Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs
AI 与机器学习心理与人性技术与编程物理与宇宙学音乐与艺术
📋 章节目录
暂无章节信息
🔑 关键词
problemsuniverselearningdatasolvegoingmodelsystemsdonnetworksgeneralnetworksimplehumansintelligencebettersolvingfuturecreativitymachine
💬 精彩语录
"they are just invisible because they are really efficient in using the energies of their own local systems"
它们只是看不见，因为它们确实有效地利用了自己本地系统的能量
— Juergen Schmidhuber （1:18:17.040）
"why are there still any visible stars left in our own galaxy, which also must have a lot of dark matter?"
为什么我们的银河系中还残留着可见的恒星，而且它也一定含有大量的暗物质？
— Juergen Schmidhuber （1:18:34.040）
"at least in our local light cone within the few hundreds of millions of light years that we can reliably observe."
至少在我们可以可靠观测到的几亿光年范围内的本地光锥中。
— Juergen Schmidhuber （1:18:54.040）
🎙️ 完整对话（1389 条）
Lex Fridman (00:00.000)
The following is a conversation with Jürgen Schmidhuber.
以下是与 Jürgen Schmidhuber 的对话。
Lex Fridman (00:03.520)
He's the co director of the CS Swiss AI Lab
他是 CS Swiss 人工智能实验室的联合主任
Lex Fridman (00:06.320)
and a co creator of long short term memory networks.
以及长期短期记忆网络的共同创造者。
Lex Fridman (00:10.360)
LSDMs are used in billions of devices today
如今，LSDM 用于数十亿台设备
Lex Fridman (00:13.720)
for speech recognition, translation, and much more.
用于语音识别、翻译等等。
Juergen Schmidhuber (00:17.400)
Over 30 years, he has proposed a lot of interesting
30多年来，他提出了很多有趣的建议
Lex Fridman (00:20.800)
out of the box ideas on meta learning, adversarial networks,
关于元学习、对抗网络的开箱即用的想法，
Juergen Schmidhuber (00:24.800)
computer vision, and even a formal theory of quote,
计算机视觉，甚至是正式的引用理论，
Lex Fridman (00:28.720)
creativity, curiosity, and fun.
创造力、好奇心和乐趣。
Juergen Schmidhuber (00:32.360)
This conversation is part of the MIT course
这段对话是麻省理工学院课程的一部分
Lex Fridman (00:34.920)
on artificial general intelligence
论通用人工智能
Lex Fridman (00:36.560)
and the artificial intelligence podcast.
和人工智能播客。
Lex Fridman (00:38.840)
If you enjoy it, subscribe on YouTube, iTunes,
如果您喜欢，请在 YouTube、iTunes、
Juergen Schmidhuber (00:41.960)
or simply connect with me on Twitter
或者直接在 Twitter 上与我联系
Lex Fridman (00:43.960)
at Lex Friedman spelled F R I D.
Lex Friedman 拼写为 F R I D。
Lex Fridman (00:47.280)
And now here's my conversation with Jürgen Schmidhuber.
现在这是我与 Jürgen Schmidhuber 的对话。
Lex Fridman (00:53.080)
Early on you dreamed of AI systems
你很早就梦想着人工智能系统
Juergen Schmidhuber (00:55.640)
that self improve recursively.
自我递归地改进。
Lex Fridman (00:58.680)
When was that dream born?
这个梦想是什么时候诞生的？
Juergen Schmidhuber (01:01.800)
When I was a baby.
当我还是个婴儿的时候。
Lex Fridman (01:03.160)
No, that's not true.
Juergen Schmidhuber (01:04.240)
When I was a teenager.
Lex Fridman (01:06.520)
And what was the catalyst for that birth?
Lex Fridman (01:09.680)
What was the thing that first inspired you?
Lex Fridman (01:12.920)
When I was a boy, I was thinking about what to do in my life
Lex Fridman (01:20.200)
and then I thought the most exciting thing
Lex Fridman (01:23.880)
is to solve the riddles of the universe.
Lex Fridman (01:28.200)
And that means you have to become a physicist.
Lex Fridman (01:30.920)
However, then I realized that there's something even grander.
Juergen Schmidhuber (01:35.840)
You can try to build a machine
Lex Fridman (01:39.880)
that isn't really a machine any longer
Juergen Schmidhuber (01:42.120)
that learns to become a much better physicist
Lex Fridman (01:44.520)
than I could ever hope to be.
Lex Fridman (01:47.080)
And that's how I thought maybe I can multiply
Lex Fridman (01:50.320)
my tiny little bit of creativity into infinity.
Lex Fridman (01:54.520)
But ultimately that creativity will be multiplied
Lex Fridman (01:57.280)
to understand the universe around us.
Juergen Schmidhuber (01:59.280)
That's the curiosity for that mystery that drove you.
Lex Fridman (02:05.800)
Yes, so if you can build a machine
Juergen Schmidhuber (02:08.440)
that learns to solve more and more complex problems
Lex Fridman (02:13.880)
and more and more general problem solver
Juergen Schmidhuber (02:16.880)
then you basically have solved all the problems,
Lex Fridman (02:22.680)
at least all the solvable problems.
Lex Fridman (02:26.080)
So how do you think, what is the mechanism
Lex Fridman (02:28.120)
for that kind of general solver look like?
Juergen Schmidhuber (02:31.640)
Obviously we don't quite yet have one
Lex Fridman (02:34.840)
or know how to build one but we have ideas
Lex Fridman (02:37.040)
and you have had throughout your career
Lex Fridman (02:39.120)
several ideas about it.
Lex Fridman (02:40.800)
So how do you think about that mechanism?
Lex Fridman (02:43.640)
So in the 80s, I thought about how to build this machine
Juergen Schmidhuber (02:48.640)
that learns to solve all these problems
Lex Fridman (02:51.040)
that I cannot solve myself.
Lex Fridman (02:54.120)
And I thought it is clear it has to be a machine
Lex Fridman (02:57.160)
that not only learns to solve this problem here
Lex Fridman (03:00.880)
and this problem here but it also has to learn
Lex Fridman (03:04.160)
to improve the learning algorithm itself.
Lex Fridman (03:09.360)
So it has to have the learning algorithm
Lex Fridman (03:12.480)
in a representation that allows it to inspect it
Lex Fridman (03:15.720)
and modify it such that it can come up
Lex Fridman (03:19.240)
with a better learning algorithm.
Lex Fridman (03:21.080)
So I call that meta learning, learning to learn
Lex Fridman (03:24.600)
and recursive self improvement
Juergen Schmidhuber (03:26.760)
that is really the pinnacle of that
Lex Fridman (03:28.760)
where you then not only learn how to improve
Juergen Schmidhuber (03:34.800)
on that problem and on that
Lex Fridman (03:36.440)
but you also improve the way the machine improves
Lex Fridman (03:40.000)
and you also improve the way it improves
Lex Fridman (03:42.160)
the way it improves itself.
Lex Fridman (03:44.600)
And that was my 1987 diploma thesis
Lex Fridman (03:47.440)
which was all about that higher education
Juergen Schmidhuber (03:50.920)
hierarchy of meta learners that have no computational limits
Lex Fridman (03:57.240)
except for the well known limits that Gödel identified
Juergen Schmidhuber (04:01.640)
in 1931 and for the limits of physics.
Lex Fridman (04:06.480)
In the recent years, meta learning has gained popularity
Juergen Schmidhuber (04:10.040)
in a specific kind of form.
Lex Fridman (04:12.760)
You've talked about how that's not really meta learning
Juergen Schmidhuber (04:16.000)
with neural networks, that's more basic transfer learning.
Lex Fridman (04:21.480)
Can you talk about the difference
Juergen Schmidhuber (04:22.720)
between the big general meta learning
Lex Fridman (04:25.460)
and a more narrow sense of meta learning
Lex Fridman (04:27.960)
the way it's used today, the way it's talked about today?
Lex Fridman (04:30.880)
Let's take the example of a deep neural network
Juergen Schmidhuber (04:33.440)
that has learned to classify images
Lex Fridman (04:37.240)
and maybe you have trained that network
Juergen Schmidhuber (04:40.060)
on 100 different databases of images.
Lex Fridman (04:43.840)
And now a new database comes along
Lex Fridman (04:48.080)
and you want to quickly learn the new thing as well.
Lex Fridman (04:53.400)
So one simple way of doing that is you take the network
Juergen Schmidhuber (04:57.720)
which already knows 100 types of databases
Lex Fridman (05:02.440)
and then you just take the top layer of that
Lex Fridman (05:06.320)
and you retrain that using the new label data
Lex Fridman (05:11.320)
that you have in the new image database.
Lex Fridman (05:14.760)
And then it turns out that it really, really quickly
Lex Fridman (05:17.360)
can learn that too, one shot basically
Juergen Schmidhuber (05:20.600)
because from the first 100 data sets,
Lex Fridman (05:24.320)
it already has learned so much about computer vision
Juergen Schmidhuber (05:27.560)
that it can reuse that and that is then almost good enough
Lex Fridman (05:31.880)
to solve the new task except you need a little bit
Juergen Schmidhuber (05:34.240)
of adjustment on the top.
Lex Fridman (05:38.400)
So that is transfer learning.
Lex Fridman (05:41.280)
And it has been done in principle for many decades.
Lex Fridman (05:44.520)
People have done similar things for decades.
Juergen Schmidhuber (05:48.520)
Meta learning too, meta learning is about
Lex Fridman (05:51.080)
having the learning algorithm itself
Juergen Schmidhuber (05:55.760)
open to introspection by the system that is using it
Lex Fridman (06:01.560)
and also open to modification such that the learning system
Juergen Schmidhuber (06:06.320)
has an opportunity to modify
Lex Fridman (06:09.680)
any part of the learning algorithm
Lex Fridman (06:12.040)
and then evaluate the consequences of that modification
Lex Fridman (06:16.840)
and then learn from that to create
Juergen Schmidhuber (06:21.000)
a better learning algorithm and so on recursively.
Lex Fridman (06:25.680)
So that's a very different animal
Juergen Schmidhuber (06:28.480)
where you are opening the space of possible learning
Lex Fridman (06:32.440)
algorithms to the learning system itself.
Juergen Schmidhuber (06:35.480)
Right, so you've, like in the 2004 paper, you described
Lex Fridman (06:40.160)
gator machines, programs that rewrite themselves, right?
Juergen Schmidhuber (06:44.480)
Philosophically and even in your paper, mathematically,
Lex Fridman (06:47.480)
these are really compelling ideas but practically,
Lex Fridman (06:52.280)
do you see these self referential programs
Lex Fridman (06:55.280)
being successful in the near term to having an impact
Juergen Schmidhuber (06:59.360)
where sort of it demonstrates to the world
Lex Fridman (07:02.960)
that this direction is a good one to pursue
Lex Fridman (07:07.400)
in the near term?
Lex Fridman (07:08.640)
Yes, we had these two different types
Juergen Schmidhuber (07:11.320)
of fundamental research,
Lex Fridman (07:13.440)
how to build a universal problem solver,
Juergen Schmidhuber (07:15.800)
one basically exploiting proof search
Lex Fridman (07:22.960)
and things like that that you need to come up with
Juergen Schmidhuber (07:25.520)
asymptotically optimal, theoretically optimal
Lex Fridman (07:30.280)
self improvers and problem solvers.
Juergen Schmidhuber (07:34.160)
However, one has to admit that through this proof search
Lex Fridman (07:40.640)
comes in an additive constant, an overhead,
Juergen Schmidhuber (07:44.480)
an additive overhead that vanishes in comparison
Lex Fridman (07:50.760)
to what you have to do to solve large problems.
Juergen Schmidhuber (07:55.160)
However, for many of the small problems
Lex Fridman (07:58.000)
that we want to solve in our everyday life,
Juergen Schmidhuber (08:00.880)
we cannot ignore this constant overhead
Lex Fridman (08:03.280)
and that's why we also have been doing other things,
Juergen Schmidhuber (08:08.120)
non universal things such as recurrent neural networks
Lex Fridman (08:12.160)
which are trained by gradient descent
Lex Fridman (08:15.400)
and local search techniques which aren't universal at all,
Lex Fridman (08:18.680)
which aren't provably optimal at all,
Juergen Schmidhuber (08:21.280)
like the other stuff that we did,
Lex Fridman (08:22.840)
but which are much more practical
Juergen Schmidhuber (08:25.400)
as long as we only want to solve the small problems
Lex Fridman (08:28.760)
that we are typically trying to solve
Juergen Schmidhuber (08:33.320)
in this environment here.
Lex Fridman (08:35.600)
So the universal problem solvers like the Gödel machine,
Lex Fridman (08:38.920)
but also Markus Hutter's fastest way
Lex Fridman (08:42.080)
of solving all possible problems,
Juergen Schmidhuber (08:44.360)
which he developed around 2002 in my lab,
Lex Fridman (08:49.040)
they are associated with these constant overheads
Juergen Schmidhuber (08:52.520)
for proof search, which guarantees that the thing
Lex Fridman (08:55.160)
that you're doing is optimal.
