OpenAI Invests in the Self-Driving Race with John Hayes, Founder of Ghost Autonomy

Full Transcript

Transcript

John Hayes: 0:00 What we're putting together is the car should tell you why it's doing what it's doing in plain language. If you've rented a car that has a bunch of features on it, it will beep at you and show you some symbol you've never seen, and you don't know why it's beeping at you. And it has this very, very narrow communication path. And so I think the next evolution is, yeah, you should just chat with your car and tell it what you want. Let's train on absolutely everything. You tend to get sensible answers, and you can ask him to explain the answers and explains the answers in terms of visual elements of the scene. You're gonna make a system that out of the box is dramatically more reliable. But hallucination is actually the best part of these models because the hallucination is where all the common sense applies. And so when you're asking to incorporate not just all the knowledge of the scene, but all sorts of the entire depth of knowledge that could be applied, you tend to get more answers that are closer to reality than what an engineer having to express their thoughts in a programming language would be able to express. I think the amount of creativity that's gonna unlock, we don't even know the limits.

Nathan Labenz: 1:08 Hello, and welcome to the Cognitive Revolution, where we interview visionary researchers, entrepreneurs, and builders working on the frontier of artificial intelligence. Each week, we'll explore their revolutionary ideas, and together, we'll build picture of how AI technology will transform work, life, and society in the coming years. I'm Nathan Labenz joined by my cohost, Erik Torenberg. Hello and welcome back to the Cognitive Revolution. Today, my guest is John Hayes, founder and CEO of Ghost Autonomy. If you thought that the OpenAI news was exhausted after Monday's Dev Day, today's episode shows that you were wrong. Just 2 days later, OpenAI and Ghost Autonomy are announcing a new partnership, including investment from OpenAI's startup fund to bring large multimodal models to the grand challenge of self driving cars. Since its founding 6 years ago, Ghost Autonomy has raised more than $200,000,000 to pursue a standalone software primary strategy, building an integrated driving system on commodity hardware and working toward long term strategic partnerships with car makers. Now today, as we all know, self driving cars are on the road but are not quite ready for prime time at scale for a mix of reasons, both technological and sociological. On the technology side, one of the major challenges has been figuring out how to handle all the unusual edge cases that humans navigate with common sense. And that's where Ghost's new partnership with OpenAI comes in. On top of the many low level, high frequency, but fundamentally very narrow components that they've already built, the plan now is to use a large multimodal model complete with general world knowledge and common sense reasoning ability to make the high level decisions and tell those lower level systems what to do. In this conversation, John shares an overview of the autonomy space, including the different kinds of technology stacks that companies have developed and the strengths and weaknesses of each. The details of their new plan to combine super specialists and highly generalist systems into a single human level AI driver, and his perspective on some of the nontechnical challenges that companies in this space still have to overcome. As noted in my recent conversation with Flo from Lindy, I am very much an accelerationist when it comes to self driving cars. And I would love not only to see this technology work, but to see society embrace it for the life and time saving value that it promises for all of us. As always, if you're finding value in the show, we'd appreciate it if you'd share it with a friend. And, of course, we welcome your feedback. Our email is tcr@turpentine.co, and I am easy to find on just about every social network. For now, I hope you enjoy this window into the historical challenges and the new opportunities in self driving with John Hayes of Ghost Autonomy. John Hayes, founder and CEO of Ghost Autonomy, welcome to the Cognitive Revolution.

John Hayes: 4:08 I'm happy to be here. Let's get into it.

Nathan Labenz: 4:11 Yeah. Excited to talk to you today. So big news coming out of your company, and we will get into this in plenty of depth. You are announcing today a partnership and an investment from OpenAI to bring multimodal models to the self driving space. Maybe give a quick high level on that, and then I want to kind of take it back to the beginning of how you got into this and started the company and the journey that you've been on. And then we'll kind of circle back to how the addition of this new technology is gonna change the game.

John Hayes: 4:41 This is the culmination we've been chasing OpenAI since March when they originally announced GPT-4 with the multimodal, and it wasn't quite ready at the time. But what we saw was a real change and almost a reset in how autonomy for vehicles, and I think most of robotics, would be built. And we can go back to the beginning about how it used to be built. I think it's gonna be very different in the future, but this is just another area where AI is basically destroying whole areas of computer science. And so I think we're seeing stuff out of Tesla that's sort of acknowledging this reality, but I think we're the first company to put a multimodal LLM directly on the driving path. And that's very exciting, and I think everyone's gonna be doing that sometime soon.

Nathan Labenz: 5:30 Cool. Well, yeah. So let's go back and kind of do a reprise of the history. Just, I guess, first of all, briefly on yourself. I understand you're a serial entrepreneur and a technology polymath who has done a few different things. Company before this was called Pure Storage, and that was a 10 year journey. And then you've moved from that into the self driving. So I'd love to Is there a connection there other than just kind of, Hey, I want to go do another 10 year super hard thing that led you from one to the other?

John Hayes: 6:01 There isn't that much of a connection, but there is a connection that we'll go into. But before Pure Storage, I first moved to California to work for a virtual world company. That was the first big startup I joined. I was at Yahoo for a couple years. We were doing talent search and then went into Pure Storage. And a lot of that was just sort of connections. I was very, very interested in the underlying technology. There is a thread that connects Pure Storage, which is a data storage company, to self driving. And the insight that we had at the beginning of Pure Storage was we're looking at Flash. And this was turn back the clock to 2009, and Flash was taking over the consumer space. So you had the MacBook Air was just released, so the most expensive computers had Flash in them. And, also, the $100 laptops had flash in them. And then all the mobile devices had flash in them. And what was obvious was that flash was just gonna completely take over the entire consumer space, and our bet was that it would also take over the enterprise space. And so we spent a lot of time taking that product out of the consumer space and putting it into the data center. And it took a lot of smart software and a lot of engineering to make that work. But, ultimately, we were proposing a product into the market that was about the same cost as the storage system you're buying, but it's 10 times faster. And it's much more reliable and much more energy efficient and all the benefits you get from just incorporating new technology. I wanted to apply some of that same formula to get into autonomy. And so I live in Mountain View, California. So I've been surrounded by autonomous vehicles for almost a decade. And there are 2 things that were sort of notable. One is that there was a lot of utopian thinking. So this is back in 2017 that people would just stop owning cars. And if you look at how people actually drive or how they actually use cars, that didn't seem very likely. People don't use taxis very often. It's about 1% of trips are taxis, and you throw in all of public transit. You get another 1%. And so you have this very, very concentrated use of taxi mainly in large cities. And then you look at the total use. It's, well, if you have a child under the age of 8, it's you have a car seat. You have stuff in your car. It's an extension of people's homes. But then the other thing was they're building very, very expensive products. They were taking and the sensors and LiDAR are very expensive computers. And so it seemed like those companies were missing 98% of the market for trips and that there wasn't gonna be a way for them to easily scale down. And so, again, we bet on 2 big technology trends that are seem almost undeniable. I'll talk about some of the objections. One of them was that we bet that mobile technology, mobile cameras, mobile CPUs, would just keep getting better because you have massive investment in those the sensors. If you look at the new release of a phone, the only thing people care about is how are the pictures better. That's it. They don't even show the front of the phone. They don't show the apps. People take that for granted. They care about that 1 sensor. And then the other thing was that AI was going to clobber all of the technology that had been developed to date. And so if you look at the people who started basically every autonomy company, look at the people who started Waymo, Cruise, Neuro, and companies that have left us. They were actually derivative of the DARPA Urban Challenge from 2004 to about 2008. And the approach that was taken in the DARPA Urban Challenge, the assumption was, well, we're gonna have a map. We're gonna map out a city. We're gonna tell you exactly where everything is. You're going to use LIDAR to look for mobile objects, like car sized mobile objects. You're going to use cameras mainly to tell what signal states are. Because in 2008, you just didn't have good image recognition. And now you just have a simple problem where you just have to solve whether you can keep going or not. And that's been the foundation of robotics. It's this idea that you have to solve what track you're going on. Well, 10 years into it, it's clearly not that easy. And when we started in 2017, we thought that they're not gonna do without a big technology reset. The prevailing opinion was that the sort of robotics approach as they sort of incorporated small amounts of AI into specific things, specific recognition, specific maneuvers, specific predictions that they were just gonna grind, and they were gonna get it done. It was gonna happen any day now. At the same time, there are a lot of criticisms of AI, which are similar to the criticisms today. One is it's opaque. It is not necessarily reliable. Today, we worry about that LLMs have hallucinations. It's, yes. Everything does. My assumption at the time is very, very large bodies of software are also opaque. No one really knows how they work. All we do is we test them, and we see whether they work or not. And models are basically the same. And so what I assumed is that AI was just getting better and better. It was just knocking down all sorts of problems, from game playing to voice recognition to image recognition that had been subject intense subjects of study in computer science for decades, like 40 or 50 years. And it was being replaced with, make a model to do it. And so our assumption was you could put these pieces together. One is you would take all of your off the shelf sensors coming back to consumer because they would just get better and better. You would connect them to AI because you would get all the information you need out of the scene. And that sort of model of building a product would be ultimately far superior to any human written algorithm for planning on where you're gonna drive.

