episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
picture of the road, let's find the yellow lane markers and,
https://karpathy.ai/lexicap/0014-large.html#00:08:00.480
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
you know, steer the wheel. And, you know, later I'd come to realize this had been
https://karpathy.ai/lexicap/0014-large.html#00:08:03.440
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
done, you know, since, since the 80s or the 70s or even
https://karpathy.ai/lexicap/0014-large.html#00:08:06.880
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
earlier, but I still wanted to do it and sort of
https://karpathy.ai/lexicap/0014-large.html#00:08:10.640
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
immediately after that trip switched from sort of BattleBots, which is more
https://karpathy.ai/lexicap/0014-large.html#00:08:13.520
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
radio controlled machines, to thinking about building,
https://karpathy.ai/lexicap/0014-large.html#00:08:16.560
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
you know, autonomous vehicles of some scale. Start off with really small
https://karpathy.ai/lexicap/0014-large.html#00:08:21.440
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
electric ones and then, you know, progress to what we're
https://karpathy.ai/lexicap/0014-large.html#00:08:24.320
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
doing now. So what was your view of artificial intelligence at that point?
https://karpathy.ai/lexicap/0014-large.html#00:08:27.680
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
What did you think? So this is before, there's been waves in artificial
https://karpathy.ai/lexicap/0014-large.html#00:08:30.640
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
intelligence, right? The current wave with deep learning
https://karpathy.ai/lexicap/0014-large.html#00:08:35.520
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
makes people believe that you can solve in a really rich deep way the computer
https://karpathy.ai/lexicap/0014-large.html#00:08:39.440
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
vision perception problem, but like in
https://karpathy.ai/lexicap/0014-large.html#00:08:44.000
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
before the deep learning craze, you know, how do you think about,
https://karpathy.ai/lexicap/0014-large.html#00:08:48.400
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
how would you even go about building a thing that perceives itself in the
https://karpathy.ai/lexicap/0014-large.html#00:08:52.720
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
world, localizes itself in the world, moves around the world?
https://karpathy.ai/lexicap/0014-large.html#00:08:56.560
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
Like when you were younger, I mean, what was your thinking about it? Well,
https://karpathy.ai/lexicap/0014-large.html#00:08:59.040
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
prior to deep neural networks or convolutional neural
https://karpathy.ai/lexicap/0014-large.html#00:09:02.320
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
analysis, these modern techniques we have, or at least
https://karpathy.ai/lexicap/0014-large.html#00:09:05.040
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
ones that are in use today, it was all a heuristic space and so like old school
https://karpathy.ai/lexicap/0014-large.html#00:09:06.960
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
image processing and I think extracting, you know, yellow lane markers out of an
https://karpathy.ai/lexicap/0014-large.html#00:09:11.680
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
image of a road is one of the problems that lends itself
https://karpathy.ai/lexicap/0014-large.html#00:09:16.800
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
reasonably well to those heuristic based methods, you know, like
https://karpathy.ai/lexicap/0014-large.html#00:09:21.200
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
just do a threshold on the color yellow and then try to
https://karpathy.ai/lexicap/0014-large.html#00:09:24.240
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
fit some lines to that using a Huff transform or something and then
https://karpathy.ai/lexicap/0014-large.html#00:09:27.440
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
go from there. Traffic light detection and stop sign detection, red, yellow, green.
https://karpathy.ai/lexicap/0014-large.html#00:09:31.040
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
And I think you could, I mean, if you wanted to do a full,
https://karpathy.ai/lexicap/0014-large.html#00:09:35.840
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
I was just trying to make something that would stay in between the lanes on a
https://karpathy.ai/lexicap/0014-large.html#00:09:39.760
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
highway, but if you wanted to do the full,
https://karpathy.ai/lexicap/0014-large.html#00:09:43.120
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
the full, you know, set of capabilities needed for a driverless car,
https://karpathy.ai/lexicap/0014-large.html#00:09:46.880
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
I think you could, and we'd done this at cruise, you know, in the very first days,
https://karpathy.ai/lexicap/0014-large.html#00:09:50.400
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
you can start off with a really simple, you know, human written heuristic just to
https://karpathy.ai/lexicap/0014-large.html#00:09:54.320
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
get the scaffolding in place for your system. Traffic light detection,
https://karpathy.ai/lexicap/0014-large.html#00:09:58.160
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
probably a really simple, you know, color thresholding on day one just to
https://karpathy.ai/lexicap/0014-large.html#00:10:01.680
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
get the system up and running before you migrate to, you know, a deep
https://karpathy.ai/lexicap/0014-large.html#00:10:04.960
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
learning based technique or something else.
