WEBVTT 00:00:00.000 --> 00:00:03.520 Imagine you're driving home from work. It's, 00:00:03.520 --> 00:00:06.740 um, it's dusk. Right. The sun is just starting 00:00:06.740 --> 00:00:09.400 to dip below the horizon. Exactly. Casting those 00:00:09.400 --> 00:00:12.240 really long shadows across the pavement. Now 00:00:12.240 --> 00:00:15.439 for you, as a human driver, this is a pretty 00:00:15.439 --> 00:00:17.640 standard, relatively safe time to be on the road. 00:00:17.800 --> 00:00:19.780 Yeah. You just flip your headlights on, adjust 00:00:19.780 --> 00:00:21.859 your visor, and you don't even think twice about 00:00:21.859 --> 00:00:24.239 it. You really don't. Yeah. But if you are a 00:00:24.239 --> 00:00:27.530 self -driving car, This exact time of day is 00:00:27.530 --> 00:00:30.149 actually the deadliest. Which is wild to think 00:00:30.149 --> 00:00:32.770 about. It is. Statistically, autonomous vehicles 00:00:32.770 --> 00:00:35.229 are like more than five times more likely to 00:00:35.229 --> 00:00:37.229 crash when the sun goes down compared to human 00:00:37.229 --> 00:00:40.170 drivers. It completely upends our intuitive understanding 00:00:40.170 --> 00:00:42.710 of safety, honestly. I mean, we tend to assume 00:00:42.710 --> 00:00:46.310 a machine equipped with infrared sensors, radar, 00:00:46.509 --> 00:00:49.229 and 360 -degree cameras would practically own 00:00:49.229 --> 00:00:51.369 the night. Right. We assume the tech is infallible. 00:00:51.590 --> 00:00:53.950 Exactly. But the data tells a drastically different, 00:00:54.090 --> 00:00:57.130 much more complicated story. Welcome to today's 00:00:57.130 --> 00:01:01.289 Deep Dive. Today is Thursday, March 26, 2026, 00:01:01.289 --> 00:01:04.870 and we are staring down a massive, comprehensive 00:01:04.870 --> 00:01:08.769 Wikipedia repository that details the exact current 00:01:08.769 --> 00:01:11.530 state of self -driving cars. There is a lot of 00:01:11.530 --> 00:01:14.469 material here. So much. And our mission for you 00:01:14.469 --> 00:01:17.030 today is to cut through the sci -fi marketing 00:01:17.030 --> 00:01:20.010 hype, bypass those terrifying news headlines, 00:01:20.189 --> 00:01:23.310 and extract the absolute reality of autonomous 00:01:23.310 --> 00:01:25.340 vehicles right now. Yeah, we're going to get 00:01:25.340 --> 00:01:27.980 into how they actually work, why they still crash. 00:01:28.239 --> 00:01:31.280 The incredibly messy ethical dilemmas they create, 00:01:31.540 --> 00:01:34.120 and what this all means for you and your morning 00:01:34.120 --> 00:01:36.840 commute. It's a lot to navigate. I mean, we're 00:01:36.840 --> 00:01:39.939 looking at a web of rapidly evolving tech, deeply 00:01:39.939 --> 00:01:42.500 complex legal frameworks, and shifting social 00:01:42.500 --> 00:01:45.180 data. Right. But understanding the actual mechanics 00:01:45.180 --> 00:01:48.099 behind the headlines is really the only way to 00:01:48.099 --> 00:01:50.239 grasp where this is all heading. Which is exactly 00:01:50.239 --> 00:01:52.319 why you are the perfect guide to help us synthesize 00:01:52.319 --> 00:01:54.519 all of this today. Okay, let's unpack this. Let's 00:01:54.519 --> 00:01:56.280 do it. Before we can even judge if self -driving 00:01:56.280 --> 00:01:58.680 cars are successful, we have to define what self 00:01:58.680 --> 00:02:01.680 -driving actually means. And surprisingly, reading 00:02:01.680 --> 00:02:03.579 through these sources, the industry doesn't even 00:02:03.579 --> 00:02:05.900 have a universally agreed -upon standard right 00:02:05.900 --> 00:02:08.000 now. Yeah, that is the core of the problem right 00:02:08.000 --> 00:02:10.639 there. For years, the technical community relied 00:02:10.639 --> 00:02:14.360 on the SAE, the Society of Automotive Engineers. 00:02:14.860 --> 00:02:16.599 Right. They created that scale of automation, 00:02:16.840 --> 00:02:18.919 right? Level zero to level five. Exactly. But 00:02:18.919 --> 00:02:21.719 for anyone who follows technology, we know this 00:02:21.719 --> 00:02:24.500 scale has basically become this convoluted marketing 00:02:24.500 --> 00:02:26.919 tool. Oh, totally. The critical part that most 00:02:26.919 --> 00:02:29.500 consumers miss is that these levels aren't just 00:02:29.500 --> 00:02:32.560 some checklist of cool features. They are strictly 00:02:32.560 --> 00:02:35.560 based on the division of legal responsibility 00:02:35.560 --> 00:02:38.169 between the human and the machine. Right, so 00:02:38.169 --> 00:02:39.710 let's breeze through those real quick because 00:02:39.710 --> 00:02:42.270 they are everywhere in the marketing. Level zero 00:02:42.270 --> 00:02:46.229 through two. You are the driver. You are legally 00:02:46.229 --> 00:02:49.110 responsible, period. Even if the car is actively 00:02:49.110 --> 00:02:50.770 steering and braking to keep you in the lane, 00:02:51.050 --> 00:02:53.449 your eyes must be on the road. Exactly. And level 00:02:53.449 --> 00:02:55.949 three is where the car genuinely drives, but 00:02:55.949 --> 00:02:57.770 you have to be ready to intervene. Right, like 00:02:57.770 --> 00:02:59.969 if the car gets confused, it pings you. Yeah, 00:03:00.289 --> 00:03:02.889 and level four means the system drives and can 00:03:02.889 --> 00:03:05.330 safely pull itself over if you ignore that alarm. 