Sebastian Thrun: Autopilot Makes Me a Safer Driver | AI Podcast Clips
00:00:00.000 | You know, it's interesting you mentioned gutsy. Let me ask some maybe unanswerable question,
00:00:07.000 | maybe edgy questions, but in terms of how much risk is required, some guts, in terms
00:00:16.520 | of leadership style, it would be good to contrast approaches. And I don't think anyone knows
00:00:22.880 | what's right. But if we compare Tesla and Waymo, for example, Elon Musk and the Waymo
00:00:29.680 | team, there's slight differences in approach. So on the Elon side, there's more, I don't
00:00:36.680 | know what the right word to use, but aggression in terms of innovation. And on Waymo side,
00:00:45.520 | there's more sort of cautious, safety focused approach to the problem. What do you think
00:00:53.640 | it takes? Which leadership at which moment is right? Which approach is right?
00:01:00.640 | Look, I don't sit in either of those teams, so I'm unable to even verify like somebody
00:01:05.760 | says correct. In the end of the day, every innovator in that space will face a fundamental
00:01:11.760 | dilemma. And I would say you could put aerospace titans into the same bucket, which is you
00:01:18.360 | have to balance public safety with your drive to innovate. And this country in particular
00:01:25.360 | in the States has a hundred plus year history of doing this very successfully. Air travel
00:01:30.360 | is what a hundred times as safe per mile than ground travel, than cars. And there's a reason
00:01:37.160 | for it because people have found ways to be very methodological about ensuring public
00:01:43.600 | safety while still being able to make progress on important aspects, for example, like air
00:01:48.800 | and noise and fuel consumption. So I think that those practices are proven and they actually
00:01:55.800 | work. We live in a world safer than ever before. And yes, there will always be the provision
00:02:00.920 | that something goes wrong. There's always the possibility that someone makes a mistake
00:02:04.240 | or there's an unexpected failure. We can never guarantee to a hundred percent absolute safety
00:02:09.960 | other than just not doing it. But I think I'm very proud of the history of the United
00:02:15.800 | States. I mean, we've dealt with much more dangerous technology like nuclear energy and
00:02:20.720 | kept that safe too. We have nuclear weapons and we keep those safe. So we have methods
00:02:27.080 | and procedures that really balance these two things very, very successfully.
00:02:32.160 | You've mentioned a lot of great autonomous vehicle companies that are taking sort of
00:02:36.280 | the level four, level five, they jump in full autonomy with a safety driver and take that
00:02:41.400 | kind of approach and also through simulation and so on. There's also the approach that
00:02:46.600 | Tesla Autopilot is doing, which is kind of incrementally taking a level two vehicle and
00:02:52.920 | using machine learning and learning from the driving of human beings and trying to creep
00:02:59.280 | up, trying to incrementally improve the system until it's able to achieve level four autonomy.
00:03:04.700 | So perfect autonomy in certain kind of geographical regions. What are your thoughts on these contrasting
00:03:10.680 | approaches?
00:03:11.680 | Well, first of all, I'm a very proud Tesla owner and I literally use the Autopilot every
00:03:16.600 | day and it literally has kept me safe. It is a beautiful technology specifically for
00:03:23.560 | highway driving when I'm slightly tired because then it turns me into a much safer driver
00:03:31.240 | and I'm 100% confident that's the case. In terms of the right approach, I think the biggest
00:03:38.080 | change I've seen since I ran the Waymo team is this thing called deep learning. I think
00:03:43.760 | deep learning was not a hot topic when I started Waymo or Google self-driving cars. It was
00:03:49.040 | there. In fact, we started Google Brain at the same time in Google X, so I invested in
00:03:53.180 | deep learning, but people didn't talk about it. It wasn't a hot topic. And now it is.
