episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
even how to get out of the house or how to make breakfast. You show this presentation of the WTF,
https://karpathy.ai/lexicap/0015-large.html#00:29:25.760
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
where's the fork of robot looking at a sink. And can you describe how we plan in this world
https://karpathy.ai/lexicap/0015-large.html#00:29:31.920
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
of this idea of hierarchical planning we've mentioned? So yeah, how can a robot hope to
https://karpathy.ai/lexicap/0015-large.html#00:29:40.640
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
plan about something with such a long horizon where the goal is quite far away?
https://karpathy.ai/lexicap/0015-large.html#00:29:47.360
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
People since probably reasoning began have thought about hierarchical reasoning,
https://karpathy.ai/lexicap/0015-large.html#00:29:54.480
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
the temporal hierarchy in particular. Well, there's spatial hierarchy, but let's talk
https://karpathy.ai/lexicap/0015-large.html#00:29:59.840
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
about temporal hierarchy. So you might say, oh, I have this long execution I have to do,
https://karpathy.ai/lexicap/0015-large.html#00:30:03.040
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
but I can divide it into some segments abstractly, right? So maybe you have to get out of the house,
https://karpathy.ai/lexicap/0015-large.html#00:30:08.960
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
I have to get in the car, I have to drive and so on. And so you can plan if you can build
https://karpathy.ai/lexicap/0015-large.html#00:30:15.360
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
abstractions. So this we started out by talking about abstractions. And we're back to that now,
https://karpathy.ai/lexicap/0015-large.html#00:30:22.720
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
if you can build abstractions in your state space, and abstractions sort of temporal abstractions,
https://karpathy.ai/lexicap/0015-large.html#00:30:26.960
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
then you can make plans at a high level. And you can say, I'm going to go to town and then I'll
https://karpathy.ai/lexicap/0015-large.html#00:30:34.560
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
have to get gas and then I can go here and I can do this other thing. And you can reason about the
https://karpathy.ai/lexicap/0015-large.html#00:30:39.760
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
dependencies and constraints among these actions, again, without thinking about the complete
https://karpathy.ai/lexicap/0015-large.html#00:30:43.520
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
details. What we do in our hierarchical planning work is then say, all right, I make a plan at a
https://karpathy.ai/lexicap/0015-large.html#00:30:50.080
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
high level of abstraction, I have to have some reason to think that it's feasible without working
https://karpathy.ai/lexicap/0015-large.html#00:30:56.800
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
it out in complete detail. And that's actually the interesting step. I always like to talk about
https://karpathy.ai/lexicap/0015-large.html#00:31:03.760
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
walking through an airport, like you can plan to go to New York and arrive at the airport, and then
https://karpathy.ai/lexicap/0015-large.html#00:31:08.720
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
find yourself an office building later. You can't even tell me in advance what your plan is for
https://karpathy.ai/lexicap/0015-large.html#00:31:15.120
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
walking through the airport, partly because you're too lazy to think about it, maybe, but partly
https://karpathy.ai/lexicap/0015-large.html#00:31:20.080
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
also because you just don't have the information, you don't know what gate you're landing in, or
https://karpathy.ai/lexicap/0015-large.html#00:31:24.960
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
what people are going to be in front of you or anything. So there's no point in planning in
https://karpathy.ai/lexicap/0015-large.html#00:31:28.800
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
detail, but you have to have, you have to make a leap of faith that you can figure it out once you
https://karpathy.ai/lexicap/0015-large.html#00:31:34.080
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
get there. And it's really interesting to me how you arrive at that. How do you, so you have learned
https://karpathy.ai/lexicap/0015-large.html#00:31:40.640
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
over your lifetime to be able to make some kinds of predictions about how hard it is to achieve some
https://karpathy.ai/lexicap/0015-large.html#00:31:50.000
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
kinds of sub goals. And that's critical. Like you would never plan to fly somewhere if you couldn't,
https://karpathy.ai/lexicap/0015-large.html#00:31:54.720
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
didn't have a model of how hard it was to do some of the intermediate steps. So one of the things
https://karpathy.ai/lexicap/0015-large.html#00:32:00.720
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
we're thinking about now is how do you do this kind of very aggressive generalization to situations
https://karpathy.ai/lexicap/0015-large.html#00:32:04.800
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
that you haven't been in and so on to predict how long will it take to walk through the Kuala Lumpur
https://karpathy.ai/lexicap/0015-large.html#00:32:12.480
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
airport. Like you could give me an estimate and it wouldn't be crazy. And you have to have an
https://karpathy.ai/lexicap/0015-large.html#00:32:16.880
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
estimate of that in order to make plans that involve walking through the Kuala Lumpur airport,
https://karpathy.ai/lexicap/0015-large.html#00:32:22.480
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
even if you don't need to know it in detail. So I'm really interested in these kinds of abstract
https://karpathy.ai/lexicap/0015-large.html#00:32:27.280
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
models and how do we acquire them. But once we have them, we can use them to do hierarchical
https://karpathy.ai/lexicap/0015-large.html#00:32:32.480
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
reasoning, which is, I think is very important. Yeah. There's this notion of goal regression and
https://karpathy.ai/lexicap/0015-large.html#00:32:37.280
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
preimage backchaining, this idea of starting at the goal and just forming these big clouds of
https://karpathy.ai/lexicap/0015-large.html#00:32:43.280
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
states. I mean, it's almost like saying to the airport, you know, once you show up to the airport
https://karpathy.ai/lexicap/0015-large.html#00:32:50.080
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
that you're like a few steps away from the goal. So like thinking of it this way, it's kind of
https://karpathy.ai/lexicap/0015-large.html#00:33:01.920
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
interesting. I don't know if you have sort of further comments on that of starting at the goal.
