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
maybe I am just biased as a human but it seems to
https://karpathy.ai/lexicap/0011-large.html#00:40:33.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
be an essential element of human level intelligence.
https://karpathy.ai/lexicap/0011-large.html#00:40:35.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Is that what you are implying?
https://karpathy.ai/lexicap/0011-large.html#00:40:41.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
To a degree?
https://karpathy.ai/lexicap/0011-large.html#00:40:44.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
If you zoom back a little bit and you just look
https://karpathy.ai/lexicap/0011-large.html#00:40:46.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
at a general problem solving machine which is
https://karpathy.ai/lexicap/0011-large.html#00:40:49.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
trying to solve arbitrary problems then this
https://karpathy.ai/lexicap/0011-large.html#00:40:51.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
machine will figure out in the course of solving
https://karpathy.ai/lexicap/0011-large.html#00:40:54.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
problems that it is good to be curious.
https://karpathy.ai/lexicap/0011-large.html#00:40:57.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So all of what I said just now about this prewired
https://karpathy.ai/lexicap/0011-large.html#00:41:00.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
curiosity and this will to invent new problems
https://karpathy.ai/lexicap/0011-large.html#00:41:04.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
that the system doesn't know how to solve yet
https://karpathy.ai/lexicap/0011-large.html#00:41:07.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
should be just a byproduct of the general search.
https://karpathy.ai/lexicap/0011-large.html#00:41:11.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
However, apparently evolution has built it into
https://karpathy.ai/lexicap/0011-large.html#00:41:15.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
us because it turned out to be so successful,
https://karpathy.ai/lexicap/0011-large.html#00:41:20.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
a prewiring, a bias, a very successful exploratory
https://karpathy.ai/lexicap/0011-large.html#00:41:24.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
bias that we are born with.
https://karpathy.ai/lexicap/0011-large.html#00:41:29.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And you have also said that consciousness in the
https://karpathy.ai/lexicap/0011-large.html#00:41:34.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
same kind of way may be a byproduct of problem solving.
https://karpathy.ai/lexicap/0011-large.html#00:41:36.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Do you find this an interesting byproduct?
https://karpathy.ai/lexicap/0011-large.html#00:41:41.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Do you think it is a useful byproduct?
https://karpathy.ai/lexicap/0011-large.html#00:41:45.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
What are your thoughts on consciousness in general?
https://karpathy.ai/lexicap/0011-large.html#00:41:47.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Or is it simply a byproduct of greater and greater
https://karpathy.ai/lexicap/0011-large.html#00:41:49.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
capabilities of problem solving that is similar
https://karpathy.ai/lexicap/0011-large.html#00:41:53.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
to creativity in that sense?
https://karpathy.ai/lexicap/0011-large.html#00:41:58.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
We never have a procedure called consciousness
https://karpathy.ai/lexicap/0011-large.html#00:42:01.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
in our machines.
https://karpathy.ai/lexicap/0011-large.html#00:42:04.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
However, we get as side effects of what these
https://karpathy.ai/lexicap/0011-large.html#00:42:05.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
machines are doing things that seem to be closely
https://karpathy.ai/lexicap/0011-large.html#00:42:09.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
related to what people call consciousness.
https://karpathy.ai/lexicap/0011-large.html#00:42:13.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So for example, already in 1990 we had simple
https://karpathy.ai/lexicap/0011-large.html#00:42:16.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
systems which were basically recurrent networks
https://karpathy.ai/lexicap/0011-large.html#00:42:20.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
and therefore universal computers trying to map
https://karpathy.ai/lexicap/0011-large.html#00:42:24.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
incoming data into actions that lead to success.
https://karpathy.ai/lexicap/0011-large.html#00:42:28.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Maximizing reward in a given environment,
https://karpathy.ai/lexicap/0011-large.html#00:42:33.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
always finding the charging station in time
https://karpathy.ai/lexicap/0011-large.html#00:42:36.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
whenever the battery is low and negative signals
https://karpathy.ai/lexicap/0011-large.html#00:42:40.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
are coming from the battery, always find the
https://karpathy.ai/lexicap/0011-large.html#00:42:42.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
charging station in time without bumping against
https://karpathy.ai/lexicap/0011-large.html#00:42:45.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
painful obstacles on the way.
https://karpathy.ai/lexicap/0011-large.html#00:42:48.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So complicated things but very easily motivated.
https://karpathy.ai/lexicap/0011-large.html#00:42:50.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And then we give these little guys a separate
https://karpathy.ai/lexicap/0011-large.html#00:42:54.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
recurrent neural network which is just predicting
https://karpathy.ai/lexicap/0011-large.html#00:43:00.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
what's happening if I do that and that.
https://karpathy.ai/lexicap/0011-large.html#00:43:02.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
What will happen as a consequence of these
https://karpathy.ai/lexicap/0011-large.html#00:43:04.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
actions that I'm executing.
https://karpathy.ai/lexicap/0011-large.html#00:43:07.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And it's just trained on the long and long history
https://karpathy.ai/lexicap/0011-large.html#00:43:09.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
of interactions with the world.
https://karpathy.ai/lexicap/0011-large.html#00:43:11.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So it becomes a predictive model of the world
https://karpathy.ai/lexicap/0011-large.html#00:43:13.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
basically.