00:03:05.710 --> 00:03:09.150 And finally, level five, which is just science 00:03:09.150 --> 00:03:11.150 fiction right now. That's a car that can drive 00:03:11.150 --> 00:03:14.449 anywhere, anytime, in a blizzard or a hurricane, 00:03:14.750 --> 00:03:17.189 without a steering wheel at all. But the confusion 00:03:17.189 --> 00:03:19.969 comes from how these levels interact with this 00:03:19.969 --> 00:03:22.789 concept called ODD. Operational Design Domain. 00:03:22.849 --> 00:03:25.770 Right. The ODD dictates the exact environmental 00:03:25.770 --> 00:03:28.590 boundaries where a car can operate safely. Yeah. 00:03:28.789 --> 00:03:31.349 Like, is it raining? How fast are we going? Are 00:03:31.349 --> 00:03:34.669 we on a pre -mapped highway or some random... 00:03:34.250 --> 00:03:37.150 dirt road. It's all about context. There's an 00:03:37.150 --> 00:03:39.189 analogy in the source material that perfectly 00:03:39.189 --> 00:03:42.129 nails this mechanism. Think of it like a human 00:03:42.129 --> 00:03:45.210 being asked to stand unassisted on one leg. Okay. 00:03:45.370 --> 00:03:48.090 That action is the dynamic driving requirement. 00:03:48.669 --> 00:03:50.830 Now, if you do that on solid ground, your designated 00:03:50.830 --> 00:03:53.030 operational design domain, you're totally fine. 00:03:53.069 --> 00:03:54.830 You can balance all day. Sure. But if you try 00:03:54.830 --> 00:03:56.770 to stand on one leg on a tightrope suspended 00:03:56.770 --> 00:03:59.129 in the air, suddenly your environment has changed. 00:03:59.330 --> 00:04:01.469 Your capability hasn't changed, right? No. But 00:04:01.469 --> 00:04:04.500 the Duane has. and you need support to keep from 00:04:04.500 --> 00:04:07.229 falling. What's fascinating here is how car makers 00:04:07.229 --> 00:04:10.710 exploit this exact confusion between a vehicle's 00:04:10.710 --> 00:04:13.409 capability and its operational domain. How do 00:04:13.409 --> 00:04:15.610 you mean? Well, a company might loudly advertise 00:04:15.610 --> 00:04:17.930 that they have a level four car. But what they 00:04:17.930 --> 00:04:20.170 actually mean is that the car has a level four 00:04:20.170 --> 00:04:23.209 feature, like, say, automated valley parking. 00:04:23.290 --> 00:04:26.870 Oh, right. But that feature only works inside 00:04:26.870 --> 00:04:30.910 a very specific geofence pre -mapped concrete 00:04:30.910 --> 00:04:34.069 garage. The moment you take that exact same car 00:04:34.069 --> 00:04:36.360 out of the garage, and onto the highway, the 00:04:36.360 --> 00:04:39.480 system downgrades to level two. Wow. Yeah, you 00:04:39.480 --> 00:04:41.480 have to keep your hands on the wheel. which is 00:04:41.480 --> 00:04:44.019 incredibly deceptive. I mean, if I buy a smartphone, 00:04:44.379 --> 00:04:46.560 it's a smartphone, whether I'm in my living room 00:04:46.560 --> 00:04:49.379 or the grocery store, it doesn't downgrade into 00:04:49.379 --> 00:04:51.399 a rotary phone when I cross the street. No, it 00:04:51.399 --> 00:04:53.199 doesn't. It's no wonder consumers are confused, 00:04:53.199 --> 00:04:56.639 which is why the CEO of Mobileye proposed a totally 00:04:56.639 --> 00:04:58.279 different way to talk about this in the sources, 00:04:58.480 --> 00:05:01.000 moving away from those rigid SAE numbers. Yes, 00:05:01.279 --> 00:05:03.560 Mobileye suggested terms that actually describe 00:05:03.560 --> 00:05:05.779 human behavior, which is so much more intuitive. 00:05:06.040 --> 00:05:08.000 Like what? They use categories like eyes on, 00:05:08.000 --> 00:05:10.980 hands on, or eyes on, hands off. Then, eyes off, 00:05:11.040 --> 00:05:15.980 hands off, and finally, no driver. clarifies 00:05:15.980 --> 00:05:18.620 the legal liability immediately. Like if you 00:05:18.620 --> 00:05:20.839 are in an eyes -on system, it doesn't matter 00:05:20.839 --> 00:05:23.300 what the steering wheel is doing. You are the 00:05:23.300 --> 00:05:25.540 driver. If you hit something, it is your fault. 