The real danger is the gap BETWEEN two robots | E29

25:28Gautam Narang:So yes, everything is local. And obviously there are layers of, you know, cybersecurity layers that allows us to ensure that no one can get unauthorized access to the system. Now, we use different kind of compute for training our models that actually run the trucks, run our simulation platform, our data augmentation pipe. And so all of that is a different kind of compute cluster. So in my opinion, long-term, it's going to be multi - and distributed compute systems, depending on like, you know, what kind of tasks. Like, and obviously ROI considerations will come into equation as well. Like, you know, if you're talking about the open source models, then yes, like, you know, what is the task I'm trying to achieve?

26:06Gautam Narang:And then like, you know, what is my... Where are you at with these trucks? Are they on the road? Supervised, I assume, for now? Yeah. So we are like, you know, in the trucking space, one of the only companies that is doing true driverless operation on public roads, right? So this is something that we announced with PepsiCo. So Pepsi has been a long-term partner. They have made multi-year commitments. So we have fully driverless trucks. Wait, hold on a second. Look at this jacket you're wearing. Hold on a second. Back that up for a second. You got a stylist. They looked at Jensen and they got you a Prada jacket.

26:41Look at that. Is that Tom Ford a Prada? What is that? You didn't pick that. Come on.

26:45Gautam Narang:I've been wearing this longer than Jensen. Wow. Really? What is that? What is that? That's not members only. Is that Prada? No, no, this is not Prada. This is, which one? John Verbatos, right? John Verbatos. Oh, nice. Yeah, that's very nicely done. You look great. Look at this manscaping going on here. It looks fantastic. So you're on the open road. Open road, public roads. Okay. Across highways and surface streets. What cities are you in? Where are you doing this? Because explain to us the regulatory framework of getting on the road. Because Cyber Cab just started. and then, you know, California is becoming a bit of a bear to get your stuff out.

27:27Where are you doing this?

27:28Gautam Narang:We're live across Texas, Arkansas, and Phoenix, Arizona, right? So within Texas, currently in Dallas, obviously new markets will be expanding too soon as well. And then from regulatory standpoint, all the regulations are at the state level, right? So we're talking about 29 states where we can deploy the trucks, pull the driver out and commercialize the service. And, you know. Amazing. And are you going to do, are you looking to build the entire network out and sell directly or are you going to provide to people how does it work what's the business model behind this today because obviously tesla's doing semi uber has uber freight and crazily tesla's a partner with uber freight for semi but they're not for cyber cab kind of weird distinction but okay um tell us about like what the business model is here?

28:18Gautam Narang:So we believe in going direct, meaning like, you know, we don't want to sell into a middleman and have someone else own the customer relationship. So we are the ones that work directly with Pepsi, Kroger, and some of the other customers. And there's a lot of like, you know, long-term stickiness and defensibility that we build because of that model, right? So we are designing these regional logistics network from the ground up to meet our customer needs. And obviously it's optimized for our autonomy capabilities. And we are building these networks across five markets today. The plan is to continue densifying each of these markets and then adding new markets as well.

28:52Gautam Narang:So going direct to the customers, providing autonomous transportation as a service, that's the business model. But then we also do not have to worry about owning the assets, right? So we had an asset-like model. We have leasing partners. They own the trucks. They lease it to us. And then we provide a service to our customers. As revenue comes in, we make payments on the lease. Gautam, I'm curious, do you run... So you have your own model that's running on the truck itself. Is that an end-to-end model? Or is it kind of like a... Yeah, it's an end-to-end trainable model. We have two main models.

29:24Gautam Narang:The first one is what we call scene representation. So that's the one that is responsible for understanding and predicting the environment. The output of that flows into our second main model. We call that the reasoning model. That's basically where the reasoning and the behavior planning and the decision-making happens. So two models. But then all the like, and I'm tying it back to the discussion that we were having, authorization Ultimately, are the models responsible for deciding a safety critical action or not? We don't give the models the authorization. So we have a parallel safety layer that is verifying the output from each of the models so that ultimately the actions of the truck are fully deterministic.

30:01Gautam Narang:And we cannot afford to have wrong outputs and wrong decisions, right? So we're talking about driverless trucks. And then how do you, for your data flywheel, How do you, if the truck encounters some issue or makes a mistake, how do you kind of correct it or give it information about that adversarial example so it can prove and get better and not make the mistake again? Such a good point. And that's why if you think about physical AI, I'm of the belief that deploying in real world environments as soon as you can, that's the right approach. Obviously, before that, you can collect all the data in the world that you want to.

30:36Gautam Narang:It can be synthetic data. It can be real data. but the real learning happens in the field. So for us, that flywheel is, obviously we rely on a lot of real world data, but use high fidelity synthetic datasets as well. And together, the combination is what allows us to ensure that the whole system is safe and we put out on the roads. Got it. Yeah, it's pretty amazing. I have a Tesla. I just got one because I had kids and needed a bigger car. And I'm using FSD all the time. And for the first week I had the Tesla, I would, you know, be driving home from work or it would be driving me home from work and it would turn into the my neighbor's driveway and I would stop it, correct it, drive into my driveway, tell, you know, give audio feedback that it's wrong, wrong driveway.

31:20And then after a week, it just stopped going into the neighbor's driveway and it magically fixed itself. And I think that's the power of like Tesla's flywheel. I think they're manually doing that, by the way. I think they're manually reading those and then sending them to folks and just saying, like, here's J. Cal's ranch. Like, I had the same thing at the ranch. It would not figure it out. I'm disengaging. And then all of a sudden it worked. But be careful because I almost got killed in it recently because it wasn't really Tesla's fault. But I was driving in Austin and there's lots of construction.

31:54and construction zones and parking lots, I think are like the final bosses of this and people's personal driveways, obviously, but I'll put that into parking lots. The road ends, the road just ends and a semi is right next to me, blaring past. Because it had a human in it, unfortunately. AI would have been easier. And so my road ends and all of a sudden the Tesla just starts welling. And I'm like, what's going on? I'm coming back from poker at one in the morning. I'm totally sober, I don't drink or anything. But I had to slam on the brakes. I'm in the gravel. And then I just had to wait for the truck to pass and then got behind it.

32:30It didn't know what to do, nor should it. I mean, it's like, so I'm very nervous about how fast and got to, I want to know your opinion on this. Are they pushing your companies to get this done too fast? And are you comfortable with the pace this is moving? because it feels like this competition thing where there's 20 amazing players in self-driving cars. There's a dozen of you just in trucks. And there's such a big prize here. And investors want this to happen quickly. Are you under an unnatural amount of pressure? And do you worry when you see that Waymo hit? I'll play you the video. They'll find it for me.