https://karpathy.ai/lexicap/0014-large.html#00:10:08.800
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
And, you know, back in when I was doing this, my first one, it was on a Pentium
https://karpathy.ai/lexicap/0014-large.html#00:10:10.960
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
203, 233 megahertz computer in it and I think I wrote the first version in
https://karpathy.ai/lexicap/0014-large.html#00:10:14.000
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
basic, which is like an interpreted language. It's
https://karpathy.ai/lexicap/0014-large.html#00:10:19.520
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
extremely slow because that's the thing I knew at the time.
https://karpathy.ai/lexicap/0014-large.html#00:10:21.680
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
And so there was no, no chance at all of using,
https://karpathy.ai/lexicap/0014-large.html#00:10:24.720
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
there was no, no computational power to do any sort of reasonable
https://karpathy.ai/lexicap/0014-large.html#00:10:27.760
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
deep nets like you have today. So I don't know what kids these days are doing. Are
https://karpathy.ai/lexicap/0014-large.html#00:10:32.480
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
kids these days, you know, at age 13 using neural networks in
https://karpathy.ai/lexicap/0014-large.html#00:10:35.360
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
their garage? I mean, that would be awesome.
https://karpathy.ai/lexicap/0014-large.html#00:10:38.720
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
I get emails all the time from, you know, like 11, 12 year olds
https://karpathy.ai/lexicap/0014-large.html#00:10:40.640
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
saying I'm having, you know, I'm trying to follow this TensorFlow tutorial
https://karpathy.ai/lexicap/0014-large.html#00:10:44.640
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
and I'm having this problem. And the general approach
https://karpathy.ai/lexicap/0014-large.html#00:10:48.640
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
in the deep learning community is of extreme optimism of, as opposed to,
https://karpathy.ai/lexicap/0014-large.html#00:10:53.200
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
you mentioned like heuristics, you can, you can, you can
https://karpathy.ai/lexicap/0014-large.html#00:11:00.000
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
separate the autonomous driving problem into modules and try to solve it sort of
https://karpathy.ai/lexicap/0014-large.html#00:11:03.040
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
rigorously, or you can just do it end to end.
https://karpathy.ai/lexicap/0014-large.html#00:11:06.560
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
And most people just kind of love the idea that, you know, us humans do it end
https://karpathy.ai/lexicap/0014-large.html#00:11:08.960
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
to end. We just perceive and act. We should be able to use that, do the
https://karpathy.ai/lexicap/0014-large.html#00:11:12.960
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
same kind of thing when you're on nets. And that,
https://karpathy.ai/lexicap/0014-large.html#00:11:17.280
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
that kind of thinking, you don't want to criticize that kind of thinking because
https://karpathy.ai/lexicap/0014-large.html#00:11:19.360
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
eventually they will be right. Yeah. And so it's exciting and especially
https://karpathy.ai/lexicap/0014-large.html#00:11:23.120
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
when they're younger to explore that as a really exciting approach. But yeah,
https://karpathy.ai/lexicap/0014-large.html#00:11:26.800
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
it's, it's changed the, the language,
https://karpathy.ai/lexicap/0014-large.html#00:11:30.960
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
the kind of stuff you're tinkering with. It's kind of exciting to see when these
https://karpathy.ai/lexicap/0014-large.html#00:11:35.360
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
teenagers grow up. Yeah. I can only imagine if you,
https://karpathy.ai/lexicap/0014-large.html#00:11:38.960
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
if your starting point is, you know, Python and TensorFlow at age 13
https://karpathy.ai/lexicap/0014-large.html#00:11:42.240
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
where you end up, you know, after 10 or 15 years of that,
https://karpathy.ai/lexicap/0014-large.html#00:11:46.640
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
that's, that's pretty cool. Because of GitHub, because the state tools for
https://karpathy.ai/lexicap/0014-large.html#00:11:49.760
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
solving most of the major problems in artificial intelligence
https://karpathy.ai/lexicap/0014-large.html#00:11:53.840
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
are within a few lines of code for most kids.
https://karpathy.ai/lexicap/0014-large.html#00:11:56.880
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
And that's incredible to think about also on the entrepreneurial side.
