1 00:00:00,160 --> 00:00:03,040 Welcome to the debate. Today we're looking at a 2 00:00:03,040 --> 00:00:07,680 controversy that sits right at the bleeding edge of 3 00:00:07,680 --> 00:00:11,120 transportation technology. It's a dispute that really 4 00:00:11,120 --> 00:00:14,360 divides the engineering world right down the middle, and the 5 00:00:14,360 --> 00:00:19,760 outcome is going to determine how and frankly if our vehicles 6 00:00:19,760 --> 00:00:21,640 drive themselves in the next decade. 7 00:00:22,000 --> 00:00:25,440 We are talking about the war between vision only systems and 8 00:00:25,440 --> 00:00:28,840 sensor fusion, specifically regarding the use of Lidar. 9 00:00:29,120 --> 00:00:31,480 And this isn't just a theoretical argument anymore, is 10 00:00:31,480 --> 00:00:33,640 it? We're looking at recent and, 11 00:00:33,640 --> 00:00:37,000 well, quite alarming reports about the deployment of Tesla's 12 00:00:37,000 --> 00:00:39,960 robo taxi fleet. The headline data suggests a 13 00:00:39,960 --> 00:00:43,080 crash rate that's significantly higher than human drivers. 14 00:00:43,320 --> 00:00:45,760 Some reports are saying up to four times higher. 15 00:00:46,360 --> 00:00:48,760 It it just raises a very uncomfortable question. 16 00:00:49,160 --> 00:00:52,840 Can a system that relies only on cameras ever truly match the 17 00:00:52,840 --> 00:00:55,920 reliability of a system that uses active laser sensors? 18 00:00:56,160 --> 00:00:58,040 And that's really the core of it. 19 00:00:58,280 --> 00:01:01,040 Can a computer see the world well enough with just video 20 00:01:01,040 --> 00:01:04,640 feeds to navigate safely? Or does excluding depth sensing 21 00:01:04,640 --> 00:01:07,920 hardware like Lidar resent an insurmountable barrier? 22 00:01:08,600 --> 00:01:11,680 I'm the advocate. My position is that visual input 23 00:01:11,680 --> 00:01:15,160 is, well, theoretically sufficient because it mimics the 24 00:01:15,160 --> 00:01:17,240 biological model. It mimics us. 25 00:01:17,640 --> 00:01:20,600 I also suspect the current crash data is being heavily 26 00:01:20,600 --> 00:01:23,920 misinterpreted because of some serious reporting biases. 27 00:01:24,360 --> 00:01:27,640 And I'm the dissenter. My position is that abandoning 28 00:01:27,640 --> 00:01:31,080 Lidar is a dangerous cost cutting measure that ignores the 29 00:01:31,080 --> 00:01:34,080 kind of redundancy you absolutely need for safety 30 00:01:34,080 --> 00:01:37,000 critical systems. We're seeing higher crash rates 31 00:01:37,000 --> 00:01:40,120 not because of a reporting bias, but because of a fundamental 32 00:01:40,120 --> 00:01:43,040 hardware deficit. When you take away the sensor 33 00:01:43,040 --> 00:01:46,080 that tells you exactly how far away an object is, you are 34 00:01:46,080 --> 00:01:48,960 introducing a level of risk that just shouldn't be on public 35 00:01:48,960 --> 00:01:50,760 roads. OK, so let's get into the 36 00:01:50,760 --> 00:01:53,160 machinery of this. We really need to start with the 37 00:01:53,160 --> 00:01:56,600 first principles argument, because this is the hill the 38 00:01:56,600 --> 00:01:59,440 vision only proponents are willing to die on. 39 00:01:59,720 --> 00:02:02,360 There's a perspective shared by a contributor to our source 40 00:02:02,360 --> 00:02:06,400 material, Retroviridae 6, that's basically an existence proof. 41 00:02:07,040 --> 00:02:09,360 Ah, the humans do it argument. Exactly. 42 00:02:09,360 --> 00:02:12,280 It's the biological argument. You and I drove here today. 43 00:02:12,280 --> 00:02:14,240 We navigated traffic. We merged. 