episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
and the other is the meta reasoning capability, which allows it to
https://karpathy.ai/lexicap/0009-large.html#00:19:53.360
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
explore some paths in the tree very deeply and to abandon other paths very quickly.
https://karpathy.ai/lexicap/0009-large.html#00:19:57.520
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
So this word meta reasoning, while technically correct, inspires perhaps the wrong degree of
https://karpathy.ai/lexicap/0009-large.html#00:20:04.400
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
power that AlphaGo has, for example, the word reasoning is a powerful word. So let me ask you,
https://karpathy.ai/lexicap/0009-large.html#00:20:14.160
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
sort of, you were part of the symbolic AI world for a while, like AI was, there's a lot of
https://karpathy.ai/lexicap/0009-large.html#00:20:19.280
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
excellent, interesting ideas there that unfortunately met a winter. And so do you think it reemerges?
https://karpathy.ai/lexicap/0009-large.html#00:20:27.760
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
So I would say, yeah, it's not quite as simple as that. So the AI winter
https://karpathy.ai/lexicap/0009-large.html#00:20:38.320
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
for the first winter that was actually named as such was the one in the late 80s.
https://karpathy.ai/lexicap/0009-large.html#00:20:44.320
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
And that came about because in the mid 80s, there was a really a concerted attempt to push AI
https://karpathy.ai/lexicap/0009-large.html#00:20:51.440
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
out into the real world using what was called expert system technology. And for the most part,
https://karpathy.ai/lexicap/0009-large.html#00:21:01.520
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
that technology was just not ready for primetime. They were trying, in many cases, to do a form of
https://karpathy.ai/lexicap/0009-large.html#00:21:09.120
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
uncertain reasoning, judgment, combinations of evidence, diagnosis, those kinds of things,
https://karpathy.ai/lexicap/0009-large.html#00:21:17.040
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
which was simply invalid. And when you try to apply invalid reasoning methods to real problems,
https://karpathy.ai/lexicap/0009-large.html#00:21:24.480
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
you can fudge it for small versions of the problem. But when it starts to get larger,
https://karpathy.ai/lexicap/0009-large.html#00:21:31.600
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
the thing just falls apart. So many companies found that the stuff just didn't work, and they
https://karpathy.ai/lexicap/0009-large.html#00:21:36.720
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
were spending tons of money on consultants to try to make it work. And there were other
https://karpathy.ai/lexicap/0009-large.html#00:21:44.240
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
practical reasons, like they were asking the companies to buy incredibly expensive
https://karpathy.ai/lexicap/0009-large.html#00:21:50.400
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
Lisp machine workstations, which were literally between $50,000 and $100,000 in 1980s money,
https://karpathy.ai/lexicap/0009-large.html#00:21:56.160
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
which would be like between $150,000 and $300,000 per workstation in current prices.
https://karpathy.ai/lexicap/0009-large.html#00:22:06.080
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
And then the bottom line, they weren't seeing a profit from it.
https://karpathy.ai/lexicap/0009-large.html#00:22:13.920
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
Yeah, in many cases. I think there were some successes, there's no doubt about that. But
https://karpathy.ai/lexicap/0009-large.html#00:22:17.280
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
people, I would say, overinvested. Every major company was starting an AI department, just like
https://karpathy.ai/lexicap/0009-large.html#00:22:23.920
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
now. And I worry a bit that we might see similar disappointments, not because the current technology
https://karpathy.ai/lexicap/0009-large.html#00:22:30.800
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
is invalid, but it's limited in its scope. And it's almost the duel of the scope problems that
https://karpathy.ai/lexicap/0009-large.html#00:22:40.000
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
expert systems had. So what have you learned from that hype cycle? And what can we do to
https://karpathy.ai/lexicap/0009-large.html#00:22:51.280
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
prevent another winter, for example? Yeah, so when I'm giving talks these days,
https://karpathy.ai/lexicap/0009-large.html#00:22:56.720
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
that's one of the warnings that I give. So this is a two part warning slide. One is that rather
https://karpathy.ai/lexicap/0009-large.html#00:23:02.800
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
than data being the new oil, data is the new snake oil. That's a good line. And then the other
https://karpathy.ai/lexicap/0009-large.html#00:23:11.120
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
is that we might see a kind of very visible failure in some of the major application areas. And I think
https://karpathy.ai/lexicap/0009-large.html#00:23:18.880
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
self driving cars would be the flagship. And I think when you look at the history,
https://karpathy.ai/lexicap/0009-large.html#00:23:30.800
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
so the first self driving car was on the freeway, driving itself, changing lanes, overtaking in 1987.
https://karpathy.ai/lexicap/0009-large.html#00:23:40.000
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
And so it's more than 30 years. And that kind of looks like where we are today, right? You know,
https://karpathy.ai/lexicap/0009-large.html#00:23:52.000
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
prototypes on the freeway, changing lanes and overtaking. Now, I think that's one of the things
https://karpathy.ai/lexicap/0009-large.html#00:23:59.040
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
that's been made, particularly on the perception side. So we worked a lot on autonomous vehicles
https://karpathy.ai/lexicap/0009-large.html#00:24:05.920
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
in the early mid 90s at Berkeley. And we had our own big demonstrations. We put congressmen into
https://karpathy.ai/lexicap/0009-large.html#00:24:12.400
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
self driving cars and had them zooming along the freeway. And the problem was clearly perception.
