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
That's the fun that you have, the scientific
https://karpathy.ai/lexicap/0011-large.html#00:29:37.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
fun, the fun in that discovery.
https://karpathy.ai/lexicap/0011-large.html#00:29:39.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And we can build artificial systems that do
https://karpathy.ai/lexicap/0011-large.html#00:29:42.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
the same thing.
https://karpathy.ai/lexicap/0011-large.html#00:29:45.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
They measure the depth of their insights as they
https://karpathy.ai/lexicap/0011-large.html#00:29:46.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
are looking at the data which is coming in
https://karpathy.ai/lexicap/0011-large.html#00:29:49.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
through their own experiments and we give
https://karpathy.ai/lexicap/0011-large.html#00:29:51.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
them a reward, an intrinsic reward in proportion
https://karpathy.ai/lexicap/0011-large.html#00:29:54.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
to this depth of insight.
https://karpathy.ai/lexicap/0011-large.html#00:29:58.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And since they are trying to maximize the
https://karpathy.ai/lexicap/0011-large.html#00:30:00.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
rewards they get they are suddenly motivated to
https://karpathy.ai/lexicap/0011-large.html#00:30:05.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
come up with new action sequences, with new
https://karpathy.ai/lexicap/0011-large.html#00:30:09.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
experiments that have the property that the data
https://karpathy.ai/lexicap/0011-large.html#00:30:13.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that is coming in as a consequence of these
https://karpathy.ai/lexicap/0011-large.html#00:30:16.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
experiments has the property that they can learn
https://karpathy.ai/lexicap/0011-large.html#00:30:19.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
something about, see a pattern in there which
https://karpathy.ai/lexicap/0011-large.html#00:30:23.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
they hadn't seen yet before.
https://karpathy.ai/lexicap/0011-large.html#00:30:25.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So there is an idea of power play that you
https://karpathy.ai/lexicap/0011-large.html#00:30:28.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
described, a training in general problem solver
https://karpathy.ai/lexicap/0011-large.html#00:30:31.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
in this kind of way of looking for the unsolved
https://karpathy.ai/lexicap/0011-large.html#00:30:34.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
problems.
https://karpathy.ai/lexicap/0011-large.html#00:30:36.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Yeah.
https://karpathy.ai/lexicap/0011-large.html#00:30:37.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Can you describe that idea a little further?
https://karpathy.ai/lexicap/0011-large.html#00:30:38.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
It's another very simple idea.
https://karpathy.ai/lexicap/0011-large.html#00:30:40.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So normally what you do in computer science,
https://karpathy.ai/lexicap/0011-large.html#00:30:42.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
you have some guy who gives you a problem and
https://karpathy.ai/lexicap/0011-large.html#00:30:45.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
then there is a huge search space of potential
https://karpathy.ai/lexicap/0011-large.html#00:30:50.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
solution candidates and you somehow try them
https://karpathy.ai/lexicap/0011-large.html#00:30:55.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
out and you have more less sophisticated ways
https://karpathy.ai/lexicap/0011-large.html#00:30:59.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
of moving around in that search space until
https://karpathy.ai/lexicap/0011-large.html#00:31:02.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
you finally found a solution which you
https://karpathy.ai/lexicap/0011-large.html#00:31:07.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
consider satisfactory.
https://karpathy.ai/lexicap/0011-large.html#00:31:10.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
That's what most of computer science is about.
https://karpathy.ai/lexicap/0011-large.html#00:31:12.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Power play just goes one little step further
https://karpathy.ai/lexicap/0011-large.html#00:31:15.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and says let's not only search for solutions
https://karpathy.ai/lexicap/0011-large.html#00:31:18.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
to a given problem but let's search to pairs of
https://karpathy.ai/lexicap/0011-large.html#00:31:23.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
problems and their solutions where the system
https://karpathy.ai/lexicap/0011-large.html#00:31:28.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
itself has the opportunity to phrase its own
https://karpathy.ai/lexicap/0011-large.html#00:31:31.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
problem.
https://karpathy.ai/lexicap/0011-large.html#00:31:35.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So we are looking suddenly at pairs of
https://karpathy.ai/lexicap/0011-large.html#00:31:37.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
problems and their solutions or modifications
https://karpathy.ai/lexicap/0011-large.html#00:31:40.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
of the problem solver that is supposed to
https://karpathy.ai/lexicap/0011-large.html#00:31:44.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
generate a solution to that new problem.
https://karpathy.ai/lexicap/0011-large.html#00:31:47.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And this additional degree of freedom allows
https://karpathy.ai/lexicap/0011-large.html#00:31:51.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
us to build career systems that are like
https://karpathy.ai/lexicap/0011-large.html#00:31:57.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
scientists in the sense that they not only
https://karpathy.ai/lexicap/0011-large.html#00:32:00.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
try to solve and try to find answers to
https://karpathy.ai/lexicap/0011-large.html#00:32:04.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
existing questions, no they are also free to
https://karpathy.ai/lexicap/0011-large.html#00:32:07.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
pose their own questions.
https://karpathy.ai/lexicap/0011-large.html#00:32:11.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So if you want to build an artificial scientist
https://karpathy.ai/lexicap/0011-large.html#00:32:13.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
you have to give it that freedom and power
https://karpathy.ai/lexicap/0011-large.html#00:32:15.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
play is exactly doing that.