44 00:02:14,400 --> 00:02:17,280 We avoided pedestrians. We did all of that using two 45 00:02:17,360 --> 00:02:20,840 passive optical sensors, our eyes, and a biological neural 46 00:02:20,840 --> 00:02:23,160 network, our brain. We don't have lidar in our 47 00:02:23,160 --> 00:02:25,960 foreheads, we don't emit laser pulses to measure time of 48 00:02:25,960 --> 00:02:29,320 flight, we don't have radar. We rely entirely on optical flow 49 00:02:29,320 --> 00:02:31,040 and pattern recognition. Sure. 50 00:02:31,160 --> 00:02:35,760 So Retrovira day six's point is that if a biological neural net 51 00:02:35,960 --> 00:02:39,960 can drive a car using only passive optical sensors, then it 52 00:02:39,960 --> 00:02:42,640 is physically possible for a synthetic neural net to do the 53 00:02:42,640 --> 00:02:44,680 same thing. The physics allows it. 54 00:02:45,120 --> 00:02:49,160 Therefore, the argument that Lidar is required is just false. 55 00:02:49,600 --> 00:02:52,760 Lidar might be a shortcut, but it isn't a necessity. 56 00:02:53,040 --> 00:02:56,480 The photons entering the camera contain all the information you 57 00:02:56,480 --> 00:02:58,920 need to drive. I'm sorry but I just don't buy 58 00:02:58,920 --> 00:03:00,960 that. Let me tell you why that is a 59 00:03:00,960 --> 00:03:04,040 huge category error. Your conflating otential with 60 00:03:04,040 --> 00:03:06,560 execution. Just because humans can drive 61 00:03:06,560 --> 00:03:08,880 with their eyes doesn't mean robots should drive without 62 00:03:08,880 --> 00:03:11,600 laser precision. But why not if the goal is to 63 00:03:11,600 --> 00:03:14,840 replicate human capability? Theology is full of flaws we're 64 00:03:14,840 --> 00:03:16,960 trying to engineer out of the system. 65 00:03:17,240 --> 00:03:20,920 The promise of autonomy isn't to drive as well as a distracted 66 00:03:20,920 --> 00:03:24,840 ape, it's to drive perfectly. And to drive perfectly, you need 67 00:03:24,840 --> 00:03:27,800 data that the human eye simply cannot provide. 68 00:03:28,000 --> 00:03:29,560 But we're not talking about fatigue. 69 00:03:29,760 --> 00:03:32,840 We're talking about the sensory input required to build a model 70 00:03:32,840 --> 00:03:34,800 of the world. But you cannot compete with 71 00:03:34,800 --> 00:03:37,880 Lidar are using only visual cameras when it comes to what 72 00:03:37,880 --> 00:03:41,560 I'd call ground truth. As another observer, Wiggly Worm 73 00:03:41,560 --> 00:03:44,840 pointed out in the materials, Lidar provides absolute depth 74 00:03:44,840 --> 00:03:46,600 data. So maybe we should break that 75 00:03:46,600 --> 00:03:48,920 down a bit for anyone who isn't a robotics engineer. 76 00:03:49,200 --> 00:03:52,240 Right. A camera is a passive sensor. 77 00:03:52,560 --> 00:03:55,640 It takes in light and it creates a flat 2D image. 78 00:03:56,200 --> 00:04:00,440 To figure out how far away a car is, the software has to analyze 79 00:04:00,440 --> 00:04:03,560 the size of that car in the image, compare it to what it 80 00:04:03,560 --> 00:04:07,000 thinks a car looks like, and then infer the distance it's 81 00:04:07,000 --> 00:04:09,600 guessing. It's a very educated guess, but 82 00:04:09,600 --> 00:04:11,920 it is a guess. It's inference based on 83 00:04:11,920 --> 00:04:16,399 perspective and parallax, yeah. Lidar is active, It shoots out a 84 00:04:16,399 --> 00:04:20,000 laser pulse, it hits a car, and it measures exactly how long it 85 00:04:20,000 --> 00:04:21,399 takes for the light to bounce back. 86 00:04:21,800 --> 00:04:24,840 It's simple physics. Distance equals time multiplied 87 00:04:24,840 --> 00:04:27,880 by the speed of light. It doesn't guess, it knows. 