Robots Outrun Usain Bolt, Skild's ChatGPT Moment & Google Bids $10M for Spirit Data | Lightwork

0:00Josh Machiz: That robot sprinting at full speed, I think faster than Usain Bolt, and then completely collapsing, hitting a wall and it caught on fire. You can't buy Boston Dynamics in the mall. Definitely feels like the chat GPT moment for robotics. What we're seeing here is quite literally that data is the new oil. It does feel like we're moving very quickly from let's beat human records to let's beat other robotic records.

0:30Josh Machiz: We're back with another big week. Robots beat three human world records at the World Humanoid Robot Games in Beijing this week. And Unitree, one of the companies in the field, went public. Skilled AI had a major breakthrough with S1, which is one AI brain the company says can run on any robot body, no matter who built it. And finally, Google bid$10 million for Spirit Airlines data to train its AI. Let's get into it. Claire, did you watch the Robot Olympics this weekend? I did not watch the full thing. I actually think it's still ongoing, but I did see the clip of that robot sprinting at full speed.

1:11Josh Machiz: I think faster than Usain Bolt and then completely collapsing, hitting a wall and it like caught on fire or something. Oh my God, it's so fun. I liked the one that was lifting the weights and then just kind of like shook and started to fall apart. Oh my gosh, I didn't even see that one. It is so interesting. This is actually the second year that they are hosting this. But it's much different this year. This year, they're beating world records. And this is a completely different animal. It's amazing how much has changed in robotics in just a year. I know. I'm kind of appreciative that they have this event because it is an industry-wide benchmark for us to understand how much of a step change we are seeing in robotics.

1:50Josh Machiz: Just for reference, just even among who participated, this year there were over 600 teams. Last year, about 280. And this year, 2 ,000 robots. Last year was 500. And then they also doubled the events from 26 events to 51 events. And actually, they also had not only sports, but also scenario-based contests such as factory assembly, hotel housekeeping, restaurant service, fire rescue even, EV charging. I thought that was really interesting. Are we that far behind China? Why don't we have our own Robot Olympics here in the U.S.? I mean, I think for sure someone should initiate hosting the Robot Olympics.

2:30The Lightspeed Robot Olympics?

2:31Josh Machiz: Yeah, maybe one day sponsored by skill, integrity, and which we'll eventually get to. But I think to answer your question, the logical answer is that, and I've spent a lot of time thinking about why, you know, China has so much of an emphasis on humanoids and robotics. And you actually see this in the numbers as well, where actually, apparently the first half of this year, global humanoid shipments hit about 19 ,000 units, which is up over 270%. But the majority of units are actually shipped into China, and the buyers are also majority China. So a lot of this excitement and activity among humanoids is actually in China.

3:13Josh Machiz: And it's partially because, one, they have the kind of financial infrastructure behind it. So there's a state-backed fund about trillion RMB, which is about 148 billion U.S. dollars. That is entirely dedicated to humanoids. I think the other part of it is that China, if you just think about their powerhouse industries, EVs, cell phones, drones, all these have given the core components, whether that's battery cells or power electronics, thermal systems, cameras, sensors, even drones. All of these are sort of building blocks that are foundational for a strong robotics economy. I think what was also notable is a couple of weeks ago, there was an announcement from the FCC that they are banning imported robots, though.

4:01Well, let's talk about some of the interesting things at the games. Obviously, there's the robot that beat Usain Bolt's record, which is incredible. But I think the interesting thing was the gate of the robot and that it's robots learning not necessarily to run the way humans run.

4:17Josh Machiz: Are you referencing the one that was like? Yes, where it's like it's like doing this like non-human run because it's like they're realizing that robots can figure out ways to run faster the way they would run rather than the way maybe we would run. Yeah, I haven't looked into that one as much, but I do think it begs a really interesting philosophical question. We spent so much time trying to mold robotics and robots into human form factors. When again, there are so many ways that maybe if you just like start from first principles, there would be other ways and like they self-develop a new running style.

