And I'm Caleb Bissinger. And this is the next big idea today. Are we approaching the ChatGPT moment for robots? And are we ready for the world they will create?
Imagine sitting down with your great-grandparents when they were your age and explaining the world we live in today. You'd have to explain that nearly two-thirds of Americans work entirely indoors. Some of our largest professions include things like home health aide and fast food worker. And millions of people now make money online as creators.
Yeah, this might require some explanation.
Yeah. Can you imagine? Like, well, great-grandpa, sometimes they film themselves opening packages so that strangers know which kind of moisturizer to buy. And then you would show them your smartphone, a small rectangle of glass and metal containing a camera, a map of the entire planet, a library, a television studio, a bank, a dating service, and nearly every song ever recorded. And not just that, it can translate languages, summon a car, deliver dinner, connect you face to face with someone on the other side of the world.
Now, reverse this thought experiment. Imagine meeting your great-grandchildren. How would they explain their world to us? They'd have a lot of questions, right? Like, wait, you used to clean your own toilets? Human beings worked inside of factories? Also, you drove the cars yourselves? That seems irresponsible.
Their world might contain millions or perhaps billions of robots. Some will look like people. Others will be arms, wheels, drones, or small machines designed to perform one unpleasant task extremely well. These machines could clean our homes, tend our gardens, build houses, assist surgeons, discover new medicines, and help older people remain independent.
Which raises a few questions, like can today's clumsy, often teleoperated robots become autonomous as quickly as large language models did? Is the human body actually the best design for a robot? Or are humanoids an expensive solution to the wrong problem? And what happens when robots begin building other robots, driving down the cost of labor and perhaps the cost of almost everything?
Yeah. Does that create spectacular abundance or a world in which a small group owns the machines and everyone else is just searching for work, for income, for purpose?
Our guest today is exceptionally well positioned to answer those questions. Stephen Witt is the author of The Thinking Machine. It's the definitive account of Jensen Huang and the rise of Nvidia, and it won the Financial Times Business Book of the Year award in 2025, and we had Stephen on the show to talk about that. It was a great episode. Love that conversation. He's also a frequent contributor to The New Yorker, where he recently published an article about the latest generation of humanoid robots. He spent months visiting labs, interviewing believers and skeptics, and he met Neo, a soft, quiet, elegantly dressed home robot that inspired in him an embarrassing emotion.
So, are we watching another group of Silicon Valley founders overpromise and underdeliver?
Or are we witnessing the awkward first steps of a new species of embodied intelligence?
Stephen Witt, welcome back to the next big idea.
Thank you. It's great to be here again.
Stephen, you've been reporting on Silicon Valley for a fairly long time. You write, "This is another way of saying that I've been served a lifetime supply of bullshit." And yet, you arrived at 1X Labs ready to call on Neo, their soft, elegant, and almost silent humanoid robot, and you felt consumed by what you call consumer lust. Can you tell us about this experience?
It's when you see the iPhone, when you see a Ferrari, when you see these extraordinary design objects and you say to yourself, I just have to have one. It might not even work that well. And in fact, it doesn't work that well, but the design is so superb and you know it will improve, right? Its brain is software and its brain is constantly being updated and getting better. So, what you want is the physical body and you'll wait for it to get smarter. You know, I was among those people who stood in line to buy an iPhone, the number one iPhone, one the day it came out, and it didn't matter that it, in fact, actually it worked great. I mean, compared to a modern iPhone, it was junk, but compared to what was on the market at the time, I was really impressed by it. And I feel like that's sort of happening with Neo and a few other humanoid robots now. These are the first that can really ship to your home and really at least attempt to respond to voice commands. We haven't seen anything like that.
You're not the only one with consumer lust. I think you report that there are 10,000 people who've put down $20,000 or at least signed up for a monthly installment plan to buy the Neo, which they're saying is supposed to come out by the end of this year. Am I getting that right?