88 00:04:27,960 --> 00:04:31,520 It says there is an object 12.4 meters away. 89 00:04:32,160 --> 00:04:35,240 Wiggly Worm's point is that when you remove that sensor, you're 90 00:04:35,240 --> 00:04:37,560 forcing the computer to hallucinate death. 91 00:04:37,960 --> 00:04:41,240 And when you look at the stats, Tesla's robo taxi is reportedly 92 00:04:41,240 --> 00:04:43,720 crashing at a rate 4 times higher than humans. 93 00:04:44,160 --> 00:04:47,040 That isn't just the learning curve, that is a failure of 94 00:04:47,040 --> 00:04:49,800 perception. I think that statistic, the four 95 00:04:49,800 --> 00:04:52,880 times higher crash rate, is doing a lot of heavy lifting in 96 00:04:52,880 --> 00:04:56,400 your argument, and I I want to contextualize it. 97 00:04:56,600 --> 00:04:59,960 We need to be very careful about comparing apples to oranges 98 00:04:59,960 --> 00:05:02,520 here. A crash is a crash, isn't it? 99 00:05:02,680 --> 00:05:07,120 Not necessarily. Another analyst, Eskrove 2 He 100 00:05:07,120 --> 00:05:10,800 noted that this specific headline is based on a really 101 00:05:10,800 --> 00:05:15,000 small sample size. 5 incidents in Austin in a single month. 102 00:05:15,240 --> 00:05:17,880 But if you look at the granularity of those incidents, 103 00:05:18,120 --> 00:05:21,760 the data includes really minor events like a tire touching a 104 00:05:21,760 --> 00:05:25,720 parking sign or bumping A curb while parking. 5 incidents in a 105 00:05:25,720 --> 00:05:27,960 month for a small fleet is still high. 106 00:05:28,120 --> 00:05:31,920 But think about human behavior. If I scrape my rim on a curb 107 00:05:31,920 --> 00:05:35,120 while I'm parallel parking, or if I tap a plastic Bullard at 108 00:05:35,120 --> 00:05:37,040 one mile per hour, do I call the police? 109 00:05:37,040 --> 00:05:39,080 Do I call my insurance company? No. 110 00:05:39,120 --> 00:05:41,240 It never enters the statistical record. 111 00:05:41,520 --> 00:05:44,520 It just vanishes. Right, it's unreported. 112 00:05:44,520 --> 00:05:49,280 Exactly, but for a robo taxi, every single sensor reading is 113 00:05:49,280 --> 00:05:51,960 logged. Every thump is a reported 114 00:05:51,960 --> 00:05:54,160 incident. The system self-reports 115 00:05:54,160 --> 00:05:57,120 everything. So you're comparing reported 116 00:05:57,160 --> 00:06:01,120 autonomous incidents where every scratch is scrutinized against 117 00:06:01,120 --> 00:06:04,800 reported human accidents, which are usually only the the one 118 00:06:04,800 --> 00:06:06,720 severe enough to require tow truck. 119 00:06:07,480 --> 00:06:11,080 Ask contributors UX Test and Jerkletos pointed out you're 120 00:06:11,080 --> 00:06:13,920 comparing A microscope to a telescope and then claiming the 121 00:06:13,920 --> 00:06:17,000 microscope sees more dirt I. Understand the reporting bias 122 00:06:17,000 --> 00:06:19,600 argument. It's valid to a point, but I 123 00:06:19,600 --> 00:06:22,440 also think it's a convenient way to wave away failure. 124 00:06:22,600 --> 00:06:25,360 You call it rubbing a curb. I call it a failure of object 125 00:06:25,360 --> 00:06:27,280 permanence. That feels like a bit of a 126 00:06:27,280 --> 00:06:29,800 stretch for a scratched rim. Is it? 127 00:06:30,240 --> 00:06:33,560 Contrast this with Waymo. We have user experiences from 128 00:06:33,560 --> 00:06:37,240 San Francisco contributors like Turbo Encapsulator and Luda lol 129 00:06:37,480 --> 00:06:41,120 who describe Waymo as flawless in the same complex urban 130 00:06:41,120 --> 00:06:43,960 environments where these vision based systems are struggling. 