Nathan Labenz: 12:08 Yeah. That's fascinating. There's a lot there, obviously, to continue to unpack. Just kind of setting the level because this is actually the first full self driving not full self driving, full first episode we've done dedicated to the self driving problem. We've just done one episode in the past on autonomy more generally, and that was with a guy from Skydio. And as you're talking, definitely one thing that really overlaps with their project is they've been in business a little longer. They had developed a very deep technology stack that operates at multiple different orders of magnitude in terms of cycle time. And this really kind of crazy mechanism for converting high level instructions down to the super lowest level control systems within their drones. And it's all fully explicit. But as you said, that doesn't mean it's transparent. It's not a learned system, it's an engineered system. But still at the end of the day, nobody can really read the code that it generates because there's just too many layers and it becomes kind of opaque in its own way. So I think that's kind of a fascinating phenomenon that kind of keeps popping up in some of these deep tech problems. I understand that there's kind of a big division and you're kind of alluding to it with the LiDAR type systems, expensive, specialized equipment versus the kind of consumer quality hardware commodity essentially cameras, right? That are kind of at the heart of some of the systems like the ones you're building. Could you kind of give us a little bit more of kind of a Zoom job, just like taxonomy of who's out there? I mean, I think people know names in the game of your cruises and your waymos, but I don't think people have a sense for the fundamentally different approaches that are kind of on the road today and how that's kind of leading into the future. Hey, we'll continue our interview in a moment after a word from our sponsors.

Nathan Labenz: (12:08) Yeah. That's fascinating. There's a lot there, obviously, to continue to unpack. Just kind of setting the level because this is actually the first full episode we've done dedicated to the self-driving problem. We've just done one episode in the past on autonomy more generally, and that was with a guy from Skydio. And as you're talking, definitely one thing that really overlaps with their project is they've been in business a little longer. They had developed a very deep technology stack that operates at multiple different orders of magnitude in terms of cycle time. And this really kind of crazy mechanism for converting high-level instructions down to the super lowest level control systems within their drones. And it's all fully explicit. But as you said, that doesn't mean it's transparent. It's not a learned system, it's an engineered system. But still at the end of the day, nobody can really read the code that it generates because there's just too many layers and it becomes kind of opaque in its own way. So I think that's kind of a fascinating phenomenon that kind of keeps popping up in some of these deep tech problems. I understand that there's kind of a big division and you're kind of alluding to it with the LiDAR type systems, expensive, specialized equipment versus the kind of consumer quality hardware commodity essentially cameras, right? That are kind of at the heart of some of the systems like the ones you're building. Could you kind of give us a little bit more of kind of a zoom out, just like taxonomy of who's out there? I mean, I think people know names in the game of your Cruises and your Waymos, but I don't think people have a sense for the fundamentally different approaches that are kind of on the road today and how that's kind of leading into the future. Hey, we'll continue our interview in a moment after a word from our sponsors.

John Hayes: (14:07) Okay. So let's start with the classic autonomy stack. And this is very derivative of the DARPA Urban Challenge, but, of course, it has become, you know, 1000x in many dimensions and elaborated. And so you start with your base is you start with a map. And the purpose of the map is to tell you the rules of the road and kind of tell you the rules that other people are expected to follow. And so when you identify, you know, another vehicle in the environment or another person in the environment, you can use their position, place it on the map, and take a guess as to where it's going to go. And so you divide that then into two steps. One is your localization. Where am I on the map? And you can use GPS, but the people who are really serious about it need really real precision use the LiDAR to position themselves on the map. So what they've done is they've gone and done a 3D scan of every single city or every single area they're going to drive so that they can measure the distances to surrounding buildings and trees and other artifacts to figure out where they are to a high degree of precision. And that's ultimately once they can identify that on the map, they can use the relative position of other actors in the scene and put them on the map as well. And that would be in combination, the perception layer. Where am I? I take all the facts on my map. Maybe I do some additional perception to double-check that my map still roughly corresponds to the environment. I look for other actors, like other mobile things that maybe there may not be there. You have to then distinguish from, like, a parked car or a moving car or part of a person or not part of a person. The second step of that perception is called prediction where you say, okay. I'm running my system maybe at about 10 hertz. That's a pretty normal condition. I'm going to take, you know, maybe up to a second. I have a pipeline where I'm accumulating facts about how the world is evolving. So what I want to do, knowing that I'm in a vehicle, I can't stop instantaneously. This is one of the things that really divides robotics from other applications is I have momentum. So I have to make decisions now that I can't reverse at a later time. So the first step in doing that is to predict where all of the other actors are going to go. And I use their position, and I use past profiles. And often, this was one of the first applications of machine learning in these stacks was, hey. I need to build a model where when I see a car, I have a pretty good guess as to where it's going to go. And some of that the initial conditions for that are also encoded in the map. So if I've driven around, I survey the same city a bunch of times to try and see how do people actually behave in that zone. The next phase I go through is planning where it's like, okay. I have all these facts that have been expressed about how I want to move through the world. Like, what is my goal? How is everyone else going to move? And now what I want to do is plot out a path in world space as to where I'm going to drive. And this has been, you know, a huge subject of research because you can easily come to conclusions, and you see this still in the world where your conclusion is just to not move. Because when you think about what everyone could do, it's like, yes, people could just suddenly change direction and drive into you. And then if you think about, you know, probably what they will do, then if you start with simple assumptions, like people just go the velocity they're going, you often will not find openings that you could otherwise drive through. And so this sort of planning has a really complex interaction with prediction where you have to rerun all of your prediction to say, well, if I move in the scene in a certain way, I expect a reaction in order to move the flow of traffic. And then once you run that, that's often the subject of more AI, lots of dynamic programming, lots of search algorithms where I'm trying to plot all the possibilities and measure all the probabilities of what's going to happen. And then controls are probably the most straightforward after that point where I've already laid out where I'm going to go, and now I just have to tell the vehicle where to go. And when you're at relatively slow speed, that's actually a pretty easy problem. You don't have a lot of dynamics to work with. And so the companies that implement like this is like Waymo is the or, you know, originally Google self-driving car is sort of the granddaddy. It's like they started their project in, like I think it's 2011 they might have run an experiment. They're doing it. They started scaling it up in 2013. I think, you know, Cruise has solidified themselves as a solid number two where they started in 2015 with basically the same stack. There are smaller companies that have really, you know, come and gone where it was easy to get a demo up, but often they died there. And then there's the trucking company. So there are companies like Kodiak. The largest, the most prominent one is Aurora. And the trucking problem is even more complex. One, you have to drive on freeways. They tend to focus on being very, very good at prediction. So very, very good at perception often seeing hundreds of meters down the road. Very good at prediction because the assumption is that the truck is a very, very high momentum device. You know, you really do want to know guess how the world's going to evolve over a number of seconds and very good at traffic dynamics. So in the whole other world, you have a company like Waive. So Waive, they probably started about the same time we did, and they've always focused on throwing out that stack and trying to build some sort of end-to-end system. And so instead of breaking it down into stages where you often have completely different teams working on those different stages and negotiating the interfaces and trying to set up performance standards at each point, what they've been focusing on is, let's just make a supermodel and just do it end-to-end. So we take, you know, all of our sensors in, we throw it into the supermodel, and then we get a vector of where I'm going to drive out. Tesla has been an interesting case because you see them actually go through multiple of these serially. And so when they started out with FSD, like their FSD beta, they first tried to build a pretty naive solution without prediction. And then you can see them changing this through their AI days. Then they moved into actually building more of a classic prediction model and more of a solver and incorporating more models into that solving process. And then they haven't done an AI day, but in the biography and in some of Elon Musk's leaks, you can see they've done another reboot where they're moving more and more towards an end-to-end model. There's sort of been these two evolutionary approaches. So one is the robotics-based companies are kind of slowly incorporating AI. If you go back and you look at their presentations on how they use AI, a lot of it is concentrated in their simulation and test environments where they're trying to create profiles and get very, very good at predicting how other cars are going to react and then rendering that environment back to the car. And I think that they're slowly moving towards this idea that that should be the same system that runs in the car. And then you have the companies that are more radical, which is we want to run everything in the car, and we don't want to contemplate these stages that were defined decades ago when you had very limited access to AI.

