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
So we have examples like that and it's only in the
https://karpathy.ai/lexicap/0011-large.html#00:56:32.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
beginning.
https://karpathy.ai/lexicap/0011-large.html#00:56:36.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
This is just the tip of the iceberg and I believe the
https://karpathy.ai/lexicap/0011-large.html#00:56:37.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
next wave of AI is going to be all about that.
https://karpathy.ai/lexicap/0011-large.html#00:56:40.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So at the moment, the current wave of AI is about
https://karpathy.ai/lexicap/0011-large.html#00:56:44.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
passive pattern observation and prediction and that's
https://karpathy.ai/lexicap/0011-large.html#00:56:48.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
what you have on your smartphone and what the major
https://karpathy.ai/lexicap/0011-large.html#00:56:53.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
companies on the Pacific Rim are using to sell you
https://karpathy.ai/lexicap/0011-large.html#00:56:56.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
ads to do marketing.
https://karpathy.ai/lexicap/0011-large.html#00:57:00.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
That's the current sort of profit in AI and that's
https://karpathy.ai/lexicap/0011-large.html#00:57:02.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
only one or two percent of the world economy.
https://karpathy.ai/lexicap/0011-large.html#00:57:05.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Which is big enough to make these companies pretty
https://karpathy.ai/lexicap/0011-large.html#00:57:08.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
much the most valuable companies in the world.
https://karpathy.ai/lexicap/0011-large.html#00:57:12.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
But there's a much, much bigger fraction of the
https://karpathy.ai/lexicap/0011-large.html#00:57:15.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
economy going to be affected by the next wave which
https://karpathy.ai/lexicap/0011-large.html#00:57:19.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
is really about machines that shape the data through
https://karpathy.ai/lexicap/0011-large.html#00:57:22.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
their own actions.
https://karpathy.ai/lexicap/0011-large.html#00:57:26.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Do you think simulation is ultimately the biggest
https://karpathy.ai/lexicap/0011-large.html#00:57:27.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
way that those methods will be successful in the next
https://karpathy.ai/lexicap/0011-large.html#00:57:31.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
10, 20 years?
https://karpathy.ai/lexicap/0011-large.html#00:57:35.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
We're not talking about 100 years from now.
https://karpathy.ai/lexicap/0011-large.html#00:57:36.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
We're talking about sort of the near term impact of
https://karpathy.ai/lexicap/0011-large.html#00:57:38.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
RL.
https://karpathy.ai/lexicap/0011-large.html#00:57:41.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Do you think really good simulation is required or
https://karpathy.ai/lexicap/0011-large.html#00:57:42.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
is there other techniques like imitation learning,
https://karpathy.ai/lexicap/0011-large.html#00:57:45.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
observing other humans operating in the real world?
https://karpathy.ai/lexicap/0011-large.html#00:57:48.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Where do you think the success will come from?
https://karpathy.ai/lexicap/0011-large.html#00:57:53.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So at the moment, we have a tendency of using physics
https://karpathy.ai/lexicap/0011-large.html#00:57:57.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
simulations to learn behavior from machines that
https://karpathy.ai/lexicap/0011-large.html#00:58:02.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
learn to solve problems that humans also do not know
https://karpathy.ai/lexicap/0011-large.html#00:58:07.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
how to solve.
https://karpathy.ai/lexicap/0011-large.html#00:58:11.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
However, this is not the future because the future is
https://karpathy.ai/lexicap/0011-large.html#00:58:12.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
in what little babies do.
https://karpathy.ai/lexicap/0011-large.html#00:58:16.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
They don't use a physics engine to simulate the
https://karpathy.ai/lexicap/0011-large.html#00:58:18.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
world.
https://karpathy.ai/lexicap/0011-large.html#00:58:21.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
No, they learn a predictive model of the world which
https://karpathy.ai/lexicap/0011-large.html#00:58:22.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
maybe sometimes is wrong in many ways but captures
https://karpathy.ai/lexicap/0011-large.html#00:58:26.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
all kinds of important abstract high level predictions
https://karpathy.ai/lexicap/0011-large.html#00:58:31.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
which are really important to be successful.
https://karpathy.ai/lexicap/0011-large.html#00:58:34.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And that's what was the future 30 years ago when we
https://karpathy.ai/lexicap/0011-large.html#00:58:37.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
started that type of research but it's still the future
https://karpathy.ai/lexicap/0011-large.html#00:58:42.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and now we know much better how to go there to move
https://karpathy.ai/lexicap/0011-large.html#00:58:45.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
forward and to really make working systems based on
https://karpathy.ai/lexicap/0011-large.html#00:58:49.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that where you have a learning model of the world,
https://karpathy.ai/lexicap/0011-large.html#00:58:54.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
a model of the world that learns to predict what's
https://karpathy.ai/lexicap/0011-large.html#00:58:57.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
going to happen if I do that and that.
https://karpathy.ai/lexicap/0011-large.html#00:58:59.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And then the controller uses that model to more
https://karpathy.ai/lexicap/0011-large.html#00:59:01.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
quickly learn successful action sequences.
https://karpathy.ai/lexicap/0011-large.html#00:59:07.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And then of course always this curiosity thing.
https://karpathy.ai/lexicap/0011-large.html#00:59:10.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
In the beginning, the model is stupid so the
https://karpathy.ai/lexicap/0011-large.html#00:59:13.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
controller should be motivated to come up with
https://karpathy.ai/lexicap/0011-large.html#00:59:15.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
experiments with action sequences that lead to data
https://karpathy.ai/lexicap/0011-large.html#00:59:18.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that improve the model.
