, he's actually training models. He's gathering data in the game, and he has a whole army of bots that he's developed.

Then his mother is like, “Why are you playing that game so much?” He explains, of course, that it's a purely educational process and experience. But it's the kind of thing—

There's no reason autonomy should be this obscure, difficult, alchemical technology. I think not only does that have a huge impact on society, it also allows people to understand that these systems aren't magic.

If I can develop a Roomba for myself in my house on a weekend using Dana, then it's not suddenly so scary. I think that's important.

It can support people in all kinds of ways that we haven't even imagined yet.

Absolutely. Think about folks with disabilities. We always think about humanoids as having this very important task of folding laundry, which seems to be—

[Laughter.]

So we focus on the important task. But when you allow these tools to exist, I feel very strongly—because we started a tooling company—that tools are what separates advanced civilizations from less advanced civilizations.

Our first mark for the company was a monkey's head, and then we got a designer who said, “This is stupid.” I thought it was pretty good.

You were talking earlier about how, when the technology got so good in mobile, there was a wave of these companies—Uber, WhatsApp, Snap, Airbnb, and so on—that emerged in quick succession. Now the technology is getting there for the infrastructure for physical AI. What are some use cases or companies that you can—obviously, it's hard to predict the future—but what are you most excited for? What could we be talking about as the equivalent here, in quick succession?

I think in the midterm—and we want Dana, if not in the short term, to really make humanoids way more real. There are, I mean, how many? Like 1,000 core tasks in a home for humanoids.

These companies—I'm sure if you talk to people who work in them, everything is difficult. Every step of the way is difficult. Collecting data is difficult. Cleaning that data is difficult. Training those models or deploying the models is difficult.

The bar is: I want a high school kid to make a humanoid. That's our path, and we think there could be a lot there. But that's the obvious stuff. I think the truly non-obvious stuff is going to be way more interesting when we look back.

There are some core ingredients that we're bringing together in Dana. We're making it much easier to get imitation learning to work, and much easier to make reinforcement learning work in combination with that. We have pretrained models that can be used as a baseline for a lot of things, world models, and advanced simulation techniques.

All of these things come together, and then you're sort of limited by your creativity: What do I want to do? If you think about any kind of physical AI task, it's understanding the world and manipulating something, and we can build that. That can be built much more easily in this tool.

I think sometimes people ask, given that we're a tooling company—and take self-driving trucks: we deploy self-driving trucks, and many of the self-driving truck companies use our tools—“With Dana, are you going to enable all these competitors?”

That's great.

Right.

That's absolutely completely fine. If you look at Google and what Google did to web applications, there was a massive internet ecosystem. Google still succeeded through search, YouTube, and other web apps.

Other folks learned and used open-source products, then ultimately closed-source products, and ultimately venture-backed products.

So, we think the same thing can happen here.

I was at a robotics startup a while back that you guys know well. They were training one of their arms to do the particularly killer app that I thought was very appealing: picking up dog poop. [laughter] Literally training over and over again with different tools. And so, I don't know why, right? Why not have the little robot follow you around when you walk the dog and pick up the poop?

I know somebody who built a little lawn robot that would go around and pick up individual leaves.

Yeah. Because you got that problem, right? You rake your yard, it's completely clean, and then, 2 hours later, there are 14 leaves, and you're like—

Yeah, yeah.

I'm just going to send out the little bot to pick up the leaves.

It's like if development costs are 0, then people will do that. You guys remember the early iPhone apps that hit were the beer one or the fart app. [laughter] If you imagine that in '98 with the Symbian mobile OS—from, I think, Ericsson or somebody—that would be impossible. You'd need a team of 50 people to develop the beer thing for the BlackBerry. So I think there's a similar type of thing happening where we really want to be a part of that and enable that. I think it will still be a while before making a robo-taxi is super easy, but that'll happen.

Yeah.

But the number of kinds of bots that could be deployed in health care is almost endless. Health care alone is endless. Home care.

Yeah.

And then in construction, in all the physical trades.

Imagine us sitting in 2007 and saying, “We should have an app store. What type of apps?” We would come up with a list of 8. There would be a messaging one and then a camera one. Now you look at the App Store, and there's an app for the hotel you go to, to order food off the menu.

Right.

Yeah.

Yeah, makes sense.

Qasar, we were talking earlier about the differences between digital AI and physical AI. We're sort of hinting at LLMs, but world models are in vogue right now. Do you want to talk about the state of them as it relates to physical AI and how we should think about them?

So, first off, “world models” means about 100 different things. We had a team at CVPR recently, and I was joking with them about just how many different ways you can define what a world model is. When we're thinking about a world model, we're typically thinking about it in the context of a simulation. There are things that are sufficiently able to represent the real world and are reactive in a sense where you can actually have, let's say, an autonomous agent acting in this world, and the world model is behaving appropriately in response to that autonomous agent.

