WEBVTT 00:00:00.000 --> 00:00:03.140 Right now, I mean, an autonomous vehicle is statistically 00:00:03.140 --> 00:00:05.540 less likely to hit a pedestrian than you are. 00:00:05.799 --> 00:00:08.660 Which is honestly a wild statistic when you actually 00:00:08.660 --> 00:00:11.759 look at the numbers. Yeah, exactly. Yet a massive 00:00:11.759 --> 00:00:13.500 chunk of the global population, I think it's 00:00:13.500 --> 00:00:16.620 nearly 75%, refuses to even consider getting 00:00:16.620 --> 00:00:19.679 into one. Right. Which raises a really strange 00:00:19.679 --> 00:00:22.460 question right out of the gate. Like, why are 00:00:22.460 --> 00:00:26.019 we so utterly terrified of a machine that, according 00:00:26.019 --> 00:00:28.640 to the data, might actually drive better than 00:00:28.640 --> 00:00:31.440 we do? I mean, it is the ultimate tension, right, 00:00:31.559 --> 00:00:35.100 between our human instinct and this huge technological 00:00:35.100 --> 00:00:38.179 promise. We're being asked to hand over our physical 00:00:38.179 --> 00:00:41.500 safety, our lives, really to, you know, lines 00:00:41.500 --> 00:00:43.960 of code. Yeah, and trust is a very difficult 00:00:43.960 --> 00:00:46.119 thing to program. Exactly. It doesn't just happen 00:00:46.119 --> 00:00:48.479 overnight. Welcome to this deep dive, everyone. 00:00:48.679 --> 00:00:51.179 Today we are looking at a massive stack of research. 00:00:51.739 --> 00:00:54.380 primarily relying on a really comprehensive, 00:00:54.539 --> 00:00:57.640 up -to -date Wikipedia overview as of early 2026, 00:00:58.079 --> 00:01:00.600 all about self -driving cars. It's a heavy topic, 00:01:00.619 --> 00:01:03.340 but a really important one. It is. And our mission 00:01:03.340 --> 00:01:06.379 here is to cut straight through the sci -fi hype 00:01:06.379 --> 00:01:08.599 and, you know, those terrifying crash headlines 00:01:08.599 --> 00:01:10.060 you always see on your feed. Yeah, because the 00:01:10.060 --> 00:01:12.459 headlines definitely skew the reality. They really 00:01:12.459 --> 00:01:14.739 do. We want to figure out what is actually happening 00:01:14.739 --> 00:01:17.140 on our roads right now and how it is going to 00:01:17.140 --> 00:01:19.329 affect your daily commute. And it is a critical 00:01:19.329 --> 00:01:22.189 time to be looking at this because the narrative 00:01:22.189 --> 00:01:25.870 we see in those glossy car advertisements simply 00:01:25.870 --> 00:01:29.150 does not match the engineering reality on the 00:01:29.150 --> 00:01:31.170 ground. Not at all. We are in a transitional 00:01:31.170 --> 00:01:36.859 phase right now that is messy. It's highly experimental 00:01:36.859 --> 00:01:40.579 and honestly wildly misunderstood. Okay. Let's 00:01:40.579 --> 00:01:41.900 unpack this right at the start because there 00:01:41.900 --> 00:01:44.400 is a massive contradiction staring us right in 00:01:44.400 --> 00:01:46.299 the face here. Oh, absolutely. These car companies 00:01:46.299 --> 00:01:48.519 are selling us the dream of like napping in the 00:01:48.519 --> 00:01:51.819 backseat on the way to work. But that 2022 survey 00:01:51.819 --> 00:01:54.280 we mentioned. Great. The global one. Yeah. Only 00:01:54.280 --> 00:01:57.159 27 % of the global population would feel safe 00:01:57.159 --> 00:01:59.959 in a self -driving car. That is barely a quarter 00:01:59.959 --> 00:02:03.329 of the world. It's a tiny fraction. So... Why 00:02:03.329 --> 00:02:06.390 is there such a massive gap between the technology 00:02:06.390 --> 00:02:09.729 being offered and our trust in it? Well, to understand 00:02:09.729 --> 00:02:12.509 that trust gap, we first have to really understand 00:02:12.509 --> 00:02:15.430 the language being used to sell these cars. Because 00:02:15.430 --> 00:02:18.009 right now, the marketing surrounding autonomous 00:02:18.009 --> 00:02:22.080 vehicles is... quite frankly, incredibly deceptive. 00:02:22.240 --> 00:02:25.000 Yeah, deceptive is the perfect word for it. Before 00:02:25.000 --> 00:02:27.180 we can even ask if these cars are safe or if 00:02:27.180 --> 00:02:28.900 they're coming to your neighborhood, we have 00:02:28.900 --> 00:02:31.039 to agree on what a quote unquote self -driving 00:02:31.039 --> 00:02:33.780 car actually is. And legally, that definition 00:02:33.780 --> 00:02:36.699 is very, very blurred. Oh, completely. When you 00:02:36.699 --> 00:02:39.159 dig into the sources, you hit this global standard 00:02:39.159 --> 00:02:41.400 created by the Society of Automotive Engineers, 00:02:41.560 --> 00:02:44.020 or the SAE. Right. They established the official 00:02:44.020 --> 00:02:46.900 framework used in over 50 countries. Yeah. And 00:02:46.900 --> 