131 00:06:44,440 --> 00:06:47,360 Waymo uses Lidar. They have that spinning bucket 132 00:06:47,360 --> 00:06:49,520 on the roof. They aren't scraping rims. 133 00:06:49,680 --> 00:06:52,640 They aren't bumping signs. They're also driving in a 134 00:06:52,640 --> 00:06:54,760 fishbowl. You're driving in San Francisco. 135 00:06:54,760 --> 00:06:57,680 That's hardly officiable. It's a Geo fenced pre mapped 136 00:06:57,680 --> 00:07:01,480 environment that Waymo knows exactly where every curb is 137 00:07:01,480 --> 00:07:04,680 because it has a high definition map stored in its hard drive. 138 00:07:05,040 --> 00:07:07,240 It's not seeing the curb, it's remembering it. 139 00:07:07,720 --> 00:07:10,760 Tesla's vision approach is trying to do something much much 140 00:07:10,760 --> 00:07:13,560 harder. Drive anywhere on any road 141 00:07:13,600 --> 00:07:15,480 without a map, just like a human. 142 00:07:15,920 --> 00:07:17,680 Of course it's going to be clumsier in the beginning. 143 00:07:18,000 --> 00:07:19,920 It's learning general intelligence, not just 144 00:07:19,920 --> 00:07:22,640 memorizing a map. But that clumsiness has real 145 00:07:22,640 --> 00:07:25,560 world consequences. The reports of these vision only 146 00:07:25,560 --> 00:07:28,800 cars hitting stationary objects like parking signs? 147 00:07:29,000 --> 00:07:30,760 That indicates A fundamental flaw. 148 00:07:31,080 --> 00:07:33,760 If a vision system cannot calculate the distance to a 149 00:07:33,760 --> 00:07:36,800 concrete Bullard well enough to avoid hitting it, how can we 150 00:07:36,800 --> 00:07:40,000 possibly trust it to calculate the velocity of a child running 151 00:07:40,000 --> 00:07:42,520 into the street? Because the neural networks are 152 00:07:42,520 --> 00:07:45,640 weighted differently for those tasks, the system is likely 153 00:07:45,640 --> 00:07:48,520 hyper cautious around pedestrians, but has a higher 154 00:07:48,520 --> 00:07:52,640 tolerance for static objects to facilitate, say, parking. 155 00:07:52,800 --> 00:07:56,320 That is an assumption. Lidar solves the static object 156 00:07:56,320 --> 00:07:59,600 problem instantly. It doesn't need to infer or wait 157 00:07:59,600 --> 00:08:01,880 anything. It hits the Ballard with a laser 158 00:08:01,880 --> 00:08:04,800 and it knows it's there. The fact that these vision based 159 00:08:04,800 --> 00:08:07,680 cars are hitting stationary objects suggests that the 160 00:08:07,680 --> 00:08:10,760 software is hallucinating free space where there is solid 161 00:08:10,760 --> 00:08:12,880 matter. That is terrifying. 162 00:08:13,280 --> 00:08:15,920 It implies the car literally does not know the physical 163 00:08:15,920 --> 00:08:17,320 boundaries of its own environment. 164 00:08:17,560 --> 00:08:20,800 I will concede that static object detection is a hurdle 165 00:08:20,920 --> 00:08:25,600 right now, but identifying these edge cases is exactly how you 166 00:08:25,600 --> 00:08:28,120 train the network. Every time it hits a parking 167 00:08:28,120 --> 00:08:31,760 sign at 2 mph, it uploads that failure and the entire fleet 168 00:08:31,760 --> 00:08:35,039 learns not to do it again. And this brings us directly to 169 00:08:35,039 --> 00:08:38,280 the concept of systemic risk. We have to talk about the 170 00:08:38,280 --> 00:08:40,640 multiplier effect. Explain how you view that. 171 00:08:40,640 --> 00:08:43,799 This was articulated very well by the source contributor Becker 172 00:08:43,799 --> 00:08:46,000 Hollow. The argument is all about error 173 00:08:46,000 --> 00:08:48,160 scaling. When a human makes a mistake, it 174 00:08:48,160 --> 00:08:51,160 causes 1 accident. Human error is stochastic. 