Maybe, Peter, I think it's worth being super explicit here. Or we could just go one level lower into determinism in simulators, the sim-to-real gap, physics-based rendering, all the way to this generated world. Where do we fit on it? Just describe the landscape, I think, maybe.

Yeah, yeah. This is simulation broadly, right? There are so many different ways of doing simulation. The more classical approaches are very physics-based, and you can decompose physics in all different ways and at all different levels of abstraction. You can simulate with sensors or without sensors. Is it just a body simulation, or are we actually simulating, for example, the light in the environment?

It's almost like the way CGI is done. If we literally had technical artists—and we have technical artists—who would create assets that would go in the simulator, which would mimic real road signs and have reflectivity and material properties that you would see in the real world. But as you guys know, Hollywood is going through its own fundamental change. And now with generative AI, the same thing is happening in our universe as well.

Yeah, so that's sort of at the far end of physics-based simulation. The opposite end is purely neural simulation, but within that spectrum, there are many different things you can do that are each useful in their own right. One of those things is Gaussian-based simulation, where you have a representation of the real world that has a 3D representation. That 3D representation is consistent, meaning that if you have some reference point, let's say a camera, and that camera moves within that 3D world, because the Gaussian is actually representing the 3D geometry of that world, you'll get very high-quality output from that. There's a lot of value in that, and I'd say that's one type of world model.

But when you go further on that spectrum into neural simulation, then you get into systems where you're actually generating the video feeds. You can think of a neural network that's actually outputting a video. That's what's actually coming out of the neurons. And that can be reactive, which gives you some very interesting properties. Reactive as in the ego does something in the environment and the other agents respond to the ego.

Exactly.

However, you're not guaranteed in that reactivity that it's accurate, right? And now it's a question of, well, how can I align this simulation, this world model, with the real world and the way that the real world would actually react? If you have perfect alignment between the real world and the world model, I think you've just sort of solved the universe, roughly, right? That's an impossibly difficult problem. But as we make progress toward that, it makes training physically accurate models much easier because you can do more of that in simulation.

The hardest part, though, is that we're always thinking about performance, right? I like to say that the labs have it easy because they can make models that are trillions of parameters, and those models can be super slow, and that's fine. But we don't have that luxury in physical AI. We deal in real time—the actual clock, real time. We have so many milliseconds before we have to do something.

Those performance constraints actually constrain the problem in a lot of ways. We can have very large models, and we do have very large models that are used in the off-board environment. But once you go on board, all of those constraints are very real. Now we need to train a much smaller model that has these safety constraints and these determinism constraints. That's the hard part about physical AI. That's also the moat, right? It is what makes our tooling and our competencies valuable, because it's just really hard to meet all of these constraints in a physical system.

Which will we get first: a perfectly simulated real-world environment for training autonomous devices, or Grand Theft Auto 6?

You know, as long as they keep putting out great trailers, I feel like I'm getting entertained without paying a dollar. I'm being reintroduced to Tom Petty because of it. [laughter]

Will you give us some timelines?

Let's run with that for a second.

Yeah, let's go for it. [laughter] Well, no, look, the whole thing with Grand Theft Auto is that the big innovation was open-world sandbox gaming. So it's a simulated city, at least in theory.

On that spectrum, I mean, we hire so many people out of the video game world. On that spectrum, it's absolutely real.

Well, then tell us about that. What's the spectrum?

This is speculation, but I think Grand Theft Auto 6 will perhaps be the last major real-world video game that's still really developed, let's say, in that legacy era of traditional computer-graphics tooling.

Technical artists, yeah.

I think that Grand Theft Auto 7 will much more likely be a world-model-based video game.

Right.

You can imagine, as AI technology evolves here, this concept of a video-game world model. There's some sort of baseline data store that represents the real world in some way, and then you have a translation layer that's actually turning that data store into something that you can see and run around in. It's possible.

The game is a construct. It could be the real world, right? As we said, you could have a complete recreation of the real world in the game.

Well, this has kind of happened with flight simulators, hasn't it? The most recent flight simulators literally render the entire planet accurately, at least from the air, is my understanding. Is that right?

Yeah. That's where our bread and butter is. When we started the business, we hired so many people out of Microsoft Flight Simulator.

But when you fly over New York—or, you know, Duluth—in the flight simulator now, it is the real city, right?

Exactly. But there are some tricks they play there, and a lot of it is fidelity. In the real world, the more you zoom in, it stays at a certain level of fidelity. So the trick they play is that you basically downsample very aggressively, and then as you get closer, it becomes higher fidelity.

The real world isn't like that. If you were to try to rebuild the world with this level of fidelity, it would take all the energy of the universe, right? It's quite complex. That's probably, by the way, the best argument against us living in a simulation.