00:02:48.979 they came up with a scale from level zero, which 00:02:48.979 --> 00:02:51.780 is just fully manual driving, all the way up 00:02:51.780 --> 00:02:54.639 to level five. And level five means full automation 00:02:54.639 --> 00:02:57.680 anywhere, anytime, under any condition. But the 00:02:57.680 --> 00:02:59.819 catch is, most commercially available systems 00:02:59.819 --> 00:03:02.580 right now, like Tesla's Autopilot or Ford's Blue 00:03:02.580 --> 00:03:05.039 Cruise, they sit around level two. Yeah, level 00:03:05.039 --> 00:03:07.520 two, which is a huge drop from the level five 00:03:07.520 --> 00:03:09.919 dream. And we should clarify what level two actually 00:03:09.919 --> 00:03:13.479 means in practice. I mean, it means the car handles 00:03:13.479 --> 00:03:16.379 the steering and the speed, but the driver must 00:03:16.379 --> 00:03:18.780 keep their eyes on the road and hands on the 00:03:18.780 --> 00:03:21.360 wheel. Right. You can't just check out. Exactly. 00:03:21.639 --> 00:03:23.759 The human is legally responsible for everything 00:03:23.759 --> 00:03:26.139 that happens. Which directly contradicts the 00:03:26.139 --> 00:03:28.460 branding we see. I mean, Tesla literally calls 00:03:28.460 --> 00:03:32.219 its level two system full self -driving or FSD. 00:03:32.419 --> 00:03:34.819 Yeah. which has caused a lot of regulatory friction. 00:03:35.199 --> 00:03:38.199 Definitely. As the California DMV and U .S. senators 00:03:38.199 --> 00:03:40.580 have repeatedly pointed out, it is not fully 00:03:40.580 --> 00:03:44.139 self -driving at all. It requires active, constant 00:03:44.139 --> 00:03:46.800 driver supervision. Which is exactly why Mobileye, 00:03:47.000 --> 00:03:49.020 they're another major player in the space, they 00:03:49.020 --> 00:03:51.819 proposed an entirely different consumer -friendly 00:03:51.819 --> 00:03:54.159 taxonomy. Oh, right. Instead of the confusing 00:03:54.159 --> 00:03:56.900 zero to five numbers. Yeah, they use terms based 00:03:56.900 --> 00:03:59.379 on actual human engagement. So things like eyes 00:03:59.379 --> 00:04:01.300 on hands on or eyes on hands off. That makes 00:04:01.300 --> 00:04:02.900 so much more sense. And then it goes to eyes 00:04:02.900 --> 00:04:05.639 off, hands off, and finally just no driver. Which 00:04:05.639 --> 00:04:07.919 honestly changes the psychology of the consumer 00:04:07.919 --> 00:04:11.180 entirely. It sets clear, undeniable boundaries. 00:04:11.560 --> 00:04:14.699 It does. But those boundaries also rely on a 00:04:14.699 --> 00:04:17.439 crucial engineering concept. It's called the 00:04:17.439 --> 00:04:21.209 operational design domain, or... ODD. OK, yes, 00:04:21.350 --> 00:04:24.750 the ODD. This is basically the specific set of 00:04:24.750 --> 00:04:27.029 environmental conditions where the car is actually 00:04:27.029 --> 00:04:29.430 capable of driving itself. Right, the safe zone, 00:04:29.629 --> 00:04:32.509 essentially. So maybe the system only works in 00:04:32.509 --> 00:04:35.250 clear weather on a pre -mapped highway going 00:04:35.250 --> 00:04:38.470 under 40 miles per hour. Let's use an analogy 00:04:38.470 --> 00:04:40.350 here, straight from the source material. Go for 00:04:40.350 --> 00:04:43.449 it. Think of an autonomous system less like a 00:04:43.449 --> 00:04:46.939 human driver and more like a train on a set of 00:04:46.939 --> 00:04:48.959 tracks. I mean, a train can drive itself perfectly, 00:04:48.959 --> 00:04:50.839 right? Whoever, as long as it has the rail. Exactly. 00:04:51.079 --> 00:04:53.459 But only if it remains exactly on its tracks. 00:04:53.980 --> 00:04:57.060 Those tracks are the ODD. The moment that train 00:04:57.060 --> 00:04:59.920 leaves the track, say, hitting a messy dirt road 00:04:59.920 --> 00:05:02.199 or dealing with a sudden blizzard, the system 00:05:02.199 --> 00:05:05.399 completely fails. It derails, literally and figuratively. 00:05:05.560 --> 00:05:08.279 Yeah. And the car is the exact same way. If it 00:05:08.279 --> 00:05:10.360 leaves its ODD, it immediately needs the human 00:05:10.360 --> 00:05:14.100 to take over. And that specific moment, the handover 00:05:14.100 --> 00:05:16.339 moment when the machine suddenly tells the human, 00:05:17.180 --> 00:05:18.660 I don't know what to do, you take the wheel, 00:05:19.180 --> 00:05:21.399 that is one of the most dangerous complexities 00:05:21.399 --> 00:05:24.019 in the entire system. Because we just aren't 00:05:24.019 --> 00:05:27.379 ready for it. Right. Human reaction time is simply 00:05:27.379 --> 00:05:29.959 too slow if you've been zoned out watching a 00:05:29.959 --> 00:05:31.980 video for the last 30 minutes. Wait, hold on. 00:05:31.980 --> 00:05:34.199 I want to push back on this SAE scale for a second. 