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
which aren't provably optimal at all,
https://karpathy.ai/lexicap/0011-large.html#00:08:18.680
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
like the other stuff that we did,
https://karpathy.ai/lexicap/0011-large.html#00:08:21.280
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
but which are much more practical
https://karpathy.ai/lexicap/0011-large.html#00:08:22.840
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
as long as we only want to solve the small problems
https://karpathy.ai/lexicap/0011-large.html#00:08:25.400
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that we are typically trying to solve
https://karpathy.ai/lexicap/0011-large.html#00:08:28.760
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
in this environment here.
https://karpathy.ai/lexicap/0011-large.html#00:08:33.320
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So the universal problem solvers like the Gödel machine,
https://karpathy.ai/lexicap/0011-large.html#00:08:35.600
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
but also Markus Hutter's fastest way
https://karpathy.ai/lexicap/0011-large.html#00:08:38.920
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
of solving all possible problems,
https://karpathy.ai/lexicap/0011-large.html#00:08:42.080
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
which he developed around 2002 in my lab,
https://karpathy.ai/lexicap/0011-large.html#00:08:44.360
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
they are associated with these constant overheads
https://karpathy.ai/lexicap/0011-large.html#00:08:49.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
for proof search, which guarantees that the thing
https://karpathy.ai/lexicap/0011-large.html#00:08:52.520
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that you're doing is optimal.
https://karpathy.ai/lexicap/0011-large.html#00:08:55.160
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
For example, there is this fastest way
https://karpathy.ai/lexicap/0011-large.html#00:08:57.480
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
of solving all problems with a computable solution,
https://karpathy.ai/lexicap/0011-large.html#00:09:01.160
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
which is due to Markus, Markus Hutter,
https://karpathy.ai/lexicap/0011-large.html#00:09:05.280
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and to explain what's going on there,
https://karpathy.ai/lexicap/0011-large.html#00:09:08.320
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
let's take traveling salesman problems.
https://karpathy.ai/lexicap/0011-large.html#00:09:12.240
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
With traveling salesman problems,
https://karpathy.ai/lexicap/0011-large.html#00:09:15.720
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
you have a number of cities and cities
https://karpathy.ai/lexicap/0011-large.html#00:09:17.360
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and you try to find the shortest path
https://karpathy.ai/lexicap/0011-large.html#00:09:21.320
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
through all these cities without visiting any city twice.
https://karpathy.ai/lexicap/0011-large.html#00:09:23.680
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And nobody knows the fastest way
https://karpathy.ai/lexicap/0011-large.html#00:09:29.440
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
of solving traveling salesman problems, TSPs,
https://karpathy.ai/lexicap/0011-large.html#00:09:32.240
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
but let's assume there is a method of solving them
https://karpathy.ai/lexicap/0011-large.html#00:09:38.720
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
within N to the five operations
https://karpathy.ai/lexicap/0011-large.html#00:09:41.720
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
where N is the number of cities.
https://karpathy.ai/lexicap/0011-large.html#00:09:45.840
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Then the universal method of Markus
https://karpathy.ai/lexicap/0011-large.html#00:09:48.520
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
is going to solve the same traveling salesman problem
https://karpathy.ai/lexicap/0011-large.html#00:09:53.000
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
also within N to the five steps,
https://karpathy.ai/lexicap/0011-large.html#00:09:57.000
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
plus O of one, plus a constant number of steps
https://karpathy.ai/lexicap/0011-large.html#00:10:00.480
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that you need for the proof searcher,
https://karpathy.ai/lexicap/0011-large.html#00:10:04.760
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
which you need to show that this particular class
https://karpathy.ai/lexicap/0011-large.html#00:10:07.600
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
of problems, the traveling salesman problems,
https://karpathy.ai/lexicap/0011-large.html#00:10:12.600
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
can be solved within a certain time frame,
https://karpathy.ai/lexicap/0011-large.html#00:10:15.680
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
solved within a certain time bound,
https://karpathy.ai/lexicap/0011-large.html#00:10:17.760
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
within order N to the five steps, basically,
https://karpathy.ai/lexicap/0011-large.html#00:10:20.680
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and this additive constant doesn't care for N,
https://karpathy.ai/lexicap/0011-large.html#00:10:24.560
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
which means as N is getting larger and larger,
https://karpathy.ai/lexicap/0011-large.html#00:10:28.720
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
as you have more and more cities,
https://karpathy.ai/lexicap/0011-large.html#00:10:32.600
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
the constant overhead pales in comparison,
https://karpathy.ai/lexicap/0011-large.html#00:10:35.080
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and that means that almost all large problems are solved
https://karpathy.ai/lexicap/0011-large.html#00:10:38.800
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
in the best possible way.
https://karpathy.ai/lexicap/0011-large.html#00:10:44.400
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Today, we already have a universal problem solver like that.
https://karpathy.ai/lexicap/0011-large.html#00:10:45.880
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
However, it's not practical because the overhead,
https://karpathy.ai/lexicap/0011-large.html#00:10:50.520
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
the constant overhead is so large
https://karpathy.ai/lexicap/0011-large.html#00:10:54.560
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that for the small kinds of problems
https://karpathy.ai/lexicap/0011-large.html#00:10:57.480
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that we want to solve in this little biosphere.
https://karpathy.ai/lexicap/0011-large.html#00:11:00.240
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
By the way, when you say small,
https://karpathy.ai/lexicap/0011-large.html#00:11:04.600
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
you're talking about things that fall
https://karpathy.ai/lexicap/0011-large.html#00:11:06.400
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
within the constraints of our computational systems.