Juergen Schmidhuber (08:57.480)
For example, there is this fastest way
Lex Fridman (09:01.160)
of solving all problems with a computable solution,
Juergen Schmidhuber (09:05.280)
which is due to Markus, Markus Hutter,
Lex Fridman (09:08.320)
and to explain what's going on there,
Juergen Schmidhuber (09:12.240)
let's take traveling salesman problems.
Lex Fridman (09:15.720)
With traveling salesman problems,
Juergen Schmidhuber (09:17.360)
you have a number of cities and cities
Lex Fridman (09:21.320)
and you try to find the shortest path
Juergen Schmidhuber (09:23.680)
through all these cities without visiting any city twice.
Lex Fridman (09:29.440)
And nobody knows the fastest way
Juergen Schmidhuber (09:32.240)
of solving traveling salesman problems, TSPs,
Lex Fridman (09:38.720)
but let's assume there is a method of solving them
Juergen Schmidhuber (09:41.720)
within N to the five operations
Lex Fridman (09:45.840)
where N is the number of cities.
Juergen Schmidhuber (09:48.520)
Then the universal method of Markus
Lex Fridman (09:53.000)
is going to solve the same traveling salesman problem
Juergen Schmidhuber (09:57.000)
also within N to the five steps,
Lex Fridman (10:00.480)
plus O of one, plus a constant number of steps
Juergen Schmidhuber (10:04.760)
that you need for the proof searcher,
Lex Fridman (10:07.600)
which you need to show that this particular class
Juergen Schmidhuber (10:12.600)
of problems, the traveling salesman problems,
Lex Fridman (10:15.680)
can be solved within a certain time frame,
Juergen Schmidhuber (10:17.760)
solved within a certain time bound,
Lex Fridman (10:20.680)
within order N to the five steps, basically,
Lex Fridman (10:24.560)
and this additive constant doesn't care for N,
Lex Fridman (10:28.720)
which means as N is getting larger and larger,
Juergen Schmidhuber (10:32.600)
as you have more and more cities,
Lex Fridman (10:35.080)
the constant overhead pales in comparison,
Lex Fridman (10:38.800)
and that means that almost all large problems are solved
Lex Fridman (10:44.400)
in the best possible way.
Juergen Schmidhuber (10:45.880)
Today, we already have a universal problem solver like that.
Lex Fridman (10:50.520)
However, it's not practical because the overhead,
Juergen Schmidhuber (10:54.560)
the constant overhead is so large
Lex Fridman (10:57.480)
that for the small kinds of problems
Juergen Schmidhuber (11:00.240)
that we want to solve in this little biosphere.
Lex Fridman (11:04.600)
By the way, when you say small,
Juergen Schmidhuber (11:06.400)
you're talking about things that fall
Lex Fridman (11:08.640)
within the constraints of our computational systems.
Lex Fridman (11:10.880)
So they can seem quite large to us mere humans, right?
Lex Fridman (11:14.440)
That's right, yeah.
Lex Fridman (11:15.360)
So they seem large and even unsolvable
Lex Fridman (11:19.000)
in a practical sense today,
Lex Fridman (11:21.040)
but they are still small compared to almost all problems
Lex Fridman (11:24.760)
because almost all problems are large problems,
Juergen Schmidhuber (11:28.480)
which are much larger than any constant.
Lex Fridman (11:31.920)
Do you find it useful as a person
Juergen Schmidhuber (11:34.520)
who has dreamed of creating a general learning system,
Lex Fridman (11:38.600)
has worked on creating one,
Juergen Schmidhuber (11:39.840)
has done a lot of interesting ideas there,
Lex Fridman (11:42.120)
to think about P versus NP,
Juergen Schmidhuber (11:46.320)
this formalization of how hard problems are,
Lex Fridman (11:50.760)
how they scale,
Juergen Schmidhuber (11:52.360)
this kind of worst case analysis type of thinking,
Lex Fridman (11:55.160)
do you find that useful?
Juergen Schmidhuber (11:56.800)
Or is it only just a mathematical,
Lex Fridman (12:00.520)
it's a set of mathematical techniques
Juergen Schmidhuber (12:02.600)
to give you intuition about what's good and bad.
Lex Fridman (12:05.720)
So P versus NP, that's super interesting
Juergen Schmidhuber (12:09.440)
from a theoretical point of view.
Lex Fridman (12:11.760)
And in fact, as you are thinking about that problem,
Juergen Schmidhuber (12:14.560)
you can also get inspiration
Lex Fridman (12:17.280)
for better practical problem solvers.
Juergen Schmidhuber (12:21.280)
On the other hand, we have to admit
Lex Fridman (12:23.320)
that at the moment, the best practical problem solvers
Juergen Schmidhuber (12:28.360)
for all kinds of problems that we are now solving
Lex Fridman (12:31.080)
through what is called AI at the moment,
Juergen Schmidhuber (12:33.840)
they are not of the kind
Lex Fridman (12:36.240)
that is inspired by these questions.
Juergen Schmidhuber (12:38.760)
There we are using general purpose computers
Lex Fridman (12:42.680)
such as recurrent neural networks,
Lex Fridman (12:44.800)
but we have a search technique
Lex Fridman (12:46.680)
which is just local search gradient descent
Juergen Schmidhuber (12:50.280)
to try to find a program
Lex Fridman (12:51.920)
that is running on these recurrent networks,
Juergen Schmidhuber (12:54.400)
such that it can solve some interesting problems
Lex Fridman (12:58.200)
such as speech recognition or machine translation
Lex Fridman (13:01.880)
and something like that.
Lex Fridman (13:03.120)
And there is very little theory behind the best solutions
Juergen Schmidhuber (13:08.120)
that we have at the moment that can do that.
Lex Fridman (13:10.840)
Do you think that needs to change?
Lex Fridman (13:12.680)
Do you think that will change?
Lex Fridman (13:14.080)
Or can we go, can we create a general intelligent systems
Juergen Schmidhuber (13:17.160)
without ever really proving that that system is intelligent
Lex Fridman (13:20.640)
in some kind of mathematical way,
Juergen Schmidhuber (13:22.600)
solving machine translation perfectly
Lex Fridman (13:25.000)
or something like that,
Juergen Schmidhuber (13:26.320)
within some kind of syntactic definition of a language,
Lex Fridman (13:29.200)
or can we just be super impressed
Lex Fridman (13:31.160)
by the thing working extremely well and that's sufficient?
Lex Fridman (13:35.120)
There's an old saying,
Lex Fridman (13:36.760)
and I don't know who brought it up first,
Lex Fridman (13:39.360)
which says, there's nothing more practical
Juergen Schmidhuber (13:42.480)
than a good theory.
Lex Fridman (13:43.720)
And a good theory of problem solving
Juergen Schmidhuber (13:52.800)
under limited resources,
Lex Fridman (13:54.360)
like here in this universe or on this little planet,
Juergen Schmidhuber (13:58.480)
has to take into account these limited resources.
Lex Fridman (14:01.800)
And so probably there is locking a theory,
Juergen Schmidhuber (14:08.040)
which is related to what we already have,
Lex Fridman (14:10.760)
these asymptotically optimal problem solvers,
Juergen Schmidhuber (14:14.400)
which tells us what we need in addition to that
Lex Fridman (14:18.520)
to come up with a practically optimal problem solver.
Lex Fridman (14:22.640)
So I believe we will have something like that.
Lex Fridman (14:27.040)
And maybe just a few little tiny twists are necessary
Juergen Schmidhuber (14:30.520)
to change what we already have,
Lex Fridman (14:34.280)
to come up with that as well.
Juergen Schmidhuber (14:36.320)
As long as we don't have that,
Lex Fridman (14:37.720)
we admit that we are taking suboptimal ways
Lex Fridman (14:42.560)
and recurrent neural networks and long short term memory
Lex Fridman (14:45.960)
for equipped with local search techniques.
Lex Fridman (14:50.400)
And we are happy that it works better
Lex Fridman (14:53.520)
than any competing methods,
Lex Fridman (14:55.480)
but that doesn't mean that we think we are done.
Lex Fridman (15:00.800)
You've said that an AGI system
Juergen Schmidhuber (15:02.720)
will ultimately be a simple one.
Lex Fridman (15:05.040)
A general intelligence system
Juergen Schmidhuber (15:06.200)
will ultimately be a simple one.
Lex Fridman (15:08.000)
Maybe a pseudocode of a few lines
Juergen Schmidhuber (15:10.240)
will be able to describe it.
Lex Fridman (15:11.840)
Can you talk through your intuition behind this idea,
Lex Fridman (15:16.760)
why you feel that at its core,
Lex Fridman (15:21.480)
intelligence is a simple algorithm?
Juergen Schmidhuber (15:26.920)
Experience tells us that the stuff that works best
Lex Fridman (15:31.680)
is really simple.
Lex Fridman (15:33.120)
So the asymptotically optimal ways of solving problems,
Lex Fridman (15:37.680)
if you look at them,
Juergen Schmidhuber (15:38.800)
they're just a few lines of code, it's really true.
Lex Fridman (15:41.840)
Although they are these amazing properties,
Juergen Schmidhuber (15:44.000)
just a few lines of code.
Lex Fridman (15:45.800)
Then the most promising and most useful practical things,
Juergen Schmidhuber (15:53.760)
maybe don't have this proof of optimality
Lex Fridman (15:56.360)
associated with them.
Juergen Schmidhuber (15:57.800)
However, they are also just a few lines of code.
Lex Fridman (16:00.880)
The most successful recurrent neural networks,
Juergen Schmidhuber (16:05.080)
you can write them down in five lines of pseudocode.
Lex Fridman (16:08.320)
That's a beautiful, almost poetic idea,
Lex Fridman (16:10.920)
but what you're describing there
Lex Fridman (16:15.640)
is the lines of pseudocode are sitting on top
Juergen Schmidhuber (16:18.200)
of layers and layers of abstractions in a sense.
Lex Fridman (16:22.280)
So you're saying at the very top,
Juergen Schmidhuber (16:25.040)
it'll be a beautifully written sort of algorithm.
Lex Fridman (16:31.120)
But do you think that there's many layers of abstractions
Lex Fridman (16:33.960)
we have to first learn to construct?
Lex Fridman (16:36.880)
Yeah, of course, we are building on all these
Juergen Schmidhuber (16:40.400)
great abstractions that people have invented over the millennia,
Lex Fridman (16:45.080)
such as matrix multiplications and real numbers
Lex Fridman (16:50.520)
and basic arithmetics and calculus
Lex Fridman (16:56.440)
and derivations of error functions
Lex Fridman (17:00.240)
and derivatives of error functions and stuff like that.
Lex Fridman (17:04.400)
So without that language that greatly simplifies
Juergen Schmidhuber (17:09.400)
our way of thinking about these problems,
Lex Fridman (17:13.840)
we couldn't do anything.
Lex Fridman (17:14.760)
So in that sense, as always,
Lex Fridman (17:16.520)
we are standing on the shoulders of the giants
Juergen Schmidhuber (17:19.520)
who in the past simplified the problem
Lex Fridman (17:24.200)
of problem solving so much
Juergen Schmidhuber (17:26.320)
that now we have a chance to do the final step.
Lex Fridman (17:29.960)
So the final step will be a simple one.
Juergen Schmidhuber (17:33.960)
If we take a step back through all of human civilization
Lex Fridman (17:36.680)
and just the universe in general,
Lex Fridman (17:40.000)
how do you think about evolution
Lex Fridman (17:41.400)
and what if creating a universe
Lex Fridman (17:44.480)
is required to achieve this final step?
Lex Fridman (17:47.240)
What if going through the very painful
Lex Fridman (17:50.880)
and inefficient process of evolution is needed
Lex Fridman (17:53.800)
to come up with this set of abstractions
Lex Fridman (17:55.800)
that ultimately lead to intelligence?
Lex Fridman (17:57.720)
Do you think there's a shortcut
Juergen Schmidhuber (18:00.720)
or do you think we have to create something like our universe
Lex Fridman (18:04.560)
in order to create something like human level intelligence?
Lex Fridman (18:09.400)
So far, the only example we have is this one,
Lex Fridman (18:13.080)
this universe in which we are living.
Lex Fridman (18:14.880)
Do you think we can do better?
Lex Fridman (18:20.800)
Maybe not, but we are part of this whole process.
Lex Fridman (18:24.960)
So apparently, so it might be the case
Lex Fridman (18:29.920)
that the code that runs the universe
Juergen Schmidhuber (18:32.080)
is really, really simple.
Lex Fridman (18:33.600)
Everything points to that possibility
Juergen Schmidhuber (18:35.760)
because gravity and other basic forces
Lex Fridman (18:39.080)
are really simple laws that can be easily described
Juergen Schmidhuber (18:43.280)
also in just a few lines of code basically.
