Physical AI: Could Training Data Become a $45B Market?

0:00China, for example, have data collection cities. JD.com, the Amazon of China, is essentially trying to create a city where they would employ 100 ,000 people in that city to collect their everyday lives, to train robots. They're having just everyday citizens go and pick up headsets to do their daily routine.

0:40Sce Pike: Welcome to Media and the Machine. I'm Rob Kelly. AI is the biggest technology shift of our lifetime, and this show is about how to profit from it. I'm a three-time founder and CEO of content-driven businesses, and each week I talk with the people on the inside of AI and content, the ones building this new world and learning what works firsthand. I'm an optimist about AI, but I'm also a realist. So every show ends with me asking our guest what AI means for our jobs, our families, and our future. My guest is Say Pike, VP of AI Growth and Solutions at Telus Digital. She built and sold two companies.

1:15Sce Pike: Ernst & Young bought her digital consultancy, Citizen, and ADT bought her smart home company, IOTAS. Say now has her sights set on physical AI and robots. We break down Google's recent breakthrough where its new Gemini 4 Argon taught a robot to play cornhole, and it made seven shots in a row. But what makes data skills like that possible, and who's going to supply it? We get specific about what AI labs want and why robotics companies like Figure AI are getting into the data collection business. We discuss why robots might need a billion training hours, plus our forecast of a$45 billion market by 2035.

1:50Sce Pike: That's just collecting data for physical AI. Then we explore China's data collection cities. Imagine 100 ,000 people recording their everyday lives just to train robots. And why she thinks China, despite being behind the U.S. and LLMs, can be number one in physical AI. Plus, a robotic centaur with chainsaw hands. As always, we close talking about AI's impact on our families and jobs, including advice for her kids, their artists, and their rebelling against her AI optimism. Please enjoy my conversation with Zay Pike. This Gemini 4 Argon playing cornhole. Yes. And it didn't just play cornhole. It sunk seven shots in a row.

2:33I think one of the things that was very nice about that demonstration is going from reel to simulation. So taking a reel video of a human doing a shot into a cornhole. So watch that video and then put it into a simulation and simulated the physics of that, of the, what do you call it? The bag, the hacky bag or whatever, right? Going into the whole - Bean bag. The bean bag, right? And so it simulated it and then it actually took it into its own practice and learned over and over again after like how much force it needed, what's the trajectory, right? And so it watched itself do that and learned over a period of time to get to understanding how to make that shot.

3:20What it didn't show was the robot picking up the beanbag and then throwing it. Contact physics, meaning like how hard it had to pick that beanbag up, how soft it had to hold that beanbag when it threw that beanbag.

3:38Sce Pike: So it's not quite the AlphaGo, Deep Blue type moment for physical AI. The ideal scenario, the end game is like, I have a robot in my house. I show it how to like, sort my laundry and put it into the laundry machine and push the buttons and then like, take it out, put it in the washer. You show it once how you want it done, and it'll just perfectly mimic you. I think that would be the breakthrough moment. If it could mimic anything that you did. I mean, that would be the end game because it's not just household chores. You're talking factory floors. You're talking retail. You're talking hospitality, right?

4:18It's just if it could actually watch you do something once, twice, three times, whatever, and then actually do the task, that would be the end game.

4:29Sce Pike: And the overall definition just of physical AI, is it simply that we're going from using AI for digital reasoning to physical action? So physical AI, when I say physical AI, it encompasses both robotics as well as automotive, autonomous vehicles as well. So physical AI to me encompasses anything that is action oriented in the physical world. So physical models versus LLMs. I just want to contrast the two because most people now are familiar with learning language models, LLMs. My layman's take on this is LLMs learned from the internet. Let's just kind of simplify and say that's the corpus where it learned from, right?