https://karpathy.ai/lexicap/0015-large.html#00:33:08.160
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
Yeah. I mean, it's interesting that Simon, Herb Simon back in the early days of AI talked a lot
https://karpathy.ai/lexicap/0015-large.html#00:33:14.480
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
about means ends reasoning and reasoning back from the goal. There's a kind of an intuition that
https://karpathy.ai/lexicap/0015-large.html#00:33:22.000
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
people have that the number of that state space is big. The number of actions you could take is
https://karpathy.ai/lexicap/0015-large.html#00:33:26.480
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
really big. So if you say, here I sit and I want to search forward from where I am, what are all
https://karpathy.ai/lexicap/0015-large.html#00:33:34.480
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
the things I could do? That's just overwhelming. If you say, if you can reason at this other level
https://karpathy.ai/lexicap/0015-large.html#00:33:39.120
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
and say, here's what I'm hoping to achieve, what could I do to make that true? That somehow the
https://karpathy.ai/lexicap/0015-large.html#00:33:44.720
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
branching is smaller. Now what's interesting is that like in the AI planning community,
https://karpathy.ai/lexicap/0015-large.html#00:33:49.520
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
that hasn't worked out in the class of problems that they look at and the methods that they tend
https://karpathy.ai/lexicap/0015-large.html#00:33:54.880
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
to use. It hasn't turned out that it's better to go backward. It's still kind of my intuition that
https://karpathy.ai/lexicap/0015-large.html#00:33:59.120
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
it is, but I can't prove that to you right now. Right. I share your intuition, at least for us
https://karpathy.ai/lexicap/0015-large.html#00:34:04.560
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
mere humans. Speaking of which, when you maybe now we take a little step into that philosophy circle.
https://karpathy.ai/lexicap/0015-large.html#00:34:10.720
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
How hard would it, when you think about human life, you give those examples often. How hard do
https://karpathy.ai/lexicap/0015-large.html#00:34:22.400
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
you think it is to formulate human life as a planning problem or aspects of human life? So
https://karpathy.ai/lexicap/0015-large.html#00:34:28.080
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
when you look at robots, you're often trying to think about object manipulation,
https://karpathy.ai/lexicap/0015-large.html#00:34:33.360
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
tasks about moving a thing. When you take a slight step outside the room, let the robot
https://karpathy.ai/lexicap/0015-large.html#00:34:38.640
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
leave and go get lunch, or maybe try to pursue more fuzzy goals. How hard do you think is that
https://karpathy.ai/lexicap/0015-large.html#00:34:46.240
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
problem? If you were to try to maybe put another way, try to formulate human life as a planning
https://karpathy.ai/lexicap/0015-large.html#00:34:54.480
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
problem. Well, that would be a mistake. I mean, it's not all a planning problem, right? I think
https://karpathy.ai/lexicap/0015-large.html#00:35:00.800
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
it's really, really important that we understand that you have to put together pieces and parts
https://karpathy.ai/lexicap/0015-large.html#00:35:05.760
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
that have different styles of reasoning and representation and learning. I think it seems
https://karpathy.ai/lexicap/0015-large.html#00:35:11.920
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
probably clear to anybody that it can't all be this or all be that. Brains aren't all like this
https://karpathy.ai/lexicap/0015-large.html#00:35:18.640
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
or all like that, right? They have different pieces and parts and substructure and so on.