Lex Fridman (18:46.240)
And then there are these other events
Juergen Schmidhuber (18:51.360)
that the apparently random events
Lex Fridman (18:54.280)
in the history of the universe,
Juergen Schmidhuber (18:55.760)
which as far as we know at the moment
Lex Fridman (18:58.000)
don't have a compact code, but who knows?
Juergen Schmidhuber (19:00.600)
Maybe somebody in the near future
Lex Fridman (19:02.440)
is going to figure out the pseudo random generator
Juergen Schmidhuber (19:06.240)
which is computing whether the measurement
Lex Fridman (19:11.800)
of that spin up or down thing here
Juergen Schmidhuber (19:15.320)
is going to be positive or negative.
Lex Fridman (19:17.840)
Underlying quantum mechanics.
Juergen Schmidhuber (19:19.280)
Yes.
Lex Fridman (19:20.120)
Do you ultimately think quantum mechanics
Lex Fridman (19:22.600)
is a pseudo random number generator?
Lex Fridman (19:24.640)
So it's all deterministic.
Juergen Schmidhuber (19:26.320)
There's no randomness in our universe.
Lex Fridman (19:28.200)
Does God play dice?
Lex Fridman (19:31.120)
So a couple of years ago, a famous physicist,
Lex Fridman (19:34.680)
quantum physicist, Anton Zeilinger,
Juergen Schmidhuber (19:37.680)
he wrote an essay in nature
Lex Fridman (19:40.080)
and it started more or less like that.
Juergen Schmidhuber (19:45.360)
One of the fundamental insights of the 20th century
Lex Fridman (19:50.360)
was that the universe is fundamentally random
Juergen Schmidhuber (19:57.360)
on the quantum level.
Lex Fridman (1:00:00.040)
So I became aware of all of that in the 80s and
Juergen Schmidhuber (1:00:04.040)
back then logic programming was a huge thing.
Lex Fridman (1:00:08.040)
Was it inspiring to you yourself?
Lex Fridman (1:00:10.040)
Did you find it compelling?
Lex Fridman (1:00:12.040)
Because a lot of your work was not so much in that
Lex Fridman (1:00:16.040)
realm, right?
Lex Fridman (1:00:17.040)
It was more in the learning systems.
Juergen Schmidhuber (1:00:18.040)
Yes and no, but we did all of that.
Lex Fridman (1:00:20.040)
So my first publication ever actually was 1987,
Juergen Schmidhuber (1:00:27.040)
was the implementation of genetic algorithm of a
Lex Fridman (1:00:31.040)
genetic programming system in Prolog.
Lex Fridman (1:00:34.040)
So Prolog, that's what you learn back then which is
Lex Fridman (1:00:38.040)
a logic programming language and the Japanese,
Juergen Schmidhuber (1:00:41.040)
they have this huge fifth generation AI project
Lex Fridman (1:00:45.040)
which was mostly about logic programming back then.
Juergen Schmidhuber (1:00:49.040)
Although neural networks existed and were well
Lex Fridman (1:00:52.040)
known back then and deep learning has existed since
Juergen Schmidhuber (1:00:56.040)
1965, since this guy in the Ukraine,
Lex Fridman (1:01:00.040)
Iwakunenko, started it.
Lex Fridman (1:01:02.040)
But the Japanese and many other people,
Lex Fridman (1:01:05.040)
they focused really on this logic programming and I
Juergen Schmidhuber (1:01:08.040)
was influenced to the extent that I said,
Lex Fridman (1:01:10.040)
okay, let's take these biologically inspired
Juergen Schmidhuber (1:01:13.040)
algorithms like evolution, programs,
Lex Fridman (1:01:20.040)
and implement that in the language which I know,
Juergen Schmidhuber (1:01:22.040)
which was Prolog, for example, back then.
Lex Fridman (1:01:25.040)
And then in many ways this came back later because
Juergen Schmidhuber (1:01:29.040)
the Gödel machine, for example,
Lex Fridman (1:01:31.040)
has a proof searcher on board and without that it
Juergen Schmidhuber (1:01:34.040)
would not be optimal.
Lex Fridman (1:01:36.040)
Well, Markus Futter's universal algorithm for
Juergen Schmidhuber (1:01:38.040)
solving all well defined problems has a proof
Lex Fridman (1:01:41.040)
searcher on board so that's very much logic programming.
Juergen Schmidhuber (1:01:46.040)
Without that it would not be asymptotically optimal.
Lex Fridman (1:01:50.040)
But then on the other hand,
Juergen Schmidhuber (1:01:51.040)
because we are very pragmatic guys also,
Lex Fridman (1:01:54.040)
we focused on recurrent neural networks and
Juergen Schmidhuber (1:02:00.040)
suboptimal stuff such as gradient based search and
Lex Fridman (1:02:04.040)
program space rather than provably optimal things.
Juergen Schmidhuber (1:02:09.040)
The logic programming certainly has a usefulness
Lex Fridman (1:02:13.040)
when you're trying to construct something provably
Juergen Schmidhuber (1:02:16.040)
optimal or provably good or something like that.
Lex Fridman (1:02:19.040)
But is it useful for practical problems?
Juergen Schmidhuber (1:02:22.040)
It's really useful for our theorem proving.
Lex Fridman (1:02:24.040)
The best theorem provers today are not neural networks.
Juergen Schmidhuber (1:02:28.040)
No, they are logic programming systems and they
Lex Fridman (1:02:31.040)
are much better theorem provers than most math
Juergen Schmidhuber (1:02:35.040)
students in the first or second semester.
Lex Fridman (1:02:38.040)
But for reasoning, for playing games of Go or chess
Juergen Schmidhuber (1:02:43.040)
or for robots, autonomous vehicles that operate in
Lex Fridman (1:02:46.040)
the real world or object manipulation,
Juergen Schmidhuber (1:02:49.040)
you think learning.
Lex Fridman (1:02:51.040)
Yeah, as long as the problems have little to do
Juergen Schmidhuber (1:02:54.040)
with theorem proving themselves,
Lex Fridman (1:02:58.040)
then as long as that is not the case,
Juergen Schmidhuber (1:03:01.040)
you just want to have better pattern recognition.
Lex Fridman (1:03:05.040)
So to build a self driving car,
Juergen Schmidhuber (1:03:07.040)
you want to have better pattern recognition and
Lex Fridman (1:03:10.040)
pedestrian recognition and all these things.
Juergen Schmidhuber (1:03:14.040)
You want to minimize the number of false positives,
Lex Fridman (1:03:19.040)
which is currently slowing down self driving cars
Juergen Schmidhuber (1:03:21.040)
in many ways.
Lex Fridman (1:03:23.040)
All of that has very little to do with logic programming.
Lex Fridman (1:03:27.040)
What are you most excited about in terms of
Lex Fridman (1:03:32.040)
directions of artificial intelligence at this moment
Juergen Schmidhuber (1:03:35.040)
in the next few years in your own research
Lex Fridman (1:03:38.040)
and in the broader community?
Lex Fridman (1:03:41.040)
So I think in the not so distant future,
Lex Fridman (1:03:44.040)
we will have for the first time little robots
Juergen Schmidhuber (1:03:50.040)
that learn like kids.
Lex Fridman (1:03:53.040)
I will be able to say to the robot,
Juergen Schmidhuber (1:03:57.040)
look here robot, we are going to assemble a smartphone.
Lex Fridman (1:04:01.040)
Let's take this slab of plastic and the screwdriver
Lex Fridman (1:04:05.040)
and let's screw in the screw like that.
Lex Fridman (1:04:09.040)
Not like that, like that.
Juergen Schmidhuber (1:04:11.040)
Not like that, like that.
Lex Fridman (1:04:14.040)
And I don't have a data glove or something.
Juergen Schmidhuber (1:04:17.040)
He will see me and he will hear me
Lex Fridman (1:04:20.040)
and he will try to do something with his own actuators,
Juergen Schmidhuber (1:04:24.040)
which will be really different from mine,
Lex Fridman (1:04:26.040)
but he will understand the difference
Lex Fridman (1:04:28.040)
and will learn to imitate me,
Lex Fridman (1:04:31.040)
but not in the supervised way
Juergen Schmidhuber (1:04:34.040)
where a teacher is giving target signals
Lex Fridman (1:04:37.040)
for all his muscles all the time.
Juergen Schmidhuber (1:04:40.040)
No, by doing this high level imitation
Lex Fridman (1:04:43.040)
where he first has to learn to imitate me
Lex Fridman (1:04:46.040)
and then to interpret these additional noises
Lex Fridman (1:04:48.040)
coming from my mouth as helping,
Juergen Schmidhuber (1:04:51.040)
helpful signals to do that better.
Lex Fridman (1:04:54.040)
And then it will by itself come up with faster ways
Lex Fridman (1:05:00.040)
and more efficient ways of doing the same thing.
Lex Fridman (1:05:03.040)
And finally I stop his learning algorithm
Lex Fridman (1:05:07.040)
and make a million copies and sell it.
Lex Fridman (1:05:10.040)
And so at the moment this is not possible,
Lex Fridman (1:05:13.040)
but we already see how we are going to get there.
Lex Fridman (1:05:16.040)
And you can imagine to the extent
Juergen Schmidhuber (1:05:19.040)
that this works economically and cheaply,
Lex Fridman (1:05:22.040)
it's going to change everything.
Juergen Schmidhuber (1:05:25.040)
Almost all of production is going to be affected by that.
Lex Fridman (1:05:31.040)
And a much bigger wave,
Juergen Schmidhuber (1:05:34.040)
a much bigger AI wave is coming
Lex Fridman (1:05:36.040)
than the one that we are currently witnessing,
Juergen Schmidhuber (1:05:38.040)
which is mostly about passive pattern recognition
Lex Fridman (1:05:40.040)
on your smartphone.
Juergen Schmidhuber (1:05:42.040)
This is about active machines that shapes data
Lex Fridman (1:05:45.040)
through the actions they are executing
Lex Fridman (1:05:48.040)
and they learn to do that in a good way.
Lex Fridman (1:05:52.040)
So many of the traditional industries
Juergen Schmidhuber (1:05:55.040)
are going to be affected by that.
Lex Fridman (1:05:57.040)
All the companies that are building machines
Juergen Schmidhuber (1:06:01.040)
will equip these machines with cameras
Lex Fridman (1:06:04.040)
and other sensors and they are going to learn
Juergen Schmidhuber (1:06:08.040)
to solve all kinds of problems
Lex Fridman (1:06:11.040)
through interaction with humans,
Lex Fridman (1:06:13.040)
but also a lot on their own
Lex Fridman (1:06:15.040)
to improve what they already can do.
Lex Fridman (1:06:20.040)
And lots of old economy is going to be affected by that.
Lex Fridman (1:06:24.040)
And in recent years I have seen that old economy
Juergen Schmidhuber (1:06:27.040)
is actually waking up and realizing that this is the case.
Lex Fridman (1:06:32.040)
Are you optimistic about that future?
Lex Fridman (1:06:34.040)
Are you concerned?
Lex Fridman (1:06:36.040)
There is a lot of people concerned in the near term
Juergen Schmidhuber (1:06:38.040)
about the transformation of the nature of work,
Lex Fridman (1:06:43.040)
the kind of ideas that you just suggested
Juergen Schmidhuber (1:06:45.040)
would have a significant impact
Lex Fridman (1:06:47.040)
of what kind of things could be automated.
Lex Fridman (1:06:49.040)
Are you optimistic about that future?
Lex Fridman (1:06:52.040)
Are you nervous about that future?
Lex Fridman (1:06:54.040)
And looking a little bit farther into the future,
Lex Fridman (1:06:58.040)
there are people like Gila Musk, Stuart Russell,
Juergen Schmidhuber (1:07:02.040)
concerned about the existential threats of that future.
Lex Fridman (1:07:06.040)
So in the near term, job loss,
Juergen Schmidhuber (1:07:08.040)
in the long term existential threat,
Lex Fridman (1:07:10.040)
are these concerns to you or are you ultimately optimistic?
Lex Fridman (1:07:15.040)
So let's first address the near future.
Lex Fridman (1:07:22.040)
We have had predictions of job losses for many decades.
Juergen Schmidhuber (1:07:28.040)
For example, when industrial robots came along,
Lex Fridman (1:07:33.040)
many people predicted that lots of jobs are going to get lost.
Lex Fridman (1:07:38.040)
And in a sense, they were right,
Lex Fridman (1:07:42.040)
because back then there were car factories
Lex Fridman (1:07:46.040)
and hundreds of people in these factories assembled cars,
Lex Fridman (1:07:51.040)
and today the same car factories have hundreds of robots
Lex Fridman (1:07:54.040)
and maybe three guys watching the robots.
Lex Fridman (1:07:59.040)
On the other hand, those countries that have lots of robots per capita,
Juergen Schmidhuber (1:08:05.040)
Japan, Korea, Germany, Switzerland,
Lex Fridman (1:08:07.040)
and a couple of other countries,
Juergen Schmidhuber (1:08:10.040)
they have really low unemployment rates.
Lex Fridman (1:08:14.040)
Somehow, all kinds of new jobs were created.
