WEBVTT 00:00.000 --> 00:03.000 The following is a conversation with Elon Musk. 00:03.000 --> 00:06.240 He's the CEO of Tesla, SpaceX, Neuralink, 00:06.240 --> 00:09.200 and a cofounder of several other companies. 00:09.200 --> 00:10.740 This conversation is part 00:10.740 --> 00:13.200 of the artificial intelligence podcast. 00:13.200 --> 00:15.640 The series includes leading researchers 00:15.640 --> 00:19.320 in academia and industry, including CEOs and CTOs 00:19.320 --> 00:24.080 of automotive, robotics, AI, and technology companies. 00:24.080 --> 00:26.880 This conversation happened after the release of the paper 00:26.880 --> 00:30.520 from our group at MIT on driver functional vigilance 00:30.520 --> 00:32.880 during use of Tesla's autopilot. 00:32.880 --> 00:34.560 The Tesla team reached out to me, 00:34.560 --> 00:37.480 offering a podcast conversation with Mr. Musk. 00:37.480 --> 00:40.640 I accepted with full control of questions I could ask 00:40.640 --> 00:43.560 and the choice of what is released publicly. 00:43.560 --> 00:46.840 I ended up editing out nothing of substance. 00:46.840 --> 00:49.680 I've never spoken with Elon before this conversation, 00:49.680 --> 00:51.720 publicly or privately. 00:51.720 --> 00:54.360 Neither he nor his companies have any influence 00:54.360 --> 00:57.840 on my opinion, nor on the rigor and integrity 00:57.840 --> 00:59.760 of the scientific method that I practice 00:59.760 --> 01:01.840 in my position at MIT. 01:01.840 --> 01:04.640 Tesla has never financially supported my research 01:04.640 --> 01:07.320 and I've never owned a Tesla vehicle. 01:07.320 --> 01:10.160 I've never owned Tesla stock. 01:10.160 --> 01:12.800 This podcast is not a scientific paper. 01:12.800 --> 01:14.360 It is a conversation. 01:14.360 --> 01:16.720 I respect Elon as I do all other leaders 01:16.720 --> 01:18.680 and engineers I've spoken with. 01:18.680 --> 01:21.440 We agree on some things and disagree on others. 01:21.440 --> 01:23.480 My goal is always with these conversations 01:23.480 --> 01:26.920 is to understand the way the guest sees the world. 01:26.920 --> 01:28.600 One particular point of this agreement 01:28.600 --> 01:30.640 in this conversation was the extent 01:30.640 --> 01:33.240 to which camera based driver monitoring 01:33.240 --> 01:36.120 will improve outcomes and for how long 01:36.120 --> 01:39.120 it will remain relevant for AI assisted driving. 01:39.960 --> 01:42.240 As someone who works on and is fascinated 01:42.240 --> 01:45.200 by human centered artificial intelligence, 01:45.200 --> 01:48.720 I believe that if implemented and integrated effectively, 01:48.720 --> 01:51.840 camera based driver monitoring is likely to be of benefit 01:51.840 --> 01:55.640 in both the short term and the long term. 01:55.640 --> 01:59.240 In contrast, Elon and Tesla's focus 01:59.240 --> 02:01.200 is on the improvement of autopilot 02:01.200 --> 02:04.480 such that its statistical safety benefits 02:04.480 --> 02:09.040 override any concern of human behavior and psychology. 02:09.040 --> 02:12.040 Elon and I may not agree on everything 02:12.040 --> 02:13.920 but I deeply respect the engineering 02:13.920 --> 02:16.880 and innovation behind the efforts that he leads. 02:16.880 --> 02:20.560 My goal here is to catalyze a rigorous, nuanced 02:20.560 --> 02:23.520 and objective discussion in industry and academia 02:23.520 --> 02:26.240 on AI assisted driving, 02:26.240 --> 02:30.840 one that ultimately makes for a safer and better world. 02:30.840 --> 02:34.600 And now here's my conversation with Elon Musk. 02:35.600 --> 02:38.640 What was the vision, the dream of autopilot 02:38.640 --> 02:41.400 when in the beginning the big picture system level 02:41.400 --> 02:43.680 when it was first conceived 02:43.680 --> 02:45.960 and started being installed in 2014 02:45.960 --> 02:47.520 in the hardware and the cars? 02:47.520 --> 02:49.760 What was the vision, the dream? 02:49.760 --> 02:51.400 I would characterize the vision or dream 02:51.400 --> 02:54.400 simply that there are obviously two 02:54.400 --> 02:59.400 massive revolutions in the automobile industry. 03:00.120 --> 03:04.440 One is the transition to electrification 03:04.440 --> 03:06.400 and then the other is autonomy. 03:07.720 --> 03:12.720 And it became obvious to me that in the future 03:13.240 --> 03:16.240 any car that does not have autonomy 03:16.240 --> 03:19.160 I would be about as useful as a horse. 03:19.160 --> 03:22.040 Which is not to say that there's no use, it's just rare 03:22.040 --> 03:23.640 and somewhat idiosyncratic 03:23.640 --> 03:25.480 if somebody has a horse at this point. 03:25.480 --> 03:28.000 It's just obvious that cars will drive themselves completely. 03:28.000 --> 03:29.600 It's just a question of time 03:29.600 --> 03:34.600 and if we did not participate in the autonomy revolution 03:36.920 --> 03:40.840 then our cars would not be useful to people 03:40.840 --> 03:43.680 relative to cars that are autonomous. 03:43.680 --> 03:47.160 I mean an autonomous car is arguably worth 03:47.160 --> 03:52.160 five to 10 times more than a car that which is not autonomous. 