Nathan Labenz: (21:51) So I'm kind of mentally mapping this into sort of a 2D, like a 2x2 kind of grid where on the one hand you have the axis of the hardware and seemingly the evolution is from super high although, I guess, there's still an active competition, right, because you see a Cruise running around today or at least until recently you did. I'm going to ask you a little bit about that too. But you see the thing spinning on the top of it, and that's a very expensive and specialized piece of equipment. Other entrants into this space have super advanced GPS, I understand, and super advanced accelerometer type stuff. And then on the cheaper, more scalable end is we're just going to use a bunch of cell phone cameras and let software do the hard part. And then on the other axis, I guess, is on the one end, and this one is maybe more clearly a direction in time. But going back to the DARPA thing, I assume at that point, you had very little AI, certainly nothing like what we would consider to be the deep learning AI that we're familiar with today. But in that scenario, you just have to break this problem down into a bazillion little pieces and have specialized software modules that deal with each one. And obviously, it becomes a real challenge to manage that code base and to figure out what to do in any number of situations. You open yourself up to a zillion edge cases. And then you can start to chip away at bits of that with AI over time and kind of replace modules or kind of, you know, create stitch together modules with kind of an AI layer on top. But then there's also this kind of big idea that, well, hey. What if we kind of threw most or all of that away and, you know, really embraced the bitter lesson and hope that it's, you know, sweet in this case and that, you know, that could kind of deliver the sort of end-to-end thing and, you know, smooth over a lot of those edge cases? How'd I do there, or how would you complicate that?

John Hayes: (24:01) I would simplify it by saying, ultimately, the question that people want to answer is how do I know it works well enough that I can deploy this in public? And so I think I would take your 2x2 and say there's two philosophies. One is if all my parts work, does the system work? And often this is rooted in the automotive practice of making safety cases where you say, hey. If there's a series of facts that I know is true, therefore, my product is safe. And then there's the measurement is the other axis where it says, I don't know exactly how every part works. All I care is I'm not going to try and rationalize why I believe it works. I'm just going to go out in the world and measure whether it works. And these to implement that correctly, you often have to end up in the more scalable domain. And so that sort of drives what axes you can do because if you're a company like Aurora, you can't drive a million trucks and just measure it. You have the trucks you have. And so you have to spend a lot of time trying to rationalize, does this work or not, based on whether the individual components work. And then if you're Tesla, which is, like, the exact opposite end where they're driving around a million cars, their philosophy is we'll just measure. Like, we'll put stuff out there. We'll get our, you know, trusted drivers, but we'll get thousands of them, I think. I don't know what the scale of their program is. I think it's thousands, maybe low tens of thousands. But then we're just going to watch what actually happens and get a statistical measure about how well it works and actually use that to identify where the real problems are.

Nathan Labenz: (25:40) This might be a good time to just run through the kind of levels of autonomy as well and give folks kind of a sense of how in the industry people think about these gradations, and then you can maybe kind of map on to those levels, like, these different paradigms as you go to.

Nathan Labenz: 25:40 This might be a good time to just run through the kind of levels of autonomy as well and give folks kind of a sense of how in the industry people think about these gradations, and then you can maybe kind of map onto those levels these different paradigms as you go to.

John Hayes: 25:59 There's really a divide in the industry. So there are in theory 5 levels. In practice, there's 2 sets of levels. And so level 0 doesn't exist, or that's like a car that was built in the forties. Level 1 is basically every car that was built before 2000. In the consumer space, we're very much in this level 2 domain. And what I consider the difference between level 2 and other cars is that the car measures the environment, usually through radars. Radars got big in the early 2000s and now through cameras, which got big a decade later. And the car can have a reaction to the environment that is independent of what the driver is doing. So lots of features like adaptive cruise control was sort of the original level 2 feature. Automatic emergency braking is a level 2 feature. All the way up to various forms of auto steering is a level 2 feature, and sometimes the industry calls this level 2 plus or plus plus or plus plus plus to try and break that down. So the divide happens at level 3 and level 4 and level 5. So let's knock off level 5 because it may never exist. So that's like some fully autonomous car that somehow knows what you want and can handle any scenario. We're very much talking about the level 3 and level 4. Now there's a huge community that says level 3 doesn't make any sense because you have a human monitoring it. And the car has to have a good enough sense as to whether it can operate correctly that it can tell a human observer whether they should or should not pay attention at a given time. And so the first company to bring out a level 3 product or market something as a level 3 product is Mercedes. There's actually debate about whether it's a level 3 product or level 2 with a lot of pluses. BMW has announced a level 3 product. Acura in Japan has announced a level 3 product. And the main difference in consumer experience is that instead of being nagged to always watch the road or touch the steering wheel to prove you're paying attention, the car can say, I've got it, and your eyes or your hands can do whatever they're going to do. Like, you could read a book until it beeps at you and attempts to get your attention, at which point it turns back into a level 2 vehicle. So Mercedes' application of this is a traffic jam assist. So it only works under 40 miles an hour. You have to have all the lanes around you filled, has to be mapped. You have to have compatible weather. When all of those conditions are true, the car will go into a level 3 mode, and it will creep along. And I think that the product that BMW has announced is quite similar. So level 4, the only companies that have claimed to be level 4. So the main difference there is what's called an MRC, a minimal risk condition. And so a level 4 car doesn't have to handle every single scenario. But when something is not being handled correctly, it has to resort to a minimal risk condition, meaning that most of the time, it just means it stops. Like, it stops in place. That's the simplest minimal risk condition. In urban environments, that is mostly acceptable except when you stop somewhere that is disruptive to other people around you, which we've seen in the news. But that's actually pretty good because it avoids certain types of liability. On a freeway, because you can't just stop in place on a freeway, it would have to involve some form of pulling over to the shoulder. And so for a car to go from level 3 to level 4, what that means is if there's a person in the car or otherwise responsible for the car and the car may demand attention of that person, but if that person never responds, the car will still go into its minimal risk condition. And so that's the technical definition. Now colloquially, people often think of level 4 as robotaxis, meaning that they're automated enough that there's never a driver. I can get into the car. And the way that the regulators are going, like NHTSA, for example, is they're just trying to explain that divide. Is something assistive, meaning that you always have to pay attention? There's no condition under which you're not driving the car. Or is it automated, meaning there are conditions under which you don't have to look out the window if you don't want to. You don't have to touch the steering wheel if you don't want to. And so everything marketed today is assistive. The government and other agencies and companies are exploring how would you know whether an automated system was good enough to market and call an automated system.

Nathan Labenz: 30:31 We'll continue our interview in a moment after a word from our sponsors. I want to tell a little bit more about what the robotaxis are really like today. I think very few people still obviously have taken a ride in one recently. Unfortunately, I'm living in Detroit. We don't have any robotaxis on the road here, even though we are the auto capital historically. I was kind of surprised to learn in prepping for this just how limited the scope of operation still is for a robotaxi.

John Hayes: 31:03 The robotaxis are developing a new product, so it makes sense to limit the scope until you know whether it works. And so what you saw in terms of scope is they're not great at weather. There's a reason Waymo deployed to Arizona because the weather is, it's extremely dry. It's also really hot, meaning that there aren't many pedestrians out, and so they're taking measures to really limit their exposure. And early on, Cruise was only operating at night. I think Waymo also started out operating at night just to have less traffic activity and less people working around. Every person I've talked to that's been in a robotaxi said, it's actually pretty good. And to the point where, kind of like an Uber, probably even better than an Uber, you stop paying attention to it. Like, you watch it for a little bit, and it's kind of neat, and it shows you what it's seeing. But you quickly relax, and you're like, this thing isn't going to screw up too badly. And so it goes along. And so in the domain where the public is taking rides, in San Francisco, now Waymo is the only operator. They're geography limited. Like, it can't be too steep. I think they're avoiding some of the most crowded areas or some of the most crowded times of day. But the general impression is that it works really well in those situations. I think the other development that's been happening is people have been paying a lot more attention to China. So China has a very vibrant both robotaxi and autonomous driving consumer driving market. And so there, people are also pretty comfortable with how those products work. And they're kind of in a hybrid state where a lot of it is robotaxi. Almost all the companies there started with the same architecture. They're moving into more consumer, incorporating more cameras, smaller computers, and such. But there, the impression is that the automated driving is actually better than the average Chinese driver. So people feel pretty comfortable about using it. And I think that that's both positive, but it's also perceived as the biggest risk, which is that people quickly get complacent. And so if you had something that worked 99% of the time, it's going to work 99% of the time. And if you're sitting there watching it, you quickly stop watching it, and then all of a sudden, it'll do something weird. It'll get into an edge case and do something weird, and you weren't expecting it. And so that's both, you know, there's considered to be an uncanny valley in the robotaxi world where as it gets closer and closer to good enough or ready to ship, people will get more and more complacent, and you'll actually see the rate of incidents increase. And that may be a little bit of what happened to Cruise.