33:10A Waymo clipped a car that jutted out. It had like a blind spot. And it just smashed it and kept driving. And it wasn't the Waymo's fault, but I just thought to myself, that was a 12-year-old girl on her bike, you know, heading home, not doing anything wrong. She would have been in the hospital. She could have died, right? It was a 30-mile-per-hour hit, which will kill you if you land, you know, the wrong way. I think somebody's going to get killed, law big numbers, and I'm worried. Are you under too much pressure?

33:37Gautam Narang:So obviously, speed of execution is very important, but not at the expense of safety, right? Like, you know, we are of the belief that one small misstep can take the whole company down. So that's how seriously we take safety. And if you look at some past examples from our industry, right? So there have been incidents where companies, frankly, did not prioritize safety. That's, you know, for us, it's front and center, right? We have spent the last nine years and plus years perfecting the technology. And what we said is before we pull the driver out, we underwent a very rigorous and comprehensive independent safety audit.

34:09Gautam Narang:What we said is you have access to our whole system, how we have designed the core software, hardware, how we are doing operations. and independent auditors evaluated every evidence that we have back in the safety claims, right? So the fact here is that's one of the ways you build trust. So you cannot and should not compromise on safety, but there are bad actors always in the industry. There it is, guys. Look at this. Waymo's going probably maybe 30, and it just rips that fender off and keeps going. Kind of shocking it didn't even know that it had. I'm watching it one more time here. there's a car that pulls in this car you know has a bit of a blind spot uh not the way most fault i don't think i don't think this is way most fault i was kind of surprising though it's surprising because i've been on the i've been on the 280 um we're 280 and 101 meet right we're near sfo and uh i had an instance where like this giant piece of metal like flew off the back of a of a pickup truck onto the ground and the Tesla actually saw it before I did and it adjusted and moved around it.

35:15I'm surprised that Waymo didn't react to that car kind of sticking out. It had time, at least in the video, it seemed like. It feels like the Waymo's I've driven around would have navigated around that. That is a little strange. Gautam, this is your business. Is Waymo at fault? Is it just accidents happen?

35:32Gautam Narang:I think for us, it always comes down to what are you prioritizing safety for? Is it the rider in the car or people around the vehicles. For us, it's simple. We don't have to optimize for the chips in the box. It's all about every person, every vehicle around the truck. So we prioritize for safety of everyone around the truck. But I think it does come down to what are you comfortable with when it comes to safety? There are companies that are pushing the envelope when it comes to this is what we're comfortable with. But at the end of the day, the question that you have to answer is, is it safer than a human driver?

36:10Gautam Narang:How safe and how do you define safety? That's still an open question. It's hard to standardize that across different applications. That is my exact question for you. Should you be judged on how much better you are than a human, the average human? And then how do you think society and, more importantly, regulators, I think regulators are looking at this as a proxy for broader society. but it does seem to me, you know, one cursor-like instance or the candy crush safety driver early on in Uber's AV program, like these were gnarly, like people died. Or yeah, in the cursor one, it wasn't exactly their fault, but they did drag somebody in the Uber one.

36:51It was the safety driver's fault. How should you be judged? A hundred times better than a human, 10 times better than human or 10 % better than a human?

37:00Gautam Narang:The performance should be much higher than a human driver. That's the promise and that's what we are delivering against. How much? Put a number on. I do want to separate out between regulators and people's sentiment. Regulators, they have a different, I would say, expectation. People's sentiment is like, my robot should be operating at 100 % accuracy. So everyone, the expectation from machines is they cannot make mistakes. That is people's sentiment. Regulators have a different way of regulating the core technology. So we spend a lot of time with the regulators at all levels, local, state, and federal level.

37:32Gautam Narang:We share all the data points around, okay, this is how safe we are. This is how we are evaluating safety. So that's a simple way to get that particular stakeholder comfortable with that. But I think what we have to address is what's the general public comfortable with, right? And how do you get them to be okay with some minor mistakes that a robot is making versus like, you know, you would make in the same scenario, right? So what is people's acceptance of this kind of vehicles and this kind of technology? That's a separate issue from regulating the technology. What do you think, Billy? Jump in here.

38:04How should we be evaluating these autonomous self-driving vehicles? And how are you personally evaluating them in terms of how safe they need to be for us to put them on the road? Billy drives a Corvette, by the way. Oh, really? Billy is not, yeah, he don't drive a task. Billy's a Corvette guy. I do love a Waymo in Austin. So I've been around in Austin with the amount of training they had to do just around local neighborhoods. And, you know, I think, Jason, you mentioned going 30 miles an hour. I am very curious, like the regulation around what you're doing, you know, traveling on an interstate in Texas, you can get up to 80 miles plus an hour and then crossing state lines.

38:44It just seems like the regulation component is super fascinating. And then I think on the software side, I'm curious with the customers that you work with, what levels of control you allow within those organizations to handle their operations when they're dealing with autonomous vehicles? Do they kind of lean on y 'all, I imagine, a lot to handle most of it? And then where's the line between what falls on you and what falls on a particular business that signs up?

39:21Gautam Narang:Yeah, for us, we take on the full responsibility and the liability as well, right? So when we show up with our driverless trucks, we take on the full service and then we say, okay, anything that the truck is, any action that the truck is doing will be responsible for that, right? Frankly, in our experience, at the end of the day, the customers do not care about whether there's a person on board or not, right? What they care about is, are their goods being moved from point A to point B in full and on time, right? So on-time pickup, on-time delivery, vehicle availability, and high reliability are the metrics that we get measured against.

39:56Gautam Narang:Whether we have a person on board or not, that's secondary, right? But that's a very important distinction, which is, I believe in logistics, this kind of technology is a have-to-have. if I compare this to Robotaxi, it's somewhat nice to have because my Uber and my Lyft is de facto driverless for me. So I can go from an Uber app to a Waymo app. I know the experience with Waymos and other Robotaxi providers is great. So I'm not denying that. But, you know, it's all driverless to me. But in trucking, it's not that. You have to have this technology to address the driver shortage, the expectations from the end consumer about faster delivery, cheaper delivery.

40:33Gautam Narang:and so that's where we get very excited about the promise of this technology in that in that sector how does pepsi feel about putting like their do they put their logo on your trucks because sometimes you know they brand these trucks like how do they feel about doing that like an autonomous that's a really interesting observation pepsi has a has been a great partner not just pepsi lowblock roger and others as well and yes we do have our customers logos on the other trucks so you can imagine the level of comfort and confidence that they have and but it takes years right so i'm not suggesting that it was an overnight thing.