03:53.760 --> 03:55.160 In the long term. 03:55.160 --> 03:56.200 Turns out what you mean by long term, 03:56.200 --> 03:59.520 but let's say at least for the next five years 03:59.520 --> 04:00.520 perhaps 10 years. 04:01.440 --> 04:04.080 So there are a lot of very interesting design choices 04:04.080 --> 04:05.720 with autopilot early on. 04:05.720 --> 04:09.960 First is showing on the instrument cluster 04:09.960 --> 04:12.680 or in the Model 3 on the center stack display 04:12.680 --> 04:15.720 what the combined sensor suite sees. 04:15.720 --> 04:17.920 What was the thinking behind that choice? 04:17.920 --> 04:18.960 Was there a debate? 04:18.960 --> 04:20.480 What was the process? 04:20.480 --> 04:24.840 The whole point of the display is to provide a health check 04:24.840 --> 04:28.080 on the vehicle's perception of reality. 04:28.080 --> 04:31.320 So the vehicle's taking information for a bunch of sensors 04:31.320 --> 04:34.680 primarily cameras, but also radar and ultrasonics, 04:34.680 --> 04:35.960 GPS and so forth. 04:37.200 --> 04:42.200 And then that information is then rendered into vector space 04:42.200 --> 04:46.360 and that with a bunch of objects with properties 04:46.360 --> 04:49.920 like lane lines and traffic lights and other cars. 04:49.920 --> 04:54.920 And then in vector space that is re rendered onto a display 04:54.920 --> 04:57.400 so you can confirm whether the car knows 04:57.400 --> 05:00.600 what's going on or not by looking out the window. 05:01.600 --> 05:04.240 Right, I think that's an extremely powerful thing 05:04.240 --> 05:06.480 for people to get an understanding 05:06.480 --> 05:07.840 to become one with the system 05:07.840 --> 05:10.400 and understanding what the system is capable of. 05:10.400 --> 05:13.600 Now, have you considered showing more? 05:13.600 --> 05:15.400 So if we look at the computer vision, 05:16.400 --> 05:18.400 you know, like road segmentation, lane detection, 05:18.400 --> 05:21.640 vehicle detection, object detection, underlying the system, 05:21.640 --> 05:24.400 there is at the edges some uncertainty. 05:24.400 --> 05:28.400 Have you considered revealing the parts 05:28.400 --> 05:32.400 that the uncertainty in the system, the sort of problem 05:32.400 --> 05:35.000 these associated with say image recognition 05:35.000 --> 05:35.840 or something like that? 05:35.840 --> 05:37.840 Yeah, so right now it shows like the vehicles 05:37.840 --> 05:40.840 and the vicinity of very clean crisp image 05:40.840 --> 05:43.840 and people do confirm that there's a car in front of me 05:43.840 --> 05:45.840 and the system sees there's a car in front of me 05:45.840 --> 05:47.840 but to help people build an intuition 05:47.840 --> 05:51.840 of what computer vision is by showing some of the uncertainty. 05:51.840 --> 05:53.840 Well, I think it's, in my car, 05:53.840 --> 05:56.840 I always look at the sort of the debug view 05:56.840 --> 05:58.840 and there's two debug views. 05:58.840 --> 06:03.840 One is augmented vision, which I'm sure you've seen 06:03.840 --> 06:07.840 where it's basically, we draw boxes and labels 06:07.840 --> 06:10.840 around objects that are recognized. 06:10.840 --> 06:14.840 And then there's what we call the visualizer, 06:14.840 --> 06:16.840 which is basically a vector space representation 06:16.840 --> 06:21.840 summing up the input from all sensors. 06:21.840 --> 06:23.840 That does not show any pictures, 06:23.840 --> 06:26.840 but it shows all of the, 06:26.840 --> 06:32.840 it basically shows the cause view of the world in vector space. 06:32.840 --> 06:36.840 But I think this is very difficult for normal people to understand. 06:36.840 --> 06:38.840 They would not know what they're looking at. 06:38.840 --> 06:40.840 So it's almost an HMI challenge. 06:40.840 --> 06:42.840 The current things that are being displayed 06:42.840 --> 06:46.840 is optimized for the general public understanding 06:46.840 --> 06:48.840 of what the system is capable of. 06:48.840 --> 06:50.840 It's like if you have no idea how computer vision works 06:50.840 --> 06:52.840 or anything, you can still look at the screen 06:52.840 --> 06:54.840 and see if the car knows what's going on. 06:54.840 --> 06:57.840 And then if you're a development engineer 06:57.840 --> 07:01.840 or if you have the development build like I do, 07:01.840 --> 07:05.840 then you can see all the debug information. 07:05.840 --> 07:10.840 But those would just be total diverse to most people. 07:10.840 --> 07:13.840 What's your view on how to best distribute effort? 07:13.840 --> 07:16.840 So there's three, I would say, technical aspects of autopilot 07:16.840 --> 07:18.840 that are really important. 07:18.840 --> 07:19.840 So it's the underlying algorithms, 07:19.840 --> 07:21.840 like the neural network architecture. 07:21.840 --> 07:23.840 There's the data that's trained on 07:23.840 --> 07:25.840 and then there's the hardware development. 07:25.840 --> 07:26.840 There may be others. 07:26.840 --> 07:31.840 But so look, algorithm, data, hardware. 