I mean, they are sticking with that. The products that I saw on my site visit to 1X's headquarters in Silicon Valley didn't look ready to ship. Having said that, they did look extraordinary. I mean, they're just so well-designed. I saw a lot of robots in reporting this article. I probably saw about 12 or 13. And most of them just were junk. You know, they looked junky. They sounded junky. They were loud. They were clunky. They clanked when they walked. They were awkward. And they looked a little scary often. I mean, they just look like toys, you know? They look like a toy that you would buy at Target blown up to the size of a humanoid. Neo didn't look like that. It looked like something else entirely.
Yeah. Why don't you describe the design? To me, it's sort of like the Mr. Rogers of humanoid robots.
Very much so. And actually, that's very much by design. The designer, who is 28 years old, Dar Sleeper, previously worked at Tesla and before that at the fashion brand Yeezy. So his first two bosses in order were Kanye West and Elon Musk. So interesting work history. And he might be the most handsome guy in Silicon Valley. He's like 6'2. He played D1 lacrosse at the University of Michigan. As I say in the article, he puts the bro in tech bro. He's like a very kind of like fraternity guy. He's very much kind of like a rap music guy. He's like perfectly proportioned. So much so that he actually used his own body and his own gait as the model for the anatomical movements of the robot. He put on a motion capture suit and then taught this robot how to walk.
I love too, by the way, that in that he tells you he's like, "Well, I have anatomically perfect running form." Like, he knows he knows that he's the hottest man in Silicon Valley and yeah, those words.
No one has ever accused me of having perfect running form.
I don't think I have an anatomically perfect run. In fact, I know I don't.
So he's the designer also, and he's really ambitious and creative, and he's got a great sense of design. I mean, this is the thing. So Bernt Børnich, the company CEO, basically understood that he had this kind of like young rockstar designer, and he just put him in charge. I mean, he really, this guy's 28 years old. He was 26 when he started. He just promoted him to basically being second in command of this entire company. And the task was to create a robot that didn't scare children. That was their design philosophy. If we can build this thing and little kids aren't scared of it, we'll know we have done our jobs. So, your analogy with Mr. Rogers is actually spot on. That is exactly what they were trying to do. And the first iterations, I think Dar said, you know, anyone under the age of 100 was scared of this robot. I saw it. It was kind of creepy looking, you know. And the problem was, for whatever reason, sci-fi influence or maybe just a lack of creativity, basically all of the humanoid robot designers had converged on this smooth black face plate that was just a little eerie. It was cold. It didn't have an uncanny valley effect. It didn't look like a human, but it did look sinister.
And I think Dar realized, okay, it's the face. It's the face that's scaring everyone. And it's these metal bodies as well. So, what he very creatively did is kept experimenting with the face until he'd removed almost all the features except for two cute black eyes. They look like teddy bear eyes. And then, this is, I think, really the stroke of genius. He realized he was never going to make a body that was not scary in some way. And so what you needed to do, and I think this was kind of a stroke of brilliance, was actually put clothing on the robot, put clothing on the humanoid. So he designed this very calm, neutral, one-piece bodysuit out of soft textile fabric. And now the robot, it really looks like a stuffed animal come to life in a way that's not creepy, though. Somehow he avoided that uncanny valley effect. He avoided the kind of cold, sinister affect of the machine. He got it just right. And when you see this thing, it's just so compelling.
The other thing that they really figured out was, you know, this robot, the AI, is just not quite there, right? Probably when you bring it home, sooner or later, it's going to fall over at least once. And maybe it'll be able to get back up, but it's going to fall down. And so, you had to make the robot as lightweight and as quiet as possible. This thing weighs 66 pounds. I could pick this robot up in a bridal carry and carry it around. So, because it has these soft textiles on it and because it's lightweight, when inevitably it does fall down or bump into something, it's not going to create that much damage.
The final thing they think they did that was, I think, pretty brilliant, they realized that, you know, a typical robot will have 200 to 300 of what are called actuators or basically just motors that create movement. Each of those actuators is a point where the robot can fail and something that makes noise. So they stripped down the number of actuators to a very much smaller number and then really in kind of a brilliant approach to biomimicry, they took tendons. The tendon in the human body is the thing that attaches the muscle to the bone. In this case in the robot, it's the thing that attaches, it's a little cord they put in there that attaches the actuators to the joints. And in this way they were able to make this really quiet robot. So, in summary, the robot is quiet, it's soft, it's lightweight, and it's non-threatening. It feels like it should work in like a high-end fertility clinic. It feels like something that is totally safe. It could just serve you espresso and you wouldn't look up from your phone. I mean, it just is such a masterpiece of design. So much so that it actually tricks you a little because you expect it to do things and actually it can't do all the things that you would expect.