175 00:08:51,160 --> 00:08:53,240 It's random. You might get distracted by a 176 00:08:53,240 --> 00:08:55,200 text. I might drop my coffee. 177 00:08:55,200 --> 00:08:58,280 It's isolated. Sure, individual variants. 178 00:08:58,400 --> 00:09:01,040 But when a vision based software has a flaw in its programming, 179 00:09:01,040 --> 00:09:04,040 say a specific inability to distinguish a white truck 180 00:09:04,040 --> 00:09:07,440 against a bright sky, that error is replicated across every 181 00:09:07,440 --> 00:09:09,920 single device on the road. The centralized bug. 182 00:09:10,120 --> 00:09:12,720 Exactly. If the software misinterprets a 183 00:09:12,720 --> 00:09:15,760 specific shadow or glare, thousands of cars become 184 00:09:15,760 --> 00:09:18,480 dangerous simultaneously in the exact same way. 185 00:09:18,880 --> 00:09:21,320 You aren't dealing with one bad driver, you're dealing with a 186 00:09:21,320 --> 00:09:23,640 fleet of clones all sharing the same blind spot. 187 00:09:24,120 --> 00:09:27,400 That is a systemic risk profile that we have never, ever dealt 188 00:09:27,400 --> 00:09:30,080 with in automotive history. That's an interesting point, 189 00:09:30,080 --> 00:09:32,040 though I would frame it differently. 190 00:09:32,280 --> 00:09:35,720 That logic flips both ways. It's actually the strongest 191 00:09:35,720 --> 00:09:39,320 argument for autonomous systems. Yes, an error is distributed, 192 00:09:39,520 --> 00:09:42,600 but so is the solution. If they catch it in time. 193 00:09:42,720 --> 00:09:45,920 Think about it, when a human driver is bad at merging, 194 00:09:45,920 --> 00:09:47,760 they're usually bad at merging forever. 195 00:09:47,920 --> 00:09:51,080 You can't patch their brain, but if you solve the edge case in 196 00:09:51,080 --> 00:09:54,680 software, if you fix that white truck against the sky bug, you 197 00:09:54,720 --> 00:09:56,840 instantly fix every car on the road. 198 00:09:57,080 --> 00:10:00,120 You can upgrade the safety of the entire fleet overnight with 199 00:10:00,120 --> 00:10:02,920 an over the air update. The multiplier effect applies to 200 00:10:02,920 --> 00:10:05,000 safety even more than it applies to error. 201 00:10:05,200 --> 00:10:07,920 You're leveraging the collective learning of millions of miles. 202 00:10:08,080 --> 00:10:10,120 But. Until that fix arrives, the risk 203 00:10:10,120 --> 00:10:12,360 is distributed to the public without their consent. 204 00:10:12,720 --> 00:10:15,440 The public roads are becoming a beta testing environment. 205 00:10:15,960 --> 00:10:19,160 We're seeing a move fast and break things mentality applied 206 00:10:19,160 --> 00:10:23,640 to two ton metal projectiles. Beckerhollow's logic holds the 207 00:10:23,680 --> 00:10:26,760 error rate is a normal human error multiplied by the number 208 00:10:26,760 --> 00:10:29,880 of devices using the program. If the program is flawed, the 209 00:10:29,880 --> 00:10:32,600 carnage is scalable. I see why you think that, but 210 00:10:32,720 --> 00:10:34,720 let me give you a different perspective on the technical 211 00:10:34,720 --> 00:10:37,680 reliability piece. You keep going back to Lidar as 212 00:10:37,680 --> 00:10:41,560 this source of truth, but Lidar has its own failure modes. 213 00:10:41,560 --> 00:10:43,800 It does, but they are different from cameras. 