07:31.840 --> 07:34.840 You only have so much money, only have so much time. 07:34.840 --> 07:36.840 What do you think is the most important thing 07:36.840 --> 07:39.840 to allocate resources to? 07:39.840 --> 07:41.840 Do you see it as pretty evenly distributed 07:41.840 --> 07:43.840 between those three? 07:43.840 --> 07:46.840 We automatically get fast amounts of data 07:46.840 --> 07:48.840 because all of our cars have 07:50.840 --> 07:54.840 eight external facing cameras and radar 07:54.840 --> 07:59.840 and usually 12 ultrasonic sensors, GPS, obviously, 07:59.840 --> 08:03.840 and IMU. 08:03.840 --> 08:08.840 And so we basically have a fleet that has, 08:08.840 --> 08:11.840 we've got about 400,000 cars on the road 08:11.840 --> 08:13.840 that have that level of data. 08:13.840 --> 08:15.840 I think you keep quite close track of it, actually. 08:15.840 --> 08:16.840 Yes. 08:16.840 --> 08:19.840 So we're approaching half a million cars 08:19.840 --> 08:22.840 on the road that have the full sensor suite. 08:22.840 --> 08:26.840 So this is, I'm not sure how many other cars 08:26.840 --> 08:28.840 on the road have this sensor suite, 08:28.840 --> 08:31.840 but I'd be surprised if it's more than 5,000, 08:31.840 --> 08:35.840 which means that we have 99% of all the data. 08:35.840 --> 08:37.840 So there's this huge inflow of data. 08:37.840 --> 08:39.840 Absolutely, massive inflow of data. 08:39.840 --> 08:43.840 And then it's taken about three years, 08:43.840 --> 08:46.840 but now we've finally developed our full self driving computer, 08:46.840 --> 08:51.840 which can process 08:51.840 --> 08:54.840 an order of magnitude as much as the NVIDIA system 08:54.840 --> 08:56.840 that we currently have in the cars. 08:56.840 --> 08:58.840 And it's really just to use it, 08:58.840 --> 09:01.840 you unplug the NVIDIA computer and plug the Tesla computer in. 09:01.840 --> 09:03.840 And that's it. 09:03.840 --> 09:06.840 And it's, in fact, we're not even, 09:06.840 --> 09:09.840 we're still exploring the boundaries of its capabilities, 09:09.840 --> 09:11.840 but we're able to run the cameras at full frame rate, 09:11.840 --> 09:14.840 full resolution, not even crop of the images, 09:14.840 --> 09:19.840 and it's still got headroom, even on one of the systems. 09:19.840 --> 09:22.840 The full self driving computer is really two computers, 09:22.840 --> 09:25.840 two systems on a chip that are fully redundant. 09:25.840 --> 09:28.840 So you could put a bolt through basically any part of that system 09:28.840 --> 09:29.840 and it still works. 09:29.840 --> 09:32.840 The redundancy, are they perfect copies of each other? 09:32.840 --> 09:35.840 Or also it's purely for redundancy 09:35.840 --> 09:37.840 as opposed to an arguing machine kind of architecture 09:37.840 --> 09:39.840 where they're both making decisions. 09:39.840 --> 09:41.840 This is purely for redundancy. 09:41.840 --> 09:44.840 I think it's more like, if you have a twin engine aircraft, 09:44.840 --> 09:46.840 commercial aircraft, 09:46.840 --> 09:51.840 this system will operate best if both systems are operating, 09:51.840 --> 09:55.840 but it's capable of operating safely on one. 09:55.840 --> 09:59.840 So, but as it is right now, we can just run, 09:59.840 --> 10:03.840 we haven't even hit the edge of performance, 10:03.840 --> 10:08.840 so there's no need to actually distribute 10:08.840 --> 10:12.840 functionality across both SoCs. 10:12.840 --> 10:16.840 We can actually just run a full duplicate on each one. 10:16.840 --> 10:20.840 You haven't really explored or hit the limit of the system? 10:20.840 --> 10:21.840 Not yet, hit the limit now. 10:21.840 --> 10:26.840 So the magic of deep learning is that it gets better with data. 10:26.840 --> 10:28.840 You said there's a huge inflow of data, 10:28.840 --> 10:33.840 but the thing about driving the really valuable data 10:33.840 --> 10:35.840 to learn from is the edge cases. 10:35.840 --> 10:42.840 So how do you, I mean, I've heard you talk somewhere about 10:42.840 --> 10:46.840 autopilot disengagement as being an important moment of time to use. 10:46.840 --> 10:51.840 Is there other edge cases or perhaps can you speak to those edge cases, 10:51.840 --> 10:53.840 what aspects of them might be valuable, 10:53.840 --> 10:55.840 or if you have other ideas, 10:55.840 --> 10:59.840 how to discover more and more and more edge cases in driving? 10:59.840 --> 11:01.840 Well, there's a lot of things that I learned. 11:01.840 --> 11:05.840 There are certainly edge cases where I say somebody's on autopilot 11:05.840 --> 11:07.840 and they take over. 11:07.840 --> 11:12.840 And then, okay, that's a trigger that goes to a system that says, 11:12.840 --> 11:14.840 okay, do they take over for convenience 11:14.840 --> 11:18.840 or do they take over because the autopilot wasn't working properly? 11:18.840 --> 11:21.840 There's also, like let's say we're trying to figure out 11:21.840 --> 11:26.840 what is the optimal spline for traversing an intersection. 11:26.840 --> 11:30.840 Then the ones where there are no interventions 11:30.840 --> 11:32.840 and are the right ones. 11:32.840 --> 11:36.840 So you then say, okay, when it looks like this, do the following. 11:36.840 --> 11:40.840 And then you get the optimal spline for a complex, 11:40.840 --> 11:44.840 now getting a complex intersection. 