https://karpathy.ai/lexicap/0015-large.html#00:35:25.680
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
So I don't think that there's any good reason to think that there's going to be like one true
https://karpathy.ai/lexicap/0015-large.html#00:35:30.160
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
algorithmic thing that's going to do the whole job. So it's a bunch of pieces together designed
https://karpathy.ai/lexicap/0015-large.html#00:35:34.320
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
to solve a bunch of specific problems. Or maybe styles of problems. I mean, there's probably some
https://karpathy.ai/lexicap/0015-large.html#00:35:40.720
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
reasoning that needs to go on in image space. I think, again, there's this model based versus
https://karpathy.ai/lexicap/0015-large.html#00:35:49.120
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
model free idea, right? So in reinforcement learning, people talk about, oh, should I learn,
https://karpathy.ai/lexicap/0015-large.html#00:35:57.360
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
I could learn a policy, just straight up a way of behaving. I could learn it's popular
https://karpathy.ai/lexicap/0015-large.html#00:36:02.480
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
on a value function. That's some kind of weird intermediate ground. Or I could learn a transition
https://karpathy.ai/lexicap/0015-large.html#00:36:08.400
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
model, which tells me something about the dynamics of the world. If I take it, imagine that I learned
https://karpathy.ai/lexicap/0015-large.html#00:36:14.960
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
a transition model and I couple it with a planner and I draw a box around that, I have a policy
https://karpathy.ai/lexicap/0015-large.html#00:36:20.240
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
again. It's just stored a different way, right? But it's just as much of a policy as the other
https://karpathy.ai/lexicap/0015-large.html#00:36:25.600
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
policy. It's just I've made, I think the way I see it is it's a time space trade off in computation,
https://karpathy.ai/lexicap/0015-large.html#00:36:32.800
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
right? A more overt policy representation. Maybe it takes more space, but maybe I can
https://karpathy.ai/lexicap/0015-large.html#00:36:40.160
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
compute quickly what action I should take. On the other hand, maybe a very compact model of
https://karpathy.ai/lexicap/0015-large.html#00:36:46.160
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
the world dynamics plus a planner lets me compute what action to take to just more slowly. There's
https://karpathy.ai/lexicap/0015-large.html#00:36:51.200
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
no, I don't, I mean, I don't think there's no argument to be had. It's just like a question of
https://karpathy.ai/lexicap/0015-large.html#00:36:57.120
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
what form of computation is best for us for the various sub problems. Right. So, and, and so like
https://karpathy.ai/lexicap/0015-large.html#00:37:02.320
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
learning to do algebra manipulations for some reason is, I mean, that's probably gonna want
https://karpathy.ai/lexicap/0015-large.html#00:37:10.320
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
naturally a sort of a different representation than writing a unicycle at the time constraints
https://karpathy.ai/lexicap/0015-large.html#00:37:16.000
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
on the unicycle are serious. The space is maybe smaller. I don't know, but so I could be the more
https://karpathy.ai/lexicap/0015-large.html#00:37:21.280
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
human size of falling in love, having a relationship that might be another, another style of how to
https://karpathy.ai/lexicap/0015-large.html#00:37:27.760
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
model that. Yeah. Let's first solve the algebra and the object manipulation. What do you think
https://karpathy.ai/lexicap/0015-large.html#00:37:36.000
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
is harder perception or planning perception? That's why understanding that's why. So what do you think
https://karpathy.ai/lexicap/0015-large.html#00:37:43.280
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
is so hard about perception by understanding the world around you? Well, I mean, I think the big
https://karpathy.ai/lexicap/0015-large.html#00:37:52.160
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
question is representational. Hugely the question is representation. So perception has made great
https://karpathy.ai/lexicap/0015-large.html#00:37:56.480
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
strides lately, right? And we can classify images and we can play certain kinds of games and predict
https://karpathy.ai/lexicap/0015-large.html#00:38:08.160
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
how to steer the car and all this sort of stuff. Um, I don't think we have a very good idea of
https://karpathy.ai/lexicap/0015-large.html#00:38:15.360
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
what perception should deliver, right? So if you, if you believe in modularity, okay, there's,
https://karpathy.ai/lexicap/0015-large.html#00:38:24.800
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
there's a very strong view which says we shouldn't build in any modularity. We should make a giant
https://karpathy.ai/lexicap/0015-large.html#00:38:29.760
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
gigantic neural network, train it end to end to do the thing. And that's the best way forward.
https://karpathy.ai/lexicap/0015-large.html#00:38:38.000
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
And it's hard to argue with that except on a sample complexity basis, right? So you might say,
https://karpathy.ai/lexicap/0015-large.html#00:38:44.320
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
Oh, well if I want to do end to end reinforcement learning on this giant, giant neural network,
https://karpathy.ai/lexicap/0015-large.html#00:38:51.440
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
it's going to take a lot of data and a lot of like broken robots and stuff. So then the only answer
https://karpathy.ai/lexicap/0015-large.html#00:38:55.280
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
is to say, okay, we have to build something in, build in some structure or some bias. We know
https://karpathy.ai/lexicap/0015-large.html#00:39:05.520
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
from theory of machine learning, the only way to cut down the sample complexity is to kind of cut
https://karpathy.ai/lexicap/0015-large.html#00:39:11.760
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
down, somehow cut down the hypothesis space. You can do that by building in bias. There's all kinds
https://karpathy.ai/lexicap/0015-large.html#00:39:15.760
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
of reasons to think that nature built bias into humans. Um, convolution is a bias, right? It's a
https://karpathy.ai/lexicap/0015-large.html#00:39:22.480
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
very strong bias and it's a very critical bias. So my own view is that we should look for more
https://karpathy.ai/lexicap/0015-large.html#00:39:30.640
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
things that are like convolution, but the address other aspects of reasoning, right? So convolution
https://karpathy.ai/lexicap/0015-large.html#00:39:37.520
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
helps us a lot with a certain kind of spatial reasoning. That's quite close to the imaging.
https://karpathy.ai/lexicap/0015-large.html#00:39:42.880
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
I think there's other ideas like that. Maybe some amount of forward search, maybe some notions of
https://karpathy.ai/lexicap/0015-large.html#00:39:48.320