214 00:10:44,120 --> 00:10:48,000 Lidar struggles with heavy rain. The laser pulses scatter off the 215 00:10:48,000 --> 00:10:50,480 water droplets. It struggles with fog. 216 00:10:50,680 --> 00:10:53,280 It can get confused by interference from other lidar 217 00:10:53,280 --> 00:10:55,160 units. It's not magic. 218 00:10:55,160 --> 00:10:57,720 And this is where the whole sensor fusion argument gets 219 00:10:57,720 --> 00:10:58,600 tricky. Go on. 220 00:10:58,680 --> 00:11:02,600 When you have a camera an A lidar, they will often disagree. 221 00:11:02,920 --> 00:11:05,920 The camera sees a plastic bag blowing across the road and 222 00:11:05,920 --> 00:11:09,880 thinks it's nothing. Lidar sees an object and says 223 00:11:10,000 --> 00:11:13,640 obstacle emergency brake. Now the computer has to decide 224 00:11:13,640 --> 00:11:16,880 which sensor to trust. This is the sensor fusion 225 00:11:16,880 --> 00:11:20,000 conflict. By removing Lidar, Tesla is 226 00:11:20,000 --> 00:11:21,880 arguing that you remove the noise. 227 00:11:22,120 --> 00:11:25,240 You force the neural net to resolve the visual data just 228 00:11:25,240 --> 00:11:28,440 like a human does, without getting confused by conflicting 229 00:11:28,440 --> 00:11:30,240 signals. That sounds like a very 230 00:11:30,240 --> 00:11:33,080 convenient engineering rationalization for saving 231 00:11:33,080 --> 00:11:35,840 money. Cameras are cheap, Lidar is 232 00:11:35,840 --> 00:11:38,400 expensive. It's definitely cheaper, but 233 00:11:38,400 --> 00:11:41,760 retroverted Z6 mentioned. We went from the horse and buggy 234 00:11:41,760 --> 00:11:43,600 to the moon in just a few decades. 235 00:11:44,040 --> 00:11:46,600 Assuming that computer vision can't bridge the gap just 236 00:11:46,600 --> 00:11:48,320 because it hasn't yet is premature. 237 00:11:48,800 --> 00:11:51,720 The issue isn't that the camera is blind, it's that the 238 00:11:51,720 --> 00:11:54,080 processing isn't yet sophisticated enough. 239 00:11:54,560 --> 00:11:57,520 But processing power is scaling exponentially. 240 00:11:57,680 --> 00:12:01,240 Software cannot conjure photons where there are none. 241 00:12:01,360 --> 00:12:04,680 That's the physics problem. But it can interpret context. 242 00:12:04,880 --> 00:12:08,000 Let's talk about those photons. Cameras are passive. 243 00:12:08,080 --> 00:12:10,520 They need light. What happens when you drive 244 00:12:10,520 --> 00:12:12,560 directly into the sunset? We've all done it. 245 00:12:12,560 --> 00:12:14,200 The visor goes down. You squint. 246 00:12:14,200 --> 00:12:16,800 You can barely see. Cameras get blinded by sun 247 00:12:16,800 --> 00:12:18,480 glare. They get obscured by mud. 248 00:12:18,720 --> 00:12:22,080 We have reports that Tesla is having to employ trailing chase 249 00:12:22,080 --> 00:12:26,080 cars with human safety monitors for their autonomous taxis. 250 00:12:26,360 --> 00:12:29,120 If the system is so theoretically sound, why does it 251 00:12:29,120 --> 00:12:31,680 need a human babysitter in a separate vehicle? 252 00:12:31,920 --> 00:12:35,240 Every developmental technology has safety protocols during 253 00:12:35,240 --> 00:12:37,800 testing, but. This is being sold as a future 254 00:12:37,800 --> 00:12:41,080 that is just around the corner. The reliance on cameras 255 00:12:41,080 --> 00:12:43,800 introduces A fragility that Lidar solves. 256 00:12:44,280 --> 00:12:47,840 Lidar cuts through sun glare. It works in total darkness. 