00:05:26.139 --> 00:05:28.639 Exactly. It removes the ambiguity. So now that 00:05:28.639 --> 00:05:30.339 we know the playing field and the terminology, 00:05:30.439 --> 00:05:32.920 how are these companies actually trying to build 00:05:32.920 --> 00:05:35.220 these machines? How do these cars practically, 00:05:35.220 --> 00:05:38.459 you know, see the road? Oh, this is a huge debate. 00:05:38.660 --> 00:05:40.800 Because going through the research, there's a 00:05:40.800 --> 00:05:43.680 massive philosophical civil war happening in 00:05:43.680 --> 00:05:45.889 the tech world. regarding the sensors. There 00:05:45.889 --> 00:05:48.790 really is. It essentially boils down to two main 00:05:48.790 --> 00:05:51.209 sensory approaches. On one side, you have companies 00:05:51.209 --> 00:05:54.410 like Waymo. They use an incredibly hardware -heavy 00:05:54.410 --> 00:05:56.009 approach. Right, the spinning things on the roof. 00:05:56.110 --> 00:05:58.649 Yeah, they rely on LiDAR, which stands for light 00:05:58.649 --> 00:06:01.449 detection and ranging. LiDAR basically bounces 00:06:01.449 --> 00:06:04.149 millions of light lasers off surrounding objects 00:06:04.149 --> 00:06:07.290 every single second to create this hyper -accurate 00:06:07.290 --> 00:06:10.149 3D point cloud of the environment. Sounds intense. 00:06:10.389 --> 00:06:13.529 It is. But Waymo pairs that live LiDAR feed with 00:06:13.870 --> 00:06:16.970 incredibly detailed, pre -recorded maps. We're 00:06:16.970 --> 00:06:19.129 talking maps that know exactly where every lane 00:06:19.129 --> 00:06:22.129 line, stop sign, and curb is, down to the centimeter. 00:06:22.350 --> 00:06:24.990 Before the car even turns onto the street. Exactly. 00:06:25.490 --> 00:06:27.589 But the mechanism there seems totally flawed 00:06:27.589 --> 00:06:30.910 for scalability. I mean, those maps require massive, 00:06:31.110 --> 00:06:33.769 constant upkeep. If a construction crew moves 00:06:33.769 --> 00:06:36.250 a traffic cone overnight, or a storm knocks down 00:06:36.250 --> 00:06:38.649 a stop sign, the pre -recorded map is suddenly 00:06:38.649 --> 00:06:41.129 out of date. And the car's logic could just break. 00:06:41.230 --> 00:06:43.769 Yeah. Although I did read in the sources about 00:06:43.769 --> 00:06:46.879 systems like MIT's MeepLite. They tried to fix 00:06:46.879 --> 00:06:50.199 this by using just basic 2D GPS maps paired with 00:06:50.199 --> 00:06:52.240 live sensors, so it doesn't need to be updated 00:06:52.240 --> 00:06:54.560 constantly by a whole fleet of mapping vans. 00:06:54.620 --> 00:06:57.019 That's true. But then on the totally opposite 00:06:57.019 --> 00:06:59.399 end of the spectrum, you have the vision -only 00:06:59.399 --> 00:07:01.620 approach. Championed most famously by Tesla, 00:07:01.800 --> 00:07:04.480 right? Yes. They have completely stripped out 00:07:04.480 --> 00:07:06.959 the expensive LiDAR arrays, and they've even 00:07:06.959 --> 00:07:09.899 removed traditional radar. They rely entirely 00:07:09.899 --> 00:07:12.439 on optical cameras feeding visual data into what's 00:07:12.439 --> 00:07:15.180 called an end -to -end neural network. OK, let's 00:07:15.180 --> 00:07:17.560 pause right there, because end -to -end neural 00:07:17.560 --> 00:07:21.459 network is one of those massive tech buzzwords 00:07:21.459 --> 00:07:23.759 that gets thrown around a lot. It really does. 00:07:24.060 --> 00:07:26.019 If I understand the source material correctly, 00:07:26.540 --> 00:07:29.980 this means the AI isn't given explicit hard -coded 00:07:29.980 --> 00:07:34.459 rules by a programmer. Right? Like, a human didn't 00:07:34.459 --> 00:07:36.740 write a line of code saying, if you see a red 00:07:36.740 --> 00:07:38.819 octagon with white letters, apply the brakes. 00:07:39.019 --> 00:07:41.500 No, not at all. Instead, they feed the computer 00:07:41.500 --> 00:07:45.019 millions of hours of video footage of real humans 00:07:45.019 --> 00:07:47.540 driving, and the computer essentially learns 00:07:47.540 --> 00:07:50.540 to mimic that behavior organically. That's right. 00:07:50.699 --> 00:07:52.420 It's like learning a language by living in a 00:07:52.420 --> 00:07:54.180 foreign country rather than reading a grammar 00:07:54.180 --> 00:07:56.500 textbook. That is a perfect way to describe it. 00:07:56.569 --> 00:07:59.230 The system organically learns the concept of 00:07:59.230 --> 00:08:01.870 stopping at an intersection by processing how 00:08:01.870 --> 00:08:03.930 human drivers navigate thousands of different 00:08:03.930 --> 00:08:06.610 intersections. Oh, wow. And Tesla's rationale 00:08:06.610 --> 00:08:09.009 for this is simple. Human beings navigate the 00:08:09.009 --> 00:08:11.689 world using only our eyes and our brain. Therefore, 00:08:11.990 --> 00:08:14.189 a car should be able to do the exact same thing 00:08:14.189 --> 00:08:16.949 using cameras and a neural network. Plus optical 00:08:16.949 --> 00:08:19.550 cameras are vastly cheaper than lidar rigs. Oh, 00:08:19.930 --> 00:08:21.870 significantly cheaper. But if you're listening 00:08:21.870 --> 00:08:23.959 to this and thinking, wait a minute... I rely 00:08:23.959 --> 00:08:26.639 on my rear -view mirrors, my peripheral vision, 00:08:26.839 --> 00:08:29.860 and my hearing to merge onto a highway. I listen 00:08:29.860 --> 00:08:33.139 for sirens. I use intuition to guess that a parked 00:08:33.139 --> 00:08:36.159 car is about to pull out. Yeah. You aren't crazy. 