episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
that tells you a lot about what's gonna be the case
https://karpathy.ai/lexicap/0010-large.html#00:18:02.560
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
in the future.
https://karpathy.ai/lexicap/0010-large.html#00:18:04.640
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Because that's what it means to take a high level action.
https://karpathy.ai/lexicap/0010-large.html#00:18:05.480
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I say okay, I decide I'm gonna navigate
https://karpathy.ai/lexicap/0010-large.html#00:18:09.400
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
to the gas station because I need to get gas for my car.
https://karpathy.ai/lexicap/0010-large.html#00:18:13.040
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Well, that'll now take five minutes to get there.
https://karpathy.ai/lexicap/0010-large.html#00:18:15.480
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
But the fact that I get there,
https://karpathy.ai/lexicap/0010-large.html#00:18:17.880
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I could already tell that from the high level action
https://karpathy.ai/lexicap/0010-large.html#00:18:19.280
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I took much earlier.
https://karpathy.ai/lexicap/0010-large.html#00:18:22.320
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
That we had a very hard time getting success with.
https://karpathy.ai/lexicap/0010-large.html#00:18:24.480
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Not saying it's a dead end necessarily,
https://karpathy.ai/lexicap/0010-large.html#00:18:28.440
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
but we had a lot of trouble getting that to work.
https://karpathy.ai/lexicap/0010-large.html#00:18:30.640
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And then we started revisiting the notion
https://karpathy.ai/lexicap/0010-large.html#00:18:33.120
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
of what are we really trying to achieve?
https://karpathy.ai/lexicap/0010-large.html#00:18:34.720
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
What we're trying to achieve is not necessarily hierarchy
https://karpathy.ai/lexicap/0010-large.html#00:18:37.800
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
per se, but you could think about
https://karpathy.ai/lexicap/0010-large.html#00:18:40.680
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
what does hierarchy give us?
https://karpathy.ai/lexicap/0010-large.html#00:18:41.720
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
What we hope it would give us is better credit assignment.
https://karpathy.ai/lexicap/0010-large.html#00:18:44.280
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
What is better credit assignment?
https://karpathy.ai/lexicap/0010-large.html#00:18:49.120
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
It's giving us, it gives us faster learning, right?
https://karpathy.ai/lexicap/0010-large.html#00:18:51.240
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And so faster learning is ultimately maybe what we're after.
https://karpathy.ai/lexicap/0010-large.html#00:18:55.760
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And so that's where we ended up with the RL squared paper
https://karpathy.ai/lexicap/0010-large.html#00:18:59.800
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
on learning to reinforcement learn,
https://karpathy.ai/lexicap/0010-large.html#00:19:03.400
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
which at a time Rocky Dwan led.
https://karpathy.ai/lexicap/0010-large.html#00:19:06.040
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And that's exactly the meta learning approach
https://karpathy.ai/lexicap/0010-large.html#00:19:08.840
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
where you say, okay, we don't know how to design hierarchy.
https://karpathy.ai/lexicap/0010-large.html#00:19:11.080
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
We know what we want to get from it.
https://karpathy.ai/lexicap/0010-large.html#00:19:14.240
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Let's just enter and optimize for what we want to get
https://karpathy.ai/lexicap/0010-large.html#00:19:15.760
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
from it and see if it might emerge.
https://karpathy.ai/lexicap/0010-large.html#00:19:18.240
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And we saw things emerge.
https://karpathy.ai/lexicap/0010-large.html#00:19:20.200
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
The maze navigation had consistent motion down hallways,
https://karpathy.ai/lexicap/0010-large.html#00:19:21.240
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
which is what you want.
https://karpathy.ai/lexicap/0010-large.html#00:19:26.120
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
A hierarchical control should say,
https://karpathy.ai/lexicap/0010-large.html#00:19:27.160
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I want to go down this hallway.
https://karpathy.ai/lexicap/0010-large.html#00:19:28.320
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And then when there is an option to take a turn,
https://karpathy.ai/lexicap/0010-large.html#00:19:29.720
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I can decide whether to take a turn or not and repeat.
https://karpathy.ai/lexicap/0010-large.html#00:19:31.640
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Even had the notion of where have you been before or not
https://karpathy.ai/lexicap/0010-large.html#00:19:33.840
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
to not revisit places you've been before.
https://karpathy.ai/lexicap/0010-large.html#00:19:37.280
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
It still didn't scale yet
https://karpathy.ai/lexicap/0010-large.html#00:19:39.960
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
to the real world kind of scenarios I think you had in mind,
https://karpathy.ai/lexicap/0010-large.html#00:19:42.520
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
but it was some sign of life
https://karpathy.ai/lexicap/0010-large.html#00:19:46.000
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
that maybe you can meta learn these hierarchical concepts.
https://karpathy.ai/lexicap/0010-large.html#00:19:47.200
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I mean, it seems like through these meta learning concepts,
https://karpathy.ai/lexicap/0010-large.html#00:19:51.160
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
get at the, what I think is one of the hardest
https://karpathy.ai/lexicap/0010-large.html#00:19:56.160
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and most important problems of AI,
https://karpathy.ai/lexicap/0010-large.html#00:19:59.800
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
which is transfer learning.
https://karpathy.ai/lexicap/0010-large.html#00:20:02.360
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So it's generalization.
https://karpathy.ai/lexicap/0010-large.html#00:20:04.040
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
How far along this journey
https://karpathy.ai/lexicap/0010-large.html#00:20:06.280
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
towards building general systems are we?
https://karpathy.ai/lexicap/0010-large.html#00:20:08.480
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Being able to do transfer learning well.