4:53Josh Machiz: But yeah, I think one of these robots actually beat Usain Bolt's 9.58 second 100 meter record. Last year's winning time for reference was 21.50 seconds. So huge jump. Another one that actually also beat the Usain Bolt record was one called Lightning from Honor, which is actually a phone company in China, which is insane to think about. Because I'm just like, can you imagine a world where Apple is building humanoids that are beating Usain Bolt? And that they're all getting into it like it's F1, but for robotics. Yeah. And it's not just running. There's also high jump. There is a massive leap there too, from standing high jump going from 0.96 meters to 2.88 meters, and then long jump going from 1.25 meters to over seven meters.

5:40It does feel like we're moving very quickly from let's beat human records to let's beat other robotic records.

5:46Josh Machiz: I will say though, actually, for me, the most impressive part wasn't even running. I actually think that, yes, is from a demo standpoint, very cool to see. And there's just a lot of fanfare around that. But I actually think the scenarios were a lot more interesting technical challenges. You have these scenarios where a lot of robots can actually beat Usain Bolt, but they can't open a simple jar or they can't fold socks or they can't stack cardboard. So it's just interesting seeing this sort of jagged frontier where, yes, that is one of the benchmarks. But I actually think the more interesting challenges are some of the other ones.

6:23Josh Machiz: A couple that I thought were really interesting were some of these more contact-rich, uncertain situations. Like one of them was grasping beans with tweezers. And that is a really good test of force modulation and a system not just being able to grasp an item, but using a tool to grasp an item. So that's a really interesting benchmark. I think that's really different, that feels harder than just running. Robots always struggle with the lid. Exactly, because it requires bimanual coordination, meaning that you constantly have to be detecting two different states and your two different arms are doing very different jobs and they constantly have to react together.

7:03Josh Machiz: There's another one which actually researchers notoriously call the peg and hole insertion problem, which is connecting cables, which is also really difficult because as humans, a lot of our ability to connect two things is based off of force sensors and our ability for our hands to identify, you know, this is not aligning perfectly. You touch an edge or you feel resistance and you adapt accordingly. All of that you have to teach a robot to do because it's not a simple pickup and, you know, put somewhere else. It is like fitting something into something else and constantly adapting based off of the force that you're receiving.

7:41Josh Machiz: And so that requires a ton of precision and there's just like a very low margin of error. And then there was also, I think, another really interesting scenario, which was on assembling power tools. And what I thought that was really cool was a demonstration of something called long horizon sequencing and specifically the ability to recover. So this idea that you have a chain of events and if they succeed on the first step at 95 percent, then how does that subsequently lead to success of a full sequence? And I think recovery is so important because, you know, when we think about robotics, I'd actually rather a robot that has the ability to recover, to notice when it's failed, to diagnose it, to back out and retry than a robot that is just like only very good at one specific task.

8:26Josh Machiz: I think the recovery part is like where you get to some of this generalizability. And that's such a big unlock. So, yeah, I think as much as I loved seeing the funny memes around robots outperforming humans on a lot of these like sports functions, I actually think a lot of the more technically difficult challenges were some of these basic scenarios that humans take for granted every single day. Like we do all of these things on a daily basis. I know we forget how hard our mundane activities are for robots to accomplish. And so I totally agree with you. Those are the things that are just a little bit more interesting because we kind of know that they're going to beat us at a whole bunch of normal things like running.

9:04Josh Machiz: Yeah. And it actually helped me a little bit reframe that like it's cool to see a really fast robot, but it's way cooler to see a robot do something super mundane like twisting open a jar. All right. So from a competition in Beijing to the bell ringing in Shanghai, we had a humanoid company go public with Unitree. This is a really big deal because it feels like this is the household name in robotics in China. Claire, do you want to break it down for us? Yeah, Unitree's been around for a while. They're sort of been comped as like the Boston Dynamics of China. Yeah, they're Boston Dynamics of China.