I've started a few companies over the years, most recently The Next Big Idea Club, which puts out this podcast. And here's one thing I've learned. The glamorous part is generating the idea. The less glamorous part is setting up all the machinery that turns the idea into a real business. But that stuff is crucial, and that's where Northwest Registered Agent comes in. They don't just file your LLC. They give you a complete business identity from day one: a business address, operating agreement, domain, website, professional email, phone number, and built-in privacy without making you cobble together half a dozen different vendors. They've been helping entrepreneurs for nearly 30 years. And their corporate guides are real people who understand local laws. They also offer thousands of free guides and attorney-drafted documents with no upsells and no selling your data. Don't pay hundreds or thousands of dollars for what you can get from Northwest for free. Visit northwestregisteredagent.com/nbifree and start using free resources to build something amazing. Get more with Northwest Registered Agent at northwestregisteredagent.com/nbifree.
If you're in meetings all day, you know it can be a struggle, right? You're nodding along, you're contributing, you're trying to stay present, but the whole time you're like, "Wait, wait, wait. What did that person just say? Oh my gosh, wait. Who's following up on that? Am I in charge?" You get to the end of the day and you realize meetings are a mess. Granola fixes that. Here's how it works. You take rough notes just like you normally would, but in the background, Granola securely transcribes the meeting and then it turns everything into a clean, structured, actually useful notes doc when the meeting ends. And the best part is that it works through your device's audio, which means you can use it without any setup and there's none of those awkward bots lurking on the video calls. I used to have a stack of legal pads on my desk with illegible meeting notes I'd made over the weeks and months. Now with Granola, everything is always organized and any answer I might need is at my fingertips. If meetings are eating up your day, Granola is a no-brainer. You can try it totally free for 3 months. Just head to granola.ai/idea. That's granola.ai/idea to get your time back. Get 3 months free at granola.ai/idea.
And so was it the form more than the functionality that created this consumer lust in you?
Absolutely. And I mean part of 1X's gamble is we're going to ship this form. We know the AI is not quite there. We know that the AI is going to improve over time. You know, this thing is going to get smarter. It's going to become more nimble, less clumsy, more capable over time. And so you will almost rather than making a purchase, it'll feel almost more like an adoption. You're going to adopt a puppy, you're going to adopt a child, you're going to adopt this robot, and you're going to watch it grow and learn with you, and then you're going to feel a tremendous sense of attachment to it. So, I think that was pretty brilliant, actually. I think their business plan is clever. They're way out on a limb. None of the other humanoid manufacturers basically are willing to put a robot in the home right now because they know the AI is not there. And some of the roboticists I talked to thought that they still think that 1X is nuts, that these robots are going to fall down, that they're going to create accidents and injuries, and that 1X is going to be sued out of existence. That may happen. I don't know.
Part of Bernt's gamble, and he's the CEO of 1X, is that robots basically have a hive mind. So when one robot learns something, every other copy of that robot learns the same thing at once. It uploads that data to the cloud and then later overnight its brain gets updated with this new information. So they have a hive mind that's updated once every 24 hours or so. The play here is to put as many, therefore to put as many robots, to deploy as many robots as you can into useful real-world settings. They'll all gather data and they'll all teach each other new things, and so there's a huge first-mover advantage here from a business perspective.
And they're currently largely teleoperated, I understand, and this can either be interpreted as kind of fraudulent, like there's, you know, like a lot of the most impressive things we see in the videos, which are all over YouTube, are actually the result of controls through teleoperators, but this is also a process of gathering data, building the necessary data, because there's an absence of the adequate physical data that is necessary in the real world.
That's precisely right. If you look at ChatGPT and the amount of data that it took to train this large language model, you know, I think in terms of real-world trajectories of joint action data, there's no such database. ChatGPT had the entire internet of text to train on as well as a bunch of cracked ebooks that they pirated. There's nothing like that for movement. We don't have that. It doesn't exist. And so they have to build that data set. That's actually probably the biggest sticking point.