Juergen Schmidhuber (1:08:18.040)
Back then, nobody anticipated those jobs.
Lex Fridman (1:08:23.040)
And decades ago, I always said,
Juergen Schmidhuber (1:08:27.040)
it's really easy to say which jobs are going to get lost,
Lex Fridman (1:08:32.040)
but it's really hard to predict the new ones.
Juergen Schmidhuber (1:08:36.040)
200 years ago, who would have predicted all these people
Lex Fridman (1:08:40.040)
making money as YouTube bloggers, for example?
Juergen Schmidhuber (1:08:46.040)
200 years ago, 60% of all people used to work in agriculture.
Lex Fridman (1:08:54.040)
Today, maybe 1%.
Lex Fridman (1:08:57.040)
But still, only, I don't know, 5% unemployment.
Lex Fridman (1:09:02.040)
Lots of new jobs were created, and Homo Ludens, the playing man,
Juergen Schmidhuber (1:09:08.040)
is inventing new jobs all the time.
Lex Fridman (1:09:11.040)
Most of these jobs are not existentially necessary
Juergen Schmidhuber (1:09:16.040)
for the survival of our species.
Lex Fridman (1:09:19.040)
There are only very few existentially necessary jobs,
Juergen Schmidhuber (1:09:23.040)
such as farming and building houses and warming up the houses,
Lex Fridman (1:09:28.040)
but less than 10% of the population is doing that.
Lex Fridman (1:09:31.040)
And most of these newly invented jobs are about
Lex Fridman (1:09:35.040)
interacting with other people in new ways,
Juergen Schmidhuber (1:09:38.040)
through new media and so on,
Lex Fridman (1:09:41.040)
getting new types of kudos and forms of likes and whatever,
Lex Fridman (1:09:46.040)
and even making money through that.
Lex Fridman (1:09:48.040)
So, Homo Ludens, the playing man, doesn't want to be unemployed,
Lex Fridman (1:09:53.040)
and that's why he's inventing new jobs all the time.
Lex Fridman (1:09:57.040)
And he keeps considering these jobs as really important
Lex Fridman (1:10:02.040)
and is investing a lot of energy and hours of work into those new jobs.
Lex Fridman (1:10:08.040)
That's quite beautifully put.
Juergen Schmidhuber (1:10:10.040)
We're really nervous about the future because we can't predict
Lex Fridman (1:10:13.040)
what kind of new jobs will be created.
Lex Fridman (1:10:15.040)
But you're ultimately optimistic that we humans are so restless
Lex Fridman (1:10:21.040)
that we create and give meaning to newer and newer jobs,
Juergen Schmidhuber (1:10:25.040)
totally new, things that get likes on Facebook
Lex Fridman (1:10:29.040)
or whatever the social platform is.
Lex Fridman (1:10:32.040)
So what about long term existential threat of AI,
Lex Fridman (1:10:36.040)
where our whole civilization may be swallowed up
Lex Fridman (1:10:41.040)
by these ultra super intelligent systems?
Lex Fridman (1:10:45.040)
Maybe it's not going to be swallowed up,
Lex Fridman (1:10:48.040)
but I'd be surprised if we humans were the last step
Lex Fridman (1:10:55.040)
in the evolution of the universe.
Juergen Schmidhuber (1:10:58.040)
You've actually had this beautiful comment somewhere that I've seen
Lex Fridman (1:11:05.040)
saying that, quite insightful, artificial general intelligence systems,
Juergen Schmidhuber (1:11:12.040)
just like us humans, will likely not want to interact with humans,
Lex Fridman (1:11:16.040)
they'll just interact amongst themselves.
Juergen Schmidhuber (1:11:18.040)
Just like ants interact amongst themselves
Lex Fridman (1:11:21.040)
and only tangentially interact with humans.
Lex Fridman (1:11:25.040)
And it's quite an interesting idea that once we create AGI,
Lex Fridman (1:11:29.040)
they will lose interest in humans and compete for their own Facebook likes
Lex Fridman (1:11:34.040)
and their own social platforms.
Lex Fridman (1:11:36.040)
So within that quite elegant idea, how do we know in a hypothetical sense
Lex Fridman (1:11:45.040)
that there's not already intelligence systems out there?
Lex Fridman (1:11:49.040)
How do you think broadly of general intelligence greater than us?
Lex Fridman (1:11:54.040)
How do we know it's out there?
Lex Fridman (1:11:56.040)
How do we know it's around us?
Lex Fridman (1:11:59.040)
And could it already be?
Lex Fridman (1:12:01.040)
I'd be surprised if within the next few decades or something like that,
Juergen Schmidhuber (1:12:07.040)
we won't have AIs that are truly smart in every single way
Lex Fridman (1:12:13.040)
and better problem solvers in almost every single important way.
Lex Fridman (1:12:17.040)
And I'd be surprised if they wouldn't realize what we have realized a long time ago,
Lex Fridman (1:12:25.040)
which is that almost all physical resources are not here in this biosphere,
Lex Fridman (1:12:31.040)
but further out, the rest of the solar system gets 2 billion times more solar energy
Lex Fridman (1:12:41.040)
than our little planet.
Juergen Schmidhuber (1:12:43.040)
There's lots of material out there that you can use to build robots
Lex Fridman (1:12:47.040)
and self replicating robot factories and all this stuff.
Lex Fridman (1:12:51.040)
And they are going to do that and they will be scientists and curious
Lex Fridman (1:12:56.040)
and they will explore what they can do.
Lex Fridman (1:12:59.040)
And in the beginning, they will be fascinated by life
Lex Fridman (1:13:04.040)
and by their own origins in our civilization.
Juergen Schmidhuber (1:13:07.040)
They will want to understand that completely, just like people today
Juergen Schmidhuber (1:13:11.040)
would like to understand how life works and also the history of our own existence
Lex Fridman (1:13:21.040)
and civilization, but then also the physical laws that created all of that.
Lex Fridman (1:13:27.040)
So in the beginning, they will be fascinated by life.
Juergen Schmidhuber (1:13:30.040)
Once they understand it, they lose interest.
Lex Fridman (1:13:34.040)
Like anybody who loses interest in things he understands.
Lex Fridman (1:13:40.040)
And then, as you said, the most interesting sources of information for them
Lex Fridman (1:13:50.040)
will be others of their own kind.
Lex Fridman (1:13:58.040)
So at least in the long run, there seems to be some sort of protection
Lex Fridman (1:14:06.040)
through lack of interest on the other side.
Lex Fridman (1:14:11.040)
And now it seems also clear, as far as we understand physics,
Juergen Schmidhuber (1:14:17.040)
you need matter and energy to compute and to build more robots and infrastructure
Juergen Schmidhuber (1:14:23.040)
for AI civilization and EIEI ecologies consisting of trillions of different types of AIs.
Lex Fridman (1:14:31.040)
And so it seems inconceivable to me that this thing is not going to expand.
Juergen Schmidhuber (1:14:37.040)
Some AI ecology not controlled by one AI, but trillions of different types of AIs
Lex Fridman (1:14:44.040)
competing in all kinds of quickly evolving and disappearing ecological niches
Juergen Schmidhuber (1:14:50.040)
in ways that we cannot fathom at the moment.
Lex Fridman (1:14:52.040)
But it's going to expand, limited by light speed and physics,
Lex Fridman (1:14:57.040)
but it's going to expand and now we realize that the universe is still young.
Juergen Schmidhuber (1:15:03.040)
It's only 13.8 billion years old and it's going to be a thousand times older than that.
Lex Fridman (1:15:10.040)
So there's plenty of time to conquer the entire universe
Lex Fridman (1:15:17.040)
and to fill it with intelligence and senders and receivers
Juergen Schmidhuber (1:15:21.040)
such that AIs can travel the way they are traveling in our labs today,
Lex Fridman (1:15:27.040)
which is by radio from sender to receiver.
Lex Fridman (1:15:31.040)
And let's call the current age of the universe one eon, one eon.
Lex Fridman (1:15:39.040)
Now it will take just a few eons from now and the entire visible universe
Juergen Schmidhuber (1:15:43.040)
is going to be full of that stuff.
Lex Fridman (1:15:47.040)
And let's look ahead to a time when the universe is going to be 1000 times older than it is now.
Juergen Schmidhuber (1:15:53.040)
They will look back and they will say, look, almost immediately after the Big Bang,
Lex Fridman (1:15:57.040)
only a few eons later, the entire universe started to become intelligent.
Juergen Schmidhuber (1:16:03.040)
Now to your question, how do we see whether anything like that has already happened
Lex Fridman (1:16:09.040)
or is already in a more advanced stage in some other part of the universe, of the visible universe?
Lex Fridman (1:16:16.040)
We are trying to look out there and nothing like that has happened so far or is that true?
Lex Fridman (1:16:22.040)
Do you think we would recognize it?
Lex Fridman (1:16:24.040)
How do we know it's not among us?
Lex Fridman (1:16:26.040)
How do we know planets aren't in themselves intelligent beings?
Lex Fridman (1:16:31.040)
How do we know ants seen as a collective are not much greater intelligence than our own?
Lex Fridman (1:16:40.040)
These kinds of ideas.
Juergen Schmidhuber (1:16:42.040)
When I was a boy, I was thinking about these things
Lex Fridman (1:16:45.040)
and I thought, maybe it has already happened.
Juergen Schmidhuber (1:16:48.040)
Because back then I knew, I learned from popular physics books,
Lex Fridman (1:16:53.040)
that the large scale structure of the universe is not homogeneous.
Juergen Schmidhuber (1:17:00.040)
You have these clusters of galaxies and then in between there are these huge empty spaces.
Lex Fridman (1:17:08.040)
And I thought, maybe they aren't really empty.
Juergen Schmidhuber (1:17:12.040)
It's just that in the middle of that, some AI civilization already has expanded
Lex Fridman (1:17:17.040)
and then has covered a bubble of a billion light years diameter
Lex Fridman (1:17:22.040)
and is using all the energy of all the stars within that bubble for its own unfathomable purposes.
Lex Fridman (1:17:29.040)
And so it already has happened and we just fail to interpret the signs.
Lex Fridman (1:17:35.040)
And then I learned that gravity by itself explains the large scale structure of the universe
Lex Fridman (1:17:43.040)
and that this is not a convincing explanation.
Lex Fridman (1:17:46.040)
And then I thought, maybe it's the dark matter.
Lex Fridman (1:17:51.040)
Because as far as we know today, 80% of the measurable matter is invisible.
Lex Fridman (1:18:01.040)
And we know that because otherwise our galaxy or other galaxies would fall apart.
Lex Fridman (1:18:06.040)
They are rotating too quickly.
Lex Fridman (1:18:10.040)
And then the idea was, maybe all of these AI civilizations that are already out there,
Juergen Schmidhuber (1:18:17.040)
they are just invisible because they are really efficient in using the energies of their own local systems
Lex Fridman (1:18:26.040)
and that's why they appear dark to us.
Lex Fridman (1:18:29.040)
But this is also not a convincing explanation because then the question becomes,
Lex Fridman (1:18:34.040)
why are there still any visible stars left in our own galaxy, which also must have a lot of dark matter?
Lex Fridman (1:18:44.040)
So that is also not a convincing thing.
Lex Fridman (1:18:46.040)
And today, I like to think it's quite plausible that maybe we are the first,
Juergen Schmidhuber (1:18:54.040)
at least in our local light cone within the few hundreds of millions of light years that we can reliably observe.
Lex Fridman (1:19:09.040)
Is that exciting to you that we might be the first?
Lex Fridman (1:19:12.040)
And it would make us much more important because if we mess it up through a nuclear war,
Juergen Schmidhuber (1:19:20.040)
then maybe this will have an effect on the development of the entire universe.
Lex Fridman (1:19:31.040)
So let's not mess it up.
Juergen Schmidhuber (1:19:32.040)
Let's not mess it up.
Lex Fridman (1:19:34.040)
Jürgen, thank you so much for talking today. I really appreciate it.
Juergen Schmidhuber (1:19:37.040)
It's my pleasure.
Lex Fridman (20:00.040)
And that whenever you measure spin up or down
Juergen Schmidhuber (20:04.040)
or something like that,
Lex Fridman (20:05.440)
a new bit of information enters the history of the universe.
Lex Fridman (20:12.040)
And while I was reading that,
Lex Fridman (20:13.200)
I was already typing the response
Lex Fridman (20:16.560)
and they had to publish it.
Lex Fridman (20:17.880)
Because I was right, that there is no evidence,
Juergen Schmidhuber (20:21.640)
no physical evidence for that.
Lex Fridman (20:25.040)
So there's an alternative explanation
Juergen Schmidhuber (20:27.720)
where everything that we consider random
Lex Fridman (20:30.680)
is actually pseudo random,
Juergen Schmidhuber (20:33.040)
such as the decimal expansion of pi,
Lex Fridman (20:35.960)
3.141 and so on, which looks random, but isn't.
