episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Of course, you're really trying to build these systems,
https://karpathy.ai/lexicap/0010-large.html#00:35:09.460
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
but do you think about the future of AI?
https://karpathy.ai/lexicap/0010-large.html#00:35:12.060
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
A lot of people have concern about safety.
https://karpathy.ai/lexicap/0010-large.html#00:35:13.780
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
How do you think about AI safety?
https://karpathy.ai/lexicap/0010-large.html#00:35:16.380
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
As you build robots that are operating in the physical world,
https://karpathy.ai/lexicap/0010-large.html#00:35:18.240
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
what is, yeah, how do you approach this problem
https://karpathy.ai/lexicap/0010-large.html#00:35:21.460
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
in an engineering kind of way, in a systematic way?
https://karpathy.ai/lexicap/0010-large.html#00:35:25.060
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So when a robot is doing things,
https://karpathy.ai/lexicap/0010-large.html#00:35:29.220
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
you kind of have a few notions of safety to worry about.
https://karpathy.ai/lexicap/0010-large.html#00:35:32.340
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
One is that the robot is physically strong
https://karpathy.ai/lexicap/0010-large.html#00:35:36.240
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and of course could do a lot of damage.
https://karpathy.ai/lexicap/0010-large.html#00:35:39.380
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Same for cars, which we can think of as robots too
https://karpathy.ai/lexicap/0010-large.html#00:35:42.340
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
in some way.
https://karpathy.ai/lexicap/0010-large.html#00:35:44.840
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And this could be completely unintentional.
https://karpathy.ai/lexicap/0010-large.html#00:35:46.780
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So it could be not the kind of longterm AI safety concerns
https://karpathy.ai/lexicap/0010-large.html#00:35:48.340
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
that, okay, AI is smarter than us and now what do we do?
https://karpathy.ai/lexicap/0010-large.html#00:35:51.780
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
But it could be just very practical.
https://karpathy.ai/lexicap/0010-large.html#00:35:54.380
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Okay, this robot, if it makes a mistake,
https://karpathy.ai/lexicap/0010-large.html#00:35:55.860
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
what are the results going to be?
https://karpathy.ai/lexicap/0010-large.html#00:35:58.920
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Of course, simulation comes in a lot there
https://karpathy.ai/lexicap/0010-large.html#00:36:00.700
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
to test in simulation. It's a difficult question.
https://karpathy.ai/lexicap/0010-large.html#00:36:02.280
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And I'm always wondering, like, I always wonder,
https://karpathy.ai/lexicap/0010-large.html#00:36:07.780
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
let's say you look at, let's go back to driving
https://karpathy.ai/lexicap/0010-large.html#00:36:09.540
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
because a lot of people know driving well, of course.
https://karpathy.ai/lexicap/0010-large.html#00:36:12.020
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
What do we do to test somebody for driving, right?
https://karpathy.ai/lexicap/0010-large.html#00:36:15.280
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Get a driver's license. What do they really do?
https://karpathy.ai/lexicap/0010-large.html#00:36:18.940
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I mean, you fill out some tests and then you drive.
https://karpathy.ai/lexicap/0010-large.html#00:36:21.420
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And I mean, it's suburban California.
https://karpathy.ai/lexicap/0010-large.html#00:36:26.660
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
That driving test is just you drive around the block,
https://karpathy.ai/lexicap/0010-large.html#00:36:29.500
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
pull over, you do a stop sign successfully,
https://karpathy.ai/lexicap/0010-large.html#00:36:32.940
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and then you pull over again and you're pretty much done.
https://karpathy.ai/lexicap/0010-large.html#00:36:36.500
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And you're like, okay, if a self driving car did that,
https://karpathy.ai/lexicap/0010-large.html#00:36:40.060
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
would you trust it that it can drive?
https://karpathy.ai/lexicap/0010-large.html#00:36:44.500
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And I'd be like, no, that's not enough for me to trust it.
https://karpathy.ai/lexicap/0010-large.html#00:36:46.840
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
But somehow for humans, we've figured out
https://karpathy.ai/lexicap/0010-large.html#00:36:48.900
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
that somebody being able to do that is representative
https://karpathy.ai/lexicap/0010-large.html#00:36:51.540
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
of them being able to do a lot of other things.
https://karpathy.ai/lexicap/0010-large.html#00:36:55.220
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And so I think somehow for humans,
https://karpathy.ai/lexicap/0010-large.html#00:36:57.900
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
we figured out representative tests
https://karpathy.ai/lexicap/0010-large.html#00:36:59.980
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
of what it means if you can do this, what you can really do.
https://karpathy.ai/lexicap/0010-large.html#00:37:02.660
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Of course, testing humans,
https://karpathy.ai/lexicap/0010-large.html#00:37:05.860
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
humans don't wanna be tested at all times.
https://karpathy.ai/lexicap/0010-large.html#00:37:07.380
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Self driving cars or robots
https://karpathy.ai/lexicap/0010-large.html#00:37:09.180
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
could be tested more often probably.
https://karpathy.ai/lexicap/0010-large.html#00:37:10.300
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
You can have replicas that get tested
https://karpathy.ai/lexicap/0010-large.html#00:37:11.980
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
that are known to be identical
https://karpathy.ai/lexicap/0010-large.html#00:37:13.460
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
because they use the same neural net and so forth.
https://karpathy.ai/lexicap/0010-large.html#00:37:14.820
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
But still, I feel like we don't have this kind of unit tests
https://karpathy.ai/lexicap/0010-large.html#00:37:17.140
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
or proper tests for robots.
