episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
And when I model some aspect of it in a certain way, that gives me some set of algorithms I can
https://karpathy.ai/lexicap/0015-large.html#00:19:12.880
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
use. You can model the world in all kinds of ways. Some have, some are, some are, some are
https://karpathy.ai/lexicap/0015-large.html#00:19:17.600
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
more accepting of uncertainty, more easily modeling uncertainty of the world. Some really force the
https://karpathy.ai/lexicap/0015-large.html#00:19:26.160
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
world to be deterministic. And so certainly MDPs model the uncertainty of the world. Yes. Model
https://karpathy.ai/lexicap/0015-large.html#00:19:33.360
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
some uncertainty. They model not present state uncertainty, but they model uncertainty in the
https://karpathy.ai/lexicap/0015-large.html#00:19:42.000
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
way the future will unfold. Right. So what are Markov decision processes? So Markov decision
https://karpathy.ai/lexicap/0015-large.html#00:19:47.520
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
process is a model. It's a kind of a model that you could make that says, I know completely the
https://karpathy.ai/lexicap/0015-large.html#00:19:54.560
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
current state of my system. And what it means to be a state is that I, that all the, I have all
https://karpathy.ai/lexicap/0015-large.html#00:19:59.760
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
the information right now that will let me make predictions about the future as well as I can.
https://karpathy.ai/lexicap/0015-large.html#00:20:05.760
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
So that remembering anything about my history wouldn't make my predictions any better.
https://karpathy.ai/lexicap/0015-large.html#00:20:11.120
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
And, but then it also says that then I can take some actions that might change the state of the
https://karpathy.ai/lexicap/0015-large.html#00:20:17.760
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
world. And that I don't have a deterministic model of those changes. I have a probabilistic
https://karpathy.ai/lexicap/0015-large.html#00:20:23.280
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
model of how the world might change. It's a, it's a useful model for some kinds of systems.
https://karpathy.ai/lexicap/0015-large.html#00:20:28.400
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
I think it's a, I mean, it's certainly not a good model for most problems, I think, because for most
https://karpathy.ai/lexicap/0015-large.html#00:20:34.160
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
problems you don't actually know the state. For most problems you, it's partially observed. So
https://karpathy.ai/lexicap/0015-large.html#00:20:42.480
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
that's now a different problem class. So, okay. That's where the POMDPs, the part that we observe
https://karpathy.ai/lexicap/0015-large.html#00:20:48.720
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
with the Markov decision processes step in. So how do they address the fact that you can't
https://karpathy.ai/lexicap/0015-large.html#00:20:55.600
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
observe most incomplete information about most of the world around you? Right. So now the idea is
https://karpathy.ai/lexicap/0015-large.html#00:21:01.360
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
we still kind of postulate that there exists a state. We think that there is some information
https://karpathy.ai/lexicap/0015-large.html#00:21:07.760
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
about the world out there such that if we knew that we could make good predictions, but we don't
https://karpathy.ai/lexicap/0015-large.html#00:21:12.880
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
know the state. And so then we have to think about how, but we do get observations. Maybe I get
https://karpathy.ai/lexicap/0015-large.html#00:21:18.000
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
images or I hear things or I feel things, and those might be local or noisy. And so therefore
https://karpathy.ai/lexicap/0015-large.html#00:21:23.680
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
they don't tell me everything about what's going on. And then I have to reason about given the
https://karpathy.ai/lexicap/0015-large.html#00:21:29.680
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
history of actions I've taken and observations I've gotten, what do I think is going on in the
https://karpathy.ai/lexicap/0015-large.html#00:21:34.720
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
world? And then given my own kind of uncertainty about what's going on in the world, I can decide
https://karpathy.ai/lexicap/0015-large.html#00:21:39.920
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
what actions to take. And so how difficult is this problem of planning under uncertainty in your
https://karpathy.ai/lexicap/0015-large.html#00:21:43.760
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
view and your long experience of modeling the world, trying to deal with this uncertainty in
https://karpathy.ai/lexicap/0015-large.html#00:21:50.080
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
especially in real world systems? Optimal planning for even discrete POMDPs can be undecidable
https://karpathy.ai/lexicap/0015-large.html#00:21:57.840
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
depending on how you set it up. And so lots of people say, I don't use POMDPs because they are
https://karpathy.ai/lexicap/0015-large.html#00:22:05.040
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
intractable. And I think that that's kind of a very funny thing to say because the problem you
https://karpathy.ai/lexicap/0015-large.html#00:22:12.960
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
have to solve is the problem you have to solve. So if the problem you have to solve is intractable,
https://karpathy.ai/lexicap/0015-large.html#00:22:19.600
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
that's what makes us AI people, right? So we solve, we understand that the problem we're
https://karpathy.ai/lexicap/0015-large.html#00:22:24.320
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
solving is wildly intractable that we can't, we will never be able to solve it optimally,
https://karpathy.ai/lexicap/0015-large.html#00:22:28.720
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
at least I don't. Yeah, right. So later we can come back to an idea about bounded optimality
https://karpathy.ai/lexicap/0015-large.html#00:22:34.320
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
and something. But anyway, we can't come up with optimal solutions to these problems.