Lex Fridman (20:41.680)
So pi is interesting because every three digits
Lex Fridman (20:45.400)
sequence, every sequence of three digits
Juergen Schmidhuber (20:50.400)
appears roughly one in a thousand times.
Lex Fridman (20:53.400)
And every five digit sequence
Juergen Schmidhuber (20:57.400)
appears roughly one in 10,000 times,
Lex Fridman (21:01.080)
what you would expect if it was random.
Lex Fridman (21:04.400)
But there's a very short algorithm,
Lex Fridman (21:06.680)
a short program that computes all of that.
Lex Fridman (21:09.120)
So it's extremely compressible.
Lex Fridman (21:11.320)
And who knows, maybe tomorrow,
Juergen Schmidhuber (21:12.640)
somebody, some grad student at CERN goes back
Lex Fridman (21:15.640)
over all these data points, better decay and whatever,
Lex Fridman (21:20.120)
and figures out, oh, it's the second billion digits of pi
Lex Fridman (21:24.760)
or something like that.
Juergen Schmidhuber (21:26.040)
We don't have any fundamental reason at the moment
Lex Fridman (21:29.080)
to believe that this is truly random
Lex Fridman (21:33.600)
and not just a deterministic video game.
Lex Fridman (21:36.680)
If it was a deterministic video game,
Juergen Schmidhuber (21:38.680)
it would be much more beautiful.
Lex Fridman (21:40.360)
Because beauty is simplicity.
Lex Fridman (21:43.840)
And many of the basic laws of the universe,
Lex Fridman (21:47.000)
like gravity and the other basic forces are very simple.
Lex Fridman (21:51.360)
So very short programs can explain what these are doing.
Lex Fridman (21:56.760)
And it would be awful and ugly.
Juergen Schmidhuber (22:00.560)
The universe would be ugly.
Lex Fridman (22:01.720)
The history of the universe would be ugly
Juergen Schmidhuber (22:04.000)
if for the extra things, the random,
Lex Fridman (22:06.240)
the seemingly random data points that we get all the time,
Juergen Schmidhuber (22:11.080)
that we really need a huge number of extra bits
Lex Fridman (22:15.160)
to describe all these extra bits of information.
Lex Fridman (22:22.160)
So as long as we don't have evidence
Lex Fridman (22:24.800)
that there is no short program
Juergen Schmidhuber (22:26.920)
that computes the entire history of the entire universe,
Lex Fridman (22:31.240)
we are, as scientists, compelled to look further
Juergen Schmidhuber (22:36.600)
for that shortest program.
Lex Fridman (22:39.760)
Your intuition says there exists a program
Juergen Schmidhuber (22:43.760)
that can backtrack to the creation of the universe.
Lex Fridman (22:47.760)
Yeah.
Lex Fridman (22:48.600)
So it can give the shortest path
Lex Fridman (22:49.440)
to the creation of the universe.
Juergen Schmidhuber (22:50.480)
Yes.
Lex Fridman (22:51.320)
Including all the entanglement things
Lex Fridman (22:54.480)
and all the spin up and down measures
Lex Fridman (22:57.800)
that have been taken place since 13.8 billion years ago.
Lex Fridman (23:06.840)
So we don't have a proof that it is random.
Lex Fridman (23:11.840)
We don't have a proof that it is compressible
Juergen Schmidhuber (23:15.600)
to a short program.
Lex Fridman (23:16.760)
But as long as we don't have that proof,
Juergen Schmidhuber (23:18.240)
we are obliged as scientists to keep looking
Lex Fridman (23:21.680)
for that simple explanation.
Juergen Schmidhuber (23:23.600)
Absolutely.
Lex Fridman (23:24.440)
So you said the simplicity is beautiful or beauty is simple.
Juergen Schmidhuber (23:27.880)
Either one works.
Lex Fridman (23:29.440)
But you also work on curiosity, discovery,
Juergen Schmidhuber (23:34.560)
the romantic notion of randomness, of serendipity,
Lex Fridman (23:39.000)
of being surprised by things that are about you.
Juergen Schmidhuber (23:45.920)
In our poetic notion of reality,
Lex Fridman (23:49.600)
we think it's kind of like,
Juergen Schmidhuber (23:51.640)
poetic notion of reality, we think as humans
Lex Fridman (23:54.920)
require randomness.
Lex Fridman (23:56.400)
So you don't find randomness beautiful.
Lex Fridman (23:59.000)
You find simple determinism beautiful.
Juergen Schmidhuber (24:04.880)
Yeah.
Lex Fridman (24:05.720)
Okay.
Lex Fridman (24:07.520)
So why?
Lex Fridman (24:08.560)
Why?
Juergen Schmidhuber (24:09.400)
Because the explanation becomes shorter.
Lex Fridman (24:13.040)
A universe that is compressible to a short program
Juergen Schmidhuber (24:18.040)
is much more elegant and much more beautiful
Lex Fridman (24:22.040)
than another one, which needs an almost infinite
Juergen Schmidhuber (24:25.040)
number of bits to be described.
Lex Fridman (24:28.040)
As far as we know, many things that are happening
Juergen Schmidhuber (24:32.040)
in this universe are really simple in terms of
Lex Fridman (24:35.040)
short programs that compute gravity
Lex Fridman (24:38.040)
and the interaction between elementary particles and so on.
Lex Fridman (24:43.040)
So all of that seems to be very, very simple.
Juergen Schmidhuber (24:45.040)
Every electron seems to reuse the same subprogram
Lex Fridman (24:50.040)
all the time, as it is interacting with
Juergen Schmidhuber (24:52.040)
other elementary particles.
Lex Fridman (24:58.040)
If we now require an extra oracle injecting
Juergen Schmidhuber (25:05.040)
new bits of information all the time for these
Lex Fridman (25:08.040)
extra things which are currently not understood,
Juergen Schmidhuber (25:11.040)
such as better decay, then the whole description
Lex Fridman (25:22.040)
length of the data that we can observe of the
Juergen Schmidhuber (25:26.040)
history of the universe would become much longer
Lex Fridman (25:31.040)
and therefore uglier.
Lex Fridman (25:33.040)
And uglier.
Lex Fridman (25:34.040)
Again, simplicity is elegant and beautiful.
Juergen Schmidhuber (25:38.040)
The history of science is a history of compression progress.
Lex Fridman (25:42.040)
Yes, so you've described sort of as we build up
Juergen Schmidhuber (25:47.040)
abstractions and you've talked about the idea
Lex Fridman (25:50.040)
of compression.
Lex Fridman (25:52.040)
How do you see this, the history of science,
Lex Fridman (25:55.040)
the history of humanity, our civilization,
Lex Fridman (25:58.040)
and life on Earth as some kind of path towards
Lex Fridman (26:02.040)
greater and greater compression?
Lex Fridman (26:04.040)
What do you mean by that?
Lex Fridman (26:05.040)
How do you think about that?
Juergen Schmidhuber (26:06.040)
Indeed, the history of science is a history of
Lex Fridman (26:10.040)
compression progress.
Lex Fridman (26:12.040)
What does that mean?
Lex Fridman (26:14.040)
Hundreds of years ago there was an astronomer
Juergen Schmidhuber (26:17.040)
whose name was Kepler and he looked at the data
Lex Fridman (26:21.040)
points that he got by watching planets move.
Lex Fridman (26:25.040)
And then he had all these data points and
Lex Fridman (26:27.040)
suddenly it turned out that he can greatly
Juergen Schmidhuber (26:30.040)
compress the data by predicting it through an
Lex Fridman (26:36.040)
ellipse law.
Lex Fridman (26:38.040)
So it turns out that all these data points are
Lex Fridman (26:40.040)
more or less on ellipses around the sun.
Lex Fridman (26:45.040)
And another guy came along whose name was
Lex Fridman (26:48.040)
Newton and before him Hooke.
Lex Fridman (26:51.040)
And they said the same thing that is making
Lex Fridman (26:55.040)
these planets move like that is what makes the
Juergen Schmidhuber (27:00.040)
apples fall down.
Lex Fridman (27:02.040)
And it also holds for stones and for all kinds
Juergen Schmidhuber (27:08.040)
of other objects.
Lex Fridman (27:11.040)
And suddenly many, many of these observations
Juergen Schmidhuber (27:15.040)
became much more compressible because as long
Lex Fridman (27:17.040)
as you can predict the next thing, given what
Juergen Schmidhuber (27:20.040)
you have seen so far, you can compress it.
Lex Fridman (27:23.040)
And you don't have to store that data extra.
Juergen Schmidhuber (27:25.040)
This is called predictive coding.
Lex Fridman (27:29.040)
And then there was still something wrong with
Juergen Schmidhuber (27:31.040)
that theory of the universe and you had
Lex Fridman (27:34.040)
deviations from these predictions of the theory.
Lex Fridman (27:37.040)
And 300 years later another guy came along
Lex Fridman (27:40.040)
whose name was Einstein.
Lex Fridman (27:42.040)
And he was able to explain away all these
Lex Fridman (27:46.040)
deviations from the predictions of the old
Juergen Schmidhuber (27:50.040)
theory through a new theory which was called
Lex Fridman (27:54.040)
the general theory of relativity.
Juergen Schmidhuber (27:57.040)
Which at first glance looks a little bit more
Lex Fridman (28:00.040)
complicated and you have to warp space and time
Lex Fridman (28:03.040)
but you can't phrase it within one single
Lex Fridman (28:05.040)
sentence which is no matter how fast you
Juergen Schmidhuber (28:08.040)
accelerate and how hard you decelerate and no
Lex Fridman (28:14.040)
matter what is the gravity in your local
Juergen Schmidhuber (28:18.040)
network, light speed always looks the same.
Lex Fridman (28:21.040)
And from that you can calculate all the
Juergen Schmidhuber (28:24.040)
consequences.
Lex Fridman (28:25.040)
So it's a very simple thing and it allows you
Juergen Schmidhuber (28:27.040)
to further compress all the observations
Lex Fridman (28:30.040)
because certainly there are hardly any
Juergen Schmidhuber (28:34.040)
deviations any longer that you can measure
Lex Fridman (28:37.040)
from the predictions of this new theory.
Lex Fridman (28:40.040)
So all of science is a history of compression
Lex Fridman (28:44.040)
progress.
Juergen Schmidhuber (28:45.040)
You never arrive immediately at the shortest
Lex Fridman (28:48.040)
explanation of the data but you're making
Juergen Schmidhuber (28:51.040)
progress.
Lex Fridman (28:52.040)
Whenever you are making progress you have an
Juergen Schmidhuber (28:55.040)
insight.
Lex Fridman (28:56.040)
You see oh first I needed so many bits of
Juergen Schmidhuber (28:59.040)
information to describe the data, to describe
Lex Fridman (29:02.040)
my falling apples, my video of falling apples,
Juergen Schmidhuber (29:04.040)
I need so many data, so many pixels have to be
Lex Fridman (29:07.040)
stored.
Lex Fridman (29:08.040)
But then suddenly I realize no there is a very
Lex Fridman (29:11.040)
simple way of predicting the third frame in the
Juergen Schmidhuber (29:14.040)
video from the first two.
Lex Fridman (29:16.040)
And maybe not every little detail can be
Juergen Schmidhuber (29:19.040)
predicted but more or less most of these orange
Lex Fridman (29:21.040)
blobs that are coming down they accelerate in
Juergen Schmidhuber (29:24.040)
the same way which means that I can greatly
Lex Fridman (29:27.040)
compress the video.
Lex Fridman (29:28.040)
And the amount of compression, progress, that
Lex Fridman (29:33.040)
is the depth of the insight that you have at
Juergen Schmidhuber (29:36.040)
that moment.
Lex Fridman (29:37.040)
That's the fun that you have, the scientific
Juergen Schmidhuber (29:39.040)
fun, the fun in that discovery.
Lex Fridman (29:42.040)
And we can build artificial systems that do
Juergen Schmidhuber (29:45.040)
the same thing.
Lex Fridman (29:46.040)
They measure the depth of their insights as they
Juergen Schmidhuber (29:49.040)
are looking at the data which is coming in
Lex Fridman (29:51.040)
through their own experiments and we give
Juergen Schmidhuber (29:54.040)
them a reward, an intrinsic reward in proportion
Lex Fridman (29:58.040)
to this depth of insight.
Lex Fridman (30:00.040)
And since they are trying to maximize the
Lex Fridman (30:05.040)
rewards they get they are suddenly motivated to
Juergen Schmidhuber (30:09.040)
come up with new action sequences, with new
Lex Fridman (30:13.040)
experiments that have the property that the data
Juergen Schmidhuber (30:16.040)
that is coming in as a consequence of these
Lex Fridman (30:19.040)
experiments has the property that they can learn
Juergen Schmidhuber (30:23.040)
something about, see a pattern in there which
Lex Fridman (30:25.040)
they hadn't seen yet before.
Lex Fridman (30:28.040)
So there is an idea of power play that you
Lex Fridman (30:31.040)
described, a training in general problem solver
Juergen Schmidhuber (30:34.040)
in this kind of way of looking for the unsolved
Lex Fridman (30:36.040)
problems.