But of course, then you would say, "Well, the simulator we're in doesn't follow the laws of physics that we're subject to."

Yeah, yeah.

As far as I know, everything happening outside this room doesn't even exist.

Yeah, Buddhism believes this. This is a different type of podcast. When you open your eyes, the world is rendered.

That's literally religious. I don't see why it's necessary for it to keep rendering if I'm not there.

Buddhism from first principles.

Yeah, exactly. That's what you should call this. That'll get a lot of clicks.

Well, just staying on the timeline topic, you gave us timelines on self-driving cars. What timelines do you want to give us, if any, on other interesting things that are worth tracking—like perhaps when we'll get laundry folded, or other things that emerge because of humanoids? Or maybe just touching a little bit on world models, where we see world models going, because I think it's fundamental to all the work we do.

Yeah, pretty sure. To answer the first question, laundry folding isn't terribly far from being solved, to be clear. There is a lot of interesting research being done.

And then humanity can rejoice. That's in Proverbs 4:16, I think. [Laughter.]

Well, here I do think housekeeping is a killer use case for physical AI, right?

Peter has 2 things he always talks about in the company. One is housekeeping and the other is entertainment. Peter is long on humanoid entertainment.

What is it like?

What kind of entertainment?

I 100% agree. I think entertainment is the robot's killer app. I don't think anybody else—

Yeah, these are your admissions.

I mean—

These Midwest white guys are really into this.

I'm just saying, but I think—

I just want to know when I get Westworld. That's all I want.

Yes.

No, I actually have an entertainer. I have a tiny little Chinese robot dog that's literally just a little robot dog. It just runs around a little bit.

Would you pay to see Cirque du Soleil with robots?

Yes.

I want to see kung fu trapeze swinging.

Spoken like a compiler guy.

I want Westworld. I want Westworld.

The funny thing is, I was just saying, would people in the suburbs of Detroit actually pay to see that? I bet people in Sterling Heights would actually pay to see that. It's actually probable. I stand corrected. [Laughter.]

But back on laundry for a moment, it's actually not far from being solved if you remove the time constraint. The trick they play in the latest research videos is they'll say, "Play it at 8× real time," or whatever, right? And that's for you to make it watchable. So the question is, when can you actually achieve human parity of performance? That's further off.

When you decouple models from just the hardware, the hardware can do it now. That used to be a constraint. The hardware is very fast and accurate now, which was actually—

There are still overheating issues that are still being dealt with, but it's not terribly far off. These are solved.

I mean, it's far off from when I was a mechy. That was fantasy. There was nothing—

What's the movie that has the most realistic future vision of robots?

Oh, man. Bicentennial Man.

Is it? Okay.

Yeah, he's realistic.

Why that one? I actually haven't seen that.

Well, I like that scene—I think it's I, Robot—when Will Smith jumps in the car and his accomplice is in the car. He puts the car in manual, and she's like, "What are you going to drive this thing yourself?" She's like, "Are you crazy? What are you going to drive this thing yourself?" That's the flight of intuition's goal.

Well, again, I haven't seen this movie in a long time, probably since it came out, so my recollection of it is probably a bit incorrect.

Don't worry, the internet will correct you. [Laughter.]

But I think Bicentennial Man has fully self-driving cars. It also has the housekeeping robot, which is played by Robin Williams. It's sort of the friendly robot that will clean up and also babysit your kids and stuff like that. It seems like it's in the not-terribly-distant future.

I got a different answer. You guys ever see that movie with Sam Rockwell, Moon?

Oh, yeah.

Yeah. The setup—I don't want to spoil it. It's a great movie. Don't watch the trailer; just watch the movie. The premise is, "250,000 miles from home, you find who you are." It's one guy who works on an energy harvesting base run by Pliant Intuition, run by Lunar Technologies.

I'm not—I don't want to be Weyland-Yutani. I don't want to be Tyrell Corporation from the Alien franchise or Blade Runner. No, no. I want to be Lunar Technologies in the Moon franchise. Not even a franchise. There’s one guy who works on the base, which basically runs by itself, and he's just there to monitor it when some error signal comes up.

The reason why it's so accurate, I think, is because the state of the art for AI systems is that these systems just need the occasional grounding.

Exactly. They'll just go off and do something crazy, and then you have to say, "No, no, stop doing that."

I like that, too. I mean, that's a coding bot, sort of, right?

And the other reason I think it's quite accurate—maybe it's uncouth now—is that Kevin Spacey is the AI, the smiley face. He's just there to placate the human, to assist, but also to say, "Oh, you seem like you're sad, Sam." But really, it's the AI that's running the base.

Hopefully—I shouldn't say we want to be Lunar Industries, because I don't know if they're quite a positive force of nature in that movie—but I think a massive energy farm that's completely autonomous is going to be the future. Everyone reacts to things like that with fear, and it's like, guys, that's amazing. That means energy costs go way down. That's an incredible positive thing.