9:40You can't buy Boston Dynamics in the mall, and I feel like you can.

9:44Josh Machiz: I actually think you can order a Unitree Humanoid on Amazon for like$1 ,000. It's really not that expensive and very impressive. Yeah. So I think that just spoke to the momentum behind the broader Chinese robotics market and the fact that you have a lot of market activity around the space. I think you are seeing a lot of supply and demand across thousands of companies focus on building humanoids now. So I think on the supply and a lot of these have historically been focused mostly on enterprise. I don't think it is true that you have your average person in China ordering a unitree for their home.

10:24Not everyone has a unitree in their home.

10:26Josh Machiz: Yeah. So I think largely the demand is really driven by more enterprise use cases or government use cases. I know the government is spending pretty heavily there. Well, I think for me, the interesting thing about this IPO is 10 years ago, this IPO would have happened in the U.S. This is the category-defining robotics company in China going public there and all of the value accruing there. What does that mean for American exceptionalism in robotics and this industry? I mean, I think it's a really good transition point to skill, which is in our portfolio and making huge strides here. But it's interesting.

11:07Josh Machiz: There is this great book by a guy called Dan Wang called Breakneck. And it really talks about why, you know, China has been able to scale a lot of the robotics industry and manufacturing. You know, to what I mentioned earlier, they have a lot of these foundational building blocks. A lot of the actual components of these robots are built in China. And so when you think about a fully formed robot, you know, a couple of weeks ago, I was able to go to dexterity and just you see this final end product in the form of humanoid when in reality, like so much goes into it, whether it's the actuators, exact grip, the amount of camera sensors, all of that.

11:43Josh Machiz: If you have a homegrown advantage in developing all of these components, it's going to lead you to a much more robust industry grown in China versus in the US. And so I think that's driving a lot of the excitement to reframe it in the U.S. to have some of these homegrown robotic companies so that you can build some sort of sovereignty around these parts and components, knowing that, you know, the physical embodied intelligence world is going to be so consequential. You know, we're now seeing so much activity in artificial intelligence in the form of software. But what happens when we move from these pixels to atoms?

12:17Josh Machiz: That's sort of the next big paradigm shift. The takeaway is, at least for me, that there's only going to be more excitement and hopefully more activity for some of these, you know, American-based robotics companies as well. So maybe moving to Skilled and the exciting announcement they had this week, they broke the timeline with their video where it has, you know, at present almost 3 million views on X. Talk to us about the S1 announcement with Skilled. What does this mean for robotics and why is it so different from what was available before? Yeah, they launched S1. It's a new AI model. And in short, the very basic version is that you show a robot one video of a person doing a task.

13:01Josh Machiz: So that can be potting a plant or making pancakes, and the robot is just able to do it. So it doesn't have any sort of programming, no training involved, no engineers were teleoperating it. And this is a really big deal because historically, anytime you needed to teach a robot a new task, it meant that you had to have humans collect 50 to 100 hours of puppeteering data or teleoperating this robot to do that task over and over and over again to collect training data. And then engineers would train a model, basically build the brain off of that task, and then turn all that weeks of work into a single task.

13:44Josh Machiz: So you can see the friction that that builds, right? Like, if you wanted to build a pancake flipping robot, you would have had to collect 50 to 100 hours of pancake flipping. Sure. And that robot would need to watch hundreds of hours of this instead of watching it once feels like a total unlock for pretty much any task. Yeah. And so now that is all entirely replaced with one video. And so actually a lot of people are comparing this to the chat GPT moment for robotics. Definitely feels like the chat GPT moment for robotics, certainly because, you know, if you think about the GPT moment, like in the early days, we were kind of spending so much time teaching AI.

14:26And then that flip switch where it was just all of a sudden very easy to prompt without having to do too much work. It feels like we're hitting there in robotics. So maybe to our earlier discussion, you know, skilled and some of these other cutting edge robotics companies in this country are helping us catch up.