11:44.840 --> 11:48.840 So that's for, there's kind of the common case. 11:48.840 --> 11:51.840 You're trying to capture a huge amount of samples 11:51.840 --> 11:54.840 of a particular intersection, how one thing went right. 11:54.840 --> 11:58.840 And then there's the edge case where, as you said, 11:58.840 --> 12:01.840 not for convenience, but something didn't go exactly right. 12:01.840 --> 12:04.840 Somebody took over, somebody asserted manual control from autopilot. 12:04.840 --> 12:08.840 And really, like the way to look at this is view all input is error. 12:08.840 --> 12:11.840 If the user had to do input, it does something. 12:11.840 --> 12:13.840 All input is error. 12:13.840 --> 12:15.840 That's a powerful line to think of it that way, 12:15.840 --> 12:17.840 because it may very well be error. 12:17.840 --> 12:19.840 But if you want to exit the highway, 12:19.840 --> 12:22.840 or if you want to, it's a navigation decision 12:22.840 --> 12:24.840 that all autopilot is not currently designed to do, 12:24.840 --> 12:26.840 then the driver takes over. 12:26.840 --> 12:28.840 How do you know the difference? 12:28.840 --> 12:30.840 Yeah, that's going to change with navigate and autopilot, 12:30.840 --> 12:33.840 which we've just released, and without stall confirm. 12:33.840 --> 12:36.840 So the navigation, like lane change based, 12:36.840 --> 12:39.840 like asserting control in order to do a lane change, 12:39.840 --> 12:43.840 or exit a freeway, or doing highway interchange, 12:43.840 --> 12:47.840 the vast majority of that will go away with the release 12:47.840 --> 12:49.840 that just went out. 12:49.840 --> 12:52.840 Yeah, I don't think people quite understand 12:52.840 --> 12:54.840 how big of a step that is. 12:54.840 --> 12:55.840 Yeah, they don't. 12:55.840 --> 12:57.840 If you drive the car, then you do. 12:57.840 --> 12:59.840 So you still have to keep your hands on the steering wheel 12:59.840 --> 13:02.840 currently when it does the automatic lane change? 13:02.840 --> 13:04.840 What are... 13:04.840 --> 13:07.840 So there's these big leaps through the development of autopilot 13:07.840 --> 13:09.840 through its history, 13:09.840 --> 13:12.840 and what stands out to you as the big leaps? 13:12.840 --> 13:14.840 I would say this one, 13:14.840 --> 13:19.840 navigate and autopilot without having to confirm, 13:19.840 --> 13:20.840 is a huge leap. 13:20.840 --> 13:21.840 It is a huge leap. 13:21.840 --> 13:24.840 It also automatically overtakes slow cars. 13:24.840 --> 13:30.840 So it's both navigation and seeking the fastest lane. 13:30.840 --> 13:36.840 So it'll overtake a slow cause and exit the freeway 13:36.840 --> 13:39.840 and take highway interchanges. 13:39.840 --> 13:46.840 And then we have traffic light recognition, 13:46.840 --> 13:49.840 which is introduced initially as a warning. 13:49.840 --> 13:51.840 I mean, on the development version that I'm driving, 13:51.840 --> 13:55.840 the car fully stops and goes at traffic lights. 13:55.840 --> 13:57.840 So those are the steps, right? 13:57.840 --> 13:59.840 You just mentioned something sort of 13:59.840 --> 14:02.840 including a step towards full autonomy. 14:02.840 --> 14:07.840 What would you say are the biggest technological roadblocks 14:07.840 --> 14:09.840 to full cell driving? 14:09.840 --> 14:10.840 Actually, I don't think... 14:10.840 --> 14:11.840 I think we just... 14:11.840 --> 14:13.840 the full cell driving computer that we just... 14:13.840 --> 14:14.840 that has a... 14:14.840 --> 14:16.840 what we call the FSD computer. 14:16.840 --> 14:20.840 That's now in production. 14:20.840 --> 14:25.840 So if you order any Model SRX or any Model 3 14:25.840 --> 14:28.840 that has the full cell driving package, 14:28.840 --> 14:31.840 you'll get the FSD computer. 14:31.840 --> 14:36.840 That's important to have enough base computation. 14:36.840 --> 14:40.840 Then refining the neural net and the control software. 14:40.840 --> 14:44.840 But all of that can just be provided as an over there update. 14:44.840 --> 14:46.840 The thing that's really profound, 14:46.840 --> 14:50.840 and where I'll be emphasizing at the... 14:50.840 --> 14:52.840 that investor day that we're having focused on autonomy, 14:52.840 --> 14:55.840 is that the cars currently being produced, 14:55.840 --> 14:57.840 or the hardware currently being produced, 14:57.840 --> 15:00.840 is capable of full cell driving. 15:00.840 --> 15:03.840 But capable is an interesting word because... 15:03.840 --> 15:05.840 Like the hardware is. 15:05.840 --> 15:08.840 And as we refine the software, 15:08.840 --> 15:11.840 the capabilities will increase dramatically 15:11.840 --> 15:13.840 and then the reliability will increase dramatically 15:13.840 --> 15:15.840 and then it will receive regulatory approval. 15:15.840 --> 15:18.840 So essentially buying a car today is an investment in the future. 15:18.840 --> 15:21.840 You're essentially buying... 15:21.840 --> 15:25.840 I think the most profound thing is that 15:25.840 --> 15:27.840 if you buy a Tesla today, 15:27.840 --> 15:29.840 I believe you are buying an appreciating asset, 15:29.840 --> 15:32.840 not a depreciating asset. 