00:08:36.379 --> 00:08:38.539 No, you're absolutely right. That is exactly 00:08:38.539 --> 00:08:41.059 why engineers are fighting over this. I have 00:08:41.059 --> 00:08:43.940 to push back on the vision -only logic. If a 00:08:43.940 --> 00:08:46.740 radar sensor can ping solid objects through a 00:08:46.740 --> 00:08:49.059 nighttime snowstorm that completely blinds an 00:08:49.059 --> 00:08:52.240 optical camera, Why in the world would you limit 00:08:52.240 --> 00:08:55.779 a two -ton vehicle to just eyes? It is a highly 00:08:55.779 --> 00:08:58.379 debated engineering choice, but the technical 00:08:58.379 --> 00:09:00.620 justification for avoiding multiple types of 00:09:00.620 --> 00:09:02.419 sensors comes down to the processing burden, 00:09:02.659 --> 00:09:04.980 and something called sensor fusion. Sensor fusion. 00:09:05.039 --> 00:09:07.919 Yeah. When you have cameras, radar, and lidar 00:09:07.919 --> 00:09:10.639 all feeding gigabytes of information to the car's 00:09:10.639 --> 00:09:13.299 computer simultaneously, the computer has to 00:09:13.299 --> 00:09:16.519 fuse all that data together to paint one unified 00:09:16.519 --> 00:09:18.899 picture of reality. OK, that makes sense. But 00:09:18.899 --> 00:09:20.960 what happens when the optical camera says the 00:09:20.960 --> 00:09:23.279 road ahead is perfectly clear, but the radar 00:09:23.279 --> 00:09:25.100 bounces off a metallic object? sign overhead 00:09:25.100 --> 00:09:27.860 and insists there is a stationary object dead 00:09:27.860 --> 00:09:30.960 ahead. Oh, the sensors disagree with each other. 00:09:31.120 --> 00:09:34.519 Yes. So it's like a jury where the two star witnesses 00:09:34.519 --> 00:09:37.320 are giving entirely conflicting testimonies and 00:09:37.320 --> 00:09:39.379 the car's brain just freezes trying to figure 00:09:39.379 --> 00:09:42.210 out who was lying. Exactly. Synthesizing that 00:09:42.210 --> 00:09:44.909 conflicting data takes computational time. It 00:09:44.909 --> 00:09:47.590 can cause the AI to hesitate, break suddenly, 00:09:47.789 --> 00:09:50.309 or make these dangerous jerky corrections. Thanks. 00:09:50.690 --> 00:09:52.889 It's like a human trying to read a book, text 00:09:52.889 --> 00:09:55.149 a friend, and listen to a deep dive all at the 00:09:55.149 --> 00:09:58.230 exact same time. Eventually, the processing lag 00:09:58.230 --> 00:10:01.490 causes a mistake. So by using only vision, the 00:10:01.490 --> 00:10:04.190 system theoretically processes one type of data 00:10:04.190 --> 00:10:07.750 much faster. Right. But as you pointed out, it 00:10:07.750 --> 00:10:10.470 sacrifices the superhuman penetrating capability 00:10:10.379 --> 00:10:13.460 of radar. And ultimately, the hardest part of 00:10:13.460 --> 00:10:15.759 all this isn't just seeing the road, it's behavior 00:10:15.759 --> 00:10:18.159 prediction. Meaning, it's not enough to just 00:10:18.159 --> 00:10:20.500 identify a pedestrian standing on the sidewalk. 00:10:20.659 --> 00:10:22.860 Right. An autonomous vehicle has to calculate 00:10:22.860 --> 00:10:25.379 that pedestrian's future trajectory. Like, are 00:10:25.379 --> 00:10:26.779 they waiting for the light? Are they looking 00:10:26.779 --> 00:10:29.340 at their phone? Exactly. Are they going to suddenly 00:10:29.340 --> 00:10:31.759 dart into the street to chase a rolling ball? 00:10:32.179 --> 00:10:34.600 The computer has to compute the position, speed, 00:10:34.700 --> 00:10:37.059 and likely intent of every single object around 00:10:37.059 --> 00:10:39.299 it multiple times a second. So we have these 00:10:39.299 --> 00:10:41.860 two computers. the competing philosophies, the 00:10:41.860 --> 00:10:44.200 heavy light arm mappers, and the vision only 00:10:44.200 --> 00:10:47.320 purists. But how are they actually performing 00:10:47.320 --> 00:10:49.500 out in the wild? That's the real question. Because 00:10:49.500 --> 00:10:51.960 when you read the news, it feels like autonomous 00:10:51.960 --> 00:10:54.340 vehicles are crashing into things every other 00:10:54.340 --> 00:10:57.120 day. But the statistics in the sources tell a 00:10:57.120 --> 00:11:00.360 much more nuanced story. They do. The gap between 00:11:00.360 --> 00:11:04.159 statistical safety and human trust is just staggering 00:11:04.159 --> 00:11:07.379 right now. Let's look at that massive 2024 study 00:11:07.379 --> 00:11:10.019 published in Nature Communications. OK. They 00:11:10.019 --> 00:11:12.100 compared the safety records of autonomous vehicles 00:11:12.100 --> 00:11:14.399 to human driven vehicles across tens of thousands 00:11:14.399 --> 00:11:16.480 of incident reports. The good news is actually 00:11:16.480 --> 00:11:19.259 phenomenal. Yeah. Autonomous vehicles have far 00:11:19.259 --> 00:11:22.100 fewer crashes involving pedestrians. About three 00:11:22.100 --> 00:11:24.519 percent of crashes compared to 15 percent for 00:11:24.519 --> 00:11:27.730 human drivers. That's a huge drop. And the machines 00:11:27.730 --> 00:11:30.789 are significantly safer than humans in heavy 00:11:30.789 --> 00:11:33.429 rain or fog. I mean, they don't get distracted 00:11:33.429 --> 00:11:35.610 by their phones, they don't drive drunk, and 00:11:35.610 --> 00:11:38.159 they don't get tired after a 12 hour shift. But 00:11:38.159 --> 00:11:40.200 the bad news is deeply concerning, which brings 00:11:40.200 --> 00:11:42.539 us right back to your opening point. The study 00:11:42.539 --> 00:11:44.700 found that autonomous vehicles are more than 00:11:44.700 --> 00:11:47.379 five times more vulnerable to collisions at dawn 00:11:47.379 --> 00:11:49.879 and dusk. Because of the shadows. Yeah, the rapidly 00:11:49.879 --> 00:11:52.480 changing light, the long shadows, the sun glare. 00:11:52.539 --> 00:11:54.799 It directly messes with the contrast required 00:11:54.799 --> 00:11:57.740 by cameras and optical sensors. And when these 00:11:57.740 --> 00:12:00.679 systems do fail, they fail in ways that humans 00:12:00.679 --> 00:12:03.360 find entirely illogical, which just shatters 00:12:03.360 --> 00:12:06.019 public trust. Let's look at exactly how they 00:12:06.019 --> 00:12:08.870 fail. because the mechanism is fascinating. The 00:12:08.870 --> 00:12:11.929 sources detail a 2022 incident where a cruise 00:12:11.929 --> 00:12:14.889 robotaxi completely blocked a fire engine that 00:12:14.889 --> 00:12:17.570 was responding to an emergency. Why did it do 00:12:17.570 --> 00:12:19.549 that? Why did it just freeze in the middle of 00:12:19.549 --> 00:12:22.090 the road? Well, the AI is programmed to obey 00:12:22.090 --> 00:12:25.309 traffic laws flawlessly. When a fire engine turns 00:12:25.309 --> 00:12:27.389 its sirens on and drives on the wrong side of 00:12:27.389 --> 00:12:30.549 the road to bypass traffic, it breaks the AI's 00:12:30.549 --> 00:12:33.070 programmed logic tree. It literally doesn't know 00:12:33.070 --> 00:12:35.909 what to do. Right. The AI's hard -coded rules 00:12:35.909 --> 00:12:39.529 hit a conflict. Fire trucks are vehicles. Vehicles 00:12:39.529 --> 00:12:41.950 must stay in the right lane. This vehicle is 00:12:41.950 --> 00:12:44.370 in the wrong lane moving toward me. I cannot 00:12:44.370 --> 00:12:47.389 predict its next legal move. So the system defaults 00:12:47.389 --> 00:12:50.289 to the only safe action it knows. It just halts 00:12:50.289 --> 00:12:52.429 in confusion. If we connect this to the bigger 00:12:52.429 --> 00:12:55.490 picture, we have actually been through this exact 00:12:55.490 --> 00:12:58.149 psychological and technological barrier before. 00:12:58.610 --> 00:13:00.830 Think about the automatic elevator. Elevators. 00:13:01.049 --> 00:13:04.230 Like in an office building. Yes. Automatic elevators 00:13:04.230 --> 00:13:07.509 were invented around 1900. But for decades, people 00:13:07.509 --> 00:13:10.730 were absolutely terrified to step into a metal 00:13:10.730 --> 00:13:13.929 box suspended in a dark shaft without a human 00:13:13.929 --> 00:13:16.309 operator standing there physically pushing the 00:13:16.309 --> 00:13:18.830 levers. I mean, I kind of get that. Right. They 00:13:18.830 --> 00:13:21.070 didn't trust the machine's logic to stop at the 00:13:21.070 --> 00:13:23.809 right floor. It took operator strikes, massive 00:13:23.809 --> 00:13:26.669 advertising campaigns and the invention of a 00:13:26.669 --> 00:13:29.830 giant reassuring red emergency stop button to 00:13:29.830 --> 00:13:32.190 finally build public trust. Oh that makes so 00:13:32.190 --> 00:13:33.870 much sense. We are currently in the elevator 00:13:33.870 --> 00:13:36.269 phase of self -driving cars. People want to know 00:13:36.269 --> 00:13:38.129 they can hit a button to stop the machine when 00:13:38.129 --> 00:13:40.490 its logic fails. Here's where it gets really 00:13:40.490 --> 00:13:43.070 interesting. Let's look at the data from the 00:13:43.070 --> 00:13:45.149 National Highway Traffic Safety Administration. 