episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
something more in the publication direction, I would do this other thing, which I thought about
https://karpathy.ai/lexicap/0015-large.html#00:50:10.160
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
doing the first time, which is to get together some set of people whose opinions I value and
https://karpathy.ai/lexicap/0015-large.html#00:50:14.560
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
who are pretty articulate. And I guess we would be public, although we could be private. I'm not sure.
https://karpathy.ai/lexicap/0015-large.html#00:50:19.840
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
And we would review papers. We wouldn't publish them and you wouldn't submit them. We would just
https://karpathy.ai/lexicap/0015-large.html#00:50:25.200
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
find papers and we would write reviews and we would make those reviews public. And maybe if you,
https://karpathy.ai/lexicap/0015-large.html#00:50:29.360
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
you know, so we're Leslie's friends who review papers and maybe eventually if, if we, our opinion
https://karpathy.ai/lexicap/0015-large.html#00:50:37.040
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
was sufficiently valued, like the opinion of JMLR is valued, then you'd say on your CV that Leslie's
https://karpathy.ai/lexicap/0015-large.html#00:50:42.880
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
friends gave my paper a five star rating. And that would be just as good as saying, I got it,
https://karpathy.ai/lexicap/0015-large.html#00:50:48.560
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
so, you know, accepted into this journal. So I think, I think we should have good public commentary
https://karpathy.ai/lexicap/0015-large.html#00:50:53.200
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
and organize it in some way, but I don't really know how to do it. It's interesting times.
https://karpathy.ai/lexicap/0015-large.html#00:51:01.840
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
The way you describe it actually is really interesting. I mean, we do it for movies,
https://karpathy.ai/lexicap/0015-large.html#00:51:06.320
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
imdb.com. There's experts, critics come in, they write reviews, but there's also
https://karpathy.ai/lexicap/0015-large.html#00:51:10.000
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
regular non critics, humans write reviews and they're separated.
https://karpathy.ai/lexicap/0015-large.html#00:51:16.000
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
I like open review. The iClear process I think is interesting.
https://karpathy.ai/lexicap/0015-large.html#00:51:19.840
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
It's a step in the right direction, but it's still not as compelling as reviewing movies or
https://karpathy.ai/lexicap/0015-large.html#00:51:29.280
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
video games. I mean, it sometimes almost, it might be silly, at least from my perspective to say,
https://karpathy.ai/lexicap/0015-large.html#00:51:35.760
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
but it boils down to the user interface, how fun and easy it is to actually perform the reviews,
https://karpathy.ai/lexicap/0015-large.html#00:51:41.600
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
how efficient, how much you as a reviewer get street cred for being a good reviewer.
https://karpathy.ai/lexicap/0015-large.html#00:51:46.720
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
Those elements, those human elements come into play.
https://karpathy.ai/lexicap/0015-large.html#00:51:54.400
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
No, it's a big investment to do a good review of a paper and the flood of papers is out of control.
https://karpathy.ai/lexicap/0015-large.html#00:51:57.200
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
Right. So, you know, there aren't 3000 new, I don't know how many new movies are there in a year.
https://karpathy.ai/lexicap/0015-large.html#00:52:04.000
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
I don't know, but that's probably going to be less than how many machine learning papers are
https://karpathy.ai/lexicap/0015-large.html#00:52:08.480
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
in a year now. And I'm worried, you know, I, right. So I'm like an old person. So of course,
https://karpathy.ai/lexicap/0015-large.html#00:52:12.160
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
I'm going to say, things are moving too fast. I'm a stick in the mud. So I can say that,
https://karpathy.ai/lexicap/0015-large.html#00:52:21.760
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
but my particular flavor of that is I think the horizon for researchers has gotten very short,
https://karpathy.ai/lexicap/0015-large.html#00:52:28.720
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
that students want to publish a lot of papers and there's a huge, there's value. It's exciting. And
https://karpathy.ai/lexicap/0015-large.html#00:52:35.760
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
there's value in that and you get patted on the head for it and so on. But, and some of that is
https://karpathy.ai/lexicap/0015-large.html#00:52:41.760
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
fine, but I'm worried that we're driving out people who would spend two years thinking about
https://karpathy.ai/lexicap/0015-large.html#00:52:49.200
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
something. Back in my day, when we worked on our thesis, we did not publish papers. You did your
https://karpathy.ai/lexicap/0015-large.html#00:52:59.200
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
thesis for years. You picked a hard problem and then you worked and chewed on it and did stuff
https://karpathy.ai/lexicap/0015-large.html#00:53:06.080
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
and wasted time and for a long time. And when it was roughly, when it was done, you would write
https://karpathy.ai/lexicap/0015-large.html#00:53:11.360
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
papers. And so I don't know how to, and I don't think that everybody has to work in that mode,
https://karpathy.ai/lexicap/0015-large.html#00:53:16.320
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
but I think there's some problems that are hard enough that it's important to have a long
https://karpathy.ai/lexicap/0015-large.html#00:53:23.520
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
research horizon. And I'm worried that we don't incentivize that at all at this point.