5:17Sce Pike: And what's the closest thing to the internet, in air quotes over here, What's the closest thing to the internet type corpus that robots will train on? So the closest thing doesn't exist. This is the issue with physical AI is that it has already ingested all of the internet. It can't learn from that because that's not how the real world works. So it needs data to train on. And unfortunately, that data doesn't exist. So traditionally, different companies tried lots of different things, which is like simulations, you know, synthetic data. And over and over again, research has shown that the best type of data for it to learn from is real world human collected data.

6:06And so this is why you're seeing this uptick of people maybe wearing headsets and collecting themselves, doing household chores, washing clothes, doing laundry, or in factories, in like garment factories where they're recording themselves sewing and other things. And so this is where China, for example, have government citywide initiatives and they call themselves like data collection cities. And so this is where it gets really fascinating. There's just not enough corpus, as you will, to train from.

6:41Sce Pike: And these data collection cities, are they, can you just describe what you mean by those? Like what would it look like? Yeah, so JD.com, one of the largest e-commerce groups in China. Sort of the Amazon of China. Yeah, the Amazon of China is essentially trying to create a city where they would employ 100 ,000 people in that city to collect their everyday lives, to train robots. They would also send out robots to do tasks in people's homes so that it could learn from it and train from it as well. But they're having just everyday citizens go and pick up headsets to do their daily routine. And that's a state-sponsored thing.

7:27And JD.com is part of it as well. And it's just fascinating to me how the whole government, I think last year, China's government, said, hey, physical AI is something that we're going to win at, right? because I think maybe they saw that with OpenAI, Anthropic, and other kind of US-based, what I would call like AGI companies win the race. They're like, we're going to win the physical AI race. And so they're just mobilizing all of their initiatives just the way they did with EVs when they wanted to get EVs adopted. There's tax breaks for the citizens. There's incentives. There's just a huge mobilization going on right now over there.

8:09Sce Pike: And the gear you're describing that they would wear, this is the egocentric kind of data collection devices. Yeah. So they're like headsets. It's like a halo around your head. Have several cameras that have a field of view of at least like 150 degrees or more. Sometimes when we do collection, we wear wrist cameras as well. So people are wearing cameras on their wrist that show kind of the dexterity of your fingers, because really that's the hardest thing, again, for robots is just understanding grasp, understanding contact friction. Well, it's interesting. I mean, if we kind of call it, if we won the LLM race, I mean, currently in the lead, the physical AI race, I mean, that is really up for grabs and you would kind of have to bet on China because of its robotics experience and its volume of people, right?

9:02Sce Pike: You nailed it. What are the top physical AI models right now? Physical intelligence, releasing their API. So I thought that was really fascinating today. Dyno Robotics is also very interesting to watch because they're also producing a lot of data as well. And that might get them ahead of the game. And then Figure AI, which I thought was fascinating, they are also a robotic company, but they also went into the data collection business as well. So some of these companies that are robotic companies are switching over to data collection, data creation, again, because that's where the revenue is right now, because there's just not enough ground truth data out there to advance the models for.

9:53Sce Pike: Let's talk about training data needed for physical models, physical AI. We touched upon it a little bit, but I just had a few key questions here. How many hours do you think are needed to really make this thing sing? So until recently, the number that was commonly shared by all the labs is like hundreds of millions of hours. Most recently at Iros in Pittsburgh, which is a fairly large and old conference for robotics and autonomous vehicles and whatnot. though this time around it was all robotics. I heard a billion hours being thrown around. Who said that? I can't say, but I'll say that they should know because they themselves are in the simulations and physics business.

10:43So physics engine business. And if a sim and physics engine business is telling you that, they know what it takes to make their simulations more accurate, right? And they also work with all the foundation labs, hyperscalers as well.

11:02Sce Pike: Can you list out, let's just walk through sort of the types of data that physical AI needs. So first question is, assumption here, they use the equivalent of the corpus of the internet, sort of like what ChatGPT and Claude and Gemini already know? Yes, they do. And, you know, the numbers vary all over, but by 2035, we may be looking at like a 45 billion or more market size, maybe bigger. This is for the data collection market? Yeah, just the data collection market. For physical AI, or does that include also? For physical AI, just for physical AI. Wow. Just for physical AI. So with egocentric data, which we've talked about, which is people wearing headsets and wrist cameras and whatnot, there's definitely something like that that's necessary.