00:13:45.470 --> 00:13:49.649 Out of roughly 4 ,000 autonomous vehicle incidents 00:13:49.649 --> 00:13:54.110 reported between 2019 and 2024, nearly 54 % of 00:13:54.110 --> 00:13:56.750 them involve Teslas. More than half. More than 00:13:56.750 --> 00:14:00.149 half. Yet Tesla aggressively markets its level 00:14:00.149 --> 00:14:02.769 two system under the brand name full self -driving, 00:14:03.110 --> 00:14:05.289 which has led to multiple federal investigations 00:14:05.289 --> 00:14:08.990 for deceptive marketing. Because legally, under 00:14:08.990 --> 00:14:11.110 that mobile IFA framework we talk about, you 00:14:11.110 --> 00:14:12.840 still have to have your eyes on the road. It 00:14:12.840 --> 00:14:15.139 perfectly illustrates how terminology shapes 00:14:15.139 --> 00:14:17.980 consumer behavior. If you name a software package 00:14:17.980 --> 00:14:20.179 full self -driving, people are naturally going 00:14:20.179 --> 00:14:22.399 to treat it like it's fully self -driving. Regardless 00:14:22.399 --> 00:14:24.299 of the fine print in the owner's manual telling 00:14:24.299 --> 00:14:26.320 them to keep their hands on the wheel. Exactly. 00:14:26.559 --> 00:14:29.220 And the sources bring up some really tragic incidents. 00:14:29.759 --> 00:14:32.639 Tesla's first fatal crash in 2016 happened because 00:14:32.639 --> 00:14:35.740 the car literally didn't distinguish a white 00:14:35.740 --> 00:14:38.200 tractor trailer crossing the highway against 00:14:38.200 --> 00:14:41.019 a brightly lit sky. It just didn't see it. Right. 00:14:41.279 --> 00:14:44.200 And there's that 2025 Cybertruck crash on full 00:14:44.200 --> 00:14:47.399 self -driving. Waymo hit a bus in 2016 trying 00:14:47.399 --> 00:14:49.940 to avoid sandbags in the road. And we have to 00:14:49.940 --> 00:14:52.600 mention the 2018 Uber Advanced Technologies group 00:14:52.600 --> 00:14:55.659 fatality. That was a perfect storm of human and 00:14:55.659 --> 00:14:58.440 machine failure. The safety driver inside the 00:14:58.440 --> 00:15:00.080 car was looking down at their phone, watching 00:15:00.080 --> 00:15:02.879 a show, and the pedestrian victim crossed the 00:15:02.879 --> 00:15:05.440 street illegally with meth in her system. The 00:15:05.440 --> 00:15:07.720 car's sensors detected her, but the software 00:15:07.720 --> 00:15:09.720 couldn't classify what she was until it was too 00:15:09.720 --> 00:15:12.429 late. It's just awful. This brings us to the 00:15:12.429 --> 00:15:14.809 underlying decision making of the AI, moving 00:15:14.809 --> 00:15:17.070 beyond just the mechanics of safety into the 00:15:17.070 --> 00:15:19.830 ethics and the economics. It's not just about 00:15:19.830 --> 00:15:22.250 a car freezing in front of a fire truck. It's 00:15:22.250 --> 00:15:25.570 about how the AI decides who or what to prioritize 00:15:25.570 --> 00:15:28.519 when a crash is mathematically unavoidable. This 00:15:28.519 --> 00:15:31.379 is where the engineering hits hard societal reality. 00:15:31.779 --> 00:15:33.740 For example, a study from researchers at Georgia 00:15:33.740 --> 00:15:36.679 Tech revealed a deeply troubling flaw in AI detection 00:15:36.679 --> 00:15:39.399 models. I read this part. It's wild. They found 00:15:39.399 --> 00:15:41.600 that autonomous detection systems were about 00:15:41.600 --> 00:15:44.539 5 % less effective at recognizing pedestrians 00:15:44.539 --> 00:15:46.799 with darker skin tones compared to lighter skin 00:15:46.799 --> 00:15:49.320 tones. Which is an astonishing engineering oversight. 00:15:49.360 --> 00:15:51.840 Yeah. But if we look at the mechanism of why 00:15:51.840 --> 00:15:53.980 that happens, it goes right back to those neural 00:15:53.980 --> 00:15:56.629 networks. Yeah. If an AI is trained primarily 00:15:56.629 --> 00:15:59.909 on data sets or video footage that lacks diverse 00:15:59.909 --> 00:16:03.129 demographic representation, the AI literally 00:16:03.129 --> 00:16:06.350 develops a blind spot. It only knows what it 00:16:06.350 --> 00:16:09.370 has been shown. Exactly. It forces us to confront 00:16:09.370 --> 00:16:12.490 the classic philosophical trolley problem, but 00:16:12.490 --> 00:16:14.870 programmed into real life. Oh, the trolley problem. 