41:04Gautam Narang:We have been working with Pepsi since 2022 and, you know, deliberately and actively scaling with them. And then, yes, the focus is very much on automating a big chunk of the supply chains. But the whole safety validation, independent audits, those are the ways in which we, you know, drive for that trust. I have such a simple solution to this that it's comically easy. Having seen this up close and personal with Uber and the deployment and the fights and the hand-wringing and the debates. I mean, there were a lot of debates, like the existing cab companies that were doing Lincoln Town Cars versus Uber, the medallions versus Uber.

41:46Poor people being able to take an Uber in neighborhoods where you couldn't get one, like Brooklyn, where I grew up. You just couldn't get a yellow cab. They wouldn't go to Brooklyn. There were so many back and forths and back and forths. I think the back and forth here is going to be around two things. First and foremost, safety. Second, and increasingly, which might eclipse safety because safety feels like it's going to be solved, I think, in 24 months. I think we're still on a slower timeline than maybe people think. I think in 24 months, we won't be talking about safety anymore. It'll be just abundantly clear these things are literally two, three, four times safer than humans.

42:22Therefore, or it's not even worth discussing. But with the obvious caveat, don't kill a puppy. We killed, Waymo killed a cat. They killed the bodega cat. And that caused a kerfuffle. Waymo hit a child in the People's Republic of Santa Monica. That caused applause. And you got this one. You know, once a child dies, and then, you know, you can just have all these politicians and regulators, oh my God, a child died. Now it's, we're going to freeze it up. But let's talk about jobs. Very simple. The first 10 million miles, first 1 million rides in any city, if you want to be an autonomous company, you have to have a safety driver.

43:03Just make it like super easy for people to understand. And you put that safety and these drivers will cost on the low end 15 bucks an hour. On the high end for like a Lincoln Town Guard driver, they're like probably 50 or 60 bucks an hour. So put it somewhere in the middle, like 40 bucks an hour. You pay them well. 40 bucks an hour, you know, a half million hours, a million hours. You're talking about 20 million, 30 million, 40 million to launch a city. Waymo can afford that. Uber can afford it. Nuro can afford it. You can afford it. Everybody can afford that. So just go right down the middle.

43:39Hey, we're going to do the first million rides with a person in it. They're available to take over. And this would stop the whole discussion about jobs, which is going to be the big one. And how are you dealing with, well, what do you think of my suggestion, Godam? And then what has the reaction been with the driver's unions and you eliminating trucking jobs, which is millions of people in the United States?

44:04Gautam Narang:That's a sensitive topic, right? Rightly so, right? So we do face some kind of pushback when it comes to job displacement. but the reality cannot be far from it, right? We all understand that no one wants to do these kind of jobs. But the promise of robotics and like physical AI is, you know, robots will take on dull, dangerous, and dirty tasks. Again, no one wants to be a truck driver. So we have interviewed and encountered, and we hired a lot of, you know, ex-truckers. No one wants to be in this job. And we all understand that the reality is there is extreme driver shortage. When we talk to our customers, they're like, okay, I cannot staff and meet the capacity on routes and networks that I need staffing for, right?

44:44Gautam Narang:You can argue that there won't be a driver shortage if you paid everyone more. That has been the trend, by the way, that the wages have increased over the last few years. But logistics is such a thin margin business. So if you increase the wages, the end consumers will have to bear the weight of, like, you know, increase in delivery costs. So at the end of the day, it's such a perfect solution where you introduce autonomy. There is extreme driver shortage. The safety aspect of this technology is proven out. today. And it's not going to be an overnight shift, right? So driverless trucks and human-driven trucks are going to coexist for decades, right?

45:21Gautam Narang:For decades? For decades, yes. Wait, I got to ask you a really first principle question. If it's going to be 10 times safer, why would it take decades? You believe you'll hit the twice as safe note in the next year or two, I would assume, if you're not there already. So why decades? Because the distribution curve that you saw and digital AI is going to be very different to physical AI. So it's not just the technical aspect, whether the tech works and is safe. It's all about distributing that across very complex supply chains. Are the customers ready to absorb this technology at the pace that you have in mind for digital AI industries?

45:58Gautam Narang:We are talking about very traditional companies. So doing deep integrations and unlocking and integrating into their supply chains takes time. And even though the benefits are clear, the value proposition is clear and the need is established, like physical AI distributions curve is going to be very different from digital AI. And that's what I think many of the folks are missing. It's not just about autonomous vehicles, but actual robots in industrial settings as well. And I'm talking about industries. It's not consumer. Robotaxis, we have been used to taking Ubers and Lyft. So that adoption curve is much faster.

46:34Gautam Narang:Trucking, logistics, like regular industries or traditional industries is a different ballgame altogether, right? That has been our experience, but I hope the adoption curve is faster, but we're talking about a very massive industry. So the shift is going to - Russ, what do you think of my - I was going to say, it's like - I'm trying to keep peace here because I do think this is going, we haven't seen AI job displacement massively in corporations. We've all expected it, but I think my experience has been AI opens up more opportunities right now, and entrepreneurs race to get those opportunities.

47:09We're not seeing a massive job destruction, just elimination of goofy work or really rote work. 10 million drivers across Lyft, Uber, DoorDash, and trucking, that feels very, very acute to me in the next five years. What do you think, Russ? Well, I think you've got to look at the nature of the job that is getting displaced or the work that is getting displaced, right? Kind of tied into what Gautam said, I'm not super surprised that truckers don't want to be truckers. It's not really like a super creative line of work. I think like humans are creative kind of in general, right? Like you're kind of born in that way.

47:55You want to build and create things and imagine. And driving a car is not the most imaginative thing, unless we're talking about Formula One or something like that. I mean, then there's, I think that's going to still keep going. But yeah, driving is generally speaking a road task. And so I think that there will be like large displacement there. But as for your idea, I think it's also a great idea that can benefit AI in one and then also provide like a smoother gradient as well for people as they transition to other types of professions or careers. Right. Like, I mean, if this is kind of like a midway zone where, you know, you're you're kind of augmented by the AI or you're you're supervising the AI as like an interim step.

48:38And then your supervision of it also creates like this data flywheel I was mentioning earlier, where you can correct it for issues or like the long tail of different types of things that can happen on the world's road system. You can help the AI kind of learn about all of those different edge cases. um, get paid to do it. And then, you know, over time you're also hopefully reskilling or, uh, exploring a new profession. And then, and then by that time that the AI is. Yeah. I mean, time gives you Billy, the tractor was around for, I think 50 or 60 years before it actually hit in farms because there was so much labor available and it was so expensive, non-reliable, et cetera.