https://karpathy.ai/lexicap/0011-large.html#00:43:16.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And therefore also a compressor of the observations
https://karpathy.ai/lexicap/0011-large.html#00:43:18.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
of the world because whatever you can predict
https://karpathy.ai/lexicap/0011-large.html#00:43:22.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
you don't have to store extra.
https://karpathy.ai/lexicap/0011-large.html#00:43:25.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So compression is a side effect of prediction.
https://karpathy.ai/lexicap/0011-large.html#00:43:27.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And how does this recurrent network compress?
https://karpathy.ai/lexicap/0011-large.html#00:43:30.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Well, it's inventing little subprograms, little
https://karpathy.ai/lexicap/0011-large.html#00:43:33.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
subnetworks that stand for everything that
https://karpathy.ai/lexicap/0011-large.html#00:43:36.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
frequently appears in the environment like
https://karpathy.ai/lexicap/0011-large.html#00:43:39.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
bottles and microphones and faces, maybe lots of
https://karpathy.ai/lexicap/0011-large.html#00:43:42.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
faces in my environment so I'm learning to create
https://karpathy.ai/lexicap/0011-large.html#00:43:45.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
something like a prototype face and a new face
https://karpathy.ai/lexicap/0011-large.html#00:43:50.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
comes along and all I have to encode are the
https://karpathy.ai/lexicap/0011-large.html#00:43:52.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
deviations from the prototype.
https://karpathy.ai/lexicap/0011-large.html#00:43:54.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So it's compressing all the time the stuff that
https://karpathy.ai/lexicap/0011-large.html#00:43:56.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
frequently appears.
https://karpathy.ai/lexicap/0011-large.html#00:43:58.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
There's one thing that appears all the time that
https://karpathy.ai/lexicap/0011-large.html#00:44:00.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
is present all the time when the agent is
https://karpathy.ai/lexicap/0011-large.html#00:44:05.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
interacting with its environment which is the
https://karpathy.ai/lexicap/0011-large.html#00:44:07.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
agent itself.
https://karpathy.ai/lexicap/0011-large.html#00:44:10.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
But just for data compression reasons it is
https://karpathy.ai/lexicap/0011-large.html#00:44:12.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
extremely natural for this recurrent network to
https://karpathy.ai/lexicap/0011-large.html#00:44:15.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
come up with little subnetworks that stand for
https://karpathy.ai/lexicap/0011-large.html#00:44:18.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
the properties of the agents, the hand, the other
https://karpathy.ai/lexicap/0011-large.html#00:44:21.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
actuators and all the stuff that you need to
https://karpathy.ai/lexicap/0011-large.html#00:44:26.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
better encode the data which is influenced by
https://karpathy.ai/lexicap/0011-large.html#00:44:29.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
the actions of the agent.
https://karpathy.ai/lexicap/0011-large.html#00:44:32.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So there just as a side effect of data compression
https://karpathy.ai/lexicap/0011-large.html#00:44:34.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
during problem solving you have internal self
https://karpathy.ai/lexicap/0011-large.html#00:44:39.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
models.
https://karpathy.ai/lexicap/0011-large.html#00:44:43.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Now you can use this model of the world to plan
https://karpathy.ai/lexicap/0011-large.html#00:44:45.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
your future and that's what we also have done
https://karpathy.ai/lexicap/0011-large.html#00:44:50.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
since 1990.
https://karpathy.ai/lexicap/0011-large.html#00:44:53.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
So the recurrent network which is the controller
https://karpathy.ai/lexicap/0011-large.html#00:44:54.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
which is trying to maximize reward can use this
https://karpathy.ai/lexicap/0011-large.html#00:44:57.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
model of the network of the world, this model
https://karpathy.ai/lexicap/0011-large.html#00:45:00.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
network of the world, this predictive model of
https://karpathy.ai/lexicap/0011-large.html#00:45:03.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
the world to plan ahead and say let's not do this
https://karpathy.ai/lexicap/0011-large.html#00:45:05.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
action sequence, let's do this action sequence
https://karpathy.ai/lexicap/0011-large.html#00:45:08.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
instead because it leads to more predicted
https://karpathy.ai/lexicap/0011-large.html#00:45:10.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
reward.
https://karpathy.ai/lexicap/0011-large.html#00:45:13.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
And whenever it is waking up these little
https://karpathy.ai/lexicap/0011-large.html#00:45:14.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
subnetworks that stand for itself then it is
https://karpathy.ai/lexicap/0011-large.html#00:45:17.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
thinking about itself and it is thinking about
https://karpathy.ai/lexicap/0011-large.html#00:45:20.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
itself and it is exploring mentally the
https://karpathy.ai/lexicap/0011-large.html#00:45:23.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
consequences of its own actions and now you tell
https://karpathy.ai/lexicap/0011-large.html#00:45:28.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
me what is still missing.
https://karpathy.ai/lexicap/0011-large.html#00:45:34.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Missing the next, the gap to consciousness.
https://karpathy.ai/lexicap/0011-large.html#00:45:36.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
There isn't.
https://karpathy.ai/lexicap/0011-large.html#00:45:40.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
That's a really beautiful idea that if life is
https://karpathy.ai/lexicap/0011-large.html#00:45:41.040
Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
a collection of data and life is a process of
https://karpathy.ai/lexicap/0011-large.html#00:45:45.040