00:05:34.220 --> 00:05:37.680 Okay. If it's a zero to five scale, my brain 00:05:37.680 --> 00:05:39.920 immediately thinks level five is just a beefed 00:05:39.920 --> 00:05:42.740 up version of level four. Are you saying it's 00:05:42.740 --> 00:05:45.439 not a linear upgrade? Not exactly, no. Doesn't 00:05:45.439 --> 00:05:48.339 a simple numbered scale falsely imply that more 00:05:48.339 --> 00:05:51.029 automation is always inherently better? What's 00:05:51.029 --> 00:05:53.490 fascinating here is that the SAE levels are not 00:05:53.490 --> 00:05:55.670 actually based purely on the car's technological 00:05:55.670 --> 00:05:58.610 features. They are based entirely on the driver's 00:05:58.610 --> 00:06:00.930 liability and responsibility. Oh, wow. I didn't 00:06:00.930 --> 00:06:03.750 realize that. Yeah. And because of that, a car 00:06:03.750 --> 00:06:06.230 isn't just locked into one level. It actually 00:06:06.230 --> 00:06:08.970 switches levels dynamically based on the driving 00:06:08.970 --> 00:06:11.050 mode and that ODD you just mentioned. So the 00:06:11.050 --> 00:06:13.949 level actually changes mid -drive. Exactly. You 00:06:13.949 --> 00:06:15.709 could be in a vehicle that operates at level 00:06:15.709 --> 00:06:18.290 four, meaning high automation, where the car 00:06:18.290 --> 00:06:21.050 can safely pull over and stop itself if you fall 00:06:21.050 --> 00:06:23.970 asleep while you were cruising on a sunny, perfectly 00:06:23.970 --> 00:06:26.610 mapped highway. OK, that sounds great. But the 00:06:26.610 --> 00:06:29.610 moment you exit that highway onto a chaotic urban 00:06:29.610 --> 00:06:32.470 side street, you know, with active construction 00:06:32.470 --> 00:06:35.209 and no lane markings, the system might drop down 00:06:35.209 --> 00:06:38.189 to level two. And suddenly you're back in charge. 00:06:38.490 --> 00:06:41.009 Suddenly it requires your full attention and 00:06:41.009 --> 00:06:43.959 your hands back on the wheel. The public confusion 00:06:43.959 --> 00:06:46.759 stems from treating the SAE level as a fixed 00:06:46.759 --> 00:06:49.360 feature you buy at the dealership, rather than 00:06:49.360 --> 00:06:51.459 a description of the dynamic second -by -second 00:06:51.459 --> 00:06:53.560 relationship between the human and the machine. 00:06:54.300 --> 00:06:56.879 That dynamic relationship completely relies on 00:06:56.879 --> 00:06:59.100 how the car perceives the world around it, which 00:06:59.100 --> 00:07:01.439 is a perfect pivot to the actual technology here. 00:07:01.800 --> 00:07:03.589 Yeah, perception is everything. How did these 00:07:03.589 --> 00:07:06.029 cars actually see the construction site or the 00:07:06.029 --> 00:07:08.310 missing lane markings? Because diving into the 00:07:08.310 --> 00:07:10.230 research, it turns out the biggest tech companies 00:07:10.230 --> 00:07:12.689 in the world completely fundamentally disagree 00:07:12.689 --> 00:07:14.829 on the answer. They really do. It is the great 00:07:14.829 --> 00:07:17.629 sensor debate. And it primarily comes down to 00:07:17.629 --> 00:07:21.089 two highly polarized engineering philosophies. 00:07:21.350 --> 00:07:24.470 LiDAR versus vision only. Exactly. Let's start 00:07:24.470 --> 00:07:27.329 with LiDAR. Because Waymo is the big champion 00:07:27.329 --> 00:07:30.149 here. They are, yeah. They use these heavy LiDAR 00:07:30.149 --> 00:07:32.910 arrays, shooting out millions of laser pulses 00:07:32.910 --> 00:07:35.930 every second to measure the exact time it takes 00:07:35.930 --> 00:07:37.850 for the light to bounce back. It's incredible 00:07:37.850 --> 00:07:41.230 technology. It really is, like a BATS echolocation 00:07:41.230 --> 00:07:44.930 superpower, but with light. This creates a highly 00:07:44.930 --> 00:07:47.810 detailed, high -resolution 3D point cloud of 00:07:47.810 --> 00:07:51.220 the environment. And Waymo pairs this with heavily 00:07:51.220 --> 00:07:53.759 geofenced routes. Right, meaning they only operate 00:07:53.759 --> 00:07:57.199 in specific, meticulously pre -mapped areas like 00:07:57.199 --> 00:07:59.620 certain parts of Phoenix or San Francisco? Exactly. 