11:47but there's also teleoperation, right, where somebody is wearing controllers for the actual robot model, right? And so mimicking all the tasks using controllers for the model and teaching it that way. So there's a combination there, but it's not just the data collection. There's also the quality control that's needed of that data, right? If you're going to generate a billion hours, how do you actually check that for quality, that it's actually usable for training. And then on top of that -

12:19Sce Pike: I would guess AI is what's used to check a billion hours. You would have to use some level of AI and this is actually where - Or another city in China. Yeah, but there is a lot of human annotation and labeling that goes on. So somebody has still on the video itself labeled the task of like, this is somebody putting on a shoe exactly at this time in seconds and milliseconds, right? And then tying the shoelaces next. But going back to just the data creation, you still need 3D assets, mapping of an environment. So like Niantic Spatial is doing that quite effectively out there. So just on the data collection math side of things, like the billion, you know, I had the CEO of CuriosityStream and the CEO of Trovio, two companies sitting on a bunch of video between, call it three to eight million hours that they've got, the two companies.

13:14Sce Pike: Just doing some basic math here, let's just say we need to get to a billion hours. I mean, it would seem like it would have to be setting up these data collection cities or companies fully. So I doubt the United States will be able to pull off a data collection city like China. The way we're doing it is through crowd collection as well, right? Tell me more. Yeah. So in our crowd, we have probably over a million people that we have utilized in the past to train AGI, Gen AI, LLMs. So we constantly survey the crowd to understand whether or not they are capable and willing to do physical AI data collection as well.

13:59Sce Pike: And this is TELUS. This is Telus Digital, yeah. So it's not us employing a million people or setting up a data collection city, but it's us reaching out to this community that we've built to train LLMs previously to now train physical AI or collect data for physical AI. it's actually just doing your normal tasks that you would do on a daily basis. Go to your job and wear a headset and just record what you're doing on your job. And we will essentially take care of the rest kind of thing, right? And then you pay them? Yeah, you're getting paid on your job and getting paid by us to record that job.

14:42For us, it's using the collective power of the crowd, the gig workers, to be able to generate this. So will it come to a point where, you know, I think people are dubious when Uber came out, this whole gig economy, right? Initially, when it first started happening, I don't think anyone understood how many people it would employ. This is going to be very similar. I think this is going to be very similar.

15:08Sce Pike: Help me picture what types of jobs, like common types of jobs, are folks doing where they're wearing headsets and getting paid to collect this data for physical AI? So right now, models are asking for scenarios within retail, hospitality, farms, factories, homes as well. So right now, those are the kind of highly in demand work, right? But seriously, like you could be a barista, you know, you could be in the fast food industry. You could be working in a hotel lobby, right? Or you could be working in a data center, or you could be a nurse. But I think those are the areas where we see a tremendous need.

15:56Anything that's also hazardous, you know, firefighting, like there's this really funny robot out there in the market that is designed for firefighting in the wild, right? So in the woods, so firefighting in the woods. And it's stuff like that. I think hopefully that will take off, that will make it easier for us to do some of these more hazardous jobs.

16:22Sce Pike: And are companies paying TELUS for this data right now? Is that how you help subsize it? Okay. Are there things that you don't want to show robots in terms of training data? Yeah, probably like we would never have a robot be trained on anything that a human did where, let's say, they pushed aside a cat or a dog or a pet not so gently. Like we would not want the robot to be trained on stuff of that nature, right? Because you don't want them to learn it. No, we don't want them to learn it. Yeah, right. So there'll be some sort of like constitution or values built in for the data set for robots, depending on what company is building what, right?