00:16:15.009 --> 00:16:17.409 Right. Imagine an autonomous vehicle is driving 00:16:17.409 --> 00:16:20.269 down a narrow street. A pedestrian suddenly darts 00:16:20.269 --> 00:16:22.950 out. The car computes that it cannot physically 00:16:22.950 --> 00:16:26.379 break in time. It has two choices. Hit the pedestrian, 00:16:26.700 --> 00:16:29.100 or swerve into a solid brick wall, which will 00:16:29.100 --> 00:16:30.940 almost certainly kill the passenger inside the 00:16:30.940 --> 00:16:33.940 car. How do you program that decision? The sources 00:16:33.940 --> 00:16:36.399 cite extensive surveys on this exact dilemma, 00:16:36.500 --> 00:16:39.000 and the results are so revealing about human 00:16:39.000 --> 00:16:41.220 psychology. They really are. Broadly, people 00:16:41.220 --> 00:16:43.720 say they want autonomous cars programmed to minimize 00:16:43.720 --> 00:16:47.340 overall harm. So, logically, sacrifice the one 00:16:47.340 --> 00:16:49.539 passenger to save three pedestrians on the sidewalk. 00:16:50.200 --> 00:16:52.860 It's pure utilitarianism. Unless the person taking 00:16:52.860 --> 00:16:55.379 the survey is the hypothetical passenger in the 00:16:55.379 --> 00:16:57.980 car, then suddenly they want the car programmed 00:16:57.980 --> 00:17:00.299 to protect them at all costs. Naturally. But 00:17:00.299 --> 00:17:02.879 I have to push back on that logic. If every single 00:17:02.879 --> 00:17:05.660 car on the road is programmed by its manufacturer 00:17:05.660 --> 00:17:08.900 to be entirely selfish, to prioritize its own 00:17:08.900 --> 00:17:11.859 passengers at the expense of everyone else, doesn't 00:17:11.859 --> 00:17:15.799 that make the roads vastly more chaotic and dangerous 00:17:15.799 --> 00:17:18.990 overall? It's like an arms race of selfish algorithms. 00:17:19.490 --> 00:17:22.109 This raises an important question. When an AI 00:17:22.109 --> 00:17:24.710 ultimately does make the wrong choice and someone 00:17:24.710 --> 00:17:27.130 gets hurt, who actually goes to jail? Yeah, who 00:17:27.130 --> 00:17:30.349 is at fault? Is it the human occupant who was 00:17:30.349 --> 00:17:33.150 legally eyes off and wasn't driving? Is it the 00:17:33.150 --> 00:17:35.630 corporate car manufacturer? Or is it the software 00:17:35.630 --> 00:17:37.809 developer who wrote the logic tree code five 00:17:37.809 --> 00:17:40.250 years ago? That's a legal nightmare. Our entire 00:17:40.250 --> 00:17:42.529 legal framework of liability is fundamentally 00:17:42.529 --> 00:17:45.230 built around human error, human negligence, and 00:17:45.230 --> 00:17:48.750 human intent. It is completely unprepared for 00:17:48.750 --> 00:17:51.059 machine liability. Which brings up a question 00:17:51.059 --> 00:17:53.420 for you listening right now. Would you put your 00:17:53.420 --> 00:17:55.799 family in a car that is legally programmed to 00:17:55.799 --> 00:17:58.059 sacrifice you in order to save three strangers 00:17:58.059 --> 00:18:00.079 on the street? It's chilling to think about. 00:18:00.319 --> 00:18:03.240 It really is. But beyond the ethics of the code, 00:18:03.799 --> 00:18:06.400 the economic reality of this technology is also 00:18:06.400 --> 00:18:09.170 hitting a massive wall. The sources note that 00:18:09.170 --> 00:18:11.789 Waymo robotaxes currently operate at a staggering 00:18:11.789 --> 00:18:15.369 loss. It cost them between $7 to $9 per mile 00:18:15.369 --> 00:18:18.569 to run a robotaxi, compared to about $1 a mile 00:18:18.569 --> 00:18:21.390 for a standard personal car. And we have to look 00:18:21.390 --> 00:18:23.569 at why it's so expensive. It's not just the electricity 00:18:23.569 --> 00:18:25.910 to run the car. It's the fact that those LiDAR 00:18:25.910 --> 00:18:28.650 arrays cost thousands of dollars. It's the massive 00:18:28.650 --> 00:18:31.910 data storage required. And crucially, it's the 00:18:31.910 --> 00:18:35.049 remote human assistance centers. Whenever a robotaxi 00:18:35.049 --> 00:18:37.509 gets confused by a construction zone, a human 00:18:37.509 --> 00:18:40.529 sitting in a control room miles away has to digitally 00:18:40.529 --> 00:18:43.289 step in and guide it. McKinsey estimates it's 00:18:43.289 --> 00:18:46.509 going to take until 2035 at the earliest to scale 00:18:46.509 --> 00:18:48.250 the technology enough to get those operating 00:18:48.250 --> 00:18:50.839 costs under $2 a mile. And if they do manage 00:18:50.839 --> 00:18:53.180 to get those costs down and scale up, the labor 00:18:53.180 --> 00:18:55.420 displacement is going to be historic. The research 00:18:55.420 --> 00:18:58.240 points out there are 2 .9 million driving jobs 00:18:58.240 --> 00:19:01.220 in the US alone. Long -haul truck drivers, taxi 00:19:01.220 --> 00:19:04.900 drivers, city bus operators, if those jobs are 00:19:04.900 --> 00:19:07.680 automated away by fleets of AI, the economic 00:19:07.680 --> 00:19:09.880 displacement would actually surpass the total 00:19:09.880 --> 00:19:13.619 job losses of the 2008 Great