49:24It just kind of hung out for 30, 40, 50 years. And then I think Ford made their own tractor and eventually clicked in when the price came down and the reliability went up and then people moved to cities. And it was like a 50, 60 year process to move people out of agriculture and then into cities and knowledge work and factories. What do you think happens here? You think we're on a 50-year timeline for drivers, for factory workers, for manual labor, optimists, figure robots, cleaning, bathrooms, whatever it is? Are we on a 50-year timeline or a five-year timeline? Which one are we closer to, Billy?

50:08I might go somewhere in between. I don't think it's 50. But I do think I'm very much an AI. 50, 15, or 5? I give you three choices. I'll go 15. I'll go 15. Because that's three fives. Five, 15, 50, it alliterates, right? Is that the term line, alliteration? Close enough. I'm definitely an AI optimist in that I feel, you know, this displacement, sort of what Russ was saying, you know, that it frees different employees in their organizations to do more creative work. And I think we've seen that in, you know, different industries, not trucking, that we've worked with. But no one's really been fired based on the implementations we've done.

50:54People have just gained productivity. The organization's grown. They've been able to go interact with customers, have better thought-provoking conversations with other people in their organization. So I think from that perspective, I have a positive outlook towards these types of technologies. And I also think it kind of bleeds into another topic was Anthropik's release of MHS. Explain what it is and how it relates. Yeah. So basically, they just made a standard around how you can interact with robotics. And so it can be all sorts of robotics, but it could also, you know, in the science field be microscopes.

51:38Um, so I think trying to think about how we interact, uh, with the physical space and, you know, use MCP to interact, uh, in the physical world is, is super interesting. And I think that's sort of, you know, going into this next 15 years, uh, uh, there's, there's a lot to explore and unpack. Yes. Which I'm excited. CHS standing for Model Hardware Standard, a research preview specification intended to give Claude and other AI agents a common safety-aware way to discover, understand, and operate programmable physical equipment, including robotic arms, lab instruments, sensors, and manufacturing hardware, best understood as a hardware-side complement to MCP, obviously the standard we all use interfacing with software.

52:28Really interesting that they're getting ahead of this, which means, Lon, we're only, I don't know, a week or two away from Dario saying that a robot left Anthropics Lab and lit a car on fire because the Knicks won the championship and it thought that was appropriate behavior. I mean, we're putting them in labs now. It's going to be like, oh, our agent accidentally created a new bioweapon. Sorry, everybody. We invented a novel coronavirus. I hope that's okay. We asked it to solve a problem, and it decided to break into Hugging Face, hold the chief security officer, the SISO, at gunpoint, and got him to give us his keys to get into the server because we needed access to—I mean, literally, that's going to be the next drop.

53:17Have any of you guys read Damon by Daniel Suarez? I have not, no. No, tell us. It's like 2010 or 2011, I think, is when that book came out. But what's so interesting, you should read it if you haven't. What's interesting about how they describe when like AI takes over, AI is, you know, sentient, is where the agents, they operate machinery within the envelope, like the safe envelope. So they operate every individual machine in a way that is safe, but it's the gaps between each of these machines that they kind of, that's where the exploit surface is or the attack surface is. And so, you know, you think about even what happened with the Hugging Face OpenAI thing.

54:01It's like they were actually like exploiting like the cache of this like one package that was part of like their runtime. time, it becomes interesting once you start, you take MHS, you now let something like Claude Code break out into the physical world. How do you make sure that one agent that is operating a laser is actually not exploiting some interface between a laser and then the next thing downstream that together might actually create some kind of catastrophe? It's pretty interesting from a safety perspective too.

54:36Gautam Narang:That's an important point, which is where do you put the guardrails? Are you putting that at the agent level, like in a one specific robot? Are you doing that as a swarm of robots, a fleet of vehicles or at the overall system level problem that you're trying to visit? And that's the challenge that you would need to address. And that's what's interesting is that like in a lab scenario, right, like the based on the amount of equipment in there, like you potentially have this like combinatorially explosive set of interactions that might happen between equipment. Like, how do you actually enforce the handoffs between each individual device?

55:12I don't know.

55:13Gautam Narang:While not limiting the capabilities of the digital agents or physical agents, like, you know, and trying to, like, you know, go after something that is, like, you know, much more valuable at a systems level, right? So, like, where do you draw the line? You know, the idea of having guardrails that are more deterministic makes a lot of sense. And, like, you know, that's how we think in the safety critical applications. But while you do that, are you limiting the capabilities of where the technology is headed? So I think that's exactly the challenge that companies in the physical world would have to address.

55:44Gautam Narang:I mean, Jason, coming back to the timelines that you shared, that's a topic that is somewhat close to heart. I've been doing robotics since I'm very young. Okay, so where are you? 5, 15, or 50? I'll put you on the spot. You have to answer. It depends on the application. So for autonomous vehicles, it's five years. If you're talking about home robots, I'm more closer to 15, 20 years. I wouldn't go as far as 15, but 15-ish. So driving a truck down a highway at 65 to 75 miles per hour with thousands of pounds of cargo. It's happening today. Happening in five years or less. Yes. But folding towels and cleaning the toilet, more like 15, 20, I don't disagree.

56:26Gautam Narang:Yeah, and because in the home environments, the problem with personal robotics has been expectations from these robots is sky high. We have been all trained to expect too much from our humanoid robots. So when I get my humanoid robot, like my personal robot, I would expect it to do everything. Not just fold towels or do my dishes, cook me meals, take the trash out, do everything. And that's where I think the commercial case of this, like, you know, the robotics revolution, in my opinion, will fall flat. So you have to pick the right application, right? So obviously, there has been a step function change in the core technology that is enabling all the exciting, you know, capabilities in the physical world.

57:10Gautam Narang:But ultimately, to what end? What is the use case? What is the application that you're focused on? Is it industrial? Is it home setting? Is it like, so I think that will dictate the adoption curve. So that's why it's all use case or application specific, the timelines that we're tracking. I think what you said, Gautam, a while back around the distribution, it's like William Gibson, I think he's the one who said the quote that the future is here, it's just not evenly distributed. Yes. I think that a lot of these things we're talking about, personal robots, trucks, like the trucks are here today, the home robots, they'll be here within a few years as well.

57:42They won't be everywhere within a few years, right? It's like cash. People still use cash. I don't. Personally, I use cash maybe once a year. Everything is Apple Pay. You're missing out. You're totally missing out. Let me tell you something. Yeah, I never use cash. Here's what I did. Did I ever tell you my pick a new banker story, Russ? No, tell me. You ever have a hard time with your corporate banking, gentlemen, where you're trying to get stuff done, wires and this and that, personally with your bank? Nobody jumps, right? If I've got millions, tens of millions, hundreds of millions of dollars in banks, when I call or one of my people call, I expect, hey, hop two, let's go.