00:03:57.720 | There's a shift of emphasis from a more geometric perspective where you use geometric sensors
00:04:03.280 | that give you a full 3D view and you do a geometric reasoning about, oh, this box over
00:04:07.080 | here might be a car, towards a more human-like, oh, let's just learn about it. This looks
00:04:13.620 | like the thing I've seen 10,000 times before, so maybe it's the same thing, machine learning
00:04:18.080 | perspective. And that has really put, I think, all these approaches on steroids. At Udacity,
00:04:25.680 | we teach a course in self-driving cars. In fact, I think we've graduated over 20,000
00:04:31.440 | or so people on self-driving car skills, so every self-driving car team in the world now
00:04:36.600 | uses our engineers. And in this course, the very first homework assignment is to do lane
00:04:42.360 | finding on images. And lane finding images for laymen, what this means is you put a camera
00:04:47.480 | into your car or you open your eyes and you wouldn't know where the lane is, right? So
00:04:51.680 | you can stay inside the lane with your car. Humans can do this super easily. You just
00:04:55.880 | look and you know where the lane is just intuitively. For machines, for a long time, it was super
00:05:00.880 | hard because people would write these kind of crazy rules. If there's like white lane
00:05:04.880 | markers and here's what white really means, this is not quite white enough, so it's, oh,
00:05:08.960 | it's not white. Or maybe the sun is shining, so when the sun shines and this is white and
00:05:12.880 | this is a straight line, or maybe it's quite a straight line because the road is curved.
00:05:16.720 | And do we know that there's really six feet between lane markings or not, or 12 feet,
00:05:20.320 | whatever it is. And now what the students are doing, they would take machine learning.
00:05:26.560 | So instead of like writing these crazy rules for the lane marker, they'll say, hey, let's
00:05:30.680 | take an hour of driving and label it and tell the vehicle this is actually the lane by hand.
00:05:34.880 | And then these are examples and have the machine find its own rules what lane markings are.
00:05:40.600 | And within 24 hours, now every student that's never done any programming before in this
00:05:44.200 | space can write a perfect lane finder as good as the best commercial lane finders. And that's
00:05:50.200 | completely amazing to me. We've seen progress using machine learning that completely dwarfs
00:05:56.600 | anything that I saw 10 years ago.
00:06:00.680 | - What are your thoughts on Elon Musk's statement, provocative statement, perhaps that lighter
00:06:05.080 | is a crutch. So this geometric way of thinking about the world may be holding us back if
00:06:12.360 | what we should instead be doing in this robotics, in this particular space of autonomous vehicles
00:06:17.320 | is using camera as a primary sensor and using computer vision and machine learning as the
00:06:23.200 | primary way to...
00:06:24.200 | - I think first of all, we all know that people can drive cars without lighters in their heads
00:06:31.640 | because we only have eyes and we mostly just use eyes for driving. Maybe we use some other
00:06:38.160 | perception about our bodies, accelerations, occasionally our ears, certainly not our noses.
00:06:45.600 | So the existence proof is there that eyes must be sufficient. In fact, we could even
00:06:51.680 | drive a car if someone put a camera out and then gave us the camera image with no latency,
00:06:58.320 | you would be able to drive a car that way the same way. So a camera is also sufficient.
00:07:03.160 | Secondly, I really love the idea that in the Western world, we have many, many different
00:07:08.040 | people trying different hypotheses. It's almost like an anthill, like if an anthill tries
00:07:12.800 | to forge for food, right? You can sit there as two ants and agree what the perfect path
00:07:17.120 | is and then every single ant marches for the most likely location of food is, or you can
00:07:21.440 | have them just spread out. And I promise you the spread out solution will be better because
00:07:26.360 | if the disgusting philosophical intellectual ants get it wrong and they're all moving the
00:07:31.120 | wrong direction, they're gonna waste the day. And then they're gonna discuss again for another
00:07:34.560 | week. Whereas if all these ants go in a random direction, someone's gonna succeed and they're
00:07:38.440 | gonna come back and claim victory and get the Nobel Prize or whatever the ant equivalent
00:07:43.000 | is. And then they all march in the same direction. And that's great about society. That's great
00:07:46.880 | about the Western society. We're not plan based, we're not central based, we don't have
00:07:50.640 | a Soviet Union style central government that tells us where to forge. We just forge, we
00:07:56.920 | start in C Corp. We get investor money, go out and try it out. And who knows who's gonna
00:08:01.840 | win.
00:08:02.320 | [BLANK_AUDIO]
00:08:09.320 | [BLANK_AUDIO]
00:08:19.320 | [BLANK_AUDIO]