Recession. We're 00:19:13.619 --> 00:19:16.180 talking about restructuring the entire blue collar 00:19:16.180 --> 00:19:18.759 labor force. Furthermore, it's not just a domestic 00:19:18.759 --> 00:19:21.319 economic issue. What else is there? The sources 00:19:21.319 --> 00:19:24.220 highlight objective reports of severe geopolitical 00:19:24.220 --> 00:19:26.599 tensions, specifically between the United States 00:19:26.599 --> 00:19:29.789 and China, over autonomous vehicle data. There 00:19:29.789 --> 00:19:32.309 are very real concerns currently being debated 00:19:32.309 --> 00:19:34.549 regarding imported self -driving tech. Right, 00:19:34.549 --> 00:19:37.069 because if you think about how LIDAR works, bouncing 00:19:37.069 --> 00:19:39.569 millions of lasers to create millimeter -accurate 00:19:39.569 --> 00:19:42.210 3D maps of its surroundings, a self -driving 00:19:42.210 --> 00:19:44.650 car is basically a roving internet -connected 00:19:44.650 --> 00:19:46.990 supercomputer. Exactly. The geopolitical fear 00:19:46.990 --> 00:19:49.089 is that imported economist vehicles could be 00:19:49.089 --> 00:19:51.869 used to facilitate esk denige by constantly mapping 00:19:51.869 --> 00:19:54.730 out critical infrastructure, power grids, and 00:19:54.730 --> 00:19:57.029 military bases in foreign countries and then 00:19:57.029 --> 00:19:59.440 just beaming that data back overseas. When a 00:19:59.440 --> 00:20:02.099 car is no longer just a mechanical vehicle, but 00:20:02.099 --> 00:20:04.859 a massive data harvesting node covered in high 00:20:04.859 --> 00:20:07.660 definition sensors, it naturally elevates from 00:20:07.660 --> 00:20:09.819 a transportation issue to a matter of national 00:20:09.819 --> 00:20:12.539 security. Okay, we have covered an incredible 00:20:12.539 --> 00:20:15.299 amount of ground today. We are clearly navigating 00:20:15.299 --> 00:20:19.140 a very messy transitional era. We have cars on 00:20:19.140 --> 00:20:21.799 the road right now that are statistically safer 00:20:21.799 --> 00:20:25.220 than you in a heavy rainstorm, but might be dangerously 00:20:25.220 --> 00:20:28.180 blind when the sun goes down. We have complex 00:20:28.180 --> 00:20:30.940 systems branded and sold to consumers as full 00:20:30.940 --> 00:20:33.980 self -driving that legally require you to babysit 00:20:33.980 --> 00:20:36.900 the steering wheel. And we have brilliant engineering 00:20:36.900 --> 00:20:39.480 that can predict complex traffic patterns, yet 00:20:39.480 --> 00:20:42.180 still struggles with diverse pedestrian recognition 00:20:42.180 --> 00:20:45.140 data and the basic legal frameworks of liability. 00:20:45.319 --> 00:20:47.250 So what does this all mean? It means what we 00:20:47.250 --> 00:20:49.470 always see in this show. Technology is moving 00:20:49.470 --> 00:20:52.170 exponentially faster than human trust and infinitely 00:20:52.170 --> 00:20:54.910 faster than our legal frameworks. We are building 00:20:54.910 --> 00:20:57.690 the airplane while we're flying it, or I guess 00:20:57.690 --> 00:21:00.250 programming the car while it's driving itself. 00:21:00.670 --> 00:21:02.690 And essentially, we are all the beta testers. 00:21:02.910 --> 00:21:04.910 That's a great way to put it. Before we go, I 00:21:04.910 --> 00:21:06.650 want to leave you with one final thought to chew 00:21:06.650 --> 00:21:09.609 on today. We've talked a lot about how self -driving 00:21:09.609 --> 00:21:13.390 cars struggle to predict what erratic, unpredictable 00:21:13.390 --> 00:21:15.799 human drivers are going to do next. Yeah, that's 00:21:15.799 --> 00:21:18.420 the hardest part for them. Well, if these autonomous 00:21:18.420 --> 00:21:21.160 vehicles ultimately prove to be drastically safer 00:21:21.160 --> 00:21:23.920 than humans, but only when they can perfectly 00:21:23.920 --> 00:21:25.900 communicate and predict the movements of other 00:21:25.900 --> 00:21:29.480 robot cars on a closed network, how long will 00:21:29.480 --> 00:21:32.460 it be until human driving is outright banned 00:21:32.460 --> 00:21:35.240 on public highways? Oh, wow. Will governments 00:21:35.240 --> 00:21:38.200 decide that human unpredictability is just too 00:21:38.200 --> 00:21:41.539 dangerous to the machines? Are we rapidly approaching 00:21:41.539 --> 00:21:44.500 a future where driving a car yourself is an illegal, 00:21:44.759 --> 00:21:47.079 obsolete hobby? Just something to think about 00:21:47.079 --> 00:21:49.019 next time you put your hands on the wheel. Thanks 00:21:49.019 --> 00:21:50.920 for joining us on this deep dive and as always 00:21:50.920 --> 00:21:52.640 we hope this keeps you the most well -informed 00:21:52.640 --> 00:21:53.440 person in the room.