Nathan Labenz: 34:03 So I've not been in the robotaxis, as I said, but I have had an opportunity to take a pretty long trip in a Tesla recently. A neighbor was kind enough to lend me theirs, and I took my grandmother on a what was an 8 hour round trip for me to take her back from a visit to our place back to where she lives. And it doesn't let you get too complacent. The nag system in the Tesla is honestly super aggressive. But I definitely could see that happening already in a Tesla too. I was very impressed overall by how well it worked. It seemed to me to kind of trust it. I had to watch it closely and I was also just watching it closely because I'm super fascinated by everything that's going on in AI. So I was kind of really trying to sense how good is this thing. But I was super impressed. I mean, we had 1 incident in my rides, to your point about weather, where it was a summer day and all of a sudden one of these kind of Midwest flash intense thunderstorms came up and dumped on us for a minute. And in that moment, it basically said, hey, it's on you for the time being. So I had to take over at that time. And there were a couple of other things where it did something a little bit weird, but I certainly never felt unsafe. And I definitely would rather have it driving me, even if I were going to get into that uncanny valley and really trust it as opposed to many people I've been in the passenger seat with. It did feel pretty solid. So I want to come back a little bit to the policy and societal stuff a little bit later, but I appreciate you kind of taking all this time to set the stage and give us the taxonomy and the different approaches. Where is Ghost Autonomy in this whole landscape? Obviously, we've talked about your consumer cameras being one big part of the bet, but I'd love to learn more about the architecture as it's developed and as you see it evolving. You can address that in any number of ways. One I always find really instructive is just mapping inputs to outputs. I think kind of for any AI system, like, what does it take in? What does it put out? Maybe there's some intermediate interpretable stages in between as well. But I'd love to kind of from potentially multiple angles, but definitely including that one, kind of understand the stack that you guys have built.

Nathan Labenz: 34:03 I've not been in the robotaxis, as I said, but I have had an opportunity to take a pretty long trip in a Tesla recently. A neighbor was kind enough to lend me theirs, and I took my grandmother on what was an 8 hour round trip for me to take her back from a visit to our place back to where she lives. And it doesn't let you get too complacent. The nag system in the Tesla is honestly super aggressive. But I definitely could see that happening already in a Tesla too. I was very impressed overall by how well it worked. It seemed to me to kind of trust it. I had to watch it closely and I was also just watching it closely because I'm super fascinated by everything that's going on in AI. So I was really trying to sense how good is this thing. But I was super impressed. We had one incident in my rides, to your point about weather, where it was a summer day and all of a sudden one of these Midwest flash intense thunderstorms came up and dumped on us for a minute. And in that moment, it basically said, "Hey, it's on you for the time being." So I had to take over at that time. And there were a couple of other things where it did something a little bit weird, but I certainly never felt unsafe. And I definitely would rather have it driving me, even if I were going to get into that uncanny valley and really trust it as opposed to many people I've been in the passenger seat with. It did feel pretty solid. So I want to come back a little bit to the policy and societal stuff a little bit later, but I appreciate you taking all this time to set the stage and give us the taxonomy and the different approaches. Where is Ghost Autonomy in this whole landscape? Obviously, we've talked about your consumer cameras being one big part of the bet, but I'd love to learn more about the architecture as it's developed and as you see it evolving. You can address that in any number of ways. One I always find really instructive is just mapping inputs to outputs. I think for any AI system, like, what does it take in? What does it put out? Maybe there's some intermediate interpretable stages in between as well. But I'd love to understand the stack that you guys have built.

John Hayes: 36:35 We're building on consumer cameras, but also more importantly, consumer chips. And we wanted to make sure that there was no magic hardware in there because our assumption is, you start from cameras and you start from chips, and what's nice is those just get better every year. And so then it comes down to, how are you going to solve some of the interesting problems of autonomy? And so if you look at the consumer space, there's a lot of image recognition. Does something look like a car? Can you guess the distance? And one of the really, really powerful things about LiDAR is that you get an actual measurement of distance. And so to solve that problem, we ended up building a stereo system. Now stereo systems have traditionally been built with custom chips first. There's actually quite a few companies out there that do that. And also, they have a lot of difficult calibration problems. So if you think about you have two cameras. Your pixels are literally a micron in size. If the system distorts by a micron, it won't line up anymore. And so what we spent time doing is we wanted real measurements of distance. We wanted to get sort of the very nice universal behavior that you get out of LiDAR, but we wanted to build that on cameras. And that was one of our first AI problems to solve, which is how do you just take two cameras that are connected to a computer and develop real measurements of distance? And so a lot of that was filling in what gaps does the lack of LiDAR leave behind. And so we believe that provides a lot of the same safety that you get out of a LiDAR system without having to put unique hardware on there. And actually, it's a lot better because we can run that 30 frames a second as opposed to 10, and it's much, much higher resolution. A LiDAR, a very expensive LiDAR is like 64 lasers. That's 64 vertical lines, and we have a 12 megapixel camera. So we have 1000 vertical lines. It's not even comparable. But the other thing is that we wanted to build something that was very adaptive to the environment. And so what that meant is we wanted to jettison a lot of the complex planning that goes on and a lot of the prediction. And so if you're driving on a freeway, often you're trying to predict because you're slow at decision making. And so we basically just ramped up the speed. We made it run about 10 times faster than most stacks, and a lot of that was focusing on software that would make decisions very, very quickly from pretty small amounts of information on where to go. And if you're wrong, you're only wrong by about 30 milliseconds, which is traveling a meter down the freeway. But that became a core assumption: let's make this fit into a consumer chip by getting rid of a lot of the compute, and a lot of the compute is that complex prediction and planning framework. And so what we did is we incorporated much simpler planning framework where, you know, you can only see so many cars. There's only so many lanes. You're on the freeway. And so as long as you can make decisions extremely rapidly, you should not have to predict very far in the future. And if you're predicting a short time into the future because you can't read the minds of the other drivers, there's a limited number of things they can do because they're in vehicles that have momentum, and so they can't change their acceleration that much over very short periods of time, like hundreds of milliseconds. That was sort of the core software philosophy of, how do you get started to make something work in a consumer platform without providing the same level of safety, without having to implement a lot of the complexity. What I've been looking for for the entire life of the company is you see a company like Waymo. It's like they have thousands of engineers. We have about 60. How do you build something with a different philosophy that gets rid of your need for having 1000 engineers? And speed is a huge part of that. The next thing is we built total symmetry into the system. So everything is recorded on the same hardware platform. And so the goal there was, again, to fit into the consumer form factor. We didn't want to have specialized survey vehicles that slow down your mapping of the environment. So instead, everything is recorded in the system, in compressed video, uploaded all the time. Obviously, in a real deployed system, most video is highly uninteresting, so we would not upload it. We'd be more selective about what we upload. But we forced ourselves into a single stack, a single software stack that can provide all the inputs for learning and all of the execution and do that simultaneously so that we could shorten the time between going on a drive, discovering a problem, and actually fixing it and then trying again. And in the end, we do multiple software releases a day in order to test different combinations and different algorithms. The story of the company has just been making that better and better, so that you have a really, really good highway experience. The other thing that we've incorporated from full self driving is that we don't ask a lot of questions about how you want the car to drive. And so what you see in most cars out there today is they actually literally ask you what speed you want to go and how far you want to follow. And so what we do is we determine that from the environment. So we're looking at all the cars. We're getting their profile on stereo camera. We're getting a profile on radar to estimate all their velocities. And so we look at the total traffic pattern to actually just choose a speed and choose a following distance so that you never have to adjust it. We spend time looking at competitive consumer products, and even Tesla, if you go through different types of traffic, like you go from stop and go to thick traffic to thin traffic, often you do make a lot of adjustments. And so we're building the same sort of anticipation and the same sort of self reconfiguration that you would have in a fully autonomous vehicle. And that sort of stuff is actually pretty easy, but it takes work. But it's not very compute intensive, and it dramatically improves the experience. It makes the car feel like it's floating in traffic as opposed to something where you've done some settings and it's going to slavishly execute them.

Nathan Labenz: 43:00 I'm always kind of interested in how AI ends up being kind of alien intelligence to us and often ends up doing things very differently. So I think it's really interesting right off the bat to just look at this notion of, okay, can we make the problem really small by making the performance really good and then being able to increase the frequency? That in and of itself, I think, is a really fascinating approach because it's like you're kind of leaning into a superhuman aspect of computers, which is just obviously that they run super fast. And if I understand correctly, then kind of saying like, well, knowing that we're going to be back at this again 30 milliseconds from now, you know, we don't have to do super high level stuff nearly as much because we're going to get another update really quick. So if things start to go weird, then we can kind of react to it more quickly versus having to anticipate it. Am I understanding that right?