17:12Exactly. And I think this is where LLM companies, for better or for worse, built their constitution after they built the LLM, right? Whoops. I would hope that robotic companies or physical AI models would build their constitution in before as well as into their training material, because I think this is kind of critical, especially for things that are going to be deployed to be working alongside us and have physical presence.

17:39Sce Pike: Can you talk about just use cases of what robots will do, maybe interesting ones outside of business? For instance, I just heard Kevin Ryan, founder of a lot of companies, he thought robot massage was going to be huge because people want their privacy. And it just can learn to apply the right pressure in the right spots. I'm skeptical about the robot massage, honestly. Because like, again, the contact. You're not signing up? Yeah, not going to sign up for that because the contact friction is probably one of the hardest things. I do think some better usage of it would be robots in dangerous situations that humans might not want to necessarily be in.

18:22Sce Pike: Fighting wildfires, as you talked about. Yeah, yeah. Or fires in homes. I know you mentioned the goat head centaur. Can you just describe that, by the way? Yeah, so that one again is for fighting wildfires, which is basically a quadruped. it's a centaur, but it's also instead of hands - Quadruped just means it's on four sort of legs. Four legs. Yeah, four legs. But instead of hands, it's chainsaw hands, right? So it's like your worst nightmare come to life. So half of the body is humanoid up top. Legs are like quadruped, it's a centaur. And instead of hands, it has chainsaw as a hand. Why does it have chainsaws when fighting fire?

19:04Sce Pike: Just to knock down trees? Yeah, to create a fire barrier, right? Wow. Amazing. Wouldn't want to meet one of those in a dark alley. But it probably was really useful if it could actually do the job. Any other wild stuff that you've seen in that category? I don't think it's like that wild, but it's mainly for security purposes. So a lot of quadrupeds in construction sites. Humanoids. What's a humanoid capable that far exceeds that of a human right now, if anything? I mean, a few did beat Usain Bolt's speed record recently in the Chinese Robotic Olympics, you know. So I guess in that sense. Yeah, that didn't get as much publicity as I thought it would.

19:49No, no. Because the Chinese media kind of hit that amongst robots running into walls and being silly, which is a whole different thing.

19:58Sce Pike: And you brought up that you had a theory about why they were being extra silly about those robots. Yeah, because I believe that the Chinese media, and don't hate me if you're a Chinese media controller or not, but I think really they're doing a great job of trying to make robots not seem scary by making them look silly. So I think it's a way to make sure that their population will adopt this technology versus our way of marketing something. I mean, like Optimus and figure AI, they do look scary. They're not the cutest looking things or the most approachable looking things at all. The design of them are not friendly.

20:39Sce Pike: What are your thoughts about how people will pay for robots, like on the consumer side? Yeah. So Optimus supposedly is going to retail for 20 to 30 ,000. And I'm guessing Tesla is going to do a lease model. And I think most companies will probably end up doing a lease model. So it'd be like a car. You can buy it or you can lease it. Exactly. So lease might be a couple hundred a month or something like that. Exactly. A couple hundred, 200, 300 a month, right? And if, I don't know about you, but like I pay more than that for my housekeeping on a monthly basis. So to me, if it actually could do all the tasks that my housekeeper does, it would pay off, right?

21:17Sce Pike: Are we going to see humanoids traveling in planes and cars? What are your thoughts about that? I guess why not? Definitely cars, I'm sure. You know someone's going to put it in the front seat of those places where you can drive in the fast lane to commute. Of course. If you have more than one person. And mannequins, right? So the question is like, will it be able to drive? Right. Right? That's a weird one. I mean, like, does it actually need to be able to drive or does it just take over the autonomous system kind of thing? Like a backup to the self-driving autonomous vehicles too. Yeah. Like a backup co-pilot.