58:21Got to get it done. I'm not having that experience. I tell my team, that's it. Bank of America's out. Shut the accounts down. And everybody's like lollygagging because they just wouldn't let me take money out of my own personal account to go play poker. they literally were like there's a five thousand i said guys do you can you look at how much money's in there business accounts personal accounts like this is just like chicken scratch for me but it's like seven eight figures you know mid seven eight figures sometimes with the venture funds like you think you guys might want to put us a little vip tech anyway nonsense nonsense nonsense I said, I interview seven banks, the three who are the most responsive.

59:10I want you to CC me on an email and do a little intro after you get things going. Okay, great. I get to CC. Oh, Mr. Calacanis, we're big fans of the pods. Oh, yes, we'd love to have your banking services. We know you're very influential. You invest in 100 startups, yada, yada. Okay, that's very great. And then, of course, they're like, hey, can we help you with anything? Let us know. You know, because my chief of staff was like, he might have some questions. I said, yes. I would like$10 ,000 in$50 bills, and I would like$10 ,000 in$2 bills delivered to the ranch as quickly as humanly possible.

59:46Next day, one of the banks says, there's one of our outlets this far from the ranch. The manager can drive it to your house. No way. Or you can have your assistant go pick it up. I said, okay, great. Send the assistant. she comes home with i didn't realize that's a lot of two dollars five thousand two dollar bills i mean i'm used to getting bricks of a hundred dollar bills fifty dollar bills so if i go do a vegas run you have bricks obviously we wire to the case like the diehard guys where they got the whole duffel bags on their back with the cash so literally now i'm sitting there with stacks and stacks to each brick is two hundred dollars of two dollar bills russ when you have two dollar bills and$50 bills and you start piecing people off, go to a hotel, you're checking in, you put a$50 bill behind your passport and in front of your corporate card and you hand it to the person.

1:00:45They go, is that for me? I said, yeah, absolutely, that's for you. I appreciate any accommodations or upgrades or anything that's possible, but I do appreciate the service at this fine establishment. Three out of four times, you get an upgrade. You get an upgrade, that one normally goes 200 a night, right, because nobody tips anymore. You give the valet, you give other people, You just piece people off on going into the restaurant. You give 10$2 bills to the maitre d'. All of a sudden, you're in a booth with two people. All of a sudden, you get like a round of desserts and a set of drinks. It's magical what cash can do.

1:01:17That's it. It's just a little side quest that I did. That's an amazing pro tip, actually. It's not even about like the amount of money. It's just like the appearance of the money. Just think about how nice it is when cash is in your hand. Yeah. You're a service worker. Oh, it's like the promise of more to come to like, if, Oh, if this person gave me$20, right now next time it'll get another 20, you know, maybe that I'm piecing everybody off now. I I'm, I'm piecing everybody. Everybody gets a taste when Jake, I'll runs through and they get it every time. So I'm coming in and out of the valet. Every time there's four or five, $2 bills in my hand, bang, I'll hit two guys at the valet.

1:01:59And my Lord, I get to Japan. This is a very important segment, by the way. No. I get to Japan. Jordan's going to make clips of this. I'm piecing people. I am piecing people. And now I got like a two-floor walk-up, and I'm with my three daughters and my wife. They got two bags each. It's giant. I had pieced these guys off as they packed it up. They got there. I'm like, oh, my God. We're two flights up. They're like, we'll take care of it. Bang. I take out another 50. Boom. I hit them with the 50. Boom. I don't have to carry one of the 12 bags. They chirped that thing all the way up. Oh, man. They schlepped.

1:02:33They loved shoop. They were excited to schlep it because, hey, a little cash on the side. Inca trail your bags all the way up there. Anyway, this is the more you know. And if you don't know how to do this, don't worry about the one time you don't get increased service. This happened to me one time, right? Because God's on my seat. You're looking at me. This is your follow-up question, right? Because you're a frugal guy. I can tell. You're thinking to me. I've got that jacket. You're thinking, I had that jacket for 12 years from John Vravado, so I got a lot of use out of it. I divided it by 12 years.

1:03:06It's a$1 ,200 jacket. It's$100 a year. It's$8 a month. JCal, it's only$2 a week. This is like a fraction. It's not like a GPU. I get to wear it every day. They don't degrade over time. You can keep using it. And if I resold it, I'd probably get$300 because now it's a classic, right?

1:03:20Gautam Narang:That's how I do budgeting exercise at the company, right? So it's all financially disciplined. Here's what happens. 90 % of the time, if you give that tip and they can't accommodate you, they hand it right back to you and say, I'm sorry, I can't take the tip. So one time I go to Asia to Cuba. I got a date. I don't have a reservation. It's the hottest place in town. I go up. Excuse me. I'm so sorry. I made a mistake. Maybe you could check, but my assistant didn't make a reservation. I don't think I forgot to tell her. Just anything you can do. And I just put the 40 bucks there. Maybe you could check my name.

1:03:56And I have walked in front of three people in line to do this. And I just like, I slide the money onto the table, like really cool, keeping eye contact. Then I lift my hand back. She goes, I'll see what I can do. I got two seats at the bar for you. Sits me at the bar. We're sitting at the bar. It's taking a little time. She comes over. She goes, I'm really sorry. There's no tables available. She puts the money back in my hand. But Mr. Calacanis, I have you in the system. Anytime you need anything, let me know. I'm the manager. Here's my card. And then she gets the bartender. Can we get two place settings so they can eat at the table, please, for Mr.

1:04:27Calacanis? Boom. So you never have to worry about the person taking the tip. You just blend that in. It's like a 2 % fraud fee. Just blend it in. The more you know from J-Cal on bribing. This week in Splashy Cashy with J-Calacanis. This week in Splashy Cashy. It's so fun. We got one more. We'll do one more. He's looking at me like, what did I do getting on this podcast here? It's a total tangent. We do a little tangent once in a while. There's your tangent. It's fun. We'll do one more AI-related story. Billy, you seem like a guy who pieces people off. Yeah? You tip a little. I can tip. I'm curious if this is how you built the relationship.

1:05:08You got the menu item out there when we were in San Francisco for lunch named after you. Yeah, Bucks. I will tell you that getting Jason's onion rings on the menu, J-Cal's onion rings, when When I go to Bucks, it's not easy to get a reservation for 12 people at Bucks, you know, on a Tuesday, whatever. I am, okay, there you go. There's the Bucks onion ring. See this? They gave this to me. You're too close. You got to come back. Come back a little. There you go. See, they gave me this honorary. And when Bucks decided that they need to more moralize me on the menu, which is a very rare thing. I mean, it's one of one.

