episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
there's something magical.
https://karpathy.ai/lexicap/0010-large.html#00:05:21.000
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So do you think about the psychology of interacting
https://karpathy.ai/lexicap/0010-large.html#00:05:23.240
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
with robots in the physical world?
https://karpathy.ai/lexicap/0010-large.html#00:05:27.560
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Even you just showed me the PR2, the robot,
https://karpathy.ai/lexicap/0010-large.html#00:05:29.080
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and there was a little bit something like a face,
https://karpathy.ai/lexicap/0010-large.html#00:05:33.720
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
had a little bit something like a face.
https://karpathy.ai/lexicap/0010-large.html#00:05:36.860
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
There's something that immediately draws you to it.
https://karpathy.ai/lexicap/0010-large.html#00:05:38.480
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Do you think about that aspect of the robotics problem?
https://karpathy.ai/lexicap/0010-large.html#00:05:40.600
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Well, it's very hard with Brad here.
https://karpathy.ai/lexicap/0010-large.html#00:05:45.160
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
We'll give him a name, Berkeley Robot
https://karpathy.ai/lexicap/0010-large.html#00:05:48.400
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
for the Elimination of Tedious Tasks.
https://karpathy.ai/lexicap/0010-large.html#00:05:50.680
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
It's very hard to not think of the robot as a person
https://karpathy.ai/lexicap/0010-large.html#00:05:52.200
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and it seems like everybody calls him a he
https://karpathy.ai/lexicap/0010-large.html#00:05:56.560
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
for whatever reason, but that also makes it more a person
https://karpathy.ai/lexicap/0010-large.html#00:05:58.880
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
than if it was a it, and it seems pretty natural
https://karpathy.ai/lexicap/0010-large.html#00:06:01.160
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
to think of it that way.
https://karpathy.ai/lexicap/0010-large.html#00:06:06.360
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
This past weekend really struck me.
https://karpathy.ai/lexicap/0010-large.html#00:06:07.320
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I've seen Pepper many times on videos,
https://karpathy.ai/lexicap/0010-large.html#00:06:08.680
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
but then I was at an event organized by,
https://karpathy.ai/lexicap/0010-large.html#00:06:13.360
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
this was by Fidelity, and they had scripted Pepper
https://karpathy.ai/lexicap/0010-large.html#00:06:15.360
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
to help moderate some sessions,
https://karpathy.ai/lexicap/0010-large.html#00:06:18.880
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and they had scripted Pepper
https://karpathy.ai/lexicap/0010-large.html#00:06:22.800
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
to have the personality of a child a little bit,
https://karpathy.ai/lexicap/0010-large.html#00:06:23.920
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and it was very hard to not think of it
https://karpathy.ai/lexicap/0010-large.html#00:06:26.520
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
as its own person in some sense
https://karpathy.ai/lexicap/0010-large.html#00:06:28.600
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
because it would just jump in the conversation,
https://karpathy.ai/lexicap/0010-large.html#00:06:31.920
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
making it very interactive.
https://karpathy.ai/lexicap/0010-large.html#00:06:34.560
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Moderate would be saying, Pepper would just jump in,
https://karpathy.ai/lexicap/0010-large.html#00:06:35.880
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
hold on, how about me?
https://karpathy.ai/lexicap/0010-large.html#00:06:37.960
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Can I participate in this too?
https://karpathy.ai/lexicap/0010-large.html#00:06:40.120
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And you're just like, okay, this is like a person,
https://karpathy.ai/lexicap/0010-large.html#00:06:41.360
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and that was 100% scripted, and even then it was hard
https://karpathy.ai/lexicap/0010-large.html#00:06:43.720
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
not to have that sense of somehow there is something there.
https://karpathy.ai/lexicap/0010-large.html#00:06:46.640
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So as we have robots interact in this physical world,
https://karpathy.ai/lexicap/0010-large.html#00:06:50.640
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
is that a signal that could be used
https://karpathy.ai/lexicap/0010-large.html#00:06:54.440
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
in reinforcement learning?
https://karpathy.ai/lexicap/0010-large.html#00:06:56.120
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
You've worked a little bit in this direction,
https://karpathy.ai/lexicap/0010-large.html#00:06:57.440
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
but do you think that psychology can be somehow pulled in?
https://karpathy.ai/lexicap/0010-large.html#00:07:00.240
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Yes, that's a question I would say
https://karpathy.ai/lexicap/0010-large.html#00:07:04.360
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
a lot of people ask, and I think part of why they ask it
https://karpathy.ai/lexicap/0010-large.html#00:07:07.160
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
is they're thinking about how unique
https://karpathy.ai/lexicap/0010-large.html#00:07:11.320
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
are we really still as people?
https://karpathy.ai/lexicap/0010-large.html#00:07:14.960
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Like after they see some results,
https://karpathy.ai/lexicap/0010-large.html#00:07:16.680
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
they see a computer play Go, they see a computer do this,
https://karpathy.ai/lexicap/0010-large.html#00:07:18.120
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
that, they're like, okay, but can it really have emotion?
https://karpathy.ai/lexicap/0010-large.html#00:07:21.440
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Can it really interact with us in that way?