21:57Sce Pike: That is crazy. Yeah. Yeah. Because my 16-year-old won't drive. Your 16-year-old won't drive? Will they let a humanoid drive? I mean, it might be better than my 16-year-old who hasn't learned how to drive, right? Or any 16-year-old. Have they been in an autonomous car yet? No, not yet. Okay. That'll be fine. But I would trust an autonomous vehicle more than I would trust an Uber driver. Yes. Transport my kids around. It's a weird mental thing. What do you think about when I just bring up cyborg, the word, and are we going to have this hybrid us humans with gear that is powered by things like physical AI?

22:37I'm probably an exception, but people used to ask, would you get a chip in your brain? And I'm like, yes, I would totally get a chip in my brain. So, you know, there's companies out there, again, Neuralink is something that Elon Musk has been working on for a while. Yeah, essentially putting a chip in your brain to be connected to the internet all the time.

22:56Sce Pike: I heard a guy had a chip just under his skin of his hand, and he's using it to open up his door in his San Francisco apartment. Yeah, I'm not surprised. Some of the things that we're getting asked for a lot right now in terms of data is natural language conversation. So what I mean by that is just from labs that were asking us for more audio content. because you can't go up to a robot and type into a screen to tell it what to do. You got to like yell at it or, you know, talk to it. And it has to be able to talk back to you. But that's not just for robots. That's for everything around you, right?

23:32So I think there'll be a lot more natural, physical interaction. And so user experience will move away from screens to more physical interactions in the natural environment.

23:43Sce Pike: You worked at Palm. I loved Palm, but we didn't make smartphones happen. iPhone did. And, you know, we were all so busy trying to make the keys angled certain way and competing against BlackBerry did not think that a keyless smartphone would be the thing that dominated, right? Interesting. What's your thought on the impact of physical AI on jobs and AI in general? I'm very personally conflicted about this because I flip-flop between like, oh no, this is going to be great for humanity. to moments of like, oh, this is not going to be great for humanity. I think the concern is that right now the jobs that the AI is going after are junior level jobs, entry level jobs, and it's really impacting the younger generation that are entering the market.

24:34And I have kids who are going to be entering the market too soon. And it does worry me. But I'll say this, like my kids are creatives, they're artists. And I do think that art is the one thing that AI will never actually truly replace. Like technology has never replaced artists as long as we've had technology going back to fire.

24:54Sce Pike: Who's the little person who appears to be walking on water in your LinkedIn profile? You know, the backdrop picture? Yes, that is my son when he was very little. He was about three or four. Yeah, he does look like he's walking on water. So he's not a robot. No, not a robot. Not a robot. But yes, that would be my son, Kian, who's now taller than I am. And he calls me short all the time. How old are your kids? 17 and 13. 17 and 13. What are you telling them in general about this new era of AI? Art has always defined, and artists have always defined the human experience. And that's, I think, something that's always going to be wanted and needed.

25:39even if it's using AI to define that human experience.

25:43Sce Pike: And are they on board? No. Yeah. So they're like many others of their peers. They're total Luddites. Yeah. This is how they rebel against me. Very much Luddites. Hilarious, huh? Who would have thought that the older members of the family are the ones who are more optimistic about AI than the younger? Yep. Yep. I think it's a major backlash. We'll say thank you so much for investing the time today. Yeah, no, thank you. This was fun. Well, that's it for this episode of Media and the Machine. Before you go, a few quick things. If you want to hear the next episode, hit follow on the show in your podcast app.

26:21Sce Pike: If you like something about this episode, the guest, the topic, even one part of the conversation, please leave a rating or review and tell me what. And if something didn't land, email me. I want to hear that too. Either way, it helps me make the show better for you. If you want to go deeper, head to mediaandthemachine.com and subscribe to my newsletter. You'll get my favorite highlights from the show, plus essays I write about AI. That's also where you can email me or recommend a guest. If they make it on the show, I'll give you a shout out. And if this episode made you think of someone, please pass it along.

26:51Sce Pike: That's how people find us, and I'm genuinely grateful. Thanks for listening, and see you next time.

27:08We'll see you next time.