00:07:59.740 --> 00:08:02.259 And the real advantage of LiDAR is objective 00:08:02.259 --> 00:08:04.379 mathematical certainty. I mean, a camera might 00:08:04.379 --> 00:08:07.180 get confused by a shadow, but a laser pulse physically 00:08:07.180 --> 00:08:09.360 hits an object and bounces back. It can't really 00:08:09.360 --> 00:08:11.439 be fooled by a picture on the site of a bus? 00:08:11.560 --> 00:08:13.939 No, it gives the computer undeniable proof that 00:08:13.939 --> 00:08:16.660 an object exists at a precise distance. However, 00:08:17.040 --> 00:08:19.199 LiDAR has a massive of physical limitation, which 00:08:19.199 --> 00:08:23.540 is precipitation. Ah, right, rain and snow. Yeah, 00:08:23.540 --> 00:08:26.480 if you shoot lasers into heavy rain, snow, or 00:08:26.480 --> 00:08:29.720 fog, the light scatters. The point cloud degrades 00:08:29.720 --> 00:08:32.000 entirely. And on the completely opposite end 00:08:32.000 --> 00:08:34.740 of the spectrum, you have Tesla. Tesla abandoned 00:08:34.740 --> 00:08:37.600 LIDAR and radar entirely. A very controversial 00:08:37.600 --> 00:08:40.820 move. Extremely. They adopted a vision -only 00:08:40.820 --> 00:08:43.879 approach. Their newer systems rely on an end 00:08:43.879 --> 00:08:46.279 -to -end deep learning neural network fed by 00:08:46.279 --> 00:08:48.480 just eight standard cameras positioned around 00:08:48.480 --> 00:08:51.929 the car. And their argument is that human beings 00:08:51.929 --> 00:08:54.649 drive using only our eyes and our brains. We 00:08:54.649 --> 00:08:56.570 don't have laser rays spinning on our heads. 00:08:56.769 --> 00:08:59.830 Right. We just have two eyeballs. Exactly. So 00:08:59.830 --> 00:09:01.950 a car should theoretically be able to navigate 00:09:01.950 --> 00:09:04.509 using just cameras and artificial intelligence. 00:09:04.730 --> 00:09:06.870 They also avoid the need for those centimeter 00:09:06.870 --> 00:09:09.629 -level, highly detailed maps that Waymo uses. 00:09:09.929 --> 00:09:12.190 The sources talk about approaches like MIT's 00:09:12.190 --> 00:09:15.529 maplight, which just uses simple 2D GPS maps 00:09:15.529 --> 00:09:17.950 paired with real -time sensor observation to 00:09:17.950 --> 00:09:20.269 figure out It's a much more adaptable approach. 00:09:20.409 --> 00:09:22.090 Yeah, it doesn't need to know where every stop 00:09:22.090 --> 00:09:24.269 sign was yesterday. You know, it just reads the 00:09:24.269 --> 00:09:27.409 road today. Which is a massive scalability advantage. 00:09:28.090 --> 00:09:30.909 Highly detailed HD maps are incredibly expensive 00:09:30.909 --> 00:09:33.889 to maintain because the real world changes daily. 00:09:34.210 --> 00:09:36.669 Construction pops up, lanes shift, signs fall 00:09:36.669 --> 00:09:39.690 down. Exactly. Systems that rely on fewer map 00:09:39.690 --> 00:09:43.129 details and more real -time processing are theoretically 00:09:43.129 --> 00:09:45.450 much easier to scale globally. Okay, wait, let 00:09:45.450 --> 00:09:47.879 me jump in here. If we successfully drive using 00:09:47.879 --> 00:09:51.240 just our eyes and a brain, why isn't Tesla's 00:09:51.240 --> 00:09:53.539 vision -only approach just fundamentally the 00:09:53.539 --> 00:09:56.340 correct path? That's a fair question. I mean, 00:09:56.580 --> 00:09:59.200 why bother with all the expensive fragile lasers 00:09:59.200 --> 00:10:01.740 if cheap cameras and a good enough neural network 00:10:01.740 --> 00:10:04.080 can do the job? It sounds perfectly logical until 00:10:04.080 --> 00:10:06.120 you look at the sheer computational nightmare 00:10:06.120 --> 00:10:08.240 of what our brains are actually doing when we 00:10:08.240 --> 00:10:10.070 drive. Okay, what do you mean? The challenge 00:10:10.070 --> 00:10:11.970 isn't just perception. It isn't just seeing the 00:10:11.970 --> 00:10:14.730 environment. The real insurmountable hurdle for 00:10:14.730 --> 00:10:17.590 AI right now is behavior prediction. You mean 00:10:17.590 --> 00:10:20.590 anticipating what happens next on the road? Precisely. 00:10:20.929 --> 00:10:23.730 It's not enough for the car's AI to see a pedestrian 00:10:23.730 --> 00:10:26.429 standing near a crosswalk. The system has to 00:10:26.429 --> 00:10:28.909 predict what that specific pedestrian will do 00:10:28.909 --> 00:10:31.370 in the next three seconds. Right. Are they going 00:10:31.370 --> 00:10:33.549 to walk or just stand there? Exactly. Will they 00:10:33.549 --> 00:10:35.529 step out? Are they looking at their phone? Are 00:10:35.529 --> 00:10:38.490 they distracted by a dog? To do this safely, 00:10:38.649 --> 00:10:41.470 the software has to wholly recompute the position, 00:10:41.809 --> 00:10:44.669 velocity, and trajectory of every single moving 00:10:44.669 --> 00:10:47.610 object in its environment many, many times per 00:10:47.610 --> 00:10:50.009 second. And doing that with just 2D video feed 00:10:50.009 --> 00:10:52.350 sounds incredibly difficult. It is intensely 00:10:52.350 --> 00:10:54.870 difficult. While humans do this instinctively 00:10:54.870 --> 00:10:57.470 through intuition and social cues, like making 00:10:57.470 --> 00:10:59.289 eye contact with a pedestrian to know they see 00:10:59.289 --> 00:11:03.149 you, a vision -only AI has to calculate 3D depth 00:11:03.149 --> 00:11:06.529 and velocity entirely from scratch. using just 00:11:06.529 --> 00:11:09.230 two -dimensional pixel changes over time. Yes. 00:11:09.549 --> 00:11:12.750 It is trying to infer depth from flat images. 