https://karpathy.ai/lexicap/0010-large.html#00:07:23.760
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And then once you're around robots,
https://karpathy.ai/lexicap/0010-large.html#00:07:26.760
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
you already start feeling it,
https://karpathy.ai/lexicap/0010-large.html#00:07:29.100
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and I think that kind of maybe mythologically,
https://karpathy.ai/lexicap/0010-large.html#00:07:30.120
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
the way that I think of it is
https://karpathy.ai/lexicap/0010-large.html#00:07:33.180
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
if you run something like reinforcement learning,
https://karpathy.ai/lexicap/0010-large.html#00:07:34.720
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
it's about optimizing some objective,
https://karpathy.ai/lexicap/0010-large.html#00:07:37.640
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and there's no reason that the objective
https://karpathy.ai/lexicap/0010-large.html#00:07:39.920
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
couldn't be tied into how much does a person like
https://karpathy.ai/lexicap/0010-large.html#00:07:45.360
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
interacting with this system,
https://karpathy.ai/lexicap/0010-large.html#00:07:49.380
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and why could not the reinforcement learning system
https://karpathy.ai/lexicap/0010-large.html#00:07:50.720
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
optimize for the robot being fun to be around?
https://karpathy.ai/lexicap/0010-large.html#00:07:53.220
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And why wouldn't it then naturally become
https://karpathy.ai/lexicap/0010-large.html#00:07:56.720
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
more and more interactive and more and more
https://karpathy.ai/lexicap/0010-large.html#00:07:58.940
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
maybe like a person or like a pet?
https://karpathy.ai/lexicap/0010-large.html#00:08:01.400
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I don't know what it would exactly be,
https://karpathy.ai/lexicap/0010-large.html#00:08:03.200
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
but more and more have those features
https://karpathy.ai/lexicap/0010-large.html#00:08:04.600
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and acquire them automatically.
https://karpathy.ai/lexicap/0010-large.html#00:08:06.640
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
As long as you can formalize an objective
https://karpathy.ai/lexicap/0010-large.html#00:08:08.320
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
of what it means to like something,
https://karpathy.ai/lexicap/0010-large.html#00:08:10.880
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
what, how you exhibit, what's the ground truth?
https://karpathy.ai/lexicap/0010-large.html#00:08:13.440
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
How do you get the reward from human?
https://karpathy.ai/lexicap/0010-large.html#00:08:16.800
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Because you have to somehow collect
https://karpathy.ai/lexicap/0010-large.html#00:08:19.560
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
that information within you, human.
https://karpathy.ai/lexicap/0010-large.html#00:08:20.760
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
But you're saying if you can formulate as an objective,
https://karpathy.ai/lexicap/0010-large.html#00:08:22.400
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
it can be learned.
https://karpathy.ai/lexicap/0010-large.html#00:08:26.280
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
There's no reason it couldn't emerge through learning,
https://karpathy.ai/lexicap/0010-large.html#00:08:27.240
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and maybe one way to formulate as an objective,
https://karpathy.ai/lexicap/0010-large.html#00:08:29.380
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
you wouldn't have to necessarily score it explicitly,
https://karpathy.ai/lexicap/0010-large.html#00:08:31.480
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
so standard rewards are numbers,
https://karpathy.ai/lexicap/0010-large.html#00:08:33.800
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and numbers are hard to come by.
https://karpathy.ai/lexicap/0010-large.html#00:08:36.560
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
This is a 1.5 or a 1.7 on some scale.
https://karpathy.ai/lexicap/0010-large.html#00:08:38.740
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
It's very hard to do for a person,
https://karpathy.ai/lexicap/0010-large.html#00:08:41.320
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
but much easier is for a person to say,
https://karpathy.ai/lexicap/0010-large.html#00:08:43.060
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
okay, what you did the last five minutes
https://karpathy.ai/lexicap/0010-large.html#00:08:45.420
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
was much nicer than what you did the previous five minutes,
https://karpathy.ai/lexicap/0010-large.html#00:08:47.800
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and that now gives a comparison.
https://karpathy.ai/lexicap/0010-large.html#00:08:51.160
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And in fact, there have been some results on that.
https://karpathy.ai/lexicap/0010-large.html#00:08:53.080
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
For example, Paul Christiano and collaborators at OpenAI
https://karpathy.ai/lexicap/0010-large.html#00:08:55.320
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
had the Hopper, Mojoko Hopper, a one legged robot,
https://karpathy.ai/lexicap/0010-large.html#00:08:57.880
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
going through backflips purely from feedback.
https://karpathy.ai/lexicap/0010-large.html#00:09:02.040
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I like this better than that.
https://karpathy.ai/lexicap/0010-large.html#00:09:05.600
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
That's kind of equally good,
https://karpathy.ai/lexicap/0010-large.html#00:09:06.920
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and after a bunch of interactions,
https://karpathy.ai/lexicap/0010-large.html#00:09:08.640
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
it figured out what it was the person was asking for,
https://karpathy.ai/lexicap/0010-large.html#00:09:10.920
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
namely a backflip.
https://karpathy.ai/lexicap/0010-large.html#00:09:13.080
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And so I think the same thing.
https://karpathy.ai/lexicap/0010-large.html#00:09:14.400
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Oh, it wasn't trying to do a backflip.
https://karpathy.ai/lexicap/0010-large.html#00:09:15.920
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
It was just getting a comparison score
https://karpathy.ai/lexicap/0010-large.html#00:09:18.640
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
from the person based on?
https://karpathy.ai/lexicap/0010-large.html#00:09:20.820
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Person having in mind, in their own mind,
https://karpathy.ai/lexicap/0010-large.html#00:09:23.320
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I wanted to do a backflip,
https://karpathy.ai/lexicap/0010-large.html#00:09:26.080
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
but the robot didn't know what it was supposed to be doing.
https://karpathy.ai/lexicap/0010-large.html#00:09:27.400
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
It just knew that sometimes the person said,
https://karpathy.ai/lexicap/0010-large.html#00:09:30.760
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
this is better, this is worse,
https://karpathy.ai/lexicap/0010-large.html#00:09:32.800