John Hayes: 44:06 I think that's true. I think the other thing is when you talk about inputs and outputs, part of the reason that AI is alien is because it's going through a filter of what do you think you need to train it on. And so this is where a lot of edge cases come in, is that people are not very good at guessing what are salient inputs to any model. And so for any model that gets developed, it's like our stereo model was pretty simple because the output is just a disparity between two cameras. And so you have a lot of physical parameters for how the cameras could be moved relative to each other, how they could distort relative to each other. For things like, what is the, where are the lanes? So it's like, where are people driving? There's a lot of factors in the environment that are just very hard to guess. And so, you know, part of the reason to be fast is because guessing all the things that could occur in the environment over a very short period of time is much easier than guessing what could occur over a longer period of time. And so to some extent, we've adopted a bit of a barbell strategy. It's like, well, for the models that we put in the car, we want to have a development loop that allows us to quickly sort of guess and have a reasonable number of things that don't require, you know, much more than a single person's intuition for how you should formulate your training sets. And then the other side of it is going all the way to LLMs that have incorporated all human knowledge and almost all human text and all human images so that you don't have to guess what is salient in a scene. Like, simple example, we found that you can be pretty good at guessing where lanes are, but performance is degraded by the presence of trees. It's like, okay. Why is that? And some of it was the shadows, and it depended on the time of day and depended on certain aspects of glare. But the model in some sense was fairly dumb because it assumed a type of illumination on the road that can be disturbed by things that are off the road. And so you're just constantly editing how you train the model to incorporate all of this common sense for how does a real road actually appear? How does it appear when it's in the middle of construction, or how does it appear when there's been just a mess? We found roads where there are literally white paint spills, and it's like, okay. That's just something that occurs. And how can you take all of the information in the scene and feed it into a model? And I think that kind of goes unsaid as a challenge of making any point model is it's very hard to take the input and output and extract a subset that is different from the totality of common sense that you could have around that input.

Nathan Labenz: 47:01 You mentioned radar. So is radar also cheap? I wouldn't have intuited that there would be consumer affordable radar systems, but maybe there are.

John Hayes: 47:13 Well, radar is put on like 30% of cars that are built. So it is cheap. It's been deployed on cars for many decades, and it's completely solid state. And so as far as something that is sort of a known thing that can be put on a car, and it can be made cheaper. And so it's outside of our usual profile where something that is built exclusively for the auto industry probably won't be cheap in terms of its price, but it is still a well known thing and noncontroversial to put a radar.

Nathan Labenz: 47:47 It's already standard, so be it. Is there a kind of intermediate, you know, just kind of piecing together how this whole system works. Like in the Tesla, there's the sort of, you know, 8 cameras around. Right? And then they synthesize all that into a 3D rendering of the scene. And then I'm looking as I'm riding along in it, I'm looking at that 3D rendering, and then they're kind of layering the planning onto the 3D scene, and I can kind of see all that as well. Do you also have a human interpretable visualization layer or some sort of intermediate layer where I can inspect how it's understanding the world around it? Nathan Labenz: 47:47 It's already standard, so be it. Is there a kind of intermediate, just kind of piecing together how this whole system works? Like in the Tesla, there's the sort of 8 cameras around, right? And then they synthesize all that into a 3D rendering of the scene. And then I'm looking as I'm riding along in it, I'm looking at that 3D rendering, and then they're kind of layering the planning onto the 3D scene, and I can kind of see all that as well. Do you also have a human interpretable visualization layer or some sort of intermediate layer where I can inspect how it's understanding the world around it?

John Hayes: 48:36 Yes. So the models that run in the car in some ways are pretty conventional. In other words, they tell you literally where are the other cars in the scene in the world coordinates, what is the configuration of the road in world coordinates. And so we do take that information, and we render it. And we do have a forward plan of where is the car going to go. The car mostly kind of follows the lane, so it's the same lane, unless you're making a lane change, at which point we draw the plan out for where the car is going to go. Because, ultimately, the output is a physical acceleration that has to be applied to the car, and we have to know what that goal is. So we do project that a couple of seconds into the future, knowing that it's subject to change if the facts in the world change. And we think that that's essential for anyone to trust what the system is doing is to have a rendering of the world that sort of exposes a bit of the thought process of the system that matches what people see when they put a when they look out the window.

That's another interesting part about moving more of the reasoning into LLMs because they think in terms of text. And so you can get more of a human readable reasoning. And so there's a demonstration from WAVE. But what they did was they trained a language model to directly interpret their machine learned models for driving. And so think of it as, you know, if the machine learned models are building a token space, you create that mapping into a language model so that you could ask it questions and answers about kind of what activations in the model, and can you describe those activations in terms of English. And I think that that's becoming an increasingly important tool because you can take a tool like TensorBoard, and an expert can look at what activations are causing what decisions. But this is a pretty interesting new tool that's enabled by open source where you can train a language model to directly interpret the activations as long as you can create a body of text or a body of responses that you can train that you can fine tune from.

Nathan Labenz: 50:40 Yeah. It's amazing how many interpretability breakthroughs are kind of happening right now. And also, I'm constantly fascinated by how bridgeable different latent spaces are one to the other. And so, I guess this is probably a good time to get back to the big news. So obviously, there's a lot of hard parts to this, but what is the frontier that has proven extremely difficult such that you were like, okay, I want to go use this more generalist kind of allegedly world model having system to help get over some of these humps. And how do you see that playing out now that you have this partnership underway with OpenAI?

John Hayes: 51:30 We've been looking a long time at how to do urban driving. And the problem is that everyone who's tried to do urban driving, they basically spent a decade banging their head against the wall. And maybe another decade, it'll work. But the main thing we saw is that they're incrementally incorporating a lot of common sense into their system, and that's just called edge cases. So the industry talks about edge cases when the system behaves incorrectly based on information that was in the scene but wasn't correctly incorporated into the drive planning. And so that's a wicked problem. It's like, right now, the only development path to solve that is to just do a lot of driving and continue to make your planning algorithm more and more complex. Like, you can add signals and add more recognizers.

So when we saw the original GPT-4V release, the first thought was, hey, we could really use this for labeling our planning to come up with different reasons that scenes happen to be different without creating special purpose recognizers. So when I mentioned that, hey, trees kind of matter. Well, that's a question that I can easily answer. Is there a tree in the scene? Is there a bridge in the scene? And we don't have to train a model to recognize that or build, say, a complex cross reference with someone else's database. But then what became more interesting as we experimented with it, realized it has a lot of common sense about driving. And so GPT-4 and GPT, if you start just asking it questions in text about describing scenarios and saying, what should I do? You start getting pretty sensible answers. And there's been a couple good papers of people making these textual virtual worlds where they describe how traffic is behaving in text, and then they get answers about how the car should behave. And so they do these sorts of agent like simulations to make that work.

But what was really interesting is that alone doesn't solve the problem of what are the salient features in the scene that could lead to a decision. And so we then started experimenting with the open source multimodal models, and we wanted to answer a simple question, which is, first, am I on a freeway? And it's like, you think it's an easy question, but it's not as straightforward as you think to answer from a map because GPS has a lot of unreliability. It especially has unreliability around on ramps and off ramps. So knowing precisely that you're on a freeway versus not on a freeway versus something else is something that would be really good to know precisely, especially from the visual field. And there's other facts that you could ask. And starting with, I believe we started with LLAVA and T5, you know, it would get it right about 60% of the time. But with fine tuning, that actually went to about 98% of the time. And that fine tuning set was probably a few hundred examples of showing you different street scenes and then defining a bit more about what we meant by freeways.

And so putting this together, we came to the conclusion that there was a totally viable path instead of training a bunch of special purpose recognizers that you would just go to GPT-4V and ask what the car should do. And that opens up a very interesting world because you express your navigation goals, maybe use a RAG or something to find relevant traffic laws, but then you literally just show it the images of what's around you. You augment those images with things that you can measure directly from your sensors. Like, perhaps you put boxes around cars, you tell it what lanes you're referring to. And even we found complex environments like construction scenes, there's a flagman, there's, like, when you ask it how you should respond to that scene, you get a pretty sensible answer.

And so what this does is it divides up the autonomous driving problem. Instead of being a monolithic system that every hundred milliseconds has to examine the entire world, build predictions for the entire world, and then issue a command to the car, you divide it up into two systems where you say, hey, there's a high speed system in the car that's running at 30 hertz. It does all the safety things. It prevents collisions. It kind of knows what path it's following in terms of lanes. And then you have a much slower system that has been trained on the entire universe, has a lot of common sense built into it, and can instruct the car what to do kind of at a maneuver level. And so in this way, the car becomes an API that is instructed in terms of its maneuvers from an LLM that is examining the same sensory data. And to me, this represents a huge reboot of how autonomous systems are built.

To solve the analogy of that problem, if I want to add a signal, like, I see that there's something important to see. I'm driving around. I notice I have edge cases. I discover that there's some particular sign or some particular artifact in the environment that should have been incorporated into my decision. I would then have to go to build a training set, learn when I see that, learn when it's relevant in terms of its scene position, and then I have to edit my planning algorithm to sometimes incorporate that data and sometimes not incorporate that data. And aside from being slow, it's kind of like you're just going to do that forever. And so the inversion is to say, let's train on absolutely everything. You tend to get sensible answers. And you can ask it to explain the answers, and it explains the answers in terms of visual elements of the scene. And you're going to make a system that out of the box is dramatically more reliable.