https://karpathy.ai/lexicap/0015-large.html#00:53:27.760
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
In this current structure. Yeah. So what do you see as, what are your hopes and fears about the
https://karpathy.ai/lexicap/0015-large.html#00:53:33.040
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
future of AI and continuing on this theme? So AI has gone through a few winters, ups and downs. Do
https://karpathy.ai/lexicap/0015-large.html#00:53:41.840
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
you see another winter of AI coming? Are you more hopeful about making robots work, as you said?
https://karpathy.ai/lexicap/0015-large.html#00:53:50.080
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
I think the cycles are inevitable, but I think each time we get higher, right? I mean, so, you
https://karpathy.ai/lexicap/0015-large.html#00:53:58.720
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
know, it's like climbing some kind of landscape with a noisy optimizer. So it's clear that the,
https://karpathy.ai/lexicap/0015-large.html#00:54:05.680
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
you know, the deep learning stuff has made deep and important improvements. And so the high water
https://karpathy.ai/lexicap/0015-large.html#00:54:15.280
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
mark is now higher. There's no question. But of course, I think people are overselling and
https://karpathy.ai/lexicap/0015-large.html#00:54:22.560
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
eventually investors, I guess, and other people will look around and say, well, you're not quite
https://karpathy.ai/lexicap/0015-large.html#00:54:29.360
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
delivering on this grand claim and that wild hypothesis. It's probably, it's going to crash
https://karpathy.ai/lexicap/0015-large.html#00:54:37.040
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
some amount and then it's okay. I mean, but I don't, I can't imagine that there's like
https://karpathy.ai/lexicap/0015-large.html#00:54:43.120
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
some awesome monotonic improvement from here to human level AI. So in, you know, I have to ask
https://karpathy.ai/lexicap/0015-large.html#00:54:49.760
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
this question, I probably anticipate answers, the answers, but do you have a worry short term or
https://karpathy.ai/lexicap/0015-large.html#00:54:58.320
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
long term about the existential threats of AI and maybe short term, less existential, but more
https://karpathy.ai/lexicap/0015-large.html#00:55:05.520
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
robots taking away jobs?
https://karpathy.ai/lexicap/0015-large.html#00:55:15.280
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
Well, actually, let me talk a little bit about utility. Actually, I had an interesting
https://karpathy.ai/lexicap/0015-large.html#00:55:17.120
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
conversation with some military ethicists who wanted to talk to me about autonomous weapons.
https://karpathy.ai/lexicap/0015-large.html#00:55:25.520
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
And they're, they were interesting, smart, well educated guys who didn't know too much about AI
https://karpathy.ai/lexicap/0015-large.html#00:55:30.880
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
or machine learning. And the first question they asked me was, has your robot ever done
https://karpathy.ai/lexicap/0015-large.html#00:55:37.440
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
something you didn't expect? And I like burst out laughing because anybody who's ever done
https://karpathy.ai/lexicap/0015-large.html#00:55:41.440
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
something on the robot right knows that they don't do it. And what I realized was that their
https://karpathy.ai/lexicap/0015-large.html#00:55:46.480
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
model of how we program a robot was completely wrong. Their model of how we can program a robot
https://karpathy.ai/lexicap/0015-large.html#00:55:51.680
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
was like Lego mind storms, like, Oh, go forward a meter, turn left, take a picture, do this, do
https://karpathy.ai/lexicap/0015-large.html#00:55:56.960
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
that. And so if you have that model of programming, then it's true. It's kind of weird that your robot
https://karpathy.ai/lexicap/0015-large.html#00:56:03.120
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
would do something that you didn't anticipate. But the fact is, and actually, so now this is my
https://karpathy.ai/lexicap/0015-large.html#00:56:08.800
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
new educational mission. If I have to talk to non experts, I try to teach them the idea that
https://karpathy.ai/lexicap/0015-large.html#00:56:13.760
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
we don't operate, we operate at least one or maybe many levels of abstraction about that. And we say,
https://karpathy.ai/lexicap/0015-large.html#00:56:20.640
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
Oh, here's a hypothesis class, maybe it's a space of plans, or maybe it's a space of
https://karpathy.ai/lexicap/0015-large.html#00:56:26.400
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
classifiers or whatever. But there's some set of answers and an objective function. And then we
https://karpathy.ai/lexicap/0015-large.html#00:56:31.200
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
work on some optimization method that tries to optimize a solution solution in that class.