14:43Josh Machiz: I think just to illustrate the step changes for, you know, anywhere from late 2015 to 2020, a lot of traditional machine learning and classifiers were this sort of BERT era of models. These were pre-Chat to BT language models. They could pretty much only focus on one specific task. So you would collect, again, a data set of hundreds of thousands of examples. You would run a training job. And then you would have to produce a new model with updated weights. It was sort of tedious. And there was a lot of value. Like AI was providing us as a society and as a technology so much value. But it wasn't generalizable.

15:24Josh Machiz: And so the ChatGPT moment was the fact that I could introduce an entirely new scenario without having taught it, here's how you write emails. or here's how you enter a customer service call. You just had given it a lot of data and it was able to build this kind of generalizable knowledge base. One big brain that is adaptable for many scenarios as opposed to building a highly specified brain. And we saw that with language models. This is sort of a similar breakthrough where now we won't have to just build the pancake-flipping robotic brain or the box-lifting brain. Rather, you just have one skilled brain in the form of S1 And you can apply that to many scenarios, except in the big difference here is instead of giving a prompt like write me this email, it's giving them a video.

16:10It's so amazing. Maybe just taking a step back for our audience who maybe don't understand how skilled works beyond this one incredible demo video. How does this skilled brain work? Like in the demo video, obviously they were showing one type of body, but it works across so many different bodies. does it mean that you know the skilled brain can control a whole bunch of different bodies at once

16:35Josh Machiz: how does the technology work yeah i'm not sure about multi embodied form coordination like i don't know if there's a scenario where one brain is just deployed to a swarm i'm not sure on that i'll have to look into it but i think one thing to make very clear here is that this is not teleoperation or mimicry. So this is not just you show it a video of it flipping a pancake and then it just mirrors it. For example, if you show a video of a watering can, but only a cup is available, it'll know to use the cup. If it drops something, it knows to retry, sort of the recovery thing, motion that we were talking about.

17:15How is it able to like adapt? Is it learning that that is a similar type of vessel? Like what's the mechanism?

17:23Josh Machiz: If I had to distill in very simple terms, in the same way that when I ask Chachi BT or Claude to write me an email specific to this one vertical or one workflow, I don't have to retrain it on hundreds of thousands of examples of that workflow. It is generalizable. So it's pulling from that same foundational model. That it has the context that all of these other things can also do that same thing. Yes, and so this process in robotics is called in-context learning. This is sort of similar where it is such a breakthrough because we are no longer in this paradigm of just like take a bunch of data points from an existing system and just learn to copy, but rather you've given it a brain and it can generalize to new situations.

18:07Josh Machiz: Actually, there were a couple of demos where even if the human in the video, in the prompt video was sloppy with their watering can, the robot actually might do it better. So I think that's like another sign, right? I think also for the pancake flipping example, they actually had checked their entire training data and pancake flipping wasn't in that data set. So this means that the robot essentially figured it out from seeing one video, tapped into the brain and then sort of figured out, okay, this is similar to some of the other flipping motions that I've done. How do I apply that to this specific workflow?

18:43Josh Machiz: In the same way that we see that for language now. What a breakthrough. What an exciting moment for robotics in this country. I know. It's just, yeah, I've been running around all day and yesterday and hyping it up to people that we've gotten this chat GPT moment for robotics. And it's exciting. I'm, you know, I feel very proud to be partners with the skilled team. They are an amazing group. And I think this will only just be to what we were discussing earlier around this, you know, this next frontier of physical and embodied AI. This is just such a huge unlock for that. Oh, for sure. I mean, look at the chat GPT moment and the way we've seen all of these things from that, like, develop so rapidly.

19:24It'll be interesting to see how quickly things move with physical AI.

19:29Josh Machiz: And I think it's only really, you know, compounding, right? Like the team has actually even they verbally said this was sort of like the scaling law for robotics, where historically in robotics, the bottleneck has always been data. You know, there's no Wikipedia of robotic actions available online. You know, for language, yes, we have the entire corpus of the Internet. But for robotic motions, someone has to go collect 50 to 100 hours of data. And so that's why robotics has never really kept pace with what we're seeing in software. And it's because the unit economics is so brutal and time consuming.