Juergen Schmidhuber (30:37.040)
Yeah.
Lex Fridman (30:38.040)
Can you describe that idea a little further?
Juergen Schmidhuber (30:40.040)
It's another very simple idea.
Lex Fridman (30:42.040)
So normally what you do in computer science,
Juergen Schmidhuber (30:45.040)
you have some guy who gives you a problem and
Lex Fridman (30:50.040)
then there is a huge search space of potential
Juergen Schmidhuber (30:55.040)
solution candidates and you somehow try them
Lex Fridman (30:59.040)
out and you have more less sophisticated ways
Juergen Schmidhuber (31:02.040)
of moving around in that search space until
Lex Fridman (31:07.040)
you finally found a solution which you
Juergen Schmidhuber (31:10.040)
consider satisfactory.
Lex Fridman (31:12.040)
That's what most of computer science is about.
Juergen Schmidhuber (31:15.040)
Power play just goes one little step further
Lex Fridman (31:18.040)
and says let's not only search for solutions
Juergen Schmidhuber (31:23.040)
to a given problem but let's search to pairs of
Lex Fridman (31:28.040)
problems and their solutions where the system
Juergen Schmidhuber (31:31.040)
itself has the opportunity to phrase its own
Lex Fridman (31:35.040)
problem.
Lex Fridman (31:37.040)
So we are looking suddenly at pairs of
Lex Fridman (31:40.040)
problems and their solutions or modifications
Juergen Schmidhuber (31:44.040)
of the problem solver that is supposed to
Lex Fridman (31:47.040)
generate a solution to that new problem.
Lex Fridman (31:51.040)
And this additional degree of freedom allows
Lex Fridman (31:57.040)
us to build career systems that are like
Juergen Schmidhuber (32:00.040)
scientists in the sense that they not only
Lex Fridman (32:04.040)
try to solve and try to find answers to
Juergen Schmidhuber (32:07.040)
existing questions, no they are also free to
Lex Fridman (32:11.040)
pose their own questions.
Lex Fridman (32:13.040)
So if you want to build an artificial scientist
Lex Fridman (32:15.040)
you have to give it that freedom and power
Juergen Schmidhuber (32:17.040)
play is exactly doing that.
Lex Fridman (32:19.040)
So that's a dimension of freedom that's
Juergen Schmidhuber (32:22.040)
important to have but how hard do you think
Lex Fridman (32:25.040)
that, how multidimensional and difficult the
Juergen Schmidhuber (32:31.040)
space of then coming up with your own questions
Lex Fridman (32:34.040)
is.
Lex Fridman (32:35.040)
So it's one of the things that as human beings
Lex Fridman (32:37.040)
we consider to be the thing that makes us
Juergen Schmidhuber (32:40.040)
special, the intelligence that makes us special
Lex Fridman (32:42.040)
is that brilliant insight that can create
Juergen Schmidhuber (32:46.040)
something totally new.
Lex Fridman (32:48.040)
Yes.
Lex Fridman (32:49.040)
So now let's look at the extreme case, let's
Lex Fridman (32:52.040)
look at the set of all possible problems that
Juergen Schmidhuber (32:56.040)
you can formally describe which is infinite,
Lex Fridman (33:00.040)
which should be the next problem that a scientist
Juergen Schmidhuber (33:05.040)
or power play is going to solve.
Lex Fridman (33:08.040)
Well, it should be the easiest problem that
Juergen Schmidhuber (33:14.040)
goes beyond what you already know.
Lex Fridman (33:17.040)
So it should be the simplest problem that the
Juergen Schmidhuber (33:21.040)
current problem solver that you have which can
Lex Fridman (33:23.040)
already solve 100 problems that he cannot solve
Juergen Schmidhuber (33:28.040)
yet by just generalizing.
Lex Fridman (33:30.040)
So it has to be new, so it has to require a
Juergen Schmidhuber (33:33.040)
modification of the problem solver such that the
Lex Fridman (33:36.040)
new problem solver can solve this new thing but
Juergen Schmidhuber (33:39.040)
the old problem solver cannot do it and in
Lex Fridman (33:42.040)
addition to that we have to make sure that the
Juergen Schmidhuber (33:46.040)
problem solver doesn't forget any of the
Lex Fridman (33:49.040)
previous solutions.
Juergen Schmidhuber (33:50.040)
Right.
Lex Fridman (33:51.040)
And so by definition power play is now trying
Juergen Schmidhuber (33:54.040)
always to search in this pair of, in the set of
Lex Fridman (33:58.040)
pairs of problems and problems over modifications
Juergen Schmidhuber (34:02.040)
for a combination that minimize the time to
Lex Fridman (34:06.040)
achieve these criteria.
Lex Fridman (34:08.040)
So it's always trying to find the problem which
Lex Fridman (34:11.040)
is easiest to add to the repertoire.
Lex Fridman (34:14.040)
So just like grad students and academics and
Lex Fridman (34:18.040)
researchers can spend their whole career in a
Juergen Schmidhuber (34:20.040)
local minima stuck trying to come up with
Lex Fridman (34:24.040)
interesting questions but ultimately doing very
Juergen Schmidhuber (34:26.040)
little.
Lex Fridman (34:27.040)
Do you think it's easy in this approach of
Juergen Schmidhuber (34:31.040)
looking for the simplest unsolvable problem to
Lex Fridman (34:33.040)
get stuck in a local minima?
Juergen Schmidhuber (34:35.040)
Is not never really discovering new, you know
Lex Fridman (34:40.040)
really jumping outside of the 100 problems that
Lex Fridman (34:42.040)
you've already solved in a genuine creative way?
Lex Fridman (34:47.040)
No, because that's the nature of power play that
Juergen Schmidhuber (34:50.040)
it's always trying to break its current
Lex Fridman (34:53.040)
generalization abilities by coming up with a new
Juergen Schmidhuber (34:57.040)
problem which is beyond the current horizon.
Lex Fridman (35:00.040)
Just shifting the horizon of knowledge a little
Juergen Schmidhuber (35:04.040)
bit out there, breaking the existing rules such
Lex Fridman (35:08.040)
that the new thing becomes solvable but wasn't
Juergen Schmidhuber (35:11.040)
solvable by the old thing.
Lex Fridman (35:13.040)
So like adding a new axiom like what Gödel did
Juergen Schmidhuber (35:17.040)
when he came up with these new sentences, new
Lex Fridman (35:21.040)
theorems that didn't have a proof in the formal
Juergen Schmidhuber (35:23.040)
system which means you can add them to the
Lex Fridman (35:25.040)
repertoire hoping that they are not going to
Juergen Schmidhuber (35:31.040)
damage the consistency of the whole thing.
Lex Fridman (35:35.040)
So in the paper with the amazing title,
Juergen Schmidhuber (35:39.040)
Formal Theory of Creativity, Fun and Intrinsic
Lex Fridman (35:43.040)
Motivation, you talk about discovery as intrinsic
Juergen Schmidhuber (35:46.040)
reward, so if you view humans as intelligent
Lex Fridman (35:50.040)
agents, what do you think is the purpose and
Lex Fridman (35:53.040)
meaning of life for us humans?
Lex Fridman (35:56.040)
You've talked about this discovery, do you see
Lex Fridman (35:59.040)
humans as an instance of power play, agents?
Lex Fridman (36:04.040)
Humans are curious and that means they behave
Juergen Schmidhuber (36:10.040)
like scientists, not only the official scientists
Lex Fridman (36:13.040)
but even the babies behave like scientists and
Juergen Schmidhuber (36:16.040)
they play around with their toys to figure out
Lex Fridman (36:19.040)
how the world works and how it is responding to
Juergen Schmidhuber (36:22.040)
their actions and that's how they learn about
Lex Fridman (36:25.040)
gravity and everything.
Juergen Schmidhuber (36:27.040)
In 1990 we had the first systems like that which
Lex Fridman (36:31.040)
would just try to play around with the environment
Lex Fridman (36:34.040)
and come up with situations that go beyond what
Lex Fridman (36:38.040)
they knew at that time and then get a reward for
Juergen Schmidhuber (36:41.040)
creating these situations and then becoming more
Lex Fridman (36:44.040)
general problem solvers and being able to understand
Juergen Schmidhuber (36:47.040)
more of the world.
Lex Fridman (36:49.040)
I think in principle that curiosity strategy or
Juergen Schmidhuber (37:01.040)
more sophisticated versions of what I just
Lex Fridman (37:03.040)
described, they are what we have built in as well
Juergen Schmidhuber (37:07.040)
because evolution discovered that's a good way of
Lex Fridman (37:10.040)
exploring the unknown world and a guy who explores
Juergen Schmidhuber (37:13.040)
the unknown world has a higher chance of solving
Lex Fridman (37:17.040)
the mystery that he needs to survive in this world.
Juergen Schmidhuber (37:20.040)
On the other hand, those guys who were too curious
Lex Fridman (37:24.040)
they were weeded out as well so you have to find
Juergen Schmidhuber (37:27.040)
this trade off.
Lex Fridman (37:28.040)
Evolution found a certain trade off.
Juergen Schmidhuber (37:30.040)
Apparently in our society there is a certain
Lex Fridman (37:33.040)
percentage of extremely explorative guys and it
Juergen Schmidhuber (37:37.040)
doesn't matter if they die because many of the
Lex Fridman (37:40.040)
others are more conservative.
Juergen Schmidhuber (37:45.040)
It would be surprising to me if that principle of
Lex Fridman (37:51.040)
artificial curiosity wouldn't be present in almost
Juergen Schmidhuber (37:56.040)
exactly the same form here.
Lex Fridman (37:58.040)
In our brains.
Juergen Schmidhuber (38:00.040)
You are a bit of a musician and an artist.
Lex Fridman (38:03.040)
Continuing on this topic of creativity, what do you
Lex Fridman (38:08.040)
think is the role of creativity and intelligence?
Lex Fridman (38:10.040)
So you've kind of implied that it's essential for
Juergen Schmidhuber (38:15.040)
intelligence if you think of intelligence as a
Lex Fridman (38:18.040)
problem solving system, as ability to solve problems.
Lex Fridman (38:23.040)
But do you think it's essential, this idea of
Lex Fridman (38:26.040)
creativity?
Juergen Schmidhuber (38:27.040)
We never have a program, a sub program that is
Lex Fridman (38:32.040)
called creativity or something.
Juergen Schmidhuber (38:34.040)
It's just a side effect of what our problem solvers
Lex Fridman (38:37.040)
do. They are searching a space of problems, a space
Juergen Schmidhuber (38:41.040)
of candidates, of solution candidates until they
Lex Fridman (38:45.040)
hopefully find a solution to a given problem.
Lex Fridman (38:48.040)
But then there are these two types of creativity
Lex Fridman (38:50.040)
and both of them are now present in our machines.
Juergen Schmidhuber (38:54.040)
The first one has been around for a long time,
Lex Fridman (38:56.040)
which is human gives problem to machine, machine
Juergen Schmidhuber (39:00.040)
tries to find a solution to that.
Lex Fridman (39:03.040)
And this has been happening for many decades and
Juergen Schmidhuber (39:06.040)
for many decades machines have found creative
Lex Fridman (39:09.040)
solutions to interesting problems where humans were
Juergen Schmidhuber (39:12.040)
not aware of these particularly creative solutions
Lex Fridman (39:17.040)
but then appreciated that the machine found that.
Juergen Schmidhuber (39:20.040)
The second is the pure creativity.
Lex Fridman (39:23.040)
That I would call, what I just mentioned, I would
Juergen Schmidhuber (39:26.040)
call the applied creativity, like applied art where
Lex Fridman (39:30.040)
somebody tells you now make a nice picture of this
Juergen Schmidhuber (39:34.040)
Pope and you will get money for that.
Lex Fridman (39:37.040)
So here is the artist and he makes a convincing
Juergen Schmidhuber (39:40.040)
picture of the Pope and the Pope likes it and gives
Lex Fridman (39:43.040)
him the money.
Lex Fridman (39:46.040)
And then there is the pure creativity which is
Lex Fridman (39:49.040)
more like the power play and the artificial
Juergen Schmidhuber (39:51.040)
curiosity thing where you have the freedom to
Lex Fridman (39:54.040)
select your own problem.
Juergen Schmidhuber (39:57.040)
Like a scientist who defines his own question
Lex Fridman (40:02.040)
to study and so that is the pure creativity if you
Juergen Schmidhuber (40:06.040)
will as opposed to the applied creativity which
Lex Fridman (40:11.040)
serves another.
Lex Fridman (40:14.040)
And in that distinction there is almost echoes of
Lex Fridman (40:16.040)
narrow AI versus general AI.
Lex Fridman (40:19.040)
So this kind of constrained painting of a Pope
Lex Fridman (40:22.040)
seems like the approaches of what people are
Juergen Schmidhuber (40:28.040)
calling narrow AI and pure creativity seems to be,
Lex Fridman (40:33.040)
maybe I am just biased as a human but it seems to
Juergen Schmidhuber (40:35.040)
be an essential element of human level intelligence.