https://karpathy.ai/lexicap/0009-large.html#00:24:21.680
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
At the time, the problem was perception. Yeah. So in simulation, with perfect perception,
https://karpathy.ai/lexicap/0009-large.html#00:24:30.640
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
you could actually show that you can drive safely for a long time, even if the other cars are
https://karpathy.ai/lexicap/0009-large.html#00:24:36.160
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
misbehaving and so on. But simultaneously, we worked on machine vision for detecting cars and
https://karpathy.ai/lexicap/0009-large.html#00:24:40.640
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
tracking pedestrians and so on. And we couldn't get the cars to do that. And so we had to do
https://karpathy.ai/lexicap/0009-large.html#00:24:48.800
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
that for pedestrians and so on. And we couldn't get the reliability of detection and tracking
https://karpathy.ai/lexicap/0009-large.html#00:24:56.880
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
up to a high enough level, particularly in bad weather conditions, nighttime,
https://karpathy.ai/lexicap/0009-large.html#00:25:03.120
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
rainfall. Good enough for demos, but perhaps not good enough to cover the general operation.
https://karpathy.ai/lexicap/0009-large.html#00:25:10.800
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
Yeah. So the thing about driving is, you know, suppose you're a taxi driver, you know,
https://karpathy.ai/lexicap/0009-large.html#00:25:15.920
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
and you drive every day, eight hours a day for 10 years, right? That's 100 million seconds of
https://karpathy.ai/lexicap/0009-large.html#00:25:19.680
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
driving, you know, and any one of those seconds, you can make a fatal mistake. So you're talking
https://karpathy.ai/lexicap/0009-large.html#00:25:25.200
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
about eight nines of reliability, right? Now, if your vision system only detects 98.3% of the
https://karpathy.ai/lexicap/0009-large.html#00:25:30.560
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
vehicles, right, then that's sort of, you know, one in a bit nines of reliability. So you have
https://karpathy.ai/lexicap/0009-large.html#00:25:40.080
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
another seven orders of magnitude to go. And this is what people don't understand. They think,
https://karpathy.ai/lexicap/0009-large.html#00:25:47.200
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
oh, because I had a successful demo, I'm pretty much done. But you're not even within seven orders
https://karpathy.ai/lexicap/0009-large.html#00:25:54.560
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
of magnitude of being done. And that's the difficulty. And it's not the, can I follow a
https://karpathy.ai/lexicap/0009-large.html#00:26:01.440
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
white line? That's not the problem, right? We follow a white line all the way across the country.
https://karpathy.ai/lexicap/0009-large.html#00:26:09.760
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
But it's the weird stuff that happens. It's all the edge cases, yeah.
https://karpathy.ai/lexicap/0009-large.html#00:26:16.640
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
The edge case, other drivers doing weird things. You know, so if you talk to Google, right, so
https://karpathy.ai/lexicap/0009-large.html#00:26:22.160
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
they had actually a very classical architecture where, you know, you had machine vision which
https://karpathy.ai/lexicap/0009-large.html#00:26:29.200
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
would detect all the other cars and pedestrians and the white lines and the road signs. And then
https://karpathy.ai/lexicap/0009-large.html#00:26:35.600
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
basically that was fed into a logical database. And then you had a classical 1970s rule based
https://karpathy.ai/lexicap/0009-large.html#00:26:41.440
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
expert system telling you, okay, if you're in the middle lane and there's a bicyclist in the right
https://karpathy.ai/lexicap/0009-large.html#00:26:48.880
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
lane who is signaling this, then you do that, right? And what they found was that every day
https://karpathy.ai/lexicap/0009-large.html#00:26:55.360
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
they'd go out and there'd be another situation that the rules didn't cover. You know, so they'd
https://karpathy.ai/lexicap/0009-large.html#00:27:02.640
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
come to a traffic circle and there's a little girl riding her bicycle the wrong way around
https://karpathy.ai/lexicap/0009-large.html#00:27:06.560
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
the traffic circle. Okay, what do you do? We don't have a rule. Oh my God. Okay, stop.