20:06Josh Machiz: Now you have a situation where with skilled and with this model, that cost of task acquisition can drop from many weeks of engineering to now. One video. Exactly. And now deploying a task is no longer I need to have hire a bunch of people to do this same motion over and over again. And instead, it takes 11 minutes for them to ingest the video. And then you can start to see how this flywheel goes, where they, through that, get more data. They are able to deploy vastly more tasks. That generates a lot more real-world data, which then improves the pre-training of the model, which then improves the in-context learning.

20:42Josh Machiz: So you really see how that flywheel goes. And I'm excited to see this all scale. Me too. I'm excited to have a skilled robot in the Lightworks studio making our guests pancakes for breakfast. Humanoid Olympics, sponsored by Lightworks, Skilled. Brought to you by Skilled. All right. Well, moving on to other kinds of interesting stories around data. Google is trying to buy the data that Spirit Airlines is leaving behind in their bankruptcy. Claire, what does this mean for the value of data? What does this mean to CFOs? Should they be valuing the data that they own as an asset on their balance sheet?

21:22Josh Machiz: I think that's a really good question. It's up for debate. I'm very curious what your view is, especially as you think about a lot of these like legacy businesses that have so much of this data. But what we're seeing here is quite literally that data is the new oil, especially in the case of Spirit Airlines. So what happened was, you know, as we know, a couple of months ago, Spirit Airlines, they're the ultra low cost carrier. They filed for bankruptcy twice and they finally recently halted all operations in May of this year. I myself have been subject to many budgets, spare airlines flights, especially in college.

21:59Josh Machiz: But now this business is being sold for parts in bankruptcy court. And in order to pay down roughly$8.1 billion in debt. As you would imagine, all the planes, the gates, their slots in the airports, all of that sold to other airlines. There are clear buyers for that. But there was one asset that was very different, which was the airline's entire digital estate. And Google went and has bid on it for about$10 million, beating out Mercor, which is another sort of AI data player whose bid was$7.5 million. And then I believe recently Mercor's rival, Micro One, came in after the auction closed. With a higher bid, right?

22:43Josh Machiz: Yeah. So this is all playing out. I'm not sure if, you know, how it'll close. There's also, I think, still some tension where existing employees haven't felt feel they haven't consented to all of their workplace communication, becoming trading data. So I think right now there still needs to be official judicial sign-off. Yeah, this feels like a very different moment, right? Because I feel like in bankruptcies in the past, the data was just kind of part of the package, but it was never an asset. And this moment in AI where training data is so important and you hear about the big frontier models buying out-of-print books and all of these different area lengths that they're going to to obtain data.

23:27Bankruptcy court is a great way to find it.

23:30Josh Machiz: Yeah. And to just illustrate why this data is so valuable. In this box or in this asset of Spirit's entire digital estate, you have about 100 million emails, 500 million Microsoft Teams messages, 17 million OneDrive files, 20.5 million SharePoint items, 30 million recorded customer service calls, 15 million support chat logs, 7.2 billion records of competitors' flight pricing, 7.5 billion passenger transaction records, and the list goes on endlessly. Just to make very clear, though, Google has said that when they buy this data, it will be rigorously scrubbed of all personal identifiable information by a third party before it receives anything.

24:19Josh Machiz: However, even after scrubbing it, there is so much data because this This is one of the most high quality data sets you can get. As an industry, we've sort of hit this like peak data problem where everything, as we mentioned earlier, on the Internet has sort of been scrubbed already, whether that's Yeah, everything that's out there has been taken. Exactly. X, Google, Wikipedia. And I think you're seeing that with video images as well. But I think the two data scarce resources right now is one, robotics data, because that's not very scalable and two workflow data and so if you think about the idea that we're building these agents to do knowledge work knowledge work lives in email threads it doesn't it isn't just raw text what is so valuable is understanding how decisions get made so the public internet there's no real examples of how you escalate a situation there's no real life examples of oh if you write an email this way, then someone will respond to this way.