https://karpathy.ai/lexicap/0011-large.html#00:32:17.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So that's a dimension of freedom that's
https://karpathy.ai/lexicap/0011-large.html#00:32:19.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
important to have but how hard do you think
https://karpathy.ai/lexicap/0011-large.html#00:32:22.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that, how multidimensional and difficult the
https://karpathy.ai/lexicap/0011-large.html#00:32:25.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
space of then coming up with your own questions
https://karpathy.ai/lexicap/0011-large.html#00:32:31.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
is.
https://karpathy.ai/lexicap/0011-large.html#00:32:34.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So it's one of the things that as human beings
https://karpathy.ai/lexicap/0011-large.html#00:32:35.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
we consider to be the thing that makes us
https://karpathy.ai/lexicap/0011-large.html#00:32:37.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
special, the intelligence that makes us special
https://karpathy.ai/lexicap/0011-large.html#00:32:40.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
is that brilliant insight that can create
https://karpathy.ai/lexicap/0011-large.html#00:32:42.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
something totally new.
https://karpathy.ai/lexicap/0011-large.html#00:32:46.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Yes.
https://karpathy.ai/lexicap/0011-large.html#00:32:48.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So now let's look at the extreme case, let's
https://karpathy.ai/lexicap/0011-large.html#00:32:49.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
look at the set of all possible problems that
https://karpathy.ai/lexicap/0011-large.html#00:32:52.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
you can formally describe which is infinite,
https://karpathy.ai/lexicap/0011-large.html#00:32:56.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
which should be the next problem that a scientist
https://karpathy.ai/lexicap/0011-large.html#00:33:00.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
or power play is going to solve.
https://karpathy.ai/lexicap/0011-large.html#00:33:05.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Well, it should be the easiest problem that
https://karpathy.ai/lexicap/0011-large.html#00:33:08.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
goes beyond what you already know.
https://karpathy.ai/lexicap/0011-large.html#00:33:14.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So it should be the simplest problem that the
https://karpathy.ai/lexicap/0011-large.html#00:33:17.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
current problem solver that you have which can
https://karpathy.ai/lexicap/0011-large.html#00:33:21.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
already solve 100 problems that he cannot solve
https://karpathy.ai/lexicap/0011-large.html#00:33:23.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
yet by just generalizing.
https://karpathy.ai/lexicap/0011-large.html#00:33:28.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So it has to be new, so it has to require a
https://karpathy.ai/lexicap/0011-large.html#00:33:30.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
modification of the problem solver such that the
https://karpathy.ai/lexicap/0011-large.html#00:33:33.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
new problem solver can solve this new thing but
https://karpathy.ai/lexicap/0011-large.html#00:33:36.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
the old problem solver cannot do it and in
https://karpathy.ai/lexicap/0011-large.html#00:33:39.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
addition to that we have to make sure that the
https://karpathy.ai/lexicap/0011-large.html#00:33:42.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
problem solver doesn't forget any of the
https://karpathy.ai/lexicap/0011-large.html#00:33:46.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
previous solutions.
https://karpathy.ai/lexicap/0011-large.html#00:33:49.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Right.
https://karpathy.ai/lexicap/0011-large.html#00:33:50.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And so by definition power play is now trying
https://karpathy.ai/lexicap/0011-large.html#00:33:51.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
always to search in this pair of, in the set of
https://karpathy.ai/lexicap/0011-large.html#00:33:54.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
pairs of problems and problems over modifications
https://karpathy.ai/lexicap/0011-large.html#00:33:58.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
for a combination that minimize the time to
https://karpathy.ai/lexicap/0011-large.html#00:34:02.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
achieve these criteria.
https://karpathy.ai/lexicap/0011-large.html#00:34:06.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So it's always trying to find the problem which
https://karpathy.ai/lexicap/0011-large.html#00:34:08.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
is easiest to add to the repertoire.
https://karpathy.ai/lexicap/0011-large.html#00:34:11.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So just like grad students and academics and
https://karpathy.ai/lexicap/0011-large.html#00:34:14.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
researchers can spend their whole career in a
https://karpathy.ai/lexicap/0011-large.html#00:34:18.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
local minima stuck trying to come up with
https://karpathy.ai/lexicap/0011-large.html#00:34:20.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
interesting questions but ultimately doing very
https://karpathy.ai/lexicap/0011-large.html#00:34:24.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
little.
https://karpathy.ai/lexicap/0011-large.html#00:34:26.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Do you think it's easy in this approach of
https://karpathy.ai/lexicap/0011-large.html#00:34:27.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
looking for the simplest unsolvable problem to
https://karpathy.ai/lexicap/0011-large.html#00:34:31.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
get stuck in a local minima?
https://karpathy.ai/lexicap/0011-large.html#00:34:33.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Is not never really discovering new, you know
https://karpathy.ai/lexicap/0011-large.html#00:34:35.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
really jumping outside of the 100 problems that
https://karpathy.ai/lexicap/0011-large.html#00:34:40.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
you've already solved in a genuine creative way?
https://karpathy.ai/lexicap/0011-large.html#00:34:42.040