episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
than another one, which needs an almost infinite
https://karpathy.ai/lexicap/0011-large.html#00:24:22.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
number of bits to be described.
https://karpathy.ai/lexicap/0011-large.html#00:24:25.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
As far as we know, many things that are happening
https://karpathy.ai/lexicap/0011-large.html#00:24:28.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
in this universe are really simple in terms of
https://karpathy.ai/lexicap/0011-large.html#00:24:32.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
short programs that compute gravity
https://karpathy.ai/lexicap/0011-large.html#00:24:35.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and the interaction between elementary particles and so on.
https://karpathy.ai/lexicap/0011-large.html#00:24:38.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So all of that seems to be very, very simple.
https://karpathy.ai/lexicap/0011-large.html#00:24:43.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Every electron seems to reuse the same subprogram
https://karpathy.ai/lexicap/0011-large.html#00:24:45.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
all the time, as it is interacting with
https://karpathy.ai/lexicap/0011-large.html#00:24:50.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
other elementary particles.
https://karpathy.ai/lexicap/0011-large.html#00:24:52.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
If we now require an extra oracle injecting
https://karpathy.ai/lexicap/0011-large.html#00:24:58.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
new bits of information all the time for these
https://karpathy.ai/lexicap/0011-large.html#00:25:05.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
extra things which are currently not understood,
https://karpathy.ai/lexicap/0011-large.html#00:25:08.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
such as better decay, then the whole description
https://karpathy.ai/lexicap/0011-large.html#00:25:11.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
length of the data that we can observe of the
https://karpathy.ai/lexicap/0011-large.html#00:25:22.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
history of the universe would become much longer
https://karpathy.ai/lexicap/0011-large.html#00:25:26.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and therefore uglier.
https://karpathy.ai/lexicap/0011-large.html#00:25:31.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And uglier.
https://karpathy.ai/lexicap/0011-large.html#00:25:33.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Again, simplicity is elegant and beautiful.
https://karpathy.ai/lexicap/0011-large.html#00:25:34.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
The history of science is a history of compression progress.
https://karpathy.ai/lexicap/0011-large.html#00:25:38.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Yes, so you've described sort of as we build up
https://karpathy.ai/lexicap/0011-large.html#00:25:42.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
abstractions and you've talked about the idea
https://karpathy.ai/lexicap/0011-large.html#00:25:47.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
of compression.
https://karpathy.ai/lexicap/0011-large.html#00:25:50.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
How do you see this, the history of science,
https://karpathy.ai/lexicap/0011-large.html#00:25:52.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
the history of humanity, our civilization,
https://karpathy.ai/lexicap/0011-large.html#00:25:55.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and life on Earth as some kind of path towards
https://karpathy.ai/lexicap/0011-large.html#00:25:58.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
greater and greater compression?
https://karpathy.ai/lexicap/0011-large.html#00:26:02.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
What do you mean by that?
https://karpathy.ai/lexicap/0011-large.html#00:26:04.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
How do you think about that?
https://karpathy.ai/lexicap/0011-large.html#00:26:05.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Indeed, the history of science is a history of
https://karpathy.ai/lexicap/0011-large.html#00:26:06.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
compression progress.
https://karpathy.ai/lexicap/0011-large.html#00:26:10.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
What does that mean?
https://karpathy.ai/lexicap/0011-large.html#00:26:12.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Hundreds of years ago there was an astronomer
https://karpathy.ai/lexicap/0011-large.html#00:26:14.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
whose name was Kepler and he looked at the data
https://karpathy.ai/lexicap/0011-large.html#00:26:17.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
points that he got by watching planets move.
https://karpathy.ai/lexicap/0011-large.html#00:26:21.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And then he had all these data points and
https://karpathy.ai/lexicap/0011-large.html#00:26:25.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
suddenly it turned out that he can greatly
https://karpathy.ai/lexicap/0011-large.html#00:26:27.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
compress the data by predicting it through an
https://karpathy.ai/lexicap/0011-large.html#00:26:30.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
ellipse law.
https://karpathy.ai/lexicap/0011-large.html#00:26:36.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So it turns out that all these data points are
https://karpathy.ai/lexicap/0011-large.html#00:26:38.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
more or less on ellipses around the sun.
https://karpathy.ai/lexicap/0011-large.html#00:26:40.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And another guy came along whose name was
https://karpathy.ai/lexicap/0011-large.html#00:26:45.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Newton and before him Hooke.
https://karpathy.ai/lexicap/0011-large.html#00:26:48.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And they said the same thing that is making
https://karpathy.ai/lexicap/0011-large.html#00:26:51.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
these planets move like that is what makes the
https://karpathy.ai/lexicap/0011-large.html#00:26:55.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
apples fall down.
https://karpathy.ai/lexicap/0011-large.html#00:27:00.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And it also holds for stones and for all kinds
https://karpathy.ai/lexicap/0011-large.html#00:27:02.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
of other objects.
https://karpathy.ai/lexicap/0011-large.html#00:27:08.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And suddenly many, many of these observations
https://karpathy.ai/lexicap/0011-large.html#00:27:11.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
became much more compressible because as long
https://karpathy.ai/lexicap/0011-large.html#00:27:15.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
as you can predict the next thing, given what
https://karpathy.ai/lexicap/0011-large.html#00:27:17.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
you have seen so far, you can compress it.