25:25Josh Machiz: And, you know, we need to take this offline in this way. And then we resolve the disagreement in this way. All of that lives in a corporation's digital estate. And so, for example, just to make it very tangible, the fact that you have 30 million plus support calls and 15 million chat logs means you get visibility into the arc of solving a customer service problem. That is highly valuable if You are one of the many AI chatbot agents trying to take over the customer service industry. You understand how, you know, an agent might interact with different systems. You might understand how they deal with a policy exemption.

26:03Josh Machiz: You understand how they de-escalate the exact language they use. I think even just think about like the fare records that they have. That's really interesting because you basically get a longitudinal record of how airline markets have priced different flights. So how Delta might have responded when Spirit dropped a fare or how prices change in holidays. Like that is really interesting data as well for pricing, for forecasting. And then even just looking at transaction records, 18 years of how actual real people book, how far ahead, what do they typically get upsold on. And so that's all just really interesting stuff that like, yes, when you think about the box and you think about Spirit Airlines data, it's like, oh, cool.

26:46Josh Machiz: What can possibly be in there? But there is a lot of valuable stuff. And I would agree, like, I do think there is a sort of new era where data is an asset. And I think there needs to be new kind of guardrails around how we process this data. I can imagine if you are an employee, it is somehow. Yeah, you're not happy if your data is being sold about some old Teams conversations that you had. Yeah. And I think it is a little bit of a dynamic where does this mean that the model providers only get stronger because they have so much more workflow data? I'm curious what that kind of tension looks like.

27:23Josh Machiz: It makes sense that the data brokers are trying to play in this space because as we look, you know, laid out, this is very valuable data. But yeah, in terms of open questions for me, it's where do we go from here? And does this make, you know, existing powerful data players like Google even more powerful? And I'm curious to also see if there will be more examples of this now that Spirit has sort of been a precedent. And TBD on if it closes, yes, that'll, you know, we'll keep the audience posted on that if there's more needs to come. But yeah, I think it's just interesting entering this world where data is just a raw asset.

28:02Yeah, either way, the acquirer is going to get a lot of information about how mad people are about having to pay for a bag.

28:09Josh Machiz: Do you think that C-suite executives are thinking a lot about how they're positioning this data? I think so. I think that many CFOs should start to value their human data, like all of the data they've collected as an asset on their balance sheets. They should also be protecting it in a bankruptcy, right? Like, clearly, that's why we're in this situation where this data is for sale. Like, if you're doing well, shouldn't you be protecting the potential that your future data would be sold in a sale of any kind? Yeah. I mean, I think it goes back to the Satya piece that we were talking about, this idea that data is everything.

28:52Josh Machiz: So, yeah, kind of full circle. So this week, it was all about separating intelligence from all these different things, from the body with skilled AI being able to put its intelligence into all these different types of bodies and teach them to learn skills, from removing the data from companies in bankruptcy, and then selling it for intelligence to train new models. Claire, what was your takeaway? way. Yeah, even though they were very different stories, I feel like they all sort of flowed into each other, especially on this theme of physical intelligence and embodied intelligence. You just see so clearly how it is, yes, the brain, but it's about the data that underpins that brain and makes that brain powerful.

29:40Josh Machiz: But it's also about putting that brain in a physical form where you need very physical industrialized foundations to support an entire humanoid ecosystem. So yeah, my takeaway is that things are happening in different parts. And this is sort of just the ways that we see the scaling law unlock is by seeing all the advancements across data, model, and embodied form. Yeah, I'm excited to see that skilled S1 brain next year in the Robot Olympics. well that's it for all of us on light work we'll see you next week where we continue to chronicle the chat gpt moment for robotics and so much more