00:11:12.889 --> 00:11:14.889 Right. That requires a staggering amount of processing 00:11:14.889 --> 00:11:18.309 power and flawless software logic. LiDAR, by 00:11:18.309 --> 00:11:21.009 contrast, just hands the computer the 3D math 00:11:21.009 --> 00:11:23.870 on a silver platter. So we have lasers struggling 00:11:23.870 --> 00:11:26.370 with rain and cameras struggling to do complex 00:11:26.370 --> 00:11:28.370 3D math. That's the state of the art right now. 00:11:28.549 --> 00:11:30.190 Are you and I actually safer handing over the 00:11:30.190 --> 00:11:32.350 wheel to a computer right now? The hard data 00:11:32.350 --> 00:11:35.149 on this. is incredibly revealing, actually, and 00:11:35.149 --> 00:11:37.090 it challenges a lot of our assumptions. What 00:11:37.090 --> 00:11:40.450 does it say? Well, a massive 2024 meta -alysis 00:11:40.450 --> 00:11:42.970 published in Nature Communications look at exactly 00:11:42.970 --> 00:11:46.330 this. They compared 2 ,100 autonomous vehicle 00:11:46.330 --> 00:11:50.429 incident records to over 35 ,000 human -driven 00:11:50.429 --> 00:11:53.590 incidents. Wow, that's a huge data set. Yeah. 00:11:54.149 --> 00:11:57.500 And the results are genuinely surprising. The 00:11:57.500 --> 00:12:00.240 good news is that, statistically, autonomous 00:12:00.240 --> 00:12:03.299 vehicles are significantly safer in most circumstances. 00:12:03.460 --> 00:12:05.360 Okay, so they are actually better than us in 00:12:05.360 --> 00:12:07.460 a lot of ways. They are. They're less likely 00:12:07.460 --> 00:12:10.200 to crash in heavy rain or fog than an unassisted 00:12:10.200 --> 00:12:13.299 human, surprisingly. But the really staggering 00:12:13.299 --> 00:12:15.860 number is regarding pedestrians. That's the one 00:12:15.860 --> 00:12:17.940 everyone worries about. Right. Autonomous vehicles 00:12:17.940 --> 00:12:20.860 have drastically fewer crashes involving pedestrians. 00:12:21.299 --> 00:12:24.519 Just 3 % of their crashes compared to 15 % for 00:12:24.519 --> 00:12:27.179 human drivers per mile traveled. That is a massive 00:12:27.179 --> 00:12:29.460 reduction in harm to the most vulnerable people 00:12:29.460 --> 00:12:32.100 on the road. It really is. The computers simply 00:12:32.100 --> 00:12:35.220 do not get distracted, they don't text, and they 00:12:35.220 --> 00:12:36.659 definitely don't drive drunk. Yeah, they don't 00:12:36.659 --> 00:12:39.820 get tired either. Exactly. But the data also 00:12:39.820 --> 00:12:42.500 revealed severe specific weaknesses that engineers 00:12:42.500 --> 00:12:46.000 are scrambling to fix. The bad news is that autonomous 00:12:46.000 --> 00:12:48.200 vehicles were found to be more than five times 00:12:48.200 --> 00:12:50.899 more vulnerable to collisions at dawn and dusk. 00:12:51.179 --> 00:12:53.320 Here's where it gets really interesting. Dawn 00:12:53.320 --> 00:12:55.559 and dusk. Let's talk about the mechanism behind 00:12:55.559 --> 00:12:58.179 that because it's not just some spooky coincidence. 00:12:58.500 --> 00:13:01.080 No, it's pure physics. It's about dynamic range. 00:13:01.360 --> 00:13:04.000 At dawn and dusk, the cameras are facing immense 00:13:04.000 --> 00:13:06.240 lighting challenges. Right, the contrast is just 00:13:06.240 --> 00:13:09.789 too high. They are trying to expose for a glaring 00:13:09.789 --> 00:13:13.049 low angle sun while simultaneously trying to 00:13:13.049 --> 00:13:16.110 see into deep dark shadows on the road. The sensors 00:13:16.110 --> 00:13:18.929 get overwhelmed. The A .I. hallucinates or completely 00:13:18.929 --> 00:13:22.149 misses objects and the system fails. And we have 00:13:22.149 --> 00:13:24.490 seen this technological limitation play out in 00:13:24.490 --> 00:13:27.610 high profile real world tragedies. We have. There 00:13:27.610 --> 00:13:30.990 was the fatal 2018 Uber crash in Arizona. The 00:13:30.990 --> 00:13:33.370 A .I. system failed to properly classify a pedestrian 00:13:33.370 --> 00:13:36.649 crossing a dark road outside a crosswalk. situation 00:13:36.649 --> 00:13:39.690 all around it was but crucially the human safety 00:13:39.690 --> 00:13:41.750 driver who was supposed to be the backup was 00:13:41.750 --> 00:13:44.009 distracted by her phone right the handover failed 00:13:44.009 --> 00:13:46.649 there have also been tesla autopilot crashes 00:13:46.649 --> 00:13:49.490 where the vision only system failed to distinguish 00:13:49.490 --> 00:13:51.850 the white side of a tractor trailer against a 00:13:51.850 --> 00:13:54.960 brightly lit glaring sky The camera just couldn't 00:13:54.960 --> 00:13:58.120 see the edge. Exactly. And in 2024, there were 00:13:58.120 --> 00:14:00.679 fatal incidents involving Ford's Blue Cruise, 00:14:01.080 --> 00:14:03.820 where drivers operating hands -free struck stationary 00:14:03.820 --> 00:14:06.700 cars on the freeway. In one of those Ford incidents, 00:14:07.039 --> 00:14:09.460 the driver was actually intoxicated in speeding. 