https://karpathy.ai/lexicap/0015-large.html#00:22:41.360
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
So we have to make approximations, approximations in modeling, approximations in the solution
https://karpathy.ai/lexicap/0015-large.html#00:22:45.520
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
algorithms and so on. And so I don't have a problem with saying, yeah, my problem actually,
https://karpathy.ai/lexicap/0015-large.html#00:22:50.640
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
it is POMDP in continuous space with continuous observations. And it's so computationally complex,
https://karpathy.ai/lexicap/0015-large.html#00:22:56.960
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
I can't even think about it's, you know, big O whatever. But that doesn't prevent me from,
https://karpathy.ai/lexicap/0015-large.html#00:23:02.480
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
it helps me, gives me some clarity to think about it that way and to then take steps to
https://karpathy.ai/lexicap/0015-large.html#00:23:09.600
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
make approximation after approximation to get down to something that's like computable
https://karpathy.ai/lexicap/0015-large.html#00:23:16.160
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
in some reasonable time. When you think about optimality, the community broadly has shifted on
https://karpathy.ai/lexicap/0015-large.html#00:23:20.880
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
that, I think a little bit in how much they value the idea of optimality, of chasing an optimal
https://karpathy.ai/lexicap/0015-large.html#00:23:26.720
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
solution. How has your views of chasing an optimal solution changed over the years when
https://karpathy.ai/lexicap/0015-large.html#00:23:34.880
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
you work with robots? That's interesting. I think we have a little bit of a methodological crisis
https://karpathy.ai/lexicap/0015-large.html#00:23:40.960
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
actually from the theoretical side. I mean, I do think that theory is important and that right now
https://karpathy.ai/lexicap/0015-large.html#00:23:48.880
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
we're not doing much of it. So there's lots of empirical hacking around and training this and
https://karpathy.ai/lexicap/0015-large.html#00:23:53.520
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
doing that and reporting numbers, but is it good? Is it bad? We don't know. It's very hard to say
https://karpathy.ai/lexicap/0015-large.html#00:24:00.080
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
things. And if you look at like computer science theory, so people talked for a while, everyone was
https://karpathy.ai/lexicap/0015-large.html#00:24:05.120
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
about solving problems optimally or completely. And then there were interesting relaxations. So
https://karpathy.ai/lexicap/0015-large.html#00:24:16.400
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
people look at, oh, are there regret bounds or can I do some kind of approximation? Can I prove
https://karpathy.ai/lexicap/0015-large.html#00:24:22.320
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
something that I can approximately solve this problem or that I get closer to the solution as
https://karpathy.ai/lexicap/0015-large.html#00:24:30.640
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
I spend more time and so on? What's interesting I think is that we don't have good approximate
https://karpathy.ai/lexicap/0015-large.html#00:24:34.880
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
solution concepts for very difficult problems. I like to say that I'm interested in doing a very
https://karpathy.ai/lexicap/0015-large.html#00:24:42.560
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
bad job of very big problems. Right. So very bad job, very big problems. I like to do that,
https://karpathy.ai/lexicap/0015-large.html#00:24:51.520
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
but I wish I could say something. I wish I had a, I don't know, some kind of a formal solution
https://karpathy.ai/lexicap/0015-large.html#00:25:00.400
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
concept that I could use to say, oh, this algorithm actually, it gives me something.
https://karpathy.ai/lexicap/0015-large.html#00:25:09.120
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
Like I know what I'm going to get. I can do something other than just run it and get out.
https://karpathy.ai/lexicap/0015-large.html#00:25:16.240
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
So that, that notion is still somewhere deeply compelling to you. The notion that you can say,
https://karpathy.ai/lexicap/0015-large.html#00:25:19.840
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
you can drop thing on the table says this, you can expect this, this algorithm will
https://karpathy.ai/lexicap/0015-large.html#00:25:27.520
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
give me some good results. I hope there's, I hope science will, I mean,
https://karpathy.ai/lexicap/0015-large.html#00:25:32.400
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
there's engineering and there's science. I think that they're not exactly the same.
https://karpathy.ai/lexicap/0015-large.html#00:25:37.280
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
And I think right now we're making huge engineering, like leaps and bounds. So the
https://karpathy.ai/lexicap/0015-large.html#00:25:42.320
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
engineering is running away ahead of the science, which is cool. And often how it goes, right? So
https://karpathy.ai/lexicap/0015-large.html#00:25:47.040
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
we're making things and nobody knows how and why they work roughly, but we need to turn that into
https://karpathy.ai/lexicap/0015-large.html#00:25:52.240
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
science. There's some form. It's a, yeah, there's some room for formalizing. We need to know what
https://karpathy.ai/lexicap/0015-large.html#00:25:59.680
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
the principles are. Why does this work? Why does that not work? I mean, for a while, people built
https://karpathy.ai/lexicap/0015-large.html#00:26:05.440
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
bridges by trying, but now we can often predict whether it's going to work or not without building
https://karpathy.ai/lexicap/0015-large.html#00:26:09.840
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
it. Can we do that for learning systems or for robots? So your hope is from a materialistic
https://karpathy.ai/lexicap/0015-large.html#00:26:14.480
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
perspective that intelligence, artificial intelligence systems, robots are just fancier
https://karpathy.ai/lexicap/0015-large.html#00:26:20.640
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
bridges. Belief space. What's the difference between belief space and state space? So you
https://karpathy.ai/lexicap/0015-large.html#00:26:27.600
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
mentioned MDPs, FOMDPs, reasoning about, you sense the world, there's a state.