Lex Fridman (40:41.040)
Is that what you are implying?
Lex Fridman (40:44.040)
To a degree?
Lex Fridman (40:46.040)
If you zoom back a little bit and you just look
Juergen Schmidhuber (40:49.040)
at a general problem solving machine which is
Lex Fridman (40:51.040)
trying to solve arbitrary problems then this
Juergen Schmidhuber (40:54.040)
machine will figure out in the course of solving
Lex Fridman (40:57.040)
problems that it is good to be curious.
Lex Fridman (41:00.040)
So all of what I said just now about this prewired
Lex Fridman (41:04.040)
curiosity and this will to invent new problems
Juergen Schmidhuber (41:07.040)
that the system doesn't know how to solve yet
Lex Fridman (41:11.040)
should be just a byproduct of the general search.
Juergen Schmidhuber (41:15.040)
However, apparently evolution has built it into
Lex Fridman (41:20.040)
us because it turned out to be so successful,
Juergen Schmidhuber (41:24.040)
a prewiring, a bias, a very successful exploratory
Lex Fridman (41:29.040)
bias that we are born with.
Lex Fridman (41:34.040)
And you have also said that consciousness in the
Lex Fridman (41:36.040)
same kind of way may be a byproduct of problem solving.
Lex Fridman (41:41.040)
Do you find this an interesting byproduct?
Lex Fridman (41:45.040)
Do you think it is a useful byproduct?
Lex Fridman (41:47.040)
What are your thoughts on consciousness in general?
Lex Fridman (41:49.040)
Or is it simply a byproduct of greater and greater
Juergen Schmidhuber (41:53.040)
capabilities of problem solving that is similar
Lex Fridman (41:58.040)
to creativity in that sense?
Juergen Schmidhuber (42:01.040)
We never have a procedure called consciousness
Lex Fridman (42:04.040)
in our machines.
Juergen Schmidhuber (42:05.040)
However, we get as side effects of what these
Lex Fridman (42:09.040)
machines are doing things that seem to be closely
Juergen Schmidhuber (42:13.040)
related to what people call consciousness.
Lex Fridman (42:16.040)
So for example, already in 1990 we had simple
Juergen Schmidhuber (42:20.040)
systems which were basically recurrent networks
Lex Fridman (42:24.040)
and therefore universal computers trying to map
Juergen Schmidhuber (42:28.040)
incoming data into actions that lead to success.
Lex Fridman (42:33.040)
Maximizing reward in a given environment,
Juergen Schmidhuber (42:36.040)
always finding the charging station in time
Lex Fridman (42:40.040)
whenever the battery is low and negative signals
Juergen Schmidhuber (42:42.040)
are coming from the battery, always find the
Lex Fridman (42:45.040)
charging station in time without bumping against
Juergen Schmidhuber (42:48.040)
painful obstacles on the way.
Lex Fridman (42:50.040)
So complicated things but very easily motivated.
Lex Fridman (42:54.040)
And then we give these little guys a separate
Lex Fridman (43:00.040)
recurrent neural network which is just predicting
Juergen Schmidhuber (43:02.040)
what's happening if I do that and that.
Lex Fridman (43:04.040)
What will happen as a consequence of these
Juergen Schmidhuber (43:07.040)
actions that I'm executing.
Lex Fridman (43:09.040)
And it's just trained on the long and long history
Juergen Schmidhuber (43:11.040)
of interactions with the world.
Lex Fridman (43:13.040)
So it becomes a predictive model of the world
Juergen Schmidhuber (43:16.040)
basically.
Lex Fridman (43:18.040)
And therefore also a compressor of the observations
Juergen Schmidhuber (43:22.040)
of the world because whatever you can predict
Lex Fridman (43:25.040)
you don't have to store extra.
Lex Fridman (43:27.040)
So compression is a side effect of prediction.
Lex Fridman (43:30.040)
And how does this recurrent network compress?
Juergen Schmidhuber (43:33.040)
Well, it's inventing little subprograms, little
Lex Fridman (43:36.040)
subnetworks that stand for everything that
Juergen Schmidhuber (43:39.040)
frequently appears in the environment like
Lex Fridman (43:42.040)
bottles and microphones and faces, maybe lots of
Juergen Schmidhuber (43:45.040)
faces in my environment so I'm learning to create
Lex Fridman (43:50.040)
something like a prototype face and a new face
Juergen Schmidhuber (43:52.040)
comes along and all I have to encode are the
Lex Fridman (43:54.040)
deviations from the prototype.
Lex Fridman (43:56.040)
So it's compressing all the time the stuff that
Lex Fridman (43:58.040)
frequently appears.
Juergen Schmidhuber (44:00.040)
There's one thing that appears all the time that
Lex Fridman (44:05.040)
is present all the time when the agent is
Juergen Schmidhuber (44:07.040)
interacting with its environment which is the
Lex Fridman (44:10.040)
agent itself.
Lex Fridman (44:12.040)
But just for data compression reasons it is
Lex Fridman (44:15.040)
extremely natural for this recurrent network to
Juergen Schmidhuber (44:18.040)
come up with little subnetworks that stand for
Lex Fridman (44:21.040)
the properties of the agents, the hand, the other
Juergen Schmidhuber (44:26.040)
actuators and all the stuff that you need to
Lex Fridman (44:29.040)
better encode the data which is influenced by
Juergen Schmidhuber (44:32.040)
the actions of the agent.
Lex Fridman (44:34.040)
So there just as a side effect of data compression
Juergen Schmidhuber (44:39.040)
during problem solving you have internal self
Lex Fridman (44:43.040)
models.
Juergen Schmidhuber (44:45.040)
Now you can use this model of the world to plan
Lex Fridman (44:50.040)
your future and that's what we also have done
Juergen Schmidhuber (44:53.040)
since 1990.
Lex Fridman (44:54.040)
So the recurrent network which is the controller
Juergen Schmidhuber (44:57.040)
which is trying to maximize reward can use this
Lex Fridman (45:00.040)
model of the network of the world, this model
Juergen Schmidhuber (45:03.040)
network of the world, this predictive model of
Lex Fridman (45:05.040)
the world to plan ahead and say let's not do this
Juergen Schmidhuber (45:08.040)
action sequence, let's do this action sequence
Lex Fridman (45:10.040)
instead because it leads to more predicted
Juergen Schmidhuber (45:13.040)
reward.
Lex Fridman (45:14.040)
And whenever it is waking up these little
Juergen Schmidhuber (45:17.040)
subnetworks that stand for itself then it is
Lex Fridman (45:20.040)
thinking about itself and it is thinking about
Juergen Schmidhuber (45:23.040)
itself and it is exploring mentally the
Lex Fridman (45:28.040)
consequences of its own actions and now you tell
Juergen Schmidhuber (45:34.040)
me what is still missing.
Lex Fridman (45:36.040)
Missing the next, the gap to consciousness.
Juergen Schmidhuber (45:40.040)
There isn't.
Lex Fridman (45:41.040)
That's a really beautiful idea that if life is
Juergen Schmidhuber (45:45.040)
a collection of data and life is a process of
Lex Fridman (45:48.040)
compressing that data to act efficiently in that
Juergen Schmidhuber (45:54.040)
data you yourself appear very often.
Lex Fridman (45:57.040)
So it's useful to form compressions of yourself
Lex Fridman (46:00.040)
and it's a really beautiful formulation of what
Lex Fridman (46:03.040)
consciousness is a necessary side effect.
Juergen Schmidhuber (46:05.040)
It's actually quite compelling to me.
Lex Fridman (46:11.040)
You've described RNNs, developed LSTMs, long
Juergen Schmidhuber (46:16.040)
short term memory networks that are a type of
Lex Fridman (46:20.040)
recurrent neural networks that have gotten a lot
Juergen Schmidhuber (46:23.040)
of success recently.
Lex Fridman (46:24.040)
So these are networks that model the temporal
Juergen Schmidhuber (46:27.040)
aspects in the data, temporal patterns in the
Lex Fridman (46:30.040)
data and you've called them the deepest of the
Juergen Schmidhuber (46:34.040)
neural networks.
Lex Fridman (46:35.040)
So what do you think is the value of depth in
Lex Fridman (46:38.040)
the models that we use to learn?
Lex Fridman (46:41.040)
Since you mentioned the long short term memory
Lex Fridman (46:46.040)
and the LSTM I have to mention the names of the
Lex Fridman (46:50.040)
brilliant students who made that possible.
Juergen Schmidhuber (46:52.040)
First of all my first student ever Sepp Hochreiter
Lex Fridman (46:56.040)
who had fundamental insights already in his
Juergen Schmidhuber (46:58.040)
diploma thesis.
Lex Fridman (46:59.040)
Then Felix Geers who had additional important
Juergen Schmidhuber (47:03.040)
contributions.
Lex Fridman (47:04.040)
Alex Gray is a guy from Scotland who is mostly
Juergen Schmidhuber (47:08.040)
responsible for this CTC algorithm which is now
Lex Fridman (47:11.040)
often used to train the LSTM to do the speech
Juergen Schmidhuber (47:15.040)
recognition on all the Google Android phones and
Lex Fridman (47:18.040)
whatever and Siri and so on.
Lex Fridman (47:21.040)
So these guys without these guys I would be
Lex Fridman (47:26.040)
nothing.
Juergen Schmidhuber (47:27.040)
It's a lot of incredible work.
Lex Fridman (47:29.040)
What is now the depth?
Lex Fridman (47:30.040)
What is the importance of depth?
Lex Fridman (47:32.040)
Well most problems in the real world are deep in
Juergen Schmidhuber (47:36.040)
the sense that the current input doesn't tell you
Lex Fridman (47:40.040)
all you need to know about the environment.
Lex Fridman (47:44.040)
So instead you have to have a memory of what
Lex Fridman (47:48.040)
happened in the past and often important parts of
Juergen Schmidhuber (47:51.040)
that memory are dated.
Lex Fridman (47:54.040)
They are pretty old.
Lex Fridman (47:56.040)
So when you're doing speech recognition for
Lex Fridman (47:59.040)
example and somebody says 11 then that's about
Juergen Schmidhuber (48:05.040)
half a second or something like that which means
Lex Fridman (48:09.040)
it's already 50 time steps.
Lex Fridman (48:11.040)
And another guy or the same guy says 7.
Lex Fridman (48:15.040)
So the ending is the same even but now the
Juergen Schmidhuber (48:19.040)
system has to see the distinction between 7 and
Lex Fridman (48:22.040)
11 and the only way it can see the difference is
Juergen Schmidhuber (48:25.040)
it has to store that 50 steps ago there was an
Lex Fridman (48:30.040)
S or an L, 11 or 7.
Lex Fridman (48:34.040)
So there you have already a problem of depth 50
Lex Fridman (48:37.040)
because for each time step you have something
Juergen Schmidhuber (48:41.040)
like a virtual layer in the expanded unrolled
Lex Fridman (48:44.040)
version of this recurrent network which is doing
Juergen Schmidhuber (48:46.040)
the speech recognition.
Lex Fridman (48:48.040)
So these long time lags they translate into
Juergen Schmidhuber (48:51.040)
problem depth.
Lex Fridman (48:53.040)
And most problems in this world are such that
Juergen Schmidhuber (48:57.040)
you really have to look far back in time to
Lex Fridman (49:01.040)
understand what is the problem and to solve it.
Lex Fridman (49:05.040)
But just like with LSTMs you don't necessarily
Lex Fridman (49:08.040)
need to when you look back in time remember every
Juergen Schmidhuber (49:11.040)
aspect you just need to remember the important
Lex Fridman (49:13.040)
aspects.
Juergen Schmidhuber (49:14.040)
That's right.
Lex Fridman (49:15.040)
The network has to learn to put the important
Juergen Schmidhuber (49:18.040)
stuff into memory and to ignore the unimportant
Lex Fridman (49:22.040)
noise.
Lex Fridman (49:23.040)
But in that sense deeper and deeper is better
Lex Fridman (49:28.040)
or is there a limitation?
Juergen Schmidhuber (49:30.040)
I mean LSTM is one of the great examples of
Lex Fridman (49:34.040)
architectures that do something beyond just
Juergen Schmidhuber (49:40.040)
deeper and deeper networks.
Lex Fridman (49:42.040)
There's clever mechanisms for filtering data,
Juergen Schmidhuber (49:45.040)
for remembering and forgetting.
Lex Fridman (49:47.040)
So do you think that kind of thinking is
Lex Fridman (49:50.040)
necessary?
Lex Fridman (49:51.040)
If you think about LSTMs as a leap, a big leap
Juergen Schmidhuber (49:54.040)
forward over traditional vanilla RNNs, what do
Lex Fridman (49:57.040)
you think is the next leap within this context?
Lex Fridman (50:02.040)
So LSTM is a very clever improvement but LSTM
Lex Fridman (50:06.040)
still don't have the same kind of ability to see
Juergen Schmidhuber (50:10.040)
far back in the past as us humans do.