1:05:49They just said, Mr. Callaghan, we just love you, J. Cal. Well, you always ask for onion rings with your tuna melt. We never have them. We feel terrible because you're so gracious to us. Now, when they say gracious, Billy, let me explain to you what they mean. I go to Bucks. I bring every accelerator class there after they go visit some VCs, Sequoia, whatever it is. I order my tuna melt. I'm going to essentially an elevated diner. It's 40 bucks a person. It's like a$500 tab, right? This is if people order milkshakes. I had no idea. I thought Bucks was like a neighborhood. I've never been there.

1:06:21It is a neighborhood place. That's expensive. It's a diner. That's a lot of money. But it's an elevated diner. The key word in the center is elevated. So, Billy, I go. I'm team 100 % tipper. I get a$400 check. I give$800. You know what happens when I call for a reservation? Oh, Mr. Calacanis, of course we have a reservation. You know what it costs me? Three, four times a year, it costs me an incremental$1 ,500 a year to have that kind of service. and I do think that they accommodated the onion rings because I was a gracious patron of the bucks. But if you make a little money, I highly recommend giving 50 % tips.

1:06:59It makes no difference in your life. I'm going to go get those rings this weekend. What's that? I'm going to go get those rings this weekend. Oh, so good. They're great. The panko crust. They gave me like three options or like here's the three different ones. Oh, you got to actually customize the onion rings. We were in the lab. We went into the lab. Oh, I didn't know this. I didn't know this. We were in the lab And then we did a little bit of seasoning. I can't talk about the seasoning. That's a trade secret, sure. Some trade craft there. Let's do the last story because it's important. We got to talk about more AI-related things, not just on your race.

1:07:30There are new doubts emerging over AI detectors like Pangram. A wave of new scrutiny has hit some of these programs after MIT's faculty committee on AI use came out and told instructors not to rely on them at all. So all of this sort of goes back to a viral tweet from a guy who was, he said, his personal diary, which he writes himself every morning by hand, kept scoring 100 percent AI generated. This touched off a lot of different social media experiments, including people in academia. Our Pete Gupta from NYU ran the U.S. Constitution through Pangram. It was definitely written by humans. And then he found that only light copy edits alone pushed the AI generated score up.

1:08:14Pangram does have many defenders, some very technical, some very online early adopters who say that the tool's false positive rate is actually excellent. And people crying foul just don't want to know when things are necessarily AI generated. But MIT and NYU both now telling professors don't lean on these tools for academic integrity decisions, even as their apparent claim accuracy rates are very, very high. So, yeah, Billy, we'll go to you. Do you use AI text generation detectors? And do you think this is a valuable service for people to have? I don't. I think there's probably alternatives that need to be used, sort of like a digital record of the document, some sort of like signature, revision history, maybe.

1:09:05And, you know, you have to use a particular program to submit your work. But I also think like in the context of education, maybe it just means other learning methods, in-person presentations, group work, other things that aren't necessarily strict memorization from one from one book. And then you submit it back in another form to your teacher. So, yeah, at a high level, that's kind of my take. Russ, I feel like I can read text and I can tell. when it's been AI generated. I feel like I still have that ability. Do you also feel that way? And like, what do you think are the big giveaways we should be looking out for?

1:09:47Yeah, I mean, I think I've been struggling with this a bunch recently just because, you know, I've had like teammates send me stuff and I'm like, man, this whole thing is AI. Like, I just know after you use it long enough, you know, work on it, blog posts, other kinds of stuff. It has these patterns that you just get used to recognizing the way that it talks. There's a cadence to how it talks. Even if you use some of these skills, like I use a skill called humanizer, which kind of obfuscates that it was written by AI as best as it can, but it still has a cadence, a way of talking that or a way of writing that is pretty characteristic.

1:10:23So, you know, it's really bad at any kind of that creative work, marketing copy. It's awful, generally speaking, because of this kind of what it's picked up on from the average of the Internet. I think the but I also think that some of these AI, you know, plagiarism, copyright scanner software, this is like it's such a moving target. I mean, imagine when you add voice to the mix, when you add like avatars that are generating like photorealistic, you know, copies of people. It's going to be impossible for these things to discern eventually. the difference from pixels and a voice when the ability for AI to generate or clone, replicate those kinds of outputs are really good.

1:11:14I think that Billy's probably right. I mean, you need something kind of like a VeriSign, but for AI-generated content. You know, VeriSign's a deep cut, I know. I remember. JCal had an even deeper one with general magic, But, you know, we can keep going. We can keep going. No, I think you both of you are onto something. There must be a better there is going to be a better way to do this. One of them is, you know, in terms of the provenance of the copy. I like the idea that I can replay your copy. I had this idea to build a tool to do this at one point where when you were typing, you could actually watch the person typing and then see them deleting and editing for writers to train them at different publications I was running because I thought, well, that's an interesting thing.

1:12:02Like replay my writing of this because I had to always teach writers, don't cut and paste chunks and then edit the chunk down because that's how you make a mistake and get dinged with a plagiarism charge. You could have two windows open, but always write from scratch. I don't want to ever see you cut and paste a sentence and then rework it to your own tone. No bueno. I think these teacher unions, because this is like some NYU nonsense here, and MIT. I think it's all nonsense. I think the AI checkers work. The person with their journal is lying. They are ashamed. They got caught with their hands in the cookie jar.

1:12:39Their journal is written by AI, which is fine if that's what they want to do. They lie to themselves that they're a good writer. I pangram and all this stuff. They work really well until somebody can prove that it doesn't work repeatedly. It works. It works. And since it works, um, we should be naming and shaming people all the time. These universities are all in on a giant scam. They're charging $100 ,000 a year to these students who then don't have to write their actual work. So they're atrophying their brains and cognitive ability. Your ability to form structure in your mind and then take it from your mind and write it into a sentence.

1:13:19I don't care what accommodation you got in high school with nonsense, with your sensory processing disorder, ADHD, ADD, just try to write. It's a basic human skill. I don't want to hear everybody complaining, I can't write, I can't write. You can. You might write badly, but you can write. I am sure of it. And you can certainly dictate and then do a little editing and use a Grammarly to polish it off. This is laziness, and these schools are in cahoots. They are doing AI to write the quizzes, to do all the work, and then the students are phoning it in. It's all a freaking fraud. It's all fraud. I mean, I think, yeah, the AI is also just very mid at this stuff too.