15:32.840 --> 15:34.840 So that's a really important statement there 15:34.840 --> 15:36.840 because if hardware is capable enough, 15:36.840 --> 15:39.840 that's the hard thing to upgrade usually. 15:39.840 --> 15:40.840 Exactly. 15:40.840 --> 15:43.840 So then the rest is a software problem. 15:43.840 --> 15:47.840 Yes. Software has no marginal cost, really. 15:47.840 --> 15:51.840 But what's your intuition on the software side? 15:51.840 --> 15:55.840 How hard are the remaining steps 15:55.840 --> 15:58.840 to get it to where... 15:58.840 --> 16:02.840 you know, the experience, 16:02.840 --> 16:05.840 not just the safety, but the full experience 16:05.840 --> 16:08.840 is something that people would enjoy. 16:08.840 --> 16:12.840 I think people would enjoy it very much on the highways. 16:12.840 --> 16:16.840 It's a total game changer for quality of life, 16:16.840 --> 16:20.840 for using Tesla autopilot on the highways. 16:20.840 --> 16:24.840 So it's really just extending that functionality to city streets, 16:24.840 --> 16:28.840 adding in the traffic light recognition, 16:28.840 --> 16:31.840 navigating complex intersections, 16:31.840 --> 16:36.840 and then being able to navigate complicated parking lots 16:36.840 --> 16:39.840 so the car can exit a parking space 16:39.840 --> 16:45.840 and come and find you even if it's in a complete maze of a parking lot. 16:45.840 --> 16:51.840 And then you can just drop you off and find a parking spot by itself. 16:51.840 --> 16:53.840 Yeah, in terms of enjoyability 16:53.840 --> 16:57.840 and something that people would actually find a lot of use from, 16:57.840 --> 17:00.840 the parking lot is a really... 17:00.840 --> 17:03.840 it's rich of annoyance when you have to do it manually, 17:03.840 --> 17:07.840 so there's a lot of benefit to be gained from automation there. 17:07.840 --> 17:11.840 So let me start injecting the human into this discussion a little bit. 17:11.840 --> 17:14.840 So let's talk about full autonomy. 17:14.840 --> 17:17.840 If you look at the current level four vehicles, 17:17.840 --> 17:19.840 being Tesla and road like Waymo and so on, 17:19.840 --> 17:22.840 they're only technically autonomous. 17:22.840 --> 17:25.840 They're really level two systems 17:25.840 --> 17:28.840 with just a different design philosophy 17:28.840 --> 17:31.840 because there's always a safety driver in almost all cases 17:31.840 --> 17:33.840 and they're monitoring the system. 17:33.840 --> 17:37.840 Maybe Tesla's full self driving 17:37.840 --> 17:41.840 is still for a time to come, 17:41.840 --> 17:44.840 requiring supervision of the human being. 17:44.840 --> 17:47.840 So its capabilities are powerful enough to drive, 17:47.840 --> 17:50.840 but nevertheless requires the human to still be supervising 17:50.840 --> 17:56.840 just like a safety driver is in a other fully autonomous vehicles. 17:56.840 --> 18:01.840 I think it will require detecting hands on wheel 18:01.840 --> 18:08.840 or at least six months or something like that from here. 18:08.840 --> 18:11.840 Really it's a question of like, 18:11.840 --> 18:15.840 from a regulatory standpoint, 18:15.840 --> 18:19.840 how much safer than a person does autopilot need to be 18:19.840 --> 18:24.840 for it to be okay to not monitor the car? 18:24.840 --> 18:27.840 And this is a debate that one can have. 18:27.840 --> 18:31.840 But you need a large amount of data 18:31.840 --> 18:34.840 so you can prove with high confidence, 18:34.840 --> 18:36.840 statistically speaking, 18:36.840 --> 18:39.840 that the car is dramatically safer than a person 18:39.840 --> 18:42.840 and that adding in the person monitoring 18:42.840 --> 18:45.840 does not materially affect the safety. 18:45.840 --> 18:49.840 So it might need to be like two or three hundred percent safer than a person. 18:49.840 --> 18:51.840 And how do you prove that? 18:51.840 --> 18:53.840 Incidence per mile. 18:53.840 --> 18:56.840 So crashes and fatalities. 18:56.840 --> 18:58.840 Yeah, fatalities would be a factor, 18:58.840 --> 19:00.840 but there are just not enough fatalities 19:00.840 --> 19:03.840 to be statistically significant at scale. 19:03.840 --> 19:06.840 But there are enough crashes, 19:06.840 --> 19:10.840 there are far more crashes than there are fatalities. 19:10.840 --> 19:15.840 So you can assess what is the probability of a crash, 19:15.840 --> 19:19.840 then there's another step which probability of injury 19:19.840 --> 19:21.840 and probability of permanent injury 19:21.840 --> 19:23.840 and probability of death. 19:23.840 --> 19:27.840 And all of those need to be much better than a person 19:27.840 --> 19:32.840 by at least perhaps two hundred percent. 19:32.840 --> 19:36.840 And you think there's the ability to have a healthy discourse 19:36.840 --> 19:39.840 with the regulatory bodies on this topic? 19:39.840 --> 19:43.840 I mean, there's no question that regulators pay 19:43.840 --> 19:48.840 disproportionate amount of attention to that which generates press. 19:48.840 --> 19:50.840 This is just an objective fact. 19:50.840 --> 19:52.840 And Tesla generates a lot of press. 