Now, one way to do that is to put people in motion capture suits. And the problem there is that even if you put the entire population of Earth in motion capture suits and had them go about their day for 10 years, you still wouldn't generate as much data as was used to train ChatGPT. So this data bottleneck is a big problem. In the interim, they're filling it via teleoperation, which is basically just turning the robot into a marionette. So, we have someone remotely, probably wearing a VR visor, holding VR controllers, and I've actually done this myself, and kind of moving the robot around, commanding it to move this way and that way, moving its arms, swiveling its head, seeing what the robot sees, and acting essentially as a remote operator of the robot.
Now, when 1X does this, it has these little earpieces that light up in a special color, so you will know instantly when it's being teleoperated. Mr. Rogers' ears are glowing green, so clearly someone in a call center in the Philippines is watching us right now.
Remotely puppeted by a teleoperator in the Philippines.
Look, Waymo does this too. Yeah, you know, when your car encounters a situation that it doesn't know what to do, your Waymo, it throws it to a teleoperator probably in the Philippines or in Asia who sort of assists in getting you out of these tricky situations.
Would you be willing to invite a robot into your living room that is a peeping tom for teleoperators? Would that be a blocker for you or not necessarily?
You would think it should be, but I think for now the kind of people who are buying a 1X robot are totally cool with teleoperation. The guy that I talked to is like, "Yeah, I'll know when it's being teleoperated. What I want is my robot to clean my house while I'm not at home. I don't want to be there. I want it to be around trying to clean my house while I'm at work." So, I come home to a clean house. And this is kind of the dream, right? You trash your house overnight. You leave. You go to your job or you go out with friends and you come back and your house is spotless. You don't have to do anything. So, in some ways, you don't even want to see the robot operate. It's in the background. I think that's sort of the long-term play.
And then there's fun. Let's say we want to have our robot bartend our party, right? Maybe you do want to tell the operator to do that. It'd be kind of fun.
Little dance party. All that stuff. Yeah.
Warren Buffett famously said that when it comes to investing, risk comes from not knowing what you're doing. I hate to say it, but I think that makes me a high-risk investor. Wealthfront helps you tune out the noise and earn even more on every dollar with a high-yield cash account and sophisticated, easy-to-use investing products. With a Wealthfront cash account, your cash can earn up to a 4.3% APY for a limited time, plus an array of fee-free checking features like no-fee instant withdrawals to an eligible account and up to $8 million in FDIC insurance through program banks. And you can easily transfer your funds to Wealthfront's expert-built investing products the moment you're ready. Here's how you can unlock a 4.3% variable APY for a limited time. You first start with a 3.3% base APY from Program Banks. New cash accounts can get an exclusive 0.75% APY boost in your first three months on up to $150,000 balance. Then you can add another 0.25% APY increase, no expiration date or balance limit, just by enabling a $1,000 a month direct deposit and maintaining a funded investment account. I'm a Wealthfront fan for real. But don't just take my word for it. Join the million-plus people already building long-term wealth and confidence. Start earning up to a 4.3% variable APY today by heading to wealthfront.com/nbi. Terms and conditions apply. This has been a paid testimonial from a current client of Wealthfront, which creates a conflict of interest. Client experiences will vary. Wealthfront brokerage is not a bank. The base APY is as of January 30th, 2026 and subject to change. For more information, please see the episode description.
When it comes to supplements, I'm skeptical. I think most nutrients are best received through a healthy diet. But creatine is among the exceptions. Several neuroscientists and health experts I've interviewed have convinced me that it supports not only strength and muscle recovery, but critically it improves cognitive performance. I've been taking creatine daily for more than a year now. But when you put something in your body every day, quality matters. That's why I use Momentous Signature Spec Creatine. It's manufactured in a pharmaceutical facility using reverse osmosis purified water and verified through an expanded six-stage testing process. It has 10 times fewer impurities, tighter limits on heavy metals, and screening for PFAS and microplastics. It's also an ultrafine powder, so it mixes easily. No grit, no clumping. That's important in my experience. Creatine has become a simple part of my daily routine, and I appreciate knowing that Momentous has gone to unusual lengths to make it as clean and carefully tested as possible. If you want to try Momentous Signature Spec Creatine, head to livemomentous.com and use code IDEA for up to 35% off your entire first order. That's livemomentous.com, promo code IDEA, for up to 35% off. livemomentous.com promo code IDEA.