00:14:09.899 --> 00:14:12.379 Which is horrifying. It is. And it highlights 00:14:12.379 --> 00:14:15.399 the terrifying blurred lines of level two systems. 00:14:15.879 --> 00:14:18.259 The driver abuses the tech, assuming it's level 00:14:18.259 --> 00:14:21.080 five, but the tech fails to prevent the disaster 00:14:21.080 --> 00:14:23.799 because it's only designed for level two supervision. 00:14:23.960 --> 00:14:26.299 But let me push back on the massive public fear 00:14:26.299 --> 00:14:28.539 surrounding this for a second. Sure. If the hard 00:14:28.539 --> 00:14:30.899 data from that Nature Communications study proves 00:14:30.899 --> 00:14:33.580 that these cars hit far fewer pedestrians than 00:14:33.580 --> 00:14:36.360 humans do, Why are we so terrified of them? That's 00:14:36.360 --> 00:14:39.500 a great question. A 2024 YouGov survey showed 00:14:39.500 --> 00:14:42.960 that 37 % of British adults feel very unsafe 00:14:42.960 --> 00:14:45.659 with the idea of a self -driving car. Are we 00:14:45.659 --> 00:14:48.120 holding machines to an impossibly perfect standard 00:14:48.120 --> 00:14:50.720 while completely ignoring our own human recklessness? 00:14:51.080 --> 00:14:53.740 We absolutely are. But it really comes down to 00:14:53.740 --> 00:14:57.399 the psychology of trust. You can throw all the 00:14:57.399 --> 00:14:59.759 statistical safety data in the world at someone, 00:14:59.879 --> 00:15:02.340 but consumers will not adopt this technology 00:15:02.340 --> 00:15:05.250 unless they feel safe. Right. Facts don't always 00:15:05.250 --> 00:15:08.789 change feelings. Exactly. Human error is familiar 00:15:08.789 --> 00:15:11.750 to us. We intimately understand a driver getting 00:15:11.750 --> 00:15:14.090 distracted or falling asleep or driving drunk. 00:15:14.389 --> 00:15:16.950 We know how to culturally categorize that tragedy. 00:15:17.110 --> 00:15:20.450 We have a box for it in our minds. Yes. But a 00:15:20.450 --> 00:15:24.590 machine making a fatal error like driving full 00:15:24.590 --> 00:15:27.230 speed into a highly visible stationary fire truck 00:15:27.230 --> 00:15:29.870 because the software's neural net misclassified 00:15:29.870 --> 00:15:32.809 its pixel data. That feels deeply unsettling 00:15:32.809 --> 00:15:34.809 and alien. Because we don't understand why it 00:15:34.809 --> 00:15:37.490 made that choice. Right. Trust is built on predictability. 00:15:38.190 --> 00:15:41.250 And right now, the edge case failures of AI are 00:15:41.250 --> 00:15:43.789 wildly unpredictable to the average person. And 00:15:43.789 --> 00:15:46.429 that unpredictability ripples out past the physical 00:15:46.429 --> 00:15:48.690 crashes and straight into the ethical decisions 00:15:48.690 --> 00:15:50.629 these software programs are making in a split 00:15:50.629 --> 00:15:52.509 second. Oh, the ethics are a whole other minefield. 00:15:52.730 --> 00:15:54.529 Because we aren't just talking about engineering 00:15:54.529 --> 00:15:56.669 challenges anymore. We are talking about massive 00:15:56.669 --> 00:16:00.279 societal impacts. Yes. The impact of this technology 00:16:00.279 --> 00:16:02.940 goes far beyond the chassis of the vehicle itself. 00:16:03.360 --> 00:16:05.600 Let's start with the ethics, because the source 00:16:05.600 --> 00:16:09.860 material brings up the modern AI version of the 00:16:09.860 --> 00:16:11.980 trolley problem. A classic philosophy thought 00:16:11.980 --> 00:16:14.759 experiment. Right. So imagine an autonomous car 00:16:14.759 --> 00:16:18.399 is driving down a narrow road. A pedestrian suddenly 00:16:18.399 --> 00:16:22.519 steps out. The car's AI calculates It physically 00:16:22.519 --> 00:16:25.059 cannot stop in time. It has to make a choice. 00:16:25.320 --> 00:16:27.639 Does it hit the pedestrian or does it swerve 00:16:27.639 --> 00:16:29.919 into a concrete wall potentially killing its 00:16:29.919 --> 00:16:33.399 own passenger? Who does the code choose to save? 00:16:33.759 --> 00:16:36.120 And the surveys on this are fascinating. They 00:16:36.120 --> 00:16:39.629 show a really dark slightly humorous hypocrisy 00:16:39.629 --> 00:16:42.190 and human nature here. People generally want 00:16:42.190 --> 00:16:44.509 cars to be programmed for the greater good meaning, 00:16:44.570 --> 00:16:46.809 you know, save the pedestrian, unless they are 00:16:46.809 --> 00:16:48.950 the ones sitting in the passenger seat. Then 00:16:48.950 --> 00:16:50.830 they want the car to protect them at all costs. 