19:52.840 --> 19:56.840 So that, you know, in the United States, 19:56.840 --> 20:00.840 there's I think almost 40,000 automotive deaths per year. 20:00.840 --> 20:03.840 But if there are four in Tesla, 20:03.840 --> 20:06.840 they'll probably receive a thousand times more press 20:06.840 --> 20:08.840 than anyone else. 20:08.840 --> 20:10.840 So the psychology of that is actually fascinating. 20:10.840 --> 20:12.840 I don't think we'll have enough time to talk about that, 20:12.840 --> 20:16.840 but I have to talk to you about the human side of things. 20:16.840 --> 20:20.840 So myself and our team at MIT recently released a paper 20:20.840 --> 20:24.840 on functional vigilance of drivers while using autopilot. 20:24.840 --> 20:27.840 This is work we've been doing since autopilot was first 20:27.840 --> 20:30.840 released publicly over three years ago, 20:30.840 --> 20:34.840 collecting video driver faces and driver body. 20:34.840 --> 20:38.840 So I saw that you tweeted a quote from the abstract 20:38.840 --> 20:43.840 so I can at least guess that you've glanced at it. 20:43.840 --> 20:46.840 Can I talk you through what we found? 20:46.840 --> 20:51.840 Okay, so it appears that in the data that we've collected 20:51.840 --> 20:54.840 that drivers are maintaining functional vigilance 20:54.840 --> 20:57.840 such that we're looking at 18,000 disengagement 20:57.840 --> 21:02.840 from autopilot, 18,900 and annotating were they able 21:02.840 --> 21:05.840 to take over control in a timely manner? 21:05.840 --> 21:07.840 So they were there present looking at the road 21:07.840 --> 21:09.840 to take over control. 21:09.840 --> 21:14.840 Okay, so this goes against what many would predict 21:14.840 --> 21:18.840 from the body of literature on vigilance with automation. 21:18.840 --> 21:21.840 Now the question is, do you think these results 21:21.840 --> 21:23.840 hold across the broader population? 21:23.840 --> 21:26.840 So ours is just a small subset. 21:26.840 --> 21:30.840 Do you think one of the criticism is that there's 21:30.840 --> 21:34.840 a small minority of drivers that may be highly responsible 21:34.840 --> 21:37.840 where their vigilance decrement would increase 21:37.840 --> 21:39.840 with autopilot use? 21:39.840 --> 21:41.840 I think this is all really going to be swept. 21:41.840 --> 21:46.840 I mean, the system's improving so much so fast 21:46.840 --> 21:50.840 that this is going to be a mood point very soon 21:50.840 --> 21:56.840 where vigilance is, if something's many times safer 21:56.840 --> 22:00.840 than a person, then adding a person does, 22:00.840 --> 22:04.840 the effect on safety is limited. 22:04.840 --> 22:09.840 And in fact, it could be negative. 22:09.840 --> 22:11.840 That's really interesting. 22:11.840 --> 22:16.840 So the fact that a human may, some percent of the population 22:16.840 --> 22:20.840 may exhibit a vigilance decrement will not affect 22:20.840 --> 22:22.840 overall statistics numbers of safety. 22:22.840 --> 22:27.840 No, in fact, I think it will become very, very quickly, 22:27.840 --> 22:29.840 maybe even towards the end of this year, 22:29.840 --> 22:32.840 but I'd say I'd be shocked if it's not next year, 22:32.840 --> 22:36.840 at the latest, that having a human intervene 22:36.840 --> 22:39.840 will increase safety. 22:39.840 --> 22:40.840 Decrease. 22:40.840 --> 22:42.840 I can imagine if you're an elevator. 22:42.840 --> 22:45.840 Now, it used to be that there were elevator operators 22:45.840 --> 22:47.840 and you couldn't go on an elevator by yourself 22:47.840 --> 22:51.840 and work the lever to move between floors. 22:51.840 --> 22:56.840 And now, nobody wants an elevator operator 22:56.840 --> 23:00.840 because the automated elevator that stops the floors 23:00.840 --> 23:03.840 is much safer than the elevator operator. 23:03.840 --> 23:05.840 And in fact, it would be quite dangerous 23:05.840 --> 23:07.840 if someone with a lever that can move 23:07.840 --> 23:09.840 the elevator between floors. 23:09.840 --> 23:12.840 So that's a really powerful statement 23:12.840 --> 23:14.840 and a really interesting one. 23:14.840 --> 23:16.840 But I also have to ask, from a user experience 23:16.840 --> 23:18.840 and from a safety perspective, 23:18.840 --> 23:20.840 one of the passions for me algorithmically 23:20.840 --> 23:25.840 is camera based detection of sensing the human, 23:25.840 --> 23:27.840 but detecting what the driver is looking at, 23:27.840 --> 23:29.840 cognitive load, body pose. 23:29.840 --> 23:31.840 On the computer vision side, that's a fascinating problem, 23:31.840 --> 23:34.840 but there's many in industry who believe 23:34.840 --> 23:37.840 you have to have camera based driver monitoring. 23:37.840 --> 23:39.840 Do you think this could be benefit gained 23:39.840 --> 23:41.840 from driver monitoring? 23:41.840 --> 23:45.840 If you have a system that's out or below 23:45.840 --> 23:49.840 human level reliability, then driver monitoring makes sense. 23:49.840 --> 23:51.840 But if your system is dramatically better, 23:51.840 --> 23:53.840 more reliable than a human, 23:53.840 --> 23:58.840 then driver monitoring is not help much. 23:58.840 --> 24:03.840 And like I said, you wouldn't want someone into... 24:03.840 --> 24:05.840 You wouldn't want someone in the elevator. 