https://karpathy.ai/lexicap/0011-large.html#00:11:08.640
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So they can seem quite large to us mere humans, right?
https://karpathy.ai/lexicap/0011-large.html#00:11:10.880
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
That's right, yeah.
https://karpathy.ai/lexicap/0011-large.html#00:11:14.440
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So they seem large and even unsolvable
https://karpathy.ai/lexicap/0011-large.html#00:11:15.360
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
in a practical sense today,
https://karpathy.ai/lexicap/0011-large.html#00:11:19.000
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
but they are still small compared to almost all problems
https://karpathy.ai/lexicap/0011-large.html#00:11:21.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
because almost all problems are large problems,
https://karpathy.ai/lexicap/0011-large.html#00:11:24.760
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
which are much larger than any constant.
https://karpathy.ai/lexicap/0011-large.html#00:11:28.480
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Do you find it useful as a person
https://karpathy.ai/lexicap/0011-large.html#00:11:31.920
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
who has dreamed of creating a general learning system,
https://karpathy.ai/lexicap/0011-large.html#00:11:34.520
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
has worked on creating one,
https://karpathy.ai/lexicap/0011-large.html#00:11:38.600
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
has done a lot of interesting ideas there,
https://karpathy.ai/lexicap/0011-large.html#00:11:39.840
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
to think about P versus NP,
https://karpathy.ai/lexicap/0011-large.html#00:11:42.120
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
this formalization of how hard problems are,
https://karpathy.ai/lexicap/0011-large.html#00:11:46.320
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
how they scale,
https://karpathy.ai/lexicap/0011-large.html#00:11:50.760
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
this kind of worst case analysis type of thinking,
https://karpathy.ai/lexicap/0011-large.html#00:11:52.360
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
do you find that useful?
https://karpathy.ai/lexicap/0011-large.html#00:11:55.160
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Or is it only just a mathematical,
https://karpathy.ai/lexicap/0011-large.html#00:11:56.800
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
it's a set of mathematical techniques
https://karpathy.ai/lexicap/0011-large.html#00:12:00.520
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
to give you intuition about what's good and bad.
https://karpathy.ai/lexicap/0011-large.html#00:12:02.600
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So P versus NP, that's super interesting
https://karpathy.ai/lexicap/0011-large.html#00:12:05.720
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
from a theoretical point of view.
https://karpathy.ai/lexicap/0011-large.html#00:12:09.440
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And in fact, as you are thinking about that problem,
https://karpathy.ai/lexicap/0011-large.html#00:12:11.760
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
you can also get inspiration
https://karpathy.ai/lexicap/0011-large.html#00:12:14.560
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
for better practical problem solvers.
https://karpathy.ai/lexicap/0011-large.html#00:12:17.280
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
On the other hand, we have to admit
https://karpathy.ai/lexicap/0011-large.html#00:12:21.280
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that at the moment, the best practical problem solvers
https://karpathy.ai/lexicap/0011-large.html#00:12:23.320
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
for all kinds of problems that we are now solving
https://karpathy.ai/lexicap/0011-large.html#00:12:28.360
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
through what is called AI at the moment,
https://karpathy.ai/lexicap/0011-large.html#00:12:31.080
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
they are not of the kind
https://karpathy.ai/lexicap/0011-large.html#00:12:33.840
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that is inspired by these questions.
https://karpathy.ai/lexicap/0011-large.html#00:12:36.240
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
There we are using general purpose computers
https://karpathy.ai/lexicap/0011-large.html#00:12:38.760
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
such as recurrent neural networks,
https://karpathy.ai/lexicap/0011-large.html#00:12:42.680
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
but we have a search technique
https://karpathy.ai/lexicap/0011-large.html#00:12:44.800
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
which is just local search gradient descent
https://karpathy.ai/lexicap/0011-large.html#00:12:46.680
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
to try to find a program
https://karpathy.ai/lexicap/0011-large.html#00:12:50.280
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that is running on these recurrent networks,
https://karpathy.ai/lexicap/0011-large.html#00:12:51.920
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
such that it can solve some interesting problems
https://karpathy.ai/lexicap/0011-large.html#00:12:54.400
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
such as speech recognition or machine translation
https://karpathy.ai/lexicap/0011-large.html#00:12:58.200
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and something like that.
https://karpathy.ai/lexicap/0011-large.html#00:13:01.880
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And there is very little theory behind the best solutions
https://karpathy.ai/lexicap/0011-large.html#00:13:03.120
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that we have at the moment that can do that.
https://karpathy.ai/lexicap/0011-large.html#00:13:08.120
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Do you think that needs to change?
https://karpathy.ai/lexicap/0011-large.html#00:13:10.840
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Do you think that will change?
https://karpathy.ai/lexicap/0011-large.html#00:13:12.680
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Or can we go, can we create a general intelligent systems
https://karpathy.ai/lexicap/0011-large.html#00:13:14.080
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
without ever really proving that that system is intelligent
https://karpathy.ai/lexicap/0011-large.html#00:13:17.160
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
in some kind of mathematical way,
https://karpathy.ai/lexicap/0011-large.html#00:13:20.640
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
solving machine translation perfectly
https://karpathy.ai/lexicap/0011-large.html#00:13:22.600
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
or something like that,
https://karpathy.ai/lexicap/0011-large.html#00:13:25.000
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
within some kind of syntactic definition of a language,
https://karpathy.ai/lexicap/0011-large.html#00:13:26.320