https://karpathy.ai/lexicap/0015-large.html#00:26:33.040
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
Uh, what, what's this belief space idea? That sounds so good.
https://karpathy.ai/lexicap/0015-large.html#00:26:39.840
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
It sounds good. So belief space, that is instead of thinking about what's the state of the world
https://karpathy.ai/lexicap/0015-large.html#00:26:44.400
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
and trying to control that as a robot, I think about what is the space of beliefs that I could
https://karpathy.ai/lexicap/0015-large.html#00:26:51.600
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
have about the world. What's, if I think of a belief as a probability distribution of our ways
https://karpathy.ai/lexicap/0015-large.html#00:26:58.880
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
the world could be, a belief state is a distribution. And then my control problem, if I'm reasoning
https://karpathy.ai/lexicap/0015-large.html#00:27:03.520
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
about how to move through a world I'm uncertain about, my control problem is actually the problem
https://karpathy.ai/lexicap/0015-large.html#00:27:10.080
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
of controlling my beliefs. So I think about taking actions, not just what effect they'll have on the
https://karpathy.ai/lexicap/0015-large.html#00:27:16.160
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
world outside, but what effect they'll have on my own understanding of the world outside. And so
https://karpathy.ai/lexicap/0015-large.html#00:27:21.360
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
that might compel me to ask a question or look somewhere to gather information, which may not
https://karpathy.ai/lexicap/0015-large.html#00:27:26.080
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
really change the world state, but it changes my own belief about the world. That's a powerful way
https://karpathy.ai/lexicap/0015-large.html#00:27:32.800
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
to, to empower the agent, to reason about the world, to explore the world. So what kind of
https://karpathy.ai/lexicap/0015-large.html#00:27:38.400
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
problems does it allow you to solve to, to consider belief space versus just state space?
https://karpathy.ai/lexicap/0015-large.html#00:27:46.320
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
Well, any problem that requires deliberate information gathering, right? So if in some
https://karpathy.ai/lexicap/0015-large.html#00:27:52.400
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
problems like chess, there's no uncertainty, or maybe there's uncertainty about the opponent,
https://karpathy.ai/lexicap/0015-large.html#00:27:58.400
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
there's no uncertainty about the state. And some problems, there's uncertainty,
https://karpathy.ai/lexicap/0015-large.html#00:28:05.040
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
but you gather information as you go, right? You might say, Oh, I'm driving my autonomous car down
https://karpathy.ai/lexicap/0015-large.html#00:28:10.320
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
the road and it doesn't know perfectly where it is, but the light hours are all going all the time.
https://karpathy.ai/lexicap/0015-large.html#00:28:16.080
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
So I don't have to think about whether to gather information. But if you're a human driving down
https://karpathy.ai/lexicap/0015-large.html#00:28:20.560
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
the road, you sometimes look over your shoulder to see what's going on behind you in the lane.
https://karpathy.ai/lexicap/0015-large.html#00:28:25.360
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
And you have to decide whether you should do that now. And you have to trade off the fact that
https://karpathy.ai/lexicap/0015-large.html#00:28:31.360
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
you're not seeing in front of you and you're looking behind you and how valuable is that
https://karpathy.ai/lexicap/0015-large.html#00:28:37.840
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
information and so on. And so to make choices about information gathering, you have to reasonably
https://karpathy.ai/lexicap/0015-large.html#00:28:41.520
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
space. Also, I mean, also to just take into account your own uncertainty before trying to
https://karpathy.ai/lexicap/0015-large.html#00:28:47.200
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
do things. So you might say, if I understand where I'm standing relative to the door jam,
https://karpathy.ai/lexicap/0015-large.html#00:28:56.800
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
pretty accurately, then it's okay for me to go through the door. But if I'm really
https://karpathy.ai/lexicap/0015-large.html#00:29:05.360
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
not sure where the door is, then it might be better to not do that right now.
https://karpathy.ai/lexicap/0015-large.html#00:29:08.640
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
The degree of your uncertainty about the world is actually part of the thing you're trying to
https://karpathy.ai/lexicap/0015-large.html#00:29:12.960
Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics | Lex Fridman Podcast #15
optimize in forming the plan, right? So this idea of a long horizon of planning for a PhD or just
https://karpathy.ai/lexicap/0015-large.html#00:29:17.760