00:16:50.970 --> 00:16:52.730 It is the ultimate dilemma for the programmers. 00:16:52.990 --> 00:16:56.009 I mean, do you program the car to be a utilitarian 00:16:56.009 --> 00:16:59.009 hero or a selfish bodyguard? And how do you even 00:16:59.009 --> 00:17:02.009 market that? You can't really. And it is not 00:17:02.009 --> 00:17:04.250 the only ethical failure point we are seeing 00:17:04.250 --> 00:17:07.509 in the software. There is also the glaring issue. 00:17:07.930 --> 00:17:10.630 of algorithmic bias. This part of the research 00:17:10.630 --> 00:17:13.410 was wild to me. Yeah. Research from Georgia Tech 00:17:13.410 --> 00:17:15.789 revealed that autonomous vehicle detection systems 00:17:15.789 --> 00:17:18.609 were actually five percent less effective at 00:17:18.609 --> 00:17:21.029 recognizing darker skinned individuals. And the 00:17:21.029 --> 00:17:23.970 mechanics of why that happens are deeply concerning. 00:17:24.650 --> 00:17:27.170 Facial recognition and pedestrian detection rely 00:17:27.170 --> 00:17:30.069 heavily on contrast and edge detection. Right. 00:17:30.089 --> 00:17:32.670 It's looking for shapes in the dark. If the AI 00:17:32.670 --> 00:17:35.509 was trained on datasets that primarily featured 00:17:35.509 --> 00:17:38.329 lighter skinned individuals, it literally struggles 00:17:38.329 --> 00:17:40.950 to mathematically identify darker skin tones, 00:17:41.190 --> 00:17:42.970 especially in low light. It just doesn't have 00:17:42.970 --> 00:17:45.930 the baseline data. It means the AI is functionally 00:17:45.930 --> 00:17:48.250 carrying the bias of its training data into the 00:17:48.250 --> 00:17:51.170 real world. Which raises massive equity and inclusion 00:17:51.170 --> 00:17:53.950 concerns for pedestrians in diverse cities. It 00:17:53.950 --> 00:17:56.049 is a stark reminder that technology is never 00:17:56.049 --> 00:17:58.980 truly neutral. And speaking of societal impact, 00:17:59.119 --> 00:18:00.859 we have to talk about the economic shockwave 00:18:00.859 --> 00:18:03.160 coming our way. The job market implications are 00:18:03.160 --> 00:18:06.039 huge. If level four and level five autonomy actually 00:18:06.039 --> 00:18:08.599 scales up, we are looking at the displacement 00:18:08.599 --> 00:18:12.440 of nearly 2 .9 million jobs in the U .S. alone. 00:18:12.599 --> 00:18:14.960 That is a staggering number. We are talking about 00:18:14.960 --> 00:18:17.220 tractor -trailer drivers, taxi drivers, delivery 00:18:17.220 --> 00:18:20.519 drivers. To put that in perspective, that surpasses 00:18:20.519 --> 00:18:23.759 the total job losses of the 2008 Great Recession. 00:18:24.029 --> 00:18:25.930 And it's not just the drivers themselves who 00:18:25.930 --> 00:18:28.990 will be affected. No. What happens to the highway 00:18:28.990 --> 00:18:31.670 diners, the motels, the entire economies built 00:18:31.670 --> 00:18:34.289 around human transit? They could collapse. Plus, 00:18:34.710 --> 00:18:37.250 state tax revenues heavily rely on transportation 00:18:37.250 --> 00:18:39.910 fees, speeding tickets, and licensing, which 00:18:39.910 --> 00:18:42.490 could just plummet. If we connect this to the 00:18:42.490 --> 00:18:44.730 bigger picture... This is where the narrative 00:18:44.730 --> 00:18:47.609 shifts completely. We have to stop talking about 00:18:47.609 --> 00:18:50.230 self -driving cars as just an engineering challenge 00:18:50.230 --> 00:18:53.569 and start talking about them as a massive socioeconomic 00:18:53.569 --> 00:18:55.910 shift. Which brings up a very blunt, practical 00:18:55.910 --> 00:18:58.170 question about liability. Always comes down to 00:18:58.170 --> 00:19:00.809 the lawyers, doesn't it? It really does. If an 00:19:00.809 --> 00:19:03.569 AI is constantly self -learning, updating itself 00:19:03.569 --> 00:19:05.269 over the air while it sits in your driveway and 00:19:05.269 --> 00:19:08.609 it causes a fatal crash the next day, who actually 00:19:08.609 --> 00:19:11.529 goes to jail? That is the million dollar question. 00:19:11.589 --> 00:19:13.730 Is it the person sitting in the passenger seat 00:19:13.730 --> 00:19:16.829 who didn't intervene fast enough? Is it the CEO 00:19:16.829 --> 00:19:20.670 of the car company? Or is it the software developer 00:19:20.670 --> 00:19:23.670 who wrote the base code three years ago? The 00:19:23.670 --> 00:19:26.710 infusion of AI into cars fundamentally breaks 00:19:26.710 --> 00:19:29.109 our traditional legal frameworks for liability. 