257 00:12:48,160 --> 00:12:52,040 It is a second layer of truth. If the camera sees a shadow and 258 00:12:52,040 --> 00:12:55,480 thinks it's a hole in the road, the Lidar says no, the ground is 259 00:12:55,480 --> 00:12:58,080 flat. Removing that sensor removes a 260 00:12:58,080 --> 00:13:01,080 layer of survival. It's engineering hubris to 261 00:13:01,080 --> 00:13:04,560 believe you can derive 100% certainty from a sensor that is 262 00:13:04,560 --> 00:13:06,600 susceptible to optical illusions. 263 00:13:06,840 --> 00:13:09,960 I'm not convinced by that line of reasoning because it assumes 264 00:13:09,960 --> 00:13:12,280 we can't solve optical illusions with better AI. 265 00:13:12,920 --> 00:13:15,440 But let's pivot to the consequences of this, because 266 00:13:15,440 --> 00:13:18,400 the legal aspect is fascinating. It's a nightmare. 267 00:13:18,680 --> 00:13:21,280 This leads us to the inevitable question of accountability. 268 00:13:21,840 --> 00:13:25,120 When these systems do fail, whether it's a clumsy bump or a 269 00:13:25,120 --> 00:13:27,360 serious collision, who is responsible? 270 00:13:27,480 --> 00:13:30,760 This is the question posed by Shifty Mennonite in our source 271 00:13:30,760 --> 00:13:33,000 threads. Who is going to be held 272 00:13:33,000 --> 00:13:35,480 accountable when these things mow people down? 273 00:13:35,800 --> 00:13:39,160 It is a legal quagmire. Is it the driver, which in this 274 00:13:39,160 --> 00:13:42,680 case is the software? Is it the manufacturer or is it 275 00:13:42,680 --> 00:13:44,680 the limitations of the sensor suite itself? 276 00:13:44,920 --> 00:13:48,120 I think we need to distinguish between a software bug and a 277 00:13:48,120 --> 00:13:50,680 design choice. This is crucial. 278 00:13:51,120 --> 00:13:54,400 If a car crashes because of a line of bad code, that's one 279 00:13:54,400 --> 00:13:57,040 thing. But if a manufacturer knowingly 280 00:13:57,040 --> 00:14:00,720 removes a safety sensor like Lidar, a sensor that is industry 281 00:14:00,720 --> 00:14:04,240 standard for competitors like Waymo, and that removal leads to 282 00:14:04,240 --> 00:14:07,320 a crash because the camera couldn't estimate depth, that 283 00:14:07,320 --> 00:14:09,720 feels very distinct from a mere coding error. 284 00:14:09,880 --> 00:14:13,120 You're suggesting negligence. I'm saying it borders on it. 285 00:14:13,280 --> 00:14:17,560 There's this sentiment expressed by User Beneficial Soup 3699 286 00:14:17,720 --> 00:14:21,240 regarding blatant fraud. While that is, you know, strong 287 00:14:21,240 --> 00:14:23,800 language, the core sentiment is valid. 288 00:14:24,000 --> 00:14:26,680 If you claim a camera is sufficient and the physics 289 00:14:26,680 --> 00:14:30,160 suggest it isn't, and the data shows it crashing, at what point 290 00:14:30,160 --> 00:14:33,280 does adherence to a vision only philosophy become liability? 291 00:14:33,640 --> 00:14:36,600 But that assumes Lidar was prevented that specific crash. 292 00:14:36,720 --> 00:14:40,400 We don't know that. Like I said, LIDAR isn't a magic 293 00:14:40,400 --> 00:14:43,240 bullet. It's not magic, it's redundancy. 294 00:14:43,480 --> 00:14:46,960 In aviation, we don't fly with one altimeter, we have three. 295 00:14:47,120 --> 00:14:49,280 Why on earth should we drive with one type of eye? 296 00:14:49,600 --> 00:14:53,280 If the camera fails due to glare or a bug or mud, there's nothing 297 00:14:53,280 --> 00:14:56,000 to catch the car. It is a single point of failure 298 00:14:56,000 --> 00:14:58,480 system. But there's an economic argument 299 00:14:58,480 --> 00:15:01,080 here too. If you require Lidar, you make 300 00:15:01,080 --> 00:15:05,040 autonomous cars cost $100,000. They become toys for the rich. 301 00:15:05,280 --> 00:15:09,200 If you can solve it with vision, the hardware costs 500 $100. 