https://karpathy.ai/lexicap/0015-large.html#00:56:36.400
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
And we don't know what solution is going to come out. Right. So I think it's important to
https://karpathy.ai/lexicap/0015-large.html#00:56:43.200
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
communicate that. So I mean, of course, probably people who listen to this, they, they know that
https://karpathy.ai/lexicap/0015-large.html#00:56:47.600
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
lesson. But I think it's really critical to communicate that lesson. And then lots of people
https://karpathy.ai/lexicap/0015-large.html#00:56:52.160
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
are now talking about, you know, the value alignment problem. So you want to be sure as
https://karpathy.ai/lexicap/0015-large.html#00:56:56.320
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
robots or software systems get more competent, that their objectives are aligned with your
https://karpathy.ai/lexicap/0015-large.html#00:57:01.840
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
objectives, or that our objectives are compatible in some way, or we have a good way of mediating
https://karpathy.ai/lexicap/0015-large.html#00:57:07.840
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
when they have different objectives. And so I think it is important to start thinking in terms
https://karpathy.ai/lexicap/0015-large.html#00:57:13.760
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
like, you don't have to be freaked out by the robot apocalypse, to accept that it's important
https://karpathy.ai/lexicap/0015-large.html#00:57:19.120
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
to think about objective functions of value alignment. Yes. And that you have to really
https://karpathy.ai/lexicap/0015-large.html#00:57:25.360
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
everyone who's done optimization knows that you have to be careful what you wish for that,
https://karpathy.ai/lexicap/0015-large.html#00:57:30.080
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
you know, sometimes you get the optimal solution, and you realize, man, that was that objective was
https://karpathy.ai/lexicap/0015-large.html#00:57:34.160
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
wrong. So pragmatically, in the shortest term, it seems to me that that that those are really
https://karpathy.ai/lexicap/0015-large.html#00:57:38.800
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
interesting and critical questions. And the idea that we're going to go from being people who
https://karpathy.ai/lexicap/0015-large.html#00:57:47.040
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
engineer algorithms to being people who engineer objective functions. I think that's, that's
https://karpathy.ai/lexicap/0015-large.html#00:57:50.960
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
definitely going to happen. And that's going to change our thinking and methodology. And so we're
https://karpathy.ai/lexicap/0015-large.html#00:57:56.480
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
gonna you started at Stanford philosophy, that's where she could be. And I will go back to
https://karpathy.ai/lexicap/0015-large.html#00:58:00.560
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
philosophy maybe. Well, I mean, they're mixed together, because because, as we also know,
https://karpathy.ai/lexicap/0015-large.html#00:58:05.920
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
as machine learning people, right? When you design, in fact, this is the lecture I gave in
https://karpathy.ai/lexicap/0015-large.html#00:58:11.920
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
class today, when you design an objective function, you have to wear both hats, there's
https://karpathy.ai/lexicap/0015-large.html#00:58:16.160
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
the hat that says, what do I want? And there's the hat that says, but I know what my optimizer
https://karpathy.ai/lexicap/0015-large.html#00:58:21.600
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
can do to some degree. And I have to take that into account. So it's it's always a trade off,
https://karpathy.ai/lexicap/0015-large.html#00:58:26.080
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
and we have to kind of be mindful of that. The part about taking people's jobs, I understand
https://karpathy.ai/lexicap/0015-large.html#00:58:31.840
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
that that's important. I don't understand sociology or economics or people very well. So I
https://karpathy.ai/lexicap/0015-large.html#00:58:38.560
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
don't know how to think about that. So that's Yeah, so there might be a sociological aspect
https://karpathy.ai/lexicap/0015-large.html#00:58:45.440
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
there, the economic aspect that's very difficult to think about. Okay. I mean, I think other people
https://karpathy.ai/lexicap/0015-large.html#00:58:50.160
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
should be thinking about it. But I'm just that's not my strength. So what do you think is the most
https://karpathy.ai/lexicap/0015-large.html#00:58:54.640
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
exciting area of research in the short term, for the community and for your for yourself?
https://karpathy.ai/lexicap/0015-large.html#00:58:58.480
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
Well, so I mean, there's the story I've been telling about how to engineer intelligent robots.
https://karpathy.ai/lexicap/0015-large.html#00:59:03.600
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
So that's what we want to do. We all kind of want to do well, I mean, some set of us want to do this.
https://karpathy.ai/lexicap/0015-large.html#00:59:10.160
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
And the question is, what's the most effective strategy? And we've tried it. And there's a bunch
https://karpathy.ai/lexicap/0015-large.html#00:59:16.160
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
of different things you could do at the extremes, right? One super extreme is, what's the most
https://karpathy.ai/lexicap/0015-large.html#00:59:20.560
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
effective strategy? And there's a bunch of different things you could do at the extremes,
https://karpathy.ai/lexicap/0015-large.html#00:59:25.360
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
right? One super extreme is, we do introspection, and we write a program. Okay, that has not worked
https://karpathy.ai/lexicap/0015-large.html#00:59:29.600
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
out very well. Another extreme is we take a giant bunch of neural goo, and we try and train it up to
https://karpathy.ai/lexicap/0015-large.html#00:59:35.920
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
do something. I don't think that's going to work either. So the question is, what's the middle
https://karpathy.ai/lexicap/0015-large.html#00:59:41.280
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
ground? And, and again, this isn't a theological question or anything like that. It's just like,
https://karpathy.ai/lexicap/0015-large.html#00:59:46.960
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
what's the middle ground? And I think it's clear, it's a combination of learning, to me, it's clear,
https://karpathy.ai/lexicap/0015-large.html#00:59:54.240