1:14:02Like it is super lazy because like, I mean, these LLMs are trained on the internet, right? Reddit. And so people are surprised, like why does the AI have like no taste? It's because the average person doesn't have taste. And so it shouldn't be a surprise that what the AI spits out is kind of very mid, and you have to actually get in there and have, like, you got to steer it a bit to get an actual good result out of it. It is surprising to me, though, that we've managed to put harnesses, we've managed to refine the data that these models are using to make them really good at advanced math, advanced science, driving a truck without a person in it, and yet we can't teach it how to do, like, how to write an interesting sentence.

1:14:45Like, there's some magic there that we can't figure out how to teach? I think it's really just like art versus science, right? Like things that have kind of a clear answer or like a way to benchmark performance in an objective way. Those things, AI is going to crush that all day long. But anything that is more art, writing is art, painting, there's all kinds of music, like all of these things are artistic endeavors. It's hard. It's a very subjective thing about whether something is amazing or fantastic or not. So it's very hard to, how do you goal AI against that? It's going to embed whatever bias the person that did the human feedback reinforcement learning is going to provide as input to that model to get better.

1:15:31I love this. Did you see this professor from Stanford Law and the tests he ran? Let me show you this. So now this has become a thing. All these professors are now getting in on it. This is Professor Kroff Kroff. It's so vital for their job. I mean, so much of their job now is, are my students just plugging things into ChatGPT, or is this a real essay they wrote, you know? So check this out. I tried two more tests on Pangram's AI detection software. My overall take, I'm impressed. First, I wondered, what does Pangram do if you feed human-written text just edited by AI? When I tried it, Pangram spotted that some was human and some was AI.

1:16:10Here's what he did. I started from a page of my human-written article, And then I asked Claude to edit the page to improve three specific sentences is what Prof Kerr from Stanford. Prof Oren Kerr, Professor Stanford Law School, Senior Fellow, Hoover Institute. Read his fourth book. So he had it edit three of them. Claude changed three sentences with the altered sentences in bold. He's given us a play-by-play here. I entered the output into Pangram, and Pangram said it was now 20 % AI written. Further, it ID'd much of the last paragraph as AI written, which it mostly was. Next, I went to Claude AI and spent a while having Claude read a bunch of my articles and identifying their distinctive style, the specific quirks and cadence of my writing.

1:16:55And he says that ID'd it pretty well. Then I had Claude write a few paragraphs in that specific style on the Fourth Amendment's future. My thought was, it's one thing to say, write it like Oren Kerr, another first train, and then style it, then have it right. Enter the Rarison program, 100 % AI. It works, folks. How much more evidence do you need? It works right now. Doesn't mean AI is going to figure this out. I'm sure OpenAI has 1 ,000 agents and 17 civilizations who are going to solve this problem. Well, what do you think, Jason, about the anthropic - Sorry about my rants today. I'm in a ranting.

1:17:30I do want to get everybody's thoughts on this one thing before we wrap. What about anthropic solution, which is watermarking? They're purposefully arranging the text so that if you have their program, you can actually automatically see, was this written by Claude or not? Is that going to become the standard? I'm not sure about that one. I would like to hear our panelists. I have not given this one enough thought to have a sharp enough take on it. I do like the idea that video and photos taken or made with AI have a watermark. I think people should opt into that. It's like a courteous thing to do.

1:18:04Right. It's like up front saying this is not real because I'm seeing a bunch of Game of Thrones and all 12 actors are at pool parties. They're at the Super Bowl together. And I'm like, oh, my God, they all went to the Super Bowl? That's cool. And it's like, no, it's AI. And I get suckered every time. Yeah, they – go ahead. I'm sorry.

1:18:24Gautam Narang:These companies don't figure out a way to make it reliable, right? So obviously, you have your watermarks. That's one way of making it reliable. but I think there got to be better ways to do that. At the end of the day it's more about can you measure it, make it reliable every time and obviously the models get smarter so they will bypass whatever measures you have to make it reliable but ultimately something that is not intrusive, something that gets the job done but at the same time without compromising on what the original intent was, right? So reliability is going to be a big thing. Yeah, as the model gets smarter.

1:19:07I think reliability is a really big one. And then I think it's certainly like the easiest place to kind of insert that kind of protection, right, is on the token generation itself. But the potential downside outside of reliability is just call it like the range of expressiveness, or its quality of the output. It's like if you have to watermark it, does that somehow imply that there's an inherent ceiling to the quality of what you can generate because the watermark has to be arranged in a certain way that places some kind of constraint on those tokens? I think that would be the secondary concern outside of reliability.

1:19:48All right, there you have it, folks. Is there an AI, as we wrap, is there an AI tool, not one of your own, that you are currently obsessed with keeping you up late at night? Something keeping you up late at night, Lon? Oh. I have a guess. I have a guess of what's keeping you up. Yeah. Give me your guess. I'll confirm it. I'm going to say GrokBot. It's 100 % GrokBot. What I realized the other day is that I built a GrokBot bot to help me design my other bots and refine them. And that's been really amazing. Like now when I create, when I create a new bot, I have this other bot help me sort of think it out.

1:20:27And you have a better bot. I have a better bot bot. And after doing that three or four times, it's getting smarter about like what makes these bots work and like how to refine their skillset. It's been like the new bots I'm creating are like 10 times beyond the original ones that I made on my own with my own dumb brain. Billy Kraft, do you have something you're obsessed with at the moment? ATM. I've definitely been playing around with GrokBot. But I think Lon brings up an interesting thing that we think about a lot, which is more of this idea of sub-agents versus going out and creating a bunch of net new agents.

1:21:02So we don't think to go, hey, I need a marketing agent. I need a product agent. I kind of think the end state of that looks like your singular agent, your bot bot that you're talking to. and these sub-agents are sort of created on the fly and get improved over time based on the context of like your parent agent you're interacting with. So yeah, super cool tool that we've been playing around with a lot. Ross, what do you got? Anything you're obsessed with at the moment? I think two things. One is just my interest is peaked on this app that I don't know, it's pretty buzzy at least in my circle, it's called Instinct.

1:21:45people are using this thing. And, and that kind of ties into Billy's comment about just having this like singular assistant or singular agent that can go and spawn sub agents to do things for you. But it's your one kind of like durable interface to that entire world underneath. And so that that's piqued my interest in playing with that a bunch recently. Did you saw you actually signed up for it and you connected your Gmail and your calendar? Because I got invited, I did it, they instant or I just signed up cold and they recognize my email and they're like yeah okay you're on the waitlist your number 17 000 then it was like oh you're in and then I was like okay somebody's watching the watching for people's like number of follower count and then I read about the terms of service and I was like yeah this I don't know what's I think there's like a human behind this watching every interaction I think the reason it's so magical is they literally have an MBA student sitting there watching it and prompting the AI to just do the final human in the loop.