302 00:15:09,240 --> 00:15:13,200 You can put it in every car. A vision based system that's 99% 303 00:15:13,200 --> 00:15:17,280 safe and available to everyone might save more total lives than 304 00:15:17,280 --> 00:15:21,960 a lighter system that's 99.9% safe but only 1000 people can 305 00:15:21,960 --> 00:15:24,680 afford it. That is a utilitarian calculus 306 00:15:24,680 --> 00:15:27,600 that works on a spreadsheet, but it doesn't work when you're the 307 00:15:27,600 --> 00:15:30,920 one crossing the street. The public Rd. should not be a 308 00:15:30,920 --> 00:15:34,200 testing ground for cost cutting measures, disguises innovation. 309 00:15:35,000 --> 00:15:37,880 The experiences of users in San Francisco and Austin show a 310 00:15:37,880 --> 00:15:40,720 clear divide. Waymo with Lidar is providing A 311 00:15:40,720 --> 00:15:44,200 flawless service while vision only systems are struggling with 312 00:15:44,200 --> 00:15:47,560 basic static objects. But again, Waymo is on rails. 313 00:15:47,600 --> 00:15:50,840 It's a local maximum. It's great for San Francisco, 314 00:15:50,840 --> 00:15:52,720 but it doesn't scale to the rest of the world. 315 00:15:53,040 --> 00:15:56,200 I'd rather have a safe local maximum than a dangerous global 316 00:15:56,200 --> 00:15:58,880 beta test. Until vision systems can match 317 00:15:58,880 --> 00:16:01,880 the redundancy and depth accuracy of Lidar, the safety 318 00:16:01,880 --> 00:16:04,760 consequences are real and they are statistically proven. 319 00:16:05,160 --> 00:16:08,480 We cannot verify the safety of a black box neural net without 320 00:16:08,480 --> 00:16:11,600 ground truth sensors. It ultimately comes down to that 321 00:16:11,600 --> 00:16:13,440 multiplier effect we talked about. 322 00:16:13,440 --> 00:16:15,600 It does. We are at a crossroads. 323 00:16:16,040 --> 00:16:19,600 We can take the safe, expensive route with Lidar, which might 324 00:16:19,600 --> 00:16:22,800 limit the scalability of the technology but provides that 325 00:16:22,800 --> 00:16:26,640 warm blanket of redundancy. Or we can push for the vision 326 00:16:26,640 --> 00:16:31,040 solution, which, if it works, multiply safety exponentially 327 00:16:31,040 --> 00:16:34,040 across the globe and solves general intelligence. 328 00:16:34,280 --> 00:16:36,680 But if it fails, it multiplies error. 329 00:16:36,960 --> 00:16:40,200 It multiplies the risk of a single software blind spot into 330 00:16:40,200 --> 00:16:43,480 a nationwide. Hazard, and that is the gamble. 331 00:16:44,040 --> 00:16:47,120 Is the current risk worth the future reward? 332 00:16:47,840 --> 00:16:51,640 I tend to believe that without taking that risk, we stagnate. 333 00:16:51,920 --> 00:16:55,560 We'd still be driving horses if we waited for the perfect car. 334 00:16:55,800 --> 00:16:59,160 And I would argue that safety is not a place for gambling when 335 00:16:59,160 --> 00:17:02,120 you're moving 2 tons of steel at 60 mph. 336 00:17:02,320 --> 00:17:05,359 Pretty good isn't good enough. You need absolute truth and 337 00:17:05,359 --> 00:17:08,800 cameras just don't provide that. A fundamental disagreement on 338 00:17:08,800 --> 00:17:12,280 the philosophy of engineering. Thank you for listening to the 339 00:17:12,280 --> 00:17:14,880 debate. We hope this exchange has 340 00:17:14,880 --> 00:17:17,440 illuminated the complexities behind the sensors. 341 00:17:17,599 --> 00:17:20,720 Drive safe everyone, and watch out for the robots. 342 00:17:20,920 --> 00:17:21,560 Goodbye.