In prep for this conversation, I was reading a really interesting essay by this guy named Rodney Brooks. I don't know if you know him, Stephen, but...
I haven't heard this one. Go on.
He's like kind of one of the OGs of robotics. So, he ran the MIT robotics lab that created the algorithms that make today's self-driving cars possible. That sort of created the first interface between AI and machines. He co-created the Roomba. So he's sort of a legend in the field. And his problem with the current training approaches is that we don't have a good way to translate tactile data into data that a machine can then learn from. That like our fingers are insanely powerful tools. They can gather all this information through touch about sensation and feel and texture and weight. And it's very hard to translate that tactile data, that sense of touch, into something that a machine could then ingest. I wonder what you think about that sort of tactile dexterity argument as like one of the big blockers here for why these robots may be further off than companies like 1X claim.
Well, it depends on what you want the robot to do. Certainly, he's right. We're years away from a robot that can fret a guitar and shuffle a deck of cards or tie your shoes. Years away, years away. And it's exactly for the reason he points out, we just don't gather tactile data. It's hard to gather, right? The robots that we're shipping now, they have 27 degrees of freedom in the finger, but they're not gathering that super fine data about touch. I should point out, if you give a child a pack of cards, they won't be able to shuffle it, right? Or fret a guitar or tie their shoes. Those are all learned behaviors that often take a long time to learn. So what we think of as our finger dexterity immediately are in fact learned skills that take sometimes years. So in some sense this objection can also be made about humans, right?
But I think he is right that a database of just pure tactile information I don't think exists right now, and that is going to limit what these robots can or can't do. But think about what you can do even basically with just even a single claw gripper, right? I mean, this was a point somebody made to me because I had made a similar objection. They're like, "Actually, a claw gripper can do quite a lot." You know, a little pincer, you can fold clothes, you can do laundry with a pincer. You can clean stuff off with a pincer. You don't need this super fine hand to do that. You know, the hand is one of the pinnacles of human evolution. It's underrated and it's going to take the robots a little bit of time to catch up. That doesn't mean it can't happen.
I don't think there's anything fundamentally blocking it. I mean, sure, it's not there yet, right? It's true, but it's kind of like objecting to, well, my car can't go up a flight of stairs. Okay, sure. You know, you can do a lot of other stuff. You know what I mean?
Well, and there's the physical dexterity, as you say, but maybe the more critical part here is the information processing and the brains. And as we've been talking about, there's this absence of the adequate data about how to move through the physical world. We're building it. There's kind of a race to get to this kind of critical mass of data, which is maybe one of the reasons that Neo 1X is moving so quickly. But there are these different sources, right? We talked about the first-person human video. There's also simulation, and then there's the beginnings of real-world robot data, right?
I actually, my only regret about this article is not publishing a little more on simulation. I actually had a lot in there and they took it out. I should have fought harder to keep it in.
So basically what we do is, let's say we want to teach our robot to do some household task. Let's say, let's say fold a shirt. Okay. And what we're going to do here is we're going to put the robot's brain in a digital gymnasium and then we're going to have it fold shirts in computer simulation time 10 billion times. Right? It would take 400 lifetimes to do that in the meat world, but in the digital world, we can do it in a couple days, right? Meanwhile, as we're doing this, we're changing variables, changing parameters within the simulator, right? So, we're altering the kind of shirt we're doing. We're altering the temperature in there. We're altering the touch sensitivity. We're even altering sometimes things like gravity. And we're just doing that to give the robot a resiliency to this task, the ability to do this task under a variety of different environmental conditions. Once we're done, actually the robot's usually perfect at it in the simulator, and then we have to download it into a real-world brain and deploy it to fold these shirts.