Lex Fridman (50:14.040)
The credit assignment problem across way back
Juergen Schmidhuber (50:18.040)
not just 50 time steps or 100 or 1000 but
Lex Fridman (50:22.040)
millions and billions.
Juergen Schmidhuber (50:24.040)
It's not clear what are the practical limits of
Lex Fridman (50:28.040)
the LSTM when it comes to looking back.
Juergen Schmidhuber (50:31.040)
Already in 2006 I think we had examples where
Lex Fridman (50:35.040)
it not only looked back tens of thousands of
Juergen Schmidhuber (50:38.040)
steps but really millions of steps.
Lex Fridman (50:41.040)
And Juan Perez Ortiz in my lab I think was the
Juergen Schmidhuber (50:45.040)
first author of a paper where we really, was it
Lex Fridman (50:49.040)
2006 or something, had examples where it learned
Juergen Schmidhuber (50:53.040)
to look back for more than 10 million steps.
Lex Fridman (50:57.040)
So for most problems of speech recognition it's
Juergen Schmidhuber (51:01.040)
not necessary to look that far back but there
Lex Fridman (51:05.040)
are examples where it does.
Juergen Schmidhuber (51:07.040)
Now the looking back thing, that's rather easy
Lex Fridman (51:11.040)
because there is only one past but there are
Juergen Schmidhuber (51:15.040)
many possible futures and so a reinforcement
Lex Fridman (51:19.040)
learning system which is trying to maximize its
Juergen Schmidhuber (51:22.040)
future expected reward and doesn't know yet which
Lex Fridman (51:26.040)
of these many possible futures should I select
Juergen Schmidhuber (51:29.040)
given this one single past is facing problems
Lex Fridman (51:33.040)
that the LSTM by itself cannot solve.
Lex Fridman (51:36.040)
So the LSTM is good for coming up with a compact
Lex Fridman (51:40.040)
representation of the history and observations
Lex Fridman (51:44.040)
and actions so far but now how do you plan in an
Lex Fridman (51:49.040)
efficient and good way among all these, how do
Juergen Schmidhuber (51:54.040)
you select one of these many possible action
Lex Fridman (51:57.040)
sequences that a reinforcement learning system
Juergen Schmidhuber (52:00.040)
has to consider to maximize reward in this
Lex Fridman (52:04.040)
unknown future?
Juergen Schmidhuber (52:06.040)
We have this basic setup where you have one
Lex Fridman (52:10.040)
recurrent network which gets in the video and
Juergen Schmidhuber (52:14.040)
the speech and whatever and it's executing
Lex Fridman (52:17.040)
actions and it's trying to maximize reward so
Juergen Schmidhuber (52:20.040)
there is no teacher who tells it what to do at
Lex Fridman (52:23.040)
which point in time.
Lex Fridman (52:25.040)
And then there's the other network which is
Lex Fridman (52:29.040)
just predicting what's going to happen if I do
Juergen Schmidhuber (52:32.040)
that and that and that could be an LSTM network
Lex Fridman (52:35.040)
and it learns to look back all the way to make
Juergen Schmidhuber (52:38.040)
better predictions of the next time step.
Lex Fridman (52:41.040)
So essentially although it's predicting only the
Juergen Schmidhuber (52:44.040)
next time step it is motivated to learn to put
Lex Fridman (52:48.040)
into memory something that happened maybe a
Juergen Schmidhuber (52:51.040)
million steps ago because it's important to
Lex Fridman (52:54.040)
memorize that if you want to predict that at the
Juergen Schmidhuber (52:57.040)
next time step, the next event.
Lex Fridman (52:59.040)
Now how can a model of the world like that, a
Juergen Schmidhuber (53:03.040)
predictive model of the world be used by the
Lex Fridman (53:06.040)
first guy?
Juergen Schmidhuber (53:07.040)
Let's call it the controller and the model, the
Lex Fridman (53:10.040)
controller and the model.
Lex Fridman (53:12.040)
How can the model be used by the controller to
Lex Fridman (53:15.040)
efficiently select among these many possible
Lex Fridman (53:18.040)
futures?
Lex Fridman (53:19.040)
The naive way we had about 30 years ago was
Juergen Schmidhuber (53:22.040)
let's just use the model of the world as a stand
Lex Fridman (53:26.040)
in, as a simulation of the world and millisecond
Juergen Schmidhuber (53:30.040)
by millisecond we plan the future and that means
Lex Fridman (53:33.040)
we have to roll it out really in detail and it
Juergen Schmidhuber (53:36.040)
will work only if the model is really good and
Lex Fridman (53:39.040)
it will still be inefficient because we have to
Juergen Schmidhuber (53:42.040)
look at all these possible futures and there are
Lex Fridman (53:45.040)
so many of them.
Lex Fridman (53:46.040)
So instead what we do now since 2015 in our CM
Lex Fridman (53:49.040)
systems, controller model systems, we give the
Juergen Schmidhuber (53:52.040)
controller the opportunity to learn by itself how
Lex Fridman (53:56.040)
to use the potentially relevant parts of the M,
Juergen Schmidhuber (54:00.040)
of the model network to solve new problems more
Lex Fridman (54:04.040)
quickly.
Lex Fridman (54:05.040)
And if it wants to, it can learn to ignore the M
Lex Fridman (54:09.040)
and sometimes it's a good idea to ignore the M
Juergen Schmidhuber (54:12.040)
because it's really bad, it's a bad predictor in
Lex Fridman (54:15.040)
this particular situation of life where the
Juergen Schmidhuber (54:19.040)
controller is currently trying to maximize reward.
Lex Fridman (54:22.040)
However, it can also learn to address and exploit
Juergen Schmidhuber (54:26.040)
some of the subprograms that came about in the
Lex Fridman (54:31.040)
model network through compressing the data by
Juergen Schmidhuber (54:35.040)
predicting it.
Lex Fridman (54:36.040)
So it now has an opportunity to reuse that code,
Juergen Schmidhuber (54:40.040)
the algorithmic information in the model network
Lex Fridman (54:44.040)
to reduce its own search space such that it can
Juergen Schmidhuber (54:49.040)
solve a new problem more quickly than without the
Lex Fridman (54:52.040)
model.
Juergen Schmidhuber (54:53.040)
Compression.
Lex Fridman (54:54.040)
So you're ultimately optimistic and excited about
Juergen Schmidhuber (54:59.040)
the power of RL, of reinforcement learning in the
Lex Fridman (55:03.040)
context of real systems.
Juergen Schmidhuber (55:05.040)
Absolutely, yeah.
Lex Fridman (55:06.040)
So you see RL as a potential having a huge impact
Juergen Schmidhuber (55:11.040)
beyond just sort of the M part is often developed on
Lex Fridman (55:16.040)
supervised learning methods.
Juergen Schmidhuber (55:19.040)
You see RL as a for problems of self driving cars
Lex Fridman (55:25.040)
or any kind of applied cyber robotics.
Juergen Schmidhuber (55:28.040)
That's the correct interesting direction for
Lex Fridman (55:32.040)
research in your view?
Juergen Schmidhuber (55:34.040)
I do think so.
Lex Fridman (55:35.040)
We have a company called Nasence which has applied
Juergen Schmidhuber (55:40.040)
reinforcement learning to little Audis which learn
Lex Fridman (55:45.040)
to park without a teacher.
Juergen Schmidhuber (55:47.040)
The same principles were used of course.
Lex Fridman (55:50.040)
So these little Audis, they are small, maybe like
Juergen Schmidhuber (55:54.040)
that, so much smaller than the real Audis.
Lex Fridman (55:57.040)
But they have all the sensors that you find in the
Juergen Schmidhuber (56:00.040)
real Audis.
Lex Fridman (56:01.040)
You find the cameras, the LIDAR sensors.
Juergen Schmidhuber (56:03.040)
They go up to 120 kilometers an hour if they want
Lex Fridman (56:08.040)
to.
Lex Fridman (56:09.040)
And they have pain sensors basically and they don't
Lex Fridman (56:13.040)
want to bump against obstacles and other Audis and
Lex Fridman (56:17.040)
so they must learn like little babies to park.
Lex Fridman (56:21.040)
Take the raw vision input and translate that into
Juergen Schmidhuber (56:25.040)
actions that lead to successful parking behavior
Lex Fridman (56:28.040)
which is a rewarding thing.
Lex Fridman (56:30.040)
And yes, they learn that.
Lex Fridman (56:32.040)
So we have examples like that and it's only in the
Juergen Schmidhuber (56:36.040)
beginning.
Lex Fridman (56:37.040)
This is just the tip of the iceberg and I believe the
Juergen Schmidhuber (56:40.040)
next wave of AI is going to be all about that.
Lex Fridman (56:44.040)
So at the moment, the current wave of AI is about
Juergen Schmidhuber (56:48.040)
passive pattern observation and prediction and that's
Lex Fridman (56:53.040)
what you have on your smartphone and what the major
Juergen Schmidhuber (56:56.040)
companies on the Pacific Rim are using to sell you
Lex Fridman (57:00.040)
ads to do marketing.
Juergen Schmidhuber (57:02.040)
That's the current sort of profit in AI and that's
Lex Fridman (57:05.040)
only one or two percent of the world economy.
Juergen Schmidhuber (57:08.040)
Which is big enough to make these companies pretty
Lex Fridman (57:12.040)
much the most valuable companies in the world.
Lex Fridman (57:15.040)
But there's a much, much bigger fraction of the
Lex Fridman (57:19.040)
economy going to be affected by the next wave which
Juergen Schmidhuber (57:22.040)
is really about machines that shape the data through
Lex Fridman (57:26.040)
their own actions.
Lex Fridman (57:27.040)
Do you think simulation is ultimately the biggest
Lex Fridman (57:31.040)
way that those methods will be successful in the next
Lex Fridman (57:35.040)
10, 20 years?
Lex Fridman (57:36.040)
We're not talking about 100 years from now.
Juergen Schmidhuber (57:38.040)
We're talking about sort of the near term impact of
Lex Fridman (57:41.040)
RL.
Lex Fridman (57:42.040)
Do you think really good simulation is required or
Lex Fridman (57:45.040)
is there other techniques like imitation learning,
Lex Fridman (57:48.040)
observing other humans operating in the real world?
Lex Fridman (57:53.040)
Where do you think the success will come from?
Lex Fridman (57:57.040)
So at the moment, we have a tendency of using physics
Lex Fridman (58:02.040)
simulations to learn behavior from machines that
Juergen Schmidhuber (58:07.040)
learn to solve problems that humans also do not know
Lex Fridman (58:11.040)
how to solve.
Juergen Schmidhuber (58:12.040)
However, this is not the future because the future is
Lex Fridman (58:16.040)
in what little babies do.
Juergen Schmidhuber (58:18.040)
They don't use a physics engine to simulate the
Lex Fridman (58:21.040)
world.
Juergen Schmidhuber (58:22.040)
No, they learn a predictive model of the world which
Lex Fridman (58:26.040)
maybe sometimes is wrong in many ways but captures
Juergen Schmidhuber (58:31.040)
all kinds of important abstract high level predictions
Lex Fridman (58:34.040)
which are really important to be successful.
Lex Fridman (58:37.040)
And that's what was the future 30 years ago when we
Lex Fridman (58:42.040)
started that type of research but it's still the future
Lex Fridman (58:45.040)
and now we know much better how to go there to move
Lex Fridman (58:49.040)
forward and to really make working systems based on
Juergen Schmidhuber (58:54.040)
that where you have a learning model of the world,
Lex Fridman (58:57.040)
a model of the world that learns to predict what's
Juergen Schmidhuber (58:59.040)
going to happen if I do that and that.
Lex Fridman (59:01.040)
And then the controller uses that model to more
Juergen Schmidhuber (59:07.040)
quickly learn successful action sequences.
Lex Fridman (59:10.040)
And then of course always this curiosity thing.
Juergen Schmidhuber (59:13.040)
In the beginning, the model is stupid so the
Lex Fridman (59:15.040)
controller should be motivated to come up with
Juergen Schmidhuber (59:18.040)
experiments with action sequences that lead to data
Lex Fridman (59:21.040)
that improve the model.
Lex Fridman (59:23.040)
Do you think improving the model, constructing an
Lex Fridman (59:27.040)
understanding of the world in this connection is
Juergen Schmidhuber (59:30.040)
now the popular approaches that have been successful
Lex Fridman (59:34.040)
are grounded in ideas of neural networks.
Lex Fridman (59:38.040)
But in the 80s with expert systems, there's
Lex Fridman (59:41.040)
symbolic AI approaches which to us humans are more
Juergen Schmidhuber (59:45.040)
intuitive in the sense that it makes sense that you
Lex Fridman (59:49.040)
build up knowledge in this knowledge representation.
Lex Fridman (59:52.040)
What kind of lessons can we draw into our current
Lex Fridman (59:54.040)
approaches from expert systems from symbolic AI?
🔗 相关节目