I think—I just did this commencement speech at my undergrad.

Did you get destroyed?

No. You know what? I—

Unlike Eric Schmidt.

Yeah, yeah. Listen, my wife started watching it, and she said, "I feel like you're yelling at me. I can't watch this." [Laughter.]

I'm not going to say which tech leaders basically avoided the topic by punting and saying, "I'm not going to talk about it." I talk about this stuff. Partly, it's the General Motors Institute. I don't want to pass judgment on the people we recruit out of MIT and Stanford, but I'd say GMI people are a little different. They're pragmatic people, and they understand—you don't go to a place like GMI if you believe a superficial view of what corporations do.

Corporations are just people working on projects together. And, by the way, people working on projects together in government and people working on projects together in nonprofits—they all screw up. It's too simple to say AI corporations are terrible. You also can't say the other side, which is, "It'll all be great."

So you have a role to play. That's basically what my message is, and that's the case. I think if the obvious abundance that comes from self-driving trucks and self-driving cars, and the fact that people don't die, which is amazing, doesn't satisfy your fear—and then you also get this efficiency of cheaper energy, et cetera—it's your responsibility, as a person, to really learn about that technology. You can't just say, “Well, I'm afraid of it, and my reaction is to shut it down.”

Mhm.

I don't say this just to say that we're competing with the Chinese, but there's a Confucian saying: “No hand can block the sun.”

Somebody else is going to do it.

Somebody else is going to do it. And if it's not the Chinese, who knows? Maybe it's the Uzbeks, or it's another country that is recognizing, “Hey, my citizens are suffering, and I'm going to use this technology to remove the—”

Honestly, it's because we're living in such a great society that we can have these, I would say, stupid conversations. There still are people who can't get food.

Yep.

Someone debating me would immediately say, “Well, there's plenty of food.” No, no, no. Let's be very specific. There's plenty of food, but getting that food to those people is difficult.

Yep.

So that means we should let robots get that food to them faster.

Yep.

That's just how it is. I think, as technologists, sometimes we have an inclination just to say, “Leave these people behind.” I think you have to bring them along. You have to explain it to them. But we also have to treat folks like adults and say, “If you don't get it after I explain it a couple times, then you just don't get it.” So there's a middle ground.

It's not that everyone's an idiot, and it's not that technology will just be perfect, perfect, perfect. There's a middle ground. Let's have that conversation to a point.

Yeah.

And then we just move forward and make society better. And then the results show it. I mean, there are people who still, shockingly, believe communism is the right answer. I just want to say: why? I'm a capitalist. I can't not admit that. But there's 70 years of history there.

That's not a debate anymore. I think it could be a debate if we're sitting here in 1965 and having a debate, and you say, “Okay, maybe centrally controlled systems work better.” There's no debate anymore, folks. Systems where individuals make decisions in their own interest actually work better for society.

And so that doesn't mean everything is perfect. You can't extrapolate. That's the same thing with AI. It doesn't mean everything's going to be perfect, but net-net, it's definitely going to be better. And that's roughly what my commencement speech was without the boos.

Marc and these guys were booing, so they just cut it out. Erik was booing and throwing stuff. They just edited it out.

You mentioned the Japan market earlier. When you talk briefly about the global ambitions, how do these technologies interplay, and what are we doing here?

So I think America, particularly, is still the most advanced when you take business models into account. The second thing for a company like Applied Intuition is that we're an extremely global company. We work with everybody—minus China; we don't have an office in China—but really everyone else on the globe.

We're a horizontal company. We're a technology provider. And I think more Silicon Valley companies can employ a little bit of what we do, which is work, I would say, very collaboratively with the local economies.

As sovereign AI becomes more of a real thing, we have to build businesses that take that into account. By the way, we're not the first ones to do this. If you look at the history of America, if you read the history of Standard Oil, you'll see that this was the history of companies: you'd work internationally. Aramco is not a random company, right? You build based on the real geopolitical realities of the time.

And so I think we've navigated it quite well. I've lived in Japan, Germany, and Dubai. Also, being Pakistani by birth, I think that's influenced our company. Peter's only lived in Michigan and here. But he is German.

So I think, innately, we think about the globe more. When I was at both Google and YC, I was always surprised at how almost myopic the companies are. They're always looking at the market that's just within the 30 miles between San Jose and San Francisco. I was like, actually, the market is really big.

I think physical AI, the nature of it being physical, means we have to be a very international company. And I think we've had a lot of success being very, very international.

Yeah.

Cool. All right, that's a good place to wrap.

Okay.

Peter Ludwig and Qasar Younis, thanks so much for coming on the podcast, and congrats on big lots of data.

Yeah, thanks for having us.

Awesome. Great to see you guys.

Okay, great.