Now, how good is the simulator versus reality? That's what roboticists call the sim-to-real gap. And closing that gap is a big focus of robotics right now. They're getting better at it. But as one person pointed out that I talked to, he's like, "Yeah, you know, we're great at getting the robot to do it in a simulator under all sorts of different temperatures, for example, and we're great at getting it to do it under all sorts of different fabrics. But when we combine the two, then suddenly the robot doesn't do it. So we can master any single parameter really well. But when we get into this more complex world of multiple, sometimes interacting parameters, right? Then suddenly it gets a lot harder to fold the shirt." Shirt folding, though, robots have got there. Think about the complexity of say washing a dish.
Now think of all the complexity, right? There's water, there's soap, the dish may be made of china or maybe it's made of plastic. How hard do I have to grip? If I grip it too light, it's going to fall and break. If I grip it too hard, I might shatter it. It's a tough, tough, tough problem. And again, as you point out, and I guess as Rodney points out, the tactile data just isn't there yet. I don't think it's something that can't be conquered. I just think it's something that we're just starting to build the data set for.
One thing I wonder about, Stephen, is whether or not we are right now hitting a kind of ChatGPT inflection point with robotics, which I think that your good friend Jensen Huang made a statement to this effect, I think, in January, right? That we're hitting an inflection point with robotics. Now, on the one hand, I think some would say, "No, no, no, no, no. This is teleoperated. This is a bunch of overhyped videos." On the other hand, Google DeepMind says, "Gemini Robotics can generalize across tasks and robot bodies, now plan multi-step actions and use digital tools." Figure is reporting that Helix O2 can control an entire humanoid through one neural system. If it's real advances in the software layer, it can complete a continuous four-minute dishwasher task autonomously. So what do you think? Is this a big moment in the sort of realization of the dream of robotics or is this just a sort of trailer precursor to what's coming five years from now?
I see evidence for both sides of this argument. I remember the first time I used ChatGPT and I was just blown away by it. I was just like, "Wow, this is stunning. How is this possible?" I didn't see anything like that in the robotics world. I saw some stuff I was like, "Oh, that's kind of cool." Kind of like proof of concept, but you know, they are now where maybe GPT-2 was, where if you squint, you could see it turning into something, but remember they went from GPT-2 to ChatGPT within the span of 14 months or whatever it was.
That's right. So, and not only that, the same people are now working on this, like literally the same people who designed kind of these large language models, the same crush of talent and money, and not only that, but now they have the AI to assist them to actually make it go faster when they learn and program these things. You know, I think the fundamental question is GPT is a transformer-based system, which means we can just, it's this funnel that we can just feed endless amounts of data into and it scales perfectly and it gets better. Is the transformer system enough given the relative paucity of data versus what we had in large language systems to actually build this practical, omnibus, physical, real-world intelligence, or do we need some other approach?
I would say probably if I had to guess, if you made me like stake my life on it or something, I think actually the transformer is enough. I think we will get there just with the transformer simply because the amount of computer power that we have available to us now is like already 100x what it was in 2022, 2023. So that's scaling rapidly and I think they're coming up with clever ways to generate this data that they need. I think that they actually will conquer this problem with the transformer. I have not seen ChatGPT for a robot. I haven't seen it yet. I haven't seen that thing that absolutely blows me away. But if you go and you read what the critics and skeptics were saying about large language models in early to mid 2022, you could find established tenured AI professors with decades of experience being like, ah, we're 10 years away from a general large language model, and we were two months away. So I think that's more likely to happen again than a stall.
It seems to me that there are sort of two sides of the challenge here. There's the engineering challenge of designing robots, but the characteristics for which you feel consumer lust, right? And it seems, you know, range of motion, battery life, form factor, and it seems like we're getting closer there. And then there's the software learning challenge of developing autonomous skill sets. And when I look at, on the one hand, you know, when I hear that we now have a robot doing a four-minute continuous dishwasher task, on the one hand, I think, well, that's a really long way from having an autonomous robot in my home that can do all the things I want it to do. On the other hand, you look at the last two years of progress with Claude doing autonomous coding. We had in March 2024, Claude 3 Opus could execute a coding task that would take a human four minutes with a 50% success rate. Fast forward just shy of two years later, February 2026, Claude Opus 4.6 could execute a task that would take a human 12 hours with the same level of reliability. 180x improvement of autonomous agentic coding in just under two years. Right? So the trend lines that we've seen are exponential, not linear. There's certainly a case that like we have line of sight, like these are fundamentally, these tactile information processing challenges are solvable. It's fundamentally the kind of problem that can be solved with reinforcement learning and we have all the tools. It's all there. We have line of sight to whether it's two, three, four or five years. The implications of having humanoid robots at scale are just mind-blowing.