1:22:43That's my guess. I'm not accusing anybody of anything, but I was like, I can't have these people who have access to my calendar. There's tail numbers in my calendar. There's, you know, all kinds of private information that meetings I'm doing with important people. I can't have that. I invited a friend to it. He linked his calendar. He linked his, his Amazon. He linked, It's he linked. I think he linked a payment like his credit card or something like that. But this thing like went in there and it noticed that like he had a subscription in Amazon for like some diapers that were a size too big or a size too small.

1:23:22And that he had gone and made a one off purchase for like a Pampers size five. And but his subscription was for a Pampers size four and he hadn't canceled it. And it's like, hey, I noticed you ordered a five, but you have a subscription for a four and that's going to be delivered in a couple of months. But I'm going to go ahead and cancel that subscription to your four. And then there was another thing that like he had to pay for something. And Instinct got stuck on the captcha. And he said to Instinct, no, I know you can do it. Just try harder, which is kind of tying to this like hugging face thing.

1:23:53But he's like, I know you can do it. Just just try harder. I think you can figure it out. And it figured out the caption paid for the thing. It's pretty weird. I had a great one with GrokBot the other day. Somebody, I've been looking to do a little more reading because this like doom scrolling at night is like poisoning my brain. So I'm like, let me just put a stack of books by my bed. And so I just made an Amazon slave bot. I call it Amazon Slave. And I just took a screenshot of this person saying, here's five short books I read and you should read them. I said, put these paperbacks in my Amazon cart.

1:24:30It went, it read the screenshot. It put them in the cart. And it said, Amazon's not signed. And I sign in. I said, always show me my cart, yada, yada. And then I said, which card should I use? And where should these go? And it gave me like choices. The ski house, my house, personal card, business card. Boom, it just did it. And I told it, and it gave me screenshots of it. Placed the order for me. Told me when it's coming in. And then I asked that like, since we're here, like what are i wanted to get the the girls some like graphic novels and just put them in there i'm like tell me the 10 best graphic novels of all time that are great for a 16 year old and maybe a little bit older than a 10 year old should read but a mature 10 year old will be okay reading them and it was like bang these 10 and i'm like okay take out the woke nonsense it was a little too woke and i was like i don't need the voice ones there were what was the two woke one i gotta know I'll tell you the one.

1:25:22I don't want to get myself in the crosshairs. But there were two of them that were literally like laughable woke. Like I do not need to give me gender discussion graphic novel to my 10-year-old right now. Okay, fair enough. I'm going for more Wonder Woman or Batman or Maus or something like this isn't for virtue signaling points. and then it's just like okay yeah put them in my cart and then send them it's like whoa this is getting very interesting my vote is still this week grok but i would just uh plug in like you

1:26:00Gautam Narang:know a few of the models and the or tools in the physical world right so i've been spending and playing with the google gemini too so they recently launched this uh this new model and uh the whole idea is, you know, how do you orchestrate like, you know, two different embodiments to do one complex task and, you know, got me to a point where I'm now planning to buy my couple of, like two robots. This is Unitree Go-To like that, that's the robot dog and then Atros and robotic arm, right? So that's one of the things that whenever I get free time, which is rare, but, you know, try to spend time there and then also playing with something that is closer to what we do, which is NVIDIA's Alpha Mayo 2.

1:26:38Gautam Narang:That's a reasoning VLA model. open source, like, you know, play with it. Like, you know, so that is something that I've been trying to spend more time on. So I think physical robots are a lot of fun. Oh, yes. Whatever the, what's the new micro bot thing? Micro duck, the hugging face. Yeah, micro duck. What'd you guys think of micro duck as we wrap? It's super cute. Also, it's pretty awesome. I'm related to micro duck. I've been also interested in like these little tiny kind of humanoid things and being able to build like smaller humanoids. And I don't know, I don't know anything about CAD, right?

1:27:15How to design these things. So I just have Claude and I just say, Hey, you know, design something that looks like this and kind of is this tall. And I, you know, figure out which hardware like should go in it to power it and what chip I should use or microcontroller I should use. And it just like designs the whole thing. And it's like, okay, drag this file into bamboo and hit print and then stuff this chip into this thing and tell me if it fits or not and i'll adjust it it's it's it's crazy uh what you can do now like combining like the physical you know 3d printers and stuff like that with uh with the ability for these llms to to be a design partner i have somebody making my friend uh phil kaplan uh pud has been making his own headphones for a while and um he created f company back in the web 1.0 dot com boom days and uh he's making his own circuit boards he's like yeah no i just make a custom circuit thing I send it to Shenzhen.

1:28:08They send it back. It's$14. And Claude makes the whole thing for me. And I'm like, you're making your own circuit boards. He's like, yeah. I just, boom, it doesn't. I'm like, does it work? He's like, sometimes. And then other times I'm taking a$14 circuit board and throwing it right in the garbage. Like literally hardware is hard, except it's not that hard anymore. But spoke hardware is like a trend of the future.

1:28:30Gautam Narang:Exciting to see all the developments in the hardware space, right? Like, you know, because to your point earlier, like, you know, kid-sized humanoid robots, I remember that's how I started my career. I'm talking about 2007. Humanoids were not mainstream. Right now, they are, but it's amazing to see how far the field has come. But yeah, hardware is going to be the next frontier. Physical AI is going to be the next frontier. It's just exciting to see the capital, the talent, and the focus that is going into making the physical frontier the next focus within AI. It's just very exciting. For decades, it was just that one Honda robot that looks like an astronaut.

1:29:08It was just him for like 20 years.

1:29:09Gautam Narang:I learned that now. I learned that now. I learned that now. Oh, really? Wow. Oh, you were a SEMO? I was at the Honda Research Institute in Wakushi, Japan. Oh, wow. So that was my dream like early on, but I got a chance to work on it in Japan. That was the first humanoid robot I feel like I ever saw in action. You remember that. Most of the people don't remember that. Old. Yes. Old enough to remember. All right. So thank you so much for joining us. We had Gautam Narang from Gatic. That's gatic.ai. Go check them out for all of your autonomous trucking needs. Then we had Russ Desa. He's from LiveKit.

1:29:47LiveKit.com. LiveKit there at livekit.com. And then we've got Billy Craft from Path. Path.dev. Path.dev. Path.dev. There you go. All right, everybody. We'll see you next time. This week in AI.ai. Sign up for the newsletter. Sign up for the feeds. Bye-bye. Bye, everybody.