https://karpathy.ai/lexicap/0009-large.html#00:27:10.880
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
And then, you know, they come back and add more rules and they just found that this was not really
https://karpathy.ai/lexicap/0009-large.html#00:27:14.720
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
converging. And if you think about it, right, how do you deal with an unexpected situation,
https://karpathy.ai/lexicap/0009-large.html#00:27:20.560
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
meaning one that you've never previously encountered and the sort of reasoning required
https://karpathy.ai/lexicap/0009-large.html#00:27:28.240
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
to figure out the solution for that situation has never been done. It doesn't match any previous
https://karpathy.ai/lexicap/0009-large.html#00:27:35.600
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
situation in terms of the kind of reasoning you have to do. Well, you know, in chess programs,
https://karpathy.ai/lexicap/0009-large.html#00:27:41.200
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
this happens all the time, right? You're constantly coming up with situations you haven't
https://karpathy.ai/lexicap/0009-large.html#00:27:46.560
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
seen before and you have to reason about them and you have to think about, okay, here are the
https://karpathy.ai/lexicap/0009-large.html#00:27:51.280
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
possible things I could do. Here are the outcomes. Here's how desirable the outcomes are and then
https://karpathy.ai/lexicap/0009-large.html#00:27:56.480
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
pick the right one. You know, in the 90s, we were saying, okay, this is how you're going to have to
https://karpathy.ai/lexicap/0009-large.html#00:28:01.680
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
do automated vehicles. They're going to have to have a look ahead capability, but the look ahead
https://karpathy.ai/lexicap/0009-large.html#00:28:05.440
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
for driving is more difficult than it is for chess because there's humans and they're less
https://karpathy.ai/lexicap/0009-large.html#00:28:10.880
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
predictable than chess pieces. Well, then you have an opponent in chess who's also somewhat
https://karpathy.ai/lexicap/0009-large.html#00:28:18.000
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
unpredictable. But for example, in chess, you always know the opponent's intention. They're
https://karpathy.ai/lexicap/0009-large.html#00:28:23.840
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
trying to beat you, right? Whereas in driving, you don't know is this guy trying to turn left
https://karpathy.ai/lexicap/0009-large.html#00:28:29.920
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
or has he just forgotten to turn off his turn signal or is he drunk or is he changing the
https://karpathy.ai/lexicap/0009-large.html#00:28:36.000
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
channel on his radio or whatever it might be. You've got to try and figure out the mental state,
https://karpathy.ai/lexicap/0009-large.html#00:28:42.000
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
the intent of the other drivers to forecast the possible evolutions of their trajectories.
https://karpathy.ai/lexicap/0009-large.html#00:28:47.520
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
And then you've got to figure out, okay, which is the trajectory for me that's going to be safest.
https://karpathy.ai/lexicap/0009-large.html#00:28:54.880
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
And those all interact with each other because the other drivers are going to react to your
https://karpathy.ai/lexicap/0009-large.html#00:29:00.400
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
trajectory and so on. So, you know, they've got the classic merging onto the freeway problem where
https://karpathy.ai/lexicap/0009-large.html#00:29:04.720
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
you're kind of racing a vehicle that's already on the freeway and you're going to pull ahead of
https://karpathy.ai/lexicap/0009-large.html#00:29:10.640
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
them or you're going to let them go first and pull in behind and you get this sort of uncertainty
https://karpathy.ai/lexicap/0009-large.html#00:29:15.520
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
about who's going first. So all those kinds of things mean that you need a decision making
https://karpathy.ai/lexicap/0009-large.html#00:29:19.920
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
architecture that's very different from either a rule based system or it seems to me kind of an
https://karpathy.ai/lexicap/0009-large.html#00:29:29.440
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
end to end neural network system. So just as AlphaGo is pretty good when it doesn't do any
https://karpathy.ai/lexicap/0009-large.html#00:29:37.200
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
look ahead, but it's way, way, way, way better when it does, I think the same is going to be
https://karpathy.ai/lexicap/0009-large.html#00:29:43.840
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
true for driving. You can have a driving system that's pretty good when it doesn't do any look
https://karpathy.ai/lexicap/0009-large.html#00:29:49.920
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
ahead, but that's not good enough. And we've already seen multiple deaths caused by poorly
https://karpathy.ai/lexicap/0009-large.html#00:29:55.120
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
designed machine learning algorithms that don't really understand what they're doing.
https://karpathy.ai/lexicap/0009-large.html#00:30:03.440
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
Yeah. On several levels, I think on the perception side, there's mistakes being made by those
https://karpathy.ai/lexicap/0009-large.html#00:30:09.360
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
algorithms where the perception is very shallow. On the planning side, the look ahead, like you
https://karpathy.ai/lexicap/0009-large.html#00:30:16.480
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
said, and the thing that we come up against that's really interesting when you try to deploy systems
https://karpathy.ai/lexicap/0009-large.html#00:30:21.200
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
in the real world is you can't think of an artificial intelligence system as a thing that
https://karpathy.ai/lexicap/0009-large.html#00:30:31.200
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
responds to the world always. You have to realize that it's an agent that others will respond to as
https://karpathy.ai/lexicap/0009-large.html#00:30:36.160
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
well. So in order to drive successfully, you can't just try to do obstacle avoidance.
https://karpathy.ai/lexicap/0009-large.html#00:30:41.440
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
Right. You can't pretend that you're invisible, right? You're the invisible car.
https://karpathy.ai/lexicap/0009-large.html#00:30:47.680
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
Right. It doesn't work that way.
https://karpathy.ai/lexicap/0009-large.html#00:30:51.920
Stuart Russell: Long-Term Future of Artificial Intelligence | Lex Fridman Podcast #9
I mean, but you have to assert yet others have to be scared of you. Just we're all,
https://karpathy.ai/lexicap/0009-large.html#00:30:53.440