https://karpathy.ai/lexicap/0010-large.html#00:20:11.160
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So there's some signs that you can generalize a little bit,
https://karpathy.ai/lexicap/0010-large.html#00:20:13.600
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
but do you think we're on the right path
https://karpathy.ai/lexicap/0010-large.html#00:20:17.520
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
or it's totally different breakthroughs are needed
https://karpathy.ai/lexicap/0010-large.html#00:20:19.600
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
to be able to transfer knowledge
https://karpathy.ai/lexicap/0010-large.html#00:20:23.760
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
between different learned models?
https://karpathy.ai/lexicap/0010-large.html#00:20:26.800
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Yeah, I'm pretty torn on this in that
https://karpathy.ai/lexicap/0010-large.html#00:20:31.240
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I think there are some very impressive.
https://karpathy.ai/lexicap/0010-large.html#00:20:33.840
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Well, there's just some very impressive results already.
https://karpathy.ai/lexicap/0010-large.html#00:20:35.560
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I mean, I would say when,
https://karpathy.ai/lexicap/0010-large.html#00:20:40.520
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
even with the initial kind of big breakthrough in 2012
https://karpathy.ai/lexicap/0010-large.html#00:20:44.040
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
with AlexNet, the initial thing is okay, great.
https://karpathy.ai/lexicap/0010-large.html#00:20:47.240
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
This does better on ImageNet, hence image recognition.
https://karpathy.ai/lexicap/0010-large.html#00:20:52.160
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
But then immediately thereafter,
https://karpathy.ai/lexicap/0010-large.html#00:20:55.680
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
there was of course the notion that,
https://karpathy.ai/lexicap/0010-large.html#00:20:57.840
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
wow, what was learned on ImageNet
https://karpathy.ai/lexicap/0010-large.html#00:21:00.520
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and you now wanna solve a new task,
https://karpathy.ai/lexicap/0010-large.html#00:21:03.320
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
you can fine tune AlexNet for new tasks.
https://karpathy.ai/lexicap/0010-large.html#00:21:05.000
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And that was often found to be the even bigger deal
https://karpathy.ai/lexicap/0010-large.html#00:21:09.080
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
that you learn something that was reusable,
https://karpathy.ai/lexicap/0010-large.html#00:21:12.040
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
which was not often the case before.
https://karpathy.ai/lexicap/0010-large.html#00:21:14.320
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Usually machine learning, you learn something
https://karpathy.ai/lexicap/0010-large.html#00:21:16.040
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
for one scenario and that was it.
https://karpathy.ai/lexicap/0010-large.html#00:21:17.520
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And that's really exciting.
https://karpathy.ai/lexicap/0010-large.html#00:21:19.320
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I mean, that's a huge application.
https://karpathy.ai/lexicap/0010-large.html#00:21:20.280
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
That's probably the biggest success
https://karpathy.ai/lexicap/0010-large.html#00:21:22.280
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
of transfer learning today in terms of scope and impact.
https://karpathy.ai/lexicap/0010-large.html#00:21:23.680
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
That was a huge breakthrough.
https://karpathy.ai/lexicap/0010-large.html#00:21:27.920
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And then recently, I feel like similar kind of,
https://karpathy.ai/lexicap/0010-large.html#00:21:29.040
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
by scaling things up, it seems like
https://karpathy.ai/lexicap/0010-large.html#00:21:33.040
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
this has been expanded upon.
https://karpathy.ai/lexicap/0010-large.html#00:21:34.760
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Like people training even bigger networks,
https://karpathy.ai/lexicap/0010-large.html#00:21:36.160
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
they might transfer even better.
https://karpathy.ai/lexicap/0010-large.html#00:21:37.960
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
If you looked at, for example,
https://karpathy.ai/lexicap/0010-large.html#00:21:39.480
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
some of the OpenAI results on language models
https://karpathy.ai/lexicap/0010-large.html#00:21:41.200
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and some of the recent Google results on language models,
https://karpathy.ai/lexicap/0010-large.html#00:21:43.400
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
they're learned for just prediction
https://karpathy.ai/lexicap/0010-large.html#00:21:47.560
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and then they get reused for other tasks.
https://karpathy.ai/lexicap/0010-large.html#00:21:51.040
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And so I think there is something there
https://karpathy.ai/lexicap/0010-large.html#00:21:54.960
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
where somehow if you train a big enough model
https://karpathy.ai/lexicap/0010-large.html#00:21:56.680
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
on enough things, it seems to transfer
https://karpathy.ai/lexicap/0010-large.html#00:21:58.520
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
some deep mind results that I thought were very impressive,
https://karpathy.ai/lexicap/0010-large.html#00:22:01.360
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
the Unreal results, where it was learned to navigate mazes
https://karpathy.ai/lexicap/0010-large.html#00:22:03.600
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
in ways where it wasn't just doing reinforcement learning,
https://karpathy.ai/lexicap/0010-large.html#00:22:09.240
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
but it had other objectives it was optimizing for.
https://karpathy.ai/lexicap/0010-large.html#00:22:11.240
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So I think there's a lot of interesting results already.
https://karpathy.ai/lexicap/0010-large.html#00:22:14.280
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I think maybe where it's hard to wrap my head around this,
https://karpathy.ai/lexicap/0010-large.html#00:22:17.240
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
to which extent or when do we call something generalization?
https://karpathy.ai/lexicap/0010-large.html#00:22:22.520
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Or the levels of generalization in the real world,
https://karpathy.ai/lexicap/0010-large.html#00:22:26.720
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
or the levels of generalization involved
https://karpathy.ai/lexicap/0010-large.html#00:22:29.760
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
in these different tasks, right?
https://karpathy.ai/lexicap/0010-large.html#00:22:31.880