Nathan Labenz: 57:17 It's amazing how well these visual things work. I just took my kids to Salem, Massachusetts the other day and tried it in ChatGPT and took a picture of this parking sign that had been taped to a lamppost. And it said, special temporary October parking, Salem PD or whatever. There's two stripes of tape through it and a couple different the sign itself was up there a couple of times. And I asked it to infer the context just from that image. And it was like, well, first it just reads the sign. And then it's like, Salem is known as the home of the Salem witch trials, and it's a big tourist destination in October. So in all likelihood, this refers to the fact that there's elevated traffic here in October. And so these temporary parking rules apply just for this time of the year and blah, blah, blah, blah. And it really is kind of striking to see how much value the whole kind of world model and historical, in this case, historical context can add to just a snapshot. Right? I mean, this is like, well above a what a caption might be able to do or an image segmentation type of thing. It is pretty incredible.

John Hayes: 58:41 And that's what people call hallucination. But hallucination is actually the best part of these models because the hallucination is where all the common sense applies. And so when you're asking to incorporate not just all the knowledge of the scene, but all sorts of the entire depth of knowledge that could be applied, you tend to get more answers that are closer to reality than what an engineer sort of sitting there having to express their thoughts in a programming language would be able to express.

Nathan Labenz: 59:11 So how does this then come to you kind of mentioned the two parts, right, of how this comes together into a new architecture. It's funny. I mean, I'm not a big analogy guy. I typically resist analogies to the way humans work for AI systems because again, think it's more alien than human like in many cases. I often say human level, but not human like when I describe GPT-4. But it is striking that there's kind of a system 1, system 2 thinking fast and slow kind of dynamic emerging here. How does the map fit into that? Is the map just another input into the GPT-4V? Is that how you imagine that working? And then I'm sure there's a lot of logistical issues too around, I would imagine you're not getting the weights. So I imagine that you're still making API calls, which creates latency and availability concerns. So this brings a lot to the table in terms of advanced understanding, but I'm sure it also brings some challenges. What's that whole picture look like right now? Nathan Labenz: 59:11 So how does this then come to you? You kind of mentioned the 2 parts, right, of how this comes together into a new architecture. It's funny. I mean, I'm not a big analogy guy. I typically resist analogies to the way humans work for AI systems because again, I think it's more alien than human-like in many cases. I often say human level, but not human-like when I describe GPT-4. But it is striking that there's kind of a system 1, system 2 thinking fast and slow kind of dynamic emerging here. How does the map fit into that? Is the map just another input into the GPT-4V? Is that how you imagine that working? And then I'm sure there's a lot of logistical issues too around, I would imagine you're not getting the weights. So I imagine that you're still making API calls, which creates latency and availability concerns. So this brings a lot to the table in terms of advanced understanding, but I'm sure it also brings some challenges. What's that whole picture look like right now?

John Hayes: 1:00:21 So it's interesting to ask about the map because we still do incorporate maps, and we do it in 2 ways. One is you license map data that's interesting and can tell you facts. And you think about constructing a prompt, you should include that information in the prompt because why not? It's free context. You can also think of the map as just an extended memory. It's like when you're driving around, if we're in analogies, it's like you don't go frame to frame. You have a memory of what this road connects to and how you should behave on it. And I think the map is a good storage for that. The other thing is that we use the map as a key to store media data. So if a road has been driven, we have images of that road from the car that drove it. We have the inferences that were made from those images. And so for pictures in the environment, one, it would be nice not to call GPT-4V because you pay for every single call for facts that are not changed. And so you can extend your memory and rebuild your memory as you drive down the road. And then the other thing is, like, not every picture is perfect. Maybe there's this time, there's a windshield wiper in the middle of the picture. You're probably better off using a previously stored picture than whatever you're sending. You will get some answers, but try and get the best answers that you can. And so I think that maps are part of an extended memory system of what are all the fixtures in the world and how do they connect. And when we're doing navigation, that's, you know, there's lots of APIs that do navigation, so we just want to convert the turn by turn instructions into specific maneuvers that are relevant to the scene. But then the other part of the memory is when if you're driving along the road, you see a sign or you see a fact, and then you have to retain that for a period of time, just like a person would retain that for a period of time. Because often, a big use of maps in autonomous vehicles is predicting the future, like predicting how the road is going to evolve by just reading forward. Well, the way we do it is we have a sign that tells us how the road is going to evolve in the future. And so to take advantage of that, we want to go from the picture, the observations we make in that picture. We then want to retain that for a period of time. We ideally retaining it just as text that is opaque to the program but, of course, interpretable by the language model. And then you're returning that back into your prompts over and over again. So in that sense, it's more like an agent model where you're updating a memory. Like, you have a long term memory that's your map, and then you have a short term memory that is observations that will expire as I travel along the road. And then logistical issues, it's like, well, number 1 issue is that GPT-4V is slow. You know, you make a call. It can take 10 seconds. It can take 30 seconds. But I'm extremely optimistic that those types of problems are solvable. So one is, yeah, we should put it on our own computers. We have an application that demands a real time response, and so we have to come up with a mechanism to manage the compute capacity to make that work. But then the other is, you know, I'm expecting that as well as getting better, the models will just get much faster over time. I mean, we saw GPT-3.5 got 10x faster. What we see in the GPT-4V in the alpha edition, it's like, I'm betting they were just trying to make it work. And there's certainly in the open source community a ton of research on different ways that you can make models faster. And I'm expecting OpenAI has not yet incorporated all of that. So I expect the models to get faster sort of 10x year over year, proportionately cheaper to call. We would love to just call the model more and more and more and just ask more and more questions, and it'll just get more efficient to run. So when we think about designing a product, you're on an auto company timeline. They want to make 5 year plans. What's the world going to be like in 5 years? I don't know. You don't know. Like, 5 years ago, you would not have guessed the world of today. But what we do know is that the models will get dramatic, like, Moore's Law will keep going on, but that'll be 1 axis of improvement. The software will improve. The model design will improve. And so everything will end up, like, a 100 to 1000 times faster and cheaper over that time period. And so that's what you want to aim for is let's assume we want to maximally call largest model we can to solve problems as a default. And then when we're forced to, from an architecture point of view, to solve something in custom model, for example, we're not going to embed this in a car. Like, we would embed much simpler models, or there are certain questions that we need answered very, very quickly, or there are certain questions that we think are extraordinarily simple to answer. If you're driving along on a country road and there's nothing around you, I'd rather have a model tell me there's nothing around, rather than go to, you know, call it the most expensive model in the world to answer an obvious question. And so there's a lot of optimization to make that work, but I think that problem gets easier over time. And the goal is always to make sure that you're enabling calling a really expensive model when you need to.

Nathan Labenz: 1:05:27 Just to develop my intuition for the architecture a little bit more. So I sit down in the Ghost Autonomy powered, and your kind of business strategy here is to partner in, as you just kind of mentioned briefly, to partner into the carmakers and become the kind of, I imagine it's kind of like, you need an answer to Tesla FSD, whether it's GM, Ford, whatever. We propose we'll be that answer for you. So now I sit down in one of these cars in the future when this is all developed, and I basically say, okay, I want to go to this address. Obviously, we've already got mapping applications that can plan out the high level route in terms of a line on a map at a great remove. And then the kind of second by second or ultimately has to be kind of sub-second decision making of what should I do now? The vision there is to take advantage of all these trends and also develop your own optimizations so that you can basically say, here's my high level goal. Here's where I'm at on the map. Here's what I see out the window in front of me. What should I do now? And then the language model, I guess, maybe has both a sort of reasoning and a maybe executable output. I mean, is how I would do this if I was doing an agent today. I would kind of have a chain of thought and it would say, well, I see that it's open road in front of you, or I see that there is a car in front of you and to the right. So therefore, I think what we need to do is slow down and then get over behind it or whatever. And then maybe also that it gets expressed as code, like tap the brake and then in 1 second move over to the right or whatever. And then all that stuff gets fed down. You mentioned car as API below, so that's kind of where I'm getting that from. And then on the car is where that stuff gets executed and is also subject to any number of safety overrides, where it's like, hey, you told me to go right, but I'm sensing something right, so therefore, I don't do that. And instead, I've sent some alarm back up that's like, can't execute the command because here's why. How am I doing? Am I mapping out that loop?

John Hayes: 1:07:56 That is the main loop. And but I also think it's continuous. And so you can think about the user's intention at multiple levels. So one, like the product as it works today is if you're driving the car, meaning you're sort of, you know, giving an input, it's a 100% conventional car. Like, car just does what you tell it. It doesn't fight you. If you let go, it kind of has a default plan, which is keep following the road. And so we can think about it. It's like the car is just autonomous if you're not participating in it. Now, you know, we can look in interior cameras. We can, I think there's a lot of clues for determining what does the person intend in this moment? And you can say it's like, how tall and what should the car be? Then you could think about, okay, what if I just give the car an instruction? I just say, take the next left. That's now I can just inject a turn by turn. And then there's the sort of point to point, which is, okay, you go to a mapping API. You get your set of turn by turn directions, and that's sort of fed down to the next level. And so I think you can do all of these things. And I think that you chances are you would proceed through those stages because you just sort of get going. And then when you think about it, it's like, have to look up an address. Like, there's data entry. I want the car to be driving itself already during that phase. And then I just want to edit what its intentions are. But, yes, then the car is an API. And you can be expansive about how you think about the car as the API. There's no reason that anything that is digitally accessible can't be also controlled by an agent using the total information from inside and outside the car.