24:05.840 --> 24:07.840 If you're in an elevator, do you really want someone 24:07.840 --> 24:09.840 with a big lever, some random person operating 24:09.840 --> 24:11.840 in the elevator between floors? 24:11.840 --> 24:13.840 I wouldn't trust that. 24:13.840 --> 24:16.840 I would rather have the buttons. 24:16.840 --> 24:19.840 Okay, you're optimistic about the pace 24:19.840 --> 24:21.840 of improvement of the system. 24:21.840 --> 24:23.840 From what you've seen with the full self driving car, 24:23.840 --> 24:25.840 computer. 24:25.840 --> 24:27.840 The rate of improvement is exponential. 24:27.840 --> 24:30.840 So one of the other very interesting design choices 24:30.840 --> 24:34.840 early on that connects to this is the operational 24:34.840 --> 24:37.840 design domain of autopilot. 24:37.840 --> 24:41.840 So where autopilot is able to be turned on. 24:41.840 --> 24:46.840 So contrast another vehicle system that we're studying 24:46.840 --> 24:48.840 is the Cadillac SuperCrew system. 24:48.840 --> 24:51.840 That's in terms of ODD, very constrained to this particular 24:51.840 --> 24:54.840 kinds of highways, well mapped, tested, 24:54.840 --> 24:58.840 but it's much narrower than the ODD of Tesla vehicles. 24:58.840 --> 25:00.840 What's... 25:00.840 --> 25:02.840 It's like ADD. 25:02.840 --> 25:04.840 Yeah. 25:04.840 --> 25:07.840 That's good. That's a good line. 25:07.840 --> 25:10.840 What was the design decision 25:10.840 --> 25:13.840 in that different philosophy of thinking where... 25:13.840 --> 25:15.840 There's pros and cons. 25:15.840 --> 25:20.840 What we see with a wide ODD is Tesla drivers are able 25:20.840 --> 25:23.840 to explore more the limitations of the system, 25:23.840 --> 25:26.840 at least early on, and they understand together 25:26.840 --> 25:28.840 the instrument cluster display. 25:28.840 --> 25:30.840 They start to understand what are the capabilities. 25:30.840 --> 25:32.840 So that's a benefit. 25:32.840 --> 25:37.840 The con is you're letting drivers use it basically anywhere. 25:37.840 --> 25:41.840 Well, anyways, I could detect lanes with confidence. 25:41.840 --> 25:46.840 Was there a philosophy design decisions that were challenging 25:46.840 --> 25:48.840 that were being made there? 25:48.840 --> 25:53.840 Or from the very beginning, was that done on purpose 25:53.840 --> 25:55.840 with intent? 25:55.840 --> 25:58.840 Frankly, it's pretty crazy letting people drive 25:58.840 --> 26:02.840 a two ton death machine manually. 26:02.840 --> 26:04.840 That's crazy. 26:04.840 --> 26:06.840 In the future, people will be like, 26:06.840 --> 26:09.840 I can't believe anyone was just allowed to drive 26:09.840 --> 26:12.840 one of these two ton death machines 26:12.840 --> 26:14.840 and they just drive wherever they wanted, 26:14.840 --> 26:16.840 just like elevators. 26:16.840 --> 26:18.840 You just move the elevator with the lever wherever you want. 26:18.840 --> 26:21.840 It can stop at halfway between floors if you want. 26:21.840 --> 26:24.840 It's pretty crazy. 26:24.840 --> 26:29.840 So it's going to seem like a mad thing in the future 26:29.840 --> 26:32.840 that people were driving cars. 26:32.840 --> 26:35.840 So I have a bunch of questions about the human psychology, 26:35.840 --> 26:37.840 about behavior and so on. 26:37.840 --> 26:39.840 I don't know. 26:39.840 --> 26:45.840 Because you have faith in the AI system, 26:45.840 --> 26:50.840 not faith, but both on the hardware side 26:50.840 --> 26:52.840 and the deep learning approach of learning from data 26:52.840 --> 26:55.840 will make it just far safer than humans. 26:55.840 --> 26:57.840 Yeah, exactly. 26:57.840 --> 27:00.840 Recently, there are a few hackers who tricked autopilot 27:00.840 --> 27:03.840 to act in unexpected ways with adversarial examples. 27:03.840 --> 27:06.840 So we all know that neural network systems 27:06.840 --> 27:08.840 are very sensitive to minor disturbances 27:08.840 --> 27:10.840 to these adversarial examples on input. 27:10.840 --> 27:13.840 Do you think it's possible to defend against something like this 27:13.840 --> 27:15.840 for the industry? 27:15.840 --> 27:17.840 Sure. 27:17.840 --> 27:22.840 Can you elaborate on the confidence behind that answer? 27:22.840 --> 27:27.840 Well, a neural net is just like a basic bunch of matrix math. 27:27.840 --> 27:30.840 You have to be like a very sophisticated, 27:30.840 --> 27:32.840 somebody who really understands neural nets 27:32.840 --> 27:37.840 and basically reverse engineer how the matrix is being built 27:37.840 --> 27:42.840 and then create a little thing that just exactly causes 27:42.840 --> 27:44.840 the matrix math to be slightly off. 27:44.840 --> 27:48.840 But it's very easy to then block that by having 27:48.840 --> 27:51.840 basically anti negative recognition. 