https://karpathy.ai/lexicap/0011-large.html#00:27:20.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And you don't have to store that data extra.
https://karpathy.ai/lexicap/0011-large.html#00:27:23.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
This is called predictive coding.
https://karpathy.ai/lexicap/0011-large.html#00:27:25.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And then there was still something wrong with
https://karpathy.ai/lexicap/0011-large.html#00:27:29.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that theory of the universe and you had
https://karpathy.ai/lexicap/0011-large.html#00:27:31.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
deviations from these predictions of the theory.
https://karpathy.ai/lexicap/0011-large.html#00:27:34.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And 300 years later another guy came along
https://karpathy.ai/lexicap/0011-large.html#00:27:37.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
whose name was Einstein.
https://karpathy.ai/lexicap/0011-large.html#00:27:40.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And he was able to explain away all these
https://karpathy.ai/lexicap/0011-large.html#00:27:42.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
deviations from the predictions of the old
https://karpathy.ai/lexicap/0011-large.html#00:27:46.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
theory through a new theory which was called
https://karpathy.ai/lexicap/0011-large.html#00:27:50.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
the general theory of relativity.
https://karpathy.ai/lexicap/0011-large.html#00:27:54.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Which at first glance looks a little bit more
https://karpathy.ai/lexicap/0011-large.html#00:27:57.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
complicated and you have to warp space and time
https://karpathy.ai/lexicap/0011-large.html#00:28:00.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
but you can't phrase it within one single
https://karpathy.ai/lexicap/0011-large.html#00:28:03.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
sentence which is no matter how fast you
https://karpathy.ai/lexicap/0011-large.html#00:28:05.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
accelerate and how hard you decelerate and no
https://karpathy.ai/lexicap/0011-large.html#00:28:08.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
matter what is the gravity in your local
https://karpathy.ai/lexicap/0011-large.html#00:28:14.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
network, light speed always looks the same.
https://karpathy.ai/lexicap/0011-large.html#00:28:18.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And from that you can calculate all the
https://karpathy.ai/lexicap/0011-large.html#00:28:21.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
consequences.
https://karpathy.ai/lexicap/0011-large.html#00:28:24.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So it's a very simple thing and it allows you
https://karpathy.ai/lexicap/0011-large.html#00:28:25.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
to further compress all the observations
https://karpathy.ai/lexicap/0011-large.html#00:28:27.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
because certainly there are hardly any
https://karpathy.ai/lexicap/0011-large.html#00:28:30.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
deviations any longer that you can measure
https://karpathy.ai/lexicap/0011-large.html#00:28:34.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
from the predictions of this new theory.
https://karpathy.ai/lexicap/0011-large.html#00:28:37.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So all of science is a history of compression
https://karpathy.ai/lexicap/0011-large.html#00:28:40.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
progress.
https://karpathy.ai/lexicap/0011-large.html#00:28:44.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
You never arrive immediately at the shortest
https://karpathy.ai/lexicap/0011-large.html#00:28:45.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
explanation of the data but you're making
https://karpathy.ai/lexicap/0011-large.html#00:28:48.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
progress.
https://karpathy.ai/lexicap/0011-large.html#00:28:51.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Whenever you are making progress you have an
https://karpathy.ai/lexicap/0011-large.html#00:28:52.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
insight.
https://karpathy.ai/lexicap/0011-large.html#00:28:55.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
You see oh first I needed so many bits of
https://karpathy.ai/lexicap/0011-large.html#00:28:56.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
information to describe the data, to describe
https://karpathy.ai/lexicap/0011-large.html#00:28:59.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
my falling apples, my video of falling apples,
https://karpathy.ai/lexicap/0011-large.html#00:29:02.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
I need so many data, so many pixels have to be
https://karpathy.ai/lexicap/0011-large.html#00:29:04.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
stored.
https://karpathy.ai/lexicap/0011-large.html#00:29:07.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
But then suddenly I realize no there is a very
https://karpathy.ai/lexicap/0011-large.html#00:29:08.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
simple way of predicting the third frame in the
https://karpathy.ai/lexicap/0011-large.html#00:29:11.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
video from the first two.
https://karpathy.ai/lexicap/0011-large.html#00:29:14.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And maybe not every little detail can be
https://karpathy.ai/lexicap/0011-large.html#00:29:16.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
predicted but more or less most of these orange
https://karpathy.ai/lexicap/0011-large.html#00:29:19.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
blobs that are coming down they accelerate in
https://karpathy.ai/lexicap/0011-large.html#00:29:21.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
the same way which means that I can greatly
https://karpathy.ai/lexicap/0011-large.html#00:29:24.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
compress the video.
https://karpathy.ai/lexicap/0011-large.html#00:29:27.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And the amount of compression, progress, that
https://karpathy.ai/lexicap/0011-large.html#00:29:28.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
is the depth of the insight that you have at
https://karpathy.ai/lexicap/0011-large.html#00:29:33.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that moment.
https://karpathy.ai/lexicap/0011-large.html#00:29:36.040