https://karpathy.ai/lexicap/0010-large.html#00:37:21.260
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And I think there's something very interesting
https://karpathy.ai/lexicap/0010-large.html#00:37:24.420
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
to be thought about there,
https://karpathy.ai/lexicap/0010-large.html#00:37:25.520
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
especially as you update things.
https://karpathy.ai/lexicap/0010-large.html#00:37:26.780
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Your software improves,
https://karpathy.ai/lexicap/0010-large.html#00:37:28.540
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
you have a better self driving car suite, you update it.
https://karpathy.ai/lexicap/0010-large.html#00:37:29.580
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
How do you know it's indeed more capable on everything
https://karpathy.ai/lexicap/0010-large.html#00:37:32.320
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
than what you had before,
https://karpathy.ai/lexicap/0010-large.html#00:37:35.960
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
that you didn't have any bad things creep into it?
https://karpathy.ai/lexicap/0010-large.html#00:37:37.280
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So I think that's a very interesting direction of research
https://karpathy.ai/lexicap/0010-large.html#00:37:41.500
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
that there is no real solution yet,
https://karpathy.ai/lexicap/0010-large.html#00:37:43.540
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
except that somehow for humans we do.
https://karpathy.ai/lexicap/0010-large.html#00:37:46.340
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Because we say, okay, you have a driving test, you passed,
https://karpathy.ai/lexicap/0010-large.html#00:37:47.980
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
you can go on the road now,
https://karpathy.ai/lexicap/0010-large.html#00:37:50.820
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and humans have accidents every like a million
https://karpathy.ai/lexicap/0010-large.html#00:37:51.940
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
or 10 million miles, something pretty phenomenal
https://karpathy.ai/lexicap/0010-large.html#00:37:54.900
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
compared to that short test that is being done.
https://karpathy.ai/lexicap/0010-large.html#00:37:57.860
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So let me ask, you've mentioned that Andrew Ng by example
https://karpathy.ai/lexicap/0010-large.html#00:38:01.660
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
showed you the value of kindness.
https://karpathy.ai/lexicap/0010-large.html#00:38:06.100
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Do you think the space of policies,
https://karpathy.ai/lexicap/0010-large.html#00:38:10.100
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
good policies for humans and for AI
https://karpathy.ai/lexicap/0010-large.html#00:38:14.580
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
is populated by policies that with kindness
https://karpathy.ai/lexicap/0010-large.html#00:38:17.500
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
or ones that are the opposite, exploitation, even evil?
https://karpathy.ai/lexicap/0010-large.html#00:38:22.500
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So if you just look at the sea of policies
https://karpathy.ai/lexicap/0010-large.html#00:38:28.220
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
we operate under as human beings,
https://karpathy.ai/lexicap/0010-large.html#00:38:30.300
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
or if AI system had to operate in this real world,
https://karpathy.ai/lexicap/0010-large.html#00:38:32.540
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
do you think it's really easy to find policies
https://karpathy.ai/lexicap/0010-large.html#00:38:35.300
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
that are full of kindness,
https://karpathy.ai/lexicap/0010-large.html#00:38:38.060
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
like we naturally fall into them?
https://karpathy.ai/lexicap/0010-large.html#00:38:39.580
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Or is it like a very hard optimization problem?
https://karpathy.ai/lexicap/0010-large.html#00:38:41.340
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I mean, there is kind of two optimizations
https://karpathy.ai/lexicap/0010-large.html#00:38:48.100
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
happening for humans, right?
https://karpathy.ai/lexicap/0010-large.html#00:38:50.300
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So for humans, there's kind of the very long term
https://karpathy.ai/lexicap/0010-large.html#00:38:52.300
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
optimization which evolution has done for us
https://karpathy.ai/lexicap/0010-large.html#00:38:54.140
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and we're kind of predisposed to like certain things.
https://karpathy.ai/lexicap/0010-large.html#00:38:56.900
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And that's in some sense what makes our learning easier
https://karpathy.ai/lexicap/0010-large.html#00:39:00.780
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
because I mean, we know things like pain
https://karpathy.ai/lexicap/0010-large.html#00:39:02.780
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and hunger and thirst.
https://karpathy.ai/lexicap/0010-large.html#00:39:05.420
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And the fact that we know about those
https://karpathy.ai/lexicap/0010-large.html#00:39:08.420
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
is not something that we were taught, that's kind of innate.
https://karpathy.ai/lexicap/0010-large.html#00:39:10.100
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
When we're hungry, we're unhappy.
https://karpathy.ai/lexicap/0010-large.html#00:39:12.740
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
When we're thirsty, we're unhappy.
https://karpathy.ai/lexicap/0010-large.html#00:39:14.060
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
When we have pain, we're unhappy.
https://karpathy.ai/lexicap/0010-large.html#00:39:16.220
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And ultimately evolution built that into us
https://karpathy.ai/lexicap/0010-large.html#00:39:18.420
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
to think about those things.
https://karpathy.ai/lexicap/0010-large.html#00:39:21.760
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And so I think there is a notion that
https://karpathy.ai/lexicap/0010-large.html#00:39:22.600
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
it seems somehow humans evolved in general
https://karpathy.ai/lexicap/0010-large.html#00:39:24.660
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
to prefer to get along in some ways,
https://karpathy.ai/lexicap/0010-large.html#00:39:28.220
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
but at the same time also to be very territorial
https://karpathy.ai/lexicap/0010-large.html#00:39:32.820
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and kind of centric to their own tribe.
https://karpathy.ai/lexicap/0010-large.html#00:39:36.940
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Like it seems like that's the kind of space
https://karpathy.ai/lexicap/0010-large.html#00:39:41.620
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
we converged onto.
https://karpathy.ai/lexicap/0010-large.html#00:39:43.580