Yeah. And we now, I mean, so this would be the argument that this is, we are hitting this kind of ChatGPT inflection point as... I think we're, yeah, I think we're, you know, we're early 2022. I'll be disappointed if I don't see something mind-blowing in the next year.
What's your threshold for mind-blowing? Like what would constitute mind-blowing to you?
Jim Fan talked of the robotics Turing test. And what this is, he's a research scientist at Nvidia. And what this is, is you throw a party. You throw a crazy party. Your new friends get trashed. You trash the place. And then you leave and you come back six hours later and the place is spotless. And the Turing test is, did a cleaning crew come in here or did a robot do it? This is the robotics Turing test, and right now we haven't passed it. If you ask a robot to clean up after a party, you will be able to tell that the robot did it and not a human, right? It will be very obvious to you because things will be placed in funny places. Probably won't be that clean, to be honest with you. It's not going to be able to mop the floor, etc. Right? So, that would be mind-blowing. ChatGPT passed that Turing test, the original Turing test. Right? You know, you configure it, right? And it's indistinguishable from talking. The only way you can tell you're not talking to a human is because it's smarter than a human, right? It's faster. You know, a human could never generate intelligent prose this quickly. Even the best writer.
Look, in some ways, we're already there, right? I mean, Waymo is amazing. Waymo is incredible. If you take a Waymo, it's like, "Oh my god, I'm never going to drive a car again." That test has already been passed. It's indistinguishable to me if a Waymo is driving the car or if a human teleoperator is driving the car. I can't tell the difference from the AI, right? But that's an easier problem in a lot of ways. As somebody pointed out to me, the car just never has to touch anything and it's a success.
The robot has to pick things up and manipulate them to be successful. And we're, as you point out, we're a little bit away from that. But I expect to be blown away by something soon.
You talk about the transformer inflection point, right, which was critical for AI. And as I understand it, this recent shift towards robots operating with foundation models at their core, like large vision-language models applying semantic understanding and planning, while at the same time working with these faster vision-language-action models and control systems to control physical movement. This seems like a pretty meaningful change.
Yeah, I think actually it's a good point. We are getting away a little bit from the pure transformer model. You know, I talked to Karol Hausman, the head of Google Robotics. I was like, well, you know, are we using transformers or not? He's like, basically, yes. Because what we're doing is we're taking the language that we ask it, which is the English or whatever language we're speaking to the robot, and we're translating that into a language, quote unquote, of joint trajectory actions. But I think beneath the hood there, there's now this more kind of mixture-of-experts approach where we're sort of creating like a model of what the scene looks like in the robot's brains. It seems like we do this as humans. It's kind of hard to tell, but it seems like we do have some kind of fundamental intuition about what is where in the scene that is probably in the reptile brain. It's quite an ancient part of the brain, right? And you know, if you think about how neural nets were developed, it was almost the opposite, right? We started with the outer layers of intelligence, like the most human parts of intelligence, and are now moving backwards into the more kind of like physical and lizard-brain parts of it.
These parallels between human evolution and robot evolution to me are totally fascinating, right? That when you think about, as I understand it, and you can correct me or improve on this description, the emerging standard is robots controlling balance and like immediate actions with edge-localized compute that physically carry, and then in parallel being operated with LLM, large foundation models that can help with broader comprehension that you're interacting with when you speak with them, and that can help them plan and execute larger intentions, which is precisely how humans evolved, right? On the one hand, we have these, you know, fast sensory-motor control systems that are controlled by our cerebellum that work with brainstem. We have these proprioceptive reflexes you're referring to, knowing where our body is, knowing where...
Those came first, right? All animals, all mammals have those.
Right. Right. And then we evolved this prefrontal cortex with this incredible capacity, higher-order cognitive system.