Nathan Labenz: 1:09:40 So do I basically just sit down and kind of chat with this? Like, much like I can kind of chat with my ChatGPT app. It's ultimately like a conversational interface even?

John Hayes: 1:09:49 What we're putting together is the car should tell you, like, why it's doing what it's doing in plain language. I think that one of the ways that cars have gotten worse is that they've become more opaque in terms of options and in terms of indications. And if you've rented a car that has, like, a bunch of features on it, it will, like, beep at you and show you some symbol you've never seen and you don't know why it's beeping at you. And it has this very, very narrow communication path that doesn't actually incorporate information, doesn't incorporate what you're doing. And I think that the first simplification, which is let's put screens in cars, like, that was a start in that it started getting rid of lots and lots of sort of secondary and miscellaneous buttons. But it's still unsatisfying because there's still, like, deep menus, and the car just seems, like, harder to use because you don't have everything clearly labeled. The discoverability has gone way down. And so I think the next evolution is like, yeah, you should just chat with your car and tell it what you want. And that opens up the that experience to be, one, a lot more contextual. It doesn't have to be dumb. Like, the car knows where you are. Like, it has the full context of your scene. And so you could just refer to things in your scene or just it knows if you're in a gas station, or it knows if you're in a charging port. It knows if you're on an unpaved road. It knows, like, and you can make the experience just so much smarter by eliminating both all the buttons and eliminating all the menus.

Nathan Labenz: 1:11:23 Yeah. This is, I think this is going to happen for software across the board, you know, and even just something like Gmail, you get into the settings page and it's like, good God, good luck. And Salesforce and the Adobe Creative Suite and just all of these things have just gotten so bloated as they've kind of built out every feature that a power user could possibly want. And at the end of that, it's like very inaccessible to anybody who's not already been using the tool for years. I see this kind of as a wave that's coming to sort of all of our software interactions. Cars, we don't necessarily always think of it as software, but increasingly, that's a huge part of what they are as well. It is going to be quite an evolution over the next few years, I think.

John Hayes: 1:12:08 Well, I think, and you talk about something like Gmail. It's like the first thing I thought of was, well, the difference with a car is, like, the car has lots of external sensors, so it has lots of context that it can use to do what you're doing. The thing I wonder is this also applies to basically every AR application. It's like, where am I right now? Like, basic questions about where am I? What have you observed me doing as essential context for the total interface so that I don't have to be exceedingly explicit about what I'm asking for.

Nathan Labenz: 1:12:40 In this context of this partnership and your overall strategy, is OpenAI the only game in town here? Or, you know, should we expect to see, like, Google and Anthropic or a Microsoft come out with their own things and maybe start developing? We kind of see this in general, right, of this kind of big tech foundation model partnering into different verticals. How do you think that shapes up over the next couple of years?

John Hayes: 1:13:09 So today, OpenAI is the only game in town. If the criteria is an advanced multimodal model that you have an API for that you can call today, yes, it is the only game in town. And, you know, Google has made a lot of promises about Gemini, but then they don't ship it or they delay shipping it. Anthropic, I don't know what their multimodal plans are. I think everyone has to go down that path. Maybe there's probably a lot of business available to be exploited in terms of documents because corporations exchange documents. I think that the enterprise market will be extremely happy to absorb lots and lots more text processing in terms of, enhancing or automating white collar work. But I think the consumer space very much depends on multimodal models because the experience that people want is you have essentially, you have a phone and maybe you have glasses in the future, but what they want is not documents. Like, people, you can only enter so much text on a mobile keyboard. And so the way those interfaces go is you have to have less and less text and more and more context and more and more video and more and more media in order to make the next generation of consumer apps. And so I see what we're doing as an extension of that or, you know, a subset of anything related to robotics that has to understand the environment, anything related to AR. And I think that multimodal is ultimately the future. And I'm, you know, kind of surprised I haven't seen more from Google or more from the others. I think they're very much going after the enterprise market, which has, you know, been very, very good to OpenAI and other companies, and there's a lot of interesting work. But I think the consumer market wants multimodal.

John Hayes: 1:13:09 So today, OpenAI is the only game in town. If the criteria is an advanced multimodal model that you have an API for that you can call today, yes, it is the only game in town. And, you know, Google has made a lot of promises about Gemini, but then they don't ship it or they delay shipping it. Anthropic, I don't know what their multimodal plans are. I think everyone has to go down that path. Maybe there's probably a lot of business available to be exploited in terms of documents because corporations exchange documents. I think the enterprise market will be extremely happy to absorb lots and lots more text processing in terms of enhancing or automating white collar work. But I think the consumer space very much depends on multimodal models because the experience that people want is you have essentially a phone and maybe you have glasses in the future, but what they want is not documents. People can only enter so much text on a mobile keyboard. And so the way those interfaces go is you have to have less and less text and more and more context and more and more video and more and more media in order to make the next generation of consumer apps. And so I see what we're doing as an extension of that or, you know, a subset of anything related to robotics that has to understand the environment, anything related to AR. And I think that multimodal is ultimately the future. And I'm, you know, kind of surprised I haven't seen more from Google or more from the others. I think they're very much going after the enterprise market, which has, you know, been very, very good to OpenAI and other companies, and there's a lot of interesting work. But I think the consumer market wants multimodal.

Nathan Labenz: 1:15:00 I'd love to hear a little bit about kind of how you see the self-driving technology wave in relation to society. Listeners of my podcast will know that I'm very naturally both a big AI enthusiast and love building applications with it, love using ChatGPT. Don't really code anymore. I ask GPT-4 to code for me. At the same time, I definitely have what I would call a healthy respectful fear of just what might be coming over the next few years, because it does seem like we're into some pretty uncharted territory. But when it comes to self-driving, I clearly come down with the accelerationists who are like, man, this is already a dangerous activity. And it seems like we have some pretty good systems already starting to take shape. If you believe the statistics from Cruise or from Tesla, I'd be interested to hear if you think these are for any reason not credible, but it seems like the data suggests that they're already on par or better in terms of safety, at least in the domains where they operate. And yet we're so sluggish about it. And we have these kind of seemingly lack of enthusiasm among many would-be possible consumers and sort of a lot of heavy handed kind of overreaction type stuff from regulators. What's your take on why we're not more excited about this? Just seems kind of crazy to me, to be honest.

John Hayes: 1:16:37 So what we see with consumers is that most consumers have not experienced any form of self-driving. And so when they're asked their opinion, they're doing it kind of in a vacuum. It's like, yeah, I've heard this thing in the media. It's like, I've heard about it for a while, and it's not very personal to them. When they've experienced self-driving, they actually become pretty bullish on it. You know? As we talked about earlier, it does work. And it works well enough. And in urban environments where you're driving pretty slow, the consequences, like, the ability to stop really mitigates a lot of practical issues. I think regulators are a very unusual space because regulators aren't magic. Like, when you really get down to the nuts and bolts of, well, what regulation should we actually write? No one actually knows because you don't have an existence proof. And so in the automotive space, traditionally, regulations are only written, and usually they're performance-based regulations. They don't tell you how to build it, but they tell you what is sort of the minimum ratcheting standard, years or even a decade after a product is released. So any car today that has, say, autosteering, there is no governing regulation for that. And I think that that's the right way to go because I don't think anyone's really nailed it. Like, there's no formula that says that one of them is sort of a lot better than the others. And, you know, five years ago, I would have said Tesla is dramatically better than everyone else, but sort of traditional autopilot, I think they're pretty similar. It's like you get in a 2020s car from almost any manufacturer, and you can debate the style at which it performs, but the performance is kind of in line. This also motivated us to go to consumers because the adoption of autonomous technology would not be driven through robotaxis just because so few people experience taxis that there would never be sort of a groundswell that people really want it because it would just be this thing over there. There's no personal upside to them. Maybe there's some downside because there's some bad news. And so that's kind of how they form their opinion. So our belief is that you have to create great experiences for people that lots of people can feel personal about. Now when you talked about Tesla isn't there yet. And when you talked about your experience driving with Tesla Autopilot, what we see is that people bifurcate in how they react to the limitations of Autopilot. You have one population that gets really, really good at predicting when it's going to work and when it's not going to work. Like, they know down to the, I know it doesn't do well on this exact turn. I know it doesn't do well in this exact weather, this exact bridge, or this, you know, this sort of traffic condition. And they get good at kind of almost unconsciously activating and deactivating as they go on a trip. And it's still, you know, a very good product because it's active 90-plus percent of the time, and they learn to manage it. The other population just gets scared and never uses it ever again. And so they haven't worked out yet how to onboard people in a way that teaches them limitations. I think that to make it good, there's always a part of the consumer population. It's the classic technology adoption curve where the product really has to be perfect before the majority of people will adopt it. And so I think that's reflected in sort of the truth of where consumer products are in that they don't work 100% of the time. They're not perfect. And I think that people react to that more than anything else.