https://karpathy.ai/lexicap/0014-large.html#00:12:00.160
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
And, and on that point, was there any thought about entrepreneurship before
https://karpathy.ai/lexicap/0014-large.html#00:12:04.160
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
you came to college? Is sort of doing, you're building this
https://karpathy.ai/lexicap/0014-large.html#00:12:10.560
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
into a thing that impacts the world on a large scale? Yeah. I've always
https://karpathy.ai/lexicap/0014-large.html#00:12:14.560
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
wanted to start a company. I think that's, you know, just a cool concept of
https://karpathy.ai/lexicap/0014-large.html#00:12:18.480
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
creating something and exchanging it for value or creating value, I guess.
https://karpathy.ai/lexicap/0014-large.html#00:12:22.720
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
So in high school, I was, I was trying to build like, you know, servo motor
https://karpathy.ai/lexicap/0014-large.html#00:12:28.240
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
drivers, little circuit boards and sell them online
https://karpathy.ai/lexicap/0014-large.html#00:12:32.240
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
or other, other things like that. And certainly knew at some point I wanted to
https://karpathy.ai/lexicap/0014-large.html#00:12:35.280
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
do a startup, but it wasn't really, I'd say until college, until I felt
https://karpathy.ai/lexicap/0014-large.html#00:12:39.520
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
like I had the,
https://karpathy.ai/lexicap/0014-large.html#00:12:42.720
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
I guess the right combination of the environment, the smart people around you
https://karpathy.ai/lexicap/0014-large.html#00:12:45.600
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
and some free time and a lot of free time at MIT.
https://karpathy.ai/lexicap/0014-large.html#00:12:48.960
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
So you came to MIT as an undergrad 2004. That's right. And that's when the first
https://karpathy.ai/lexicap/0014-large.html#00:12:52.240
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
DARPA Grand Challenge was happening. Yeah. The, the timing of that is
https://karpathy.ai/lexicap/0014-large.html#00:12:58.000
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
beautifully poetic. So how did you get yourself involved in that one?
https://karpathy.ai/lexicap/0014-large.html#00:13:02.080
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
Originally there wasn't a official entry. Yeah, faculty sponsored thing. And so a
https://karpathy.ai/lexicap/0014-large.html#00:13:05.600
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
bunch of undergrads, myself included, started meeting and got together and
https://karpathy.ai/lexicap/0014-large.html#00:13:09.840
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
tried to haggle together some sponsorships. We got a vehicle donated,
https://karpathy.ai/lexicap/0014-large.html#00:13:14.720
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
a bunch of sensors and tried to put something together. And so we had,
https://karpathy.ai/lexicap/0014-large.html#00:13:18.720
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
our team was probably mostly freshmen and sophomores, you know, which, which was
https://karpathy.ai/lexicap/0014-large.html#00:13:22.480
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
not really a fair, fair fight against maybe the,
https://karpathy.ai/lexicap/0014-large.html#00:13:26.320
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
you know, postdoc and faculty led teams from other schools. But
https://karpathy.ai/lexicap/0014-large.html#00:13:30.080
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
we, we got something up and running. We had our vehicle drive by wire and
https://karpathy.ai/lexicap/0014-large.html#00:13:33.280
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
you know, very, very basic control and things. But
https://karpathy.ai/lexicap/0014-large.html#00:13:36.960
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
on the day of the qualifying, sort of pre qualifying round, the one and
https://karpathy.ai/lexicap/0014-large.html#00:13:41.840
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
only steering motor that we had purchased,
https://karpathy.ai/lexicap/0014-large.html#00:13:47.520
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
the thing that we had retrofitted to turn the steering wheel on the truck
https://karpathy.ai/lexicap/0014-large.html#00:13:50.720
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
died. And so our vehicle was just dead in the water, couldn't steer.
https://karpathy.ai/lexicap/0014-large.html#00:13:55.200
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
So we didn't make it very far. On the hardware side. So was there a software
https://karpathy.ai/lexicap/0014-large.html#00:13:58.480
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
component? Was there, like, how did your view of autonomous
https://karpathy.ai/lexicap/0014-large.html#00:14:02.400
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
vehicles in terms of artificial intelligence
https://karpathy.ai/lexicap/0014-large.html#00:14:05.760
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
evolve in this moment? I mean, you know, like you said from the 80s has been
https://karpathy.ai/lexicap/0014-large.html#00:14:09.520
Kyle Vogt: Cruise Automation | Lex Fridman Podcast #14
autonomous vehicles, but really that was the birth of the modern wave.
https://karpathy.ai/lexicap/0014-large.html#00:14:13.120