https://karpathy.ai/lexicap/0011-large.html#00:59:21.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Do you think improving the model, constructing an
https://karpathy.ai/lexicap/0011-large.html#00:59:23.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
understanding of the world in this connection is
https://karpathy.ai/lexicap/0011-large.html#00:59:27.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
now the popular approaches that have been successful
https://karpathy.ai/lexicap/0011-large.html#00:59:30.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
are grounded in ideas of neural networks.
https://karpathy.ai/lexicap/0011-large.html#00:59:34.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
But in the 80s with expert systems, there's
https://karpathy.ai/lexicap/0011-large.html#00:59:38.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
symbolic AI approaches which to us humans are more
https://karpathy.ai/lexicap/0011-large.html#00:59:41.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
intuitive in the sense that it makes sense that you
https://karpathy.ai/lexicap/0011-large.html#00:59:45.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
build up knowledge in this knowledge representation.
https://karpathy.ai/lexicap/0011-large.html#00:59:49.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
What kind of lessons can we draw into our current
https://karpathy.ai/lexicap/0011-large.html#00:59:52.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
approaches from expert systems from symbolic AI?
https://karpathy.ai/lexicap/0011-large.html#00:59:54.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So I became aware of all of that in the 80s and
https://karpathy.ai/lexicap/0011-large.html#01:00:00.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
back then logic programming was a huge thing.
https://karpathy.ai/lexicap/0011-large.html#01:00:04.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Was it inspiring to you yourself?
https://karpathy.ai/lexicap/0011-large.html#01:00:08.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Did you find it compelling?
https://karpathy.ai/lexicap/0011-large.html#01:00:10.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Because a lot of your work was not so much in that
https://karpathy.ai/lexicap/0011-large.html#01:00:12.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
realm, right?
https://karpathy.ai/lexicap/0011-large.html#01:00:16.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
It was more in the learning systems.
https://karpathy.ai/lexicap/0011-large.html#01:00:17.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Yes and no, but we did all of that.
https://karpathy.ai/lexicap/0011-large.html#01:00:18.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So my first publication ever actually was 1987,
https://karpathy.ai/lexicap/0011-large.html#01:00:20.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
was the implementation of genetic algorithm of a
https://karpathy.ai/lexicap/0011-large.html#01:00:27.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
genetic programming system in Prolog.
https://karpathy.ai/lexicap/0011-large.html#01:00:31.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So Prolog, that's what you learn back then which is
https://karpathy.ai/lexicap/0011-large.html#01:00:34.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
a logic programming language and the Japanese,
https://karpathy.ai/lexicap/0011-large.html#01:00:38.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
they have this huge fifth generation AI project
https://karpathy.ai/lexicap/0011-large.html#01:00:41.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
which was mostly about logic programming back then.
https://karpathy.ai/lexicap/0011-large.html#01:00:45.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Although neural networks existed and were well
https://karpathy.ai/lexicap/0011-large.html#01:00:49.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
known back then and deep learning has existed since
https://karpathy.ai/lexicap/0011-large.html#01:00:52.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
1965, since this guy in the Ukraine,
https://karpathy.ai/lexicap/0011-large.html#01:00:56.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Iwakunenko, started it.
https://karpathy.ai/lexicap/0011-large.html#01:01:00.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
But the Japanese and many other people,
https://karpathy.ai/lexicap/0011-large.html#01:01:02.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
they focused really on this logic programming and I
https://karpathy.ai/lexicap/0011-large.html#01:01:05.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
was influenced to the extent that I said,
https://karpathy.ai/lexicap/0011-large.html#01:01:08.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
okay, let's take these biologically inspired
https://karpathy.ai/lexicap/0011-large.html#01:01:10.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
algorithms like evolution, programs,
https://karpathy.ai/lexicap/0011-large.html#01:01:13.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and implement that in the language which I know,
https://karpathy.ai/lexicap/0011-large.html#01:01:20.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
which was Prolog, for example, back then.
https://karpathy.ai/lexicap/0011-large.html#01:01:22.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And then in many ways this came back later because
https://karpathy.ai/lexicap/0011-large.html#01:01:25.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
the Gödel machine, for example,
https://karpathy.ai/lexicap/0011-large.html#01:01:29.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
has a proof searcher on board and without that it
https://karpathy.ai/lexicap/0011-large.html#01:01:31.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
would not be optimal.
https://karpathy.ai/lexicap/0011-large.html#01:01:34.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Well, Markus Futter's universal algorithm for
https://karpathy.ai/lexicap/0011-large.html#01:01:36.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
solving all well defined problems has a proof
https://karpathy.ai/lexicap/0011-large.html#01:01:38.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
searcher on board so that's very much logic programming.
https://karpathy.ai/lexicap/0011-large.html#01:01:41.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Without that it would not be asymptotically optimal.
https://karpathy.ai/lexicap/0011-large.html#01:01:46.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
But then on the other hand,
https://karpathy.ai/lexicap/0011-large.html#01:01:50.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
because we are very pragmatic guys also,
https://karpathy.ai/lexicap/0011-large.html#01:01:51.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
we focused on recurrent neural networks and
https://karpathy.ai/lexicap/0011-large.html#01:01:54.040