Nathan Labenz: 1:20:29 So how do you think companies should, in the meantime, kind of deal with this? I mean, we're talking just a number of days after, I guess it was maybe more like a month since the original incident where there was an incident with Cruise that basically got into an accident. Originally, not its fault as I understand the situation, but then it kind of made a pretty big mistake and dragged a woman under the car for 20 feet or something, and she was hurt. And so that's obviously a really bad incident. But I was kind of surprised that then they ended up voluntarily shutting down the fleet in other places. I mean, the California regulator, which seems a little heavy-handed to me, required them to shut down, but then they voluntarily shut down the rest of their fleet nationwide. And I was kind of like, where is our inner Travis here? Do we have no willingness to fight for our life's work in this space? Why is nobody kind of standing up and saying, hey, we are already safer. And yes, it's a terrible incident, and we'll do everything we can to make it better, but we're not going to shut down because of this. Are you guys crazy? We're already safer. This is a dangerous activity. We're making it safer. Am I wrong-headed for thinking that somebody kind of needs to stand up for the technology?

John Hayes: 1:21:45 They definitely need to stand up for the technology. I think that, you know, Cruise's situation was very much a government relations screw-up at every level. And I think they're trying to figure out how they screwed up so badly, which should have been the kind of incident that they could, you know, explain in public, that they could explain to the regulators. It was an unusual thing. Like, this isn't going to happen every day. They completely screwed up their media and their government relations. And so GM, their, you know, their parent corporation and by far biggest funder, is probably pretty upset about that. So I don't know what's going to happen with Cruise. I think that that's a problem that they can solve, and they'll come up with some way where, one, they fix the communication problem, and then they'll have to come up with some engineering problem they fix. And so if you look at what happened, the car was in a collision. You know? Maybe not their fault. Woman was tossed. And then it executed an MRC. And so instead of just staying where it was after something unusual happened, it decided that I'm going to pull over to get out of flow of traffic. Now Cruise has had a lot of negative press in the past about stopping in place and blocking emergency vehicles or stopping in place in really inopportune places. And so I think what's indicating is these sort of end states are an extraordinarily difficult problem to solve when you have to try and program a solution, when you have to try and predict everything that's going to happen ahead of time. And so I think they have to come up with some solution that has a lot more common sense about what is an MRC in a given complex scenario. I don't know if that's solvable with a conventional program. But to get that, it's like, they have to fight for it. It's like, it's that or shut down. I think they really do have a binary choice of we're going to try and reactivate as many jurisdictions as possible. We're going to go on our campaign, or they're going to have to just shut it down, which both are hard decisions for them to make.

Nathan Labenz: 1:24:10 Yeah. Well, shutting it down doesn't seem like, you know, it seems almost impossible to imagine from my perspective anyway. We've got the AI Safety Summit happening over in the UK, and we've got a lot of discussion of can we trust China at all? Or is it going to be a total AI arms race vis-a-vis China? You mentioned earlier that there's been a lot of progress in autonomy in China. I wonder if there's anything we could do to sort of create some friendly rivalry between the US or the West and China when it comes to some of these mundane but very big quality of life improvements, the introduction of self-driving cars. It feels to me like if we had some mojo, we'd be like, what can we do to really make this thing work? We're not going to let China get to self-driving cars and enjoy that incredible luxury before we do. We're going to go trim the trees that are blocking the road, and we're going to make sure there aren't any of these white paint spills that are kind of confusing. And we're going to, I don't know, put QR codes on all the road signs so we can more easily locate where we are. There's so many things that we could do. Right now, we're kind of like, we're not going to change the environment at all. We're not going to make anything easier. We're just going to make this work. And we're going to insist that you be at least 10 times safer than a human driver and come back when you've worked it all out. And I guess I wonder, did you think about that at all? Is there any way for us to kind of create a narrative of let's go win this? And are there any other things that you would think of that we could do as a society to try to actually make this something where there's a natural momentum toward it as opposed to just being like, well, hey, we don't care. We're not ready to listen to you until you've solved it all.

John Hayes: 1:25:56 The China situation is interesting because they have specific social forces that are pushing them down this path. So one is they're adding 20 million new first-generation drivers onto the roads every year. So, like, let's imagine we were adding 20 million effectively teenagers onto the road. And, you know, in a number of years, it'll get better, but that's actually happening. And so they're very, very concerned about are road deaths going to spike in a way that causes, you know, unrest or slows down development? Because, you know, if they're bulldozing whatever they're going to bulldoze to build that. They're also much further down the path of electric cars. You know, there's only two companies that are producing millions of electric cars. Tesla's one of them. BYD is the other one, which is a Chinese company. No one else counts. Like, you can add up all the rest. It doesn't count for anything. So what I'm seeing in the US is a sort of halting rivalry on the area of EVs. So when you talk to auto companies, they are absolutely consumed with EVs. And, like, how are we going to build it? Where are the minerals coming from? You know? I have to retool all my factories, retrain all my workers. I have to retrain my designers. I have all these mechanical engineers that maybe I don't need anymore. How am I going to get this into out of sedans, into my best-selling trucks and my best-selling SUVs? They're absolutely consumed. Whenever you go to any automotive conference, it is 1% AV and 99% EV. It'll come. It's just that the industry doesn't care about that yet because what they think is they're going to get killed on the straight-up EV front without first considering other transformations they should be incorporating. Regulators in the US are responsive to consumers. And so I think you have to go and you have to make a case to consumers first to say, this is why you want this technology. And then it can become an issue that percolates up through Congress and through the agencies as something that's worth prioritizing. A lot of that is what could you have? Right now, people don't see what's going on in China, so it's just not in their mind right now. I think we're a few years away from having that technology race.

Nathan Labenz: 1:28:28 Yeah. I must imagine you might need to just kind of go town to town across America offering free first rides to build the enthusiasm. That's like, again, the sort of VC subsidization of the early part of the adoption curve to then create that demand.

John Hayes: 1:28:46 Yeah. Was really hoping Waymo would do that. Like, I want them to go and more people go to Waymo or go on Cruise, and just experience it and then kind of ask. It's like, can I have this for me? They've been, you know, slow with their expansion for, you know, a bunch of reasons, but I'm really happy that the robotaxi companies, to the extent that there's not a Travis there who's pushing them, they still are expanding their domain, expanding their customer base, you know, giving people free rides, you know, working the press. And I think all of that is important to build an actual consumer groundswell and say we really need this.

Nathan Labenz: 1:29:28 Yeah. Well, sign me up anytime you need somebody to help write a congressman or whatever. You know, zooming out even beyond the autonomy realm, what's your outlook on the AI revolution from kind of the widest angle lens?

Nathan Labenz: 1:29:28 Yeah. Well, sign me up anytime you need somebody to help write a congressman or whatever. You know, zooming out even beyond the autonomy realm, what's your outlook on the AI revolution from kind of the widest angle lens?

John Hayes: 1:29:44 I think AI is basically retooling everything we know about computer science. We basically need a whole new generation, and our bottleneck for putting AI absolutely everywhere is just there aren't enough people who actually understand it. And the thing that makes me really excited about the generative AI is all of a sudden you've opened it up to a much, much wider audience where people can directly touch an AI system and get a response out of it without having to write a bunch of Python. I think the amount of creativity that's gonna unlock, we don't even know the limits. And right now, to get good results, you still have to be pretty good. You have to be the kind of person who's willing to write a lot of different prompts and really think hard about that. As the models get better, that'll become less and less important. And furthermore, it's like there should be multiple models where, you know, eventually, Google will ship Gemini. Like, it will happen. You will get new versions of Claude. It's gonna happen. As that occurs, these models will actually converge in their functionality just like CPUs have converged with their functionality over a long period of time. And I think that once that happens, you basically have this new computer where generative AI kind of mediates all the APIs and all the people, and it kind of doesn't matter. It becomes the totality of human knowledge and the superintelligence around that becomes itself a commodity that everyone can have access to, that there's enormous amount of competition on. And so that's a pretty exciting future because a lot of our deployment of computers has been limited to the number of people who can program a computer in a meaningful way. And we saw in the last decade, there's this idea, hey, we're gonna do no code. Well, now we have real no code. So let's actually see what happens because there's a real ability to say what you want in ordinary language and have a computer do something that was previously inaccessible.

Nathan Labenz: 1:32:01 Cool. Well, this has been a ton of fun. John Hayes, founder and CEO of Ghost Autonomy. I can't wait to ride in one myself. For now, thank you for being part of the Cognitive Revolution. It is both energizing and enlightening to hear why people listen and learn what they value about the show. So please don't hesitate to reach out via email at tcr@turpentine.co, or you can DM me on the social media platform of your choice.