27:51.840 --> 27:55.840 It's like if the system sees something that looks like a matrix hack 27:55.840 --> 28:01.840 excluded, it's such an easy thing to do. 28:01.840 --> 28:05.840 So learn both on the valid data and the invalid data. 28:05.840 --> 28:07.840 So basically learn on the adversarial examples 28:07.840 --> 28:09.840 to be able to exclude them. 28:09.840 --> 28:12.840 Yeah, you basically want to both know what is a car 28:12.840 --> 28:15.840 and what is definitely not a car. 28:15.840 --> 28:18.840 You train for this is a car and this is definitely not a car. 28:18.840 --> 28:20.840 Those are two different things. 28:20.840 --> 28:23.840 People have no idea neural nets really. 28:23.840 --> 28:25.840 They probably think neural nets involves like, you know, 28:25.840 --> 28:28.840 fishing net or something. 28:28.840 --> 28:35.840 So as you know, taking a step beyond just Tesla and autopilot, 28:35.840 --> 28:39.840 current deep learning approaches still seem in some ways 28:39.840 --> 28:44.840 to be far from general intelligence systems. 28:44.840 --> 28:49.840 Do you think the current approaches will take us to general intelligence 28:49.840 --> 28:55.840 or do totally new ideas need to be invented? 28:55.840 --> 28:59.840 I think we're missing a few key ideas for general intelligence, 28:59.840 --> 29:04.840 general, artificial general intelligence. 29:04.840 --> 29:08.840 But it's going to be upon us very quickly 29:08.840 --> 29:11.840 and then we'll need to figure out what shall we do 29:11.840 --> 29:15.840 if we even have that choice. 29:15.840 --> 29:18.840 But it's amazing how people can't differentiate between, say, 29:18.840 --> 29:22.840 the narrow AI that, you know, allows a car to figure out 29:22.840 --> 29:25.840 what a lane line is and, you know, 29:25.840 --> 29:29.840 and navigate streets versus general intelligence. 29:29.840 --> 29:32.840 Like these are just very different things. 29:32.840 --> 29:35.840 Like your toaster and your computer are both machines, 29:35.840 --> 29:38.840 but one's much more sophisticated than another. 29:38.840 --> 29:43.840 You're confident with Tesla you can create the world's best toaster. 29:43.840 --> 29:45.840 The world's best toaster, yes. 29:45.840 --> 29:48.840 The world's best self driving. 29:48.840 --> 29:51.840 I'm, yes. 29:51.840 --> 29:54.840 To me, right now, this seems game set match. 29:54.840 --> 29:57.840 I don't, I mean, that's, I don't want to be complacent or overconfident, 29:57.840 --> 29:59.840 but that's what it appears. 29:59.840 --> 30:02.840 That is just literally what it, how it appears right now. 30:02.840 --> 30:06.840 It could be wrong, but it appears to be the case 30:06.840 --> 30:10.840 that Tesla is vastly ahead of everyone. 30:10.840 --> 30:13.840 Do you think we will ever create an AI system 30:13.840 --> 30:17.840 that we can love and loves us back in a deep meaningful way 30:17.840 --> 30:20.840 like in the movie, Her? 30:20.840 --> 30:23.840 I think AI will be capable of convincing you 30:23.840 --> 30:25.840 to fall in love with it very well. 30:25.840 --> 30:28.840 And that's different than us humans? 30:28.840 --> 30:31.840 You know, we start getting into a metaphysical question 30:31.840 --> 30:35.840 and do emotions and thoughts exist in a different realm than the physical. 30:35.840 --> 30:37.840 And maybe they do, maybe they don't. 30:37.840 --> 30:39.840 I don't know, but from a physics standpoint, 30:39.840 --> 30:43.840 I tend to think of things, you know, 30:43.840 --> 30:47.840 like physics was my main sort of training. 30:47.840 --> 30:50.840 And from a physics standpoint, 30:50.840 --> 30:52.840 essentially, if it loves you in a way 30:52.840 --> 30:57.840 that you can't tell whether it's real or not, it is real. 30:57.840 --> 30:59.840 That's a physics view of love. 30:59.840 --> 31:04.840 If you cannot prove that it does not, 31:04.840 --> 31:07.840 if there's no test that you can apply 31:07.840 --> 31:14.840 that would make it allow you to tell the difference, 31:14.840 --> 31:16.840 then there is no difference. 31:16.840 --> 31:20.840 And it's similar to seeing our world as simulation. 31:20.840 --> 31:22.840 There may not be a test to tell the difference 31:22.840 --> 31:24.840 between what the real world and the simulation. 31:24.840 --> 31:26.840 And therefore, from a physics perspective, 31:26.840 --> 31:28.840 it might as well be the same thing. 31:28.840 --> 31:29.840 Yes. 31:29.840 --> 31:32.840 There may be ways to test whether it's a simulation. 31:32.840 --> 31:35.840 There might be, I'm not saying there aren't, 31:35.840 --> 31:38.840 but you could certainly imagine that a simulation could correct 31:38.840 --> 31:40.840 that once an entity in the simulation 31:40.840 --> 31:42.840 found a way to detect the simulation, 31:42.840 --> 31:44.840 it could either restart, you know, 31:44.840 --> 31:47.840 pause the simulation, start a new simulation, 31:47.840 --> 31:52.840 or do one of many other things that then corrects for that error. 31:52.840 --> 31:58.840 So when maybe you or somebody else creates an AGI system 31:58.840 --> 32:02.840 and you get to ask her one question, 32:02.840 --> 32:16.840 what would that question be? 32:16.840 --> 32:21.840 What's outside the simulation? 32:21.840 --> 32:23.840 Milan, thank you so much for talking today. 32:23.840 --> 32:52.840 All right, thank you.