00:19:29.829 --> 00:19:32.769 For a century, traffic law has assumed a human 00:19:32.769 --> 00:19:35.470 mind is in control of the vehicle. Right. There's 00:19:35.470 --> 00:19:37.940 always someone to blame. Exactly. When you remove 00:19:37.940 --> 00:19:40.779 the human mind, you fragment the legal responsibility. 00:19:41.460 --> 00:19:43.859 Regulators in the UK and the EU are scrambling 00:19:43.859 --> 00:19:46.619 right now to draft new laws, like the Automated 00:19:46.619 --> 00:19:49.240 Vehicles Bill, to figure out exactly who holds 00:19:49.240 --> 00:19:51.900 the bag when the algorithm makes a fatal choice. 00:19:52.059 --> 00:19:54.079 It's totally uncharted territory. They are trying 00:19:54.079 --> 00:19:57.220 to create legal categories for, quote, entities 00:19:57.220 --> 00:19:59.660 backing the automated system because suing a 00:19:59.660 --> 00:20:02.170 neural network just doesn't work. It is an absolute 00:20:02.170 --> 00:20:04.109 legal minefield. We have covered a massive amount 00:20:04.109 --> 00:20:06.410 of ground today. We really have. We've seen that 00:20:06.410 --> 00:20:08.809 the road to level five full autonomy is paved 00:20:08.809 --> 00:20:11.690 with incredible AI, complex neural networks, 00:20:12.210 --> 00:20:14.289 and advanced sensor tech that is mathematically 00:20:14.289 --> 00:20:18.190 safer in many ways. The tech is a marvel, undeniably. 00:20:18.319 --> 00:20:20.799 But it is also littered with complex ethical 00:20:20.799 --> 00:20:23.740 dilemmas, confusing and sometimes deceptive marketing, 00:20:24.299 --> 00:20:26.700 and massive economic implications for millions 00:20:26.700 --> 00:20:29.420 of workers. The technology is advancing at lightning 00:20:29.420 --> 00:20:32.579 speed, but the societal infrastructure required 00:20:32.579 --> 00:20:35.759 to support it – the laws, the ethics, the public 00:20:35.759 --> 00:20:39.059 trust – is honestly struggling to keep pace. 00:20:39.529 --> 00:20:41.710 So what does this all mean for your morning commute? 00:20:41.930 --> 00:20:44.430 It means that even if you never personally buy 00:20:44.430 --> 00:20:46.769 a self -driving car, you are going to be sharing 00:20:46.769 --> 00:20:48.630 the road with them. They are already out there. 00:20:48.730 --> 00:20:51.230 You'll be sharing crosswalks and your local economy 00:20:51.230 --> 00:20:54.190 with them very soon. You need to understand what's 00:20:54.190 --> 00:20:57.250 happening. For instance, in places like California 00:20:57.250 --> 00:21:00.569 and Nevada, Mercedes is already rolling out SAE 00:21:00.569 --> 00:21:03.150 Level 3 systems with special turquoise marker 00:21:03.150 --> 00:21:05.410 lights on the exterior of the car. It's a very 00:21:05.410 --> 00:21:08.059 specific color for a reason. Yeah. When you see 00:21:08.059 --> 00:21:09.980 that turquoise light pull up next to you at a 00:21:09.980 --> 00:21:11.900 stoplight, you need to know it means the car 00:21:11.900 --> 00:21:14.519 is legally driving itself, and the human inside 00:21:14.519 --> 00:21:17.400 might be legally watching a movie. And as we 00:21:17.400 --> 00:21:20.720 adapt to that new reality, there is one final 00:21:20.720 --> 00:21:22.819 crucial element to consider that we haven't even 00:21:22.819 --> 00:21:26.619 touched on yet. What's that? Privacy. Oh, of 00:21:26.619 --> 00:21:29.279 course. These advanced cars require constant 00:21:29.279 --> 00:21:31.220 high -speed internet connections to function. 00:21:31.390 --> 00:21:34.329 They are caching your destinations, your daily 00:21:34.329 --> 00:21:37.190 routes, your media preferences, and recordings 00:21:37.190 --> 00:21:39.009 from the high definition cameras looking both 00:21:39.009 --> 00:21:41.829 outside and inside the cabin. It's a lot of data. 00:21:42.289 --> 00:21:44.589 When you step into a self -driving car, are you 00:21:44.589 --> 00:21:47.730 just entering a vehicle or are you stepping into 00:21:47.730 --> 00:21:50.509 a rolling data collection machine? One that could 00:21:50.509 --> 00:21:53.289 be hacked, monetized by corporations or monitored 00:21:53.289 --> 00:21:55.450 by foreign entities. That brings us right back 00:21:55.450 --> 00:21:58.089 to where we started. That primal feeling of letting 00:21:58.089 --> 00:22:00.210 go of the steering wheel and closing your eyes. 00:22:00.589 --> 00:22:02.710 Next time you get into a smart car and let the 00:22:02.710 --> 00:22:05.650 automation take over, ask yourself who else is 00:22:05.650 --> 00:22:06.230 riding with you?