WEBVTT

00:00:16.256 --> 00:00:19.993
Intelligence, what is it?

00:00:19.993 --> 00:00:24.610
If we take a look back at the history
of how intelligence is being viewed,

00:00:24.610 --> 00:00:31.187
one seminal example has been
Edsger Dijkstra's famous quote

00:00:31.187 --> 00:00:34.898
that the question of
whether a machine can think

00:00:34.898 --> 00:00:37.779
is about as interesting as the question of

00:00:37.779 --> 00:00:40.942
whether a submarine can swim.

00:00:41.412 --> 00:00:46.832
Now, Edsger Dijkstra, when he wrote this,
intended it as a criticism

00:00:46.832 --> 00:00:51.519
of early pioneers of computer science
like Alan Turing.

00:00:52.679 --> 00:00:55.532
However, if you take a look back

00:00:55.532 --> 00:00:59.406
and think about what have been
the most empowering innovations

00:00:59.406 --> 00:01:03.007
that enable us to build
artificial machines that swim

00:01:03.507 --> 00:01:06.278
and artificial machines that think,

00:01:06.388 --> 00:01:10.700
you find that it was only through
understanding the underlying

00:01:10.810 --> 00:01:15.916
physical mechanisms of swimming
and flight that we were able

00:01:15.916 --> 00:01:18.442
to build these machines.

00:01:18.442 --> 00:01:22.019
And so, several years ago,
I undertook a program

00:01:22.019 --> 00:01:26.487
to try to understand the fundamental
physical mechanisms

00:01:26.487 --> 00:01:29.065
underlying intelligence.

00:01:30.275 --> 00:01:32.338
Let's take a step back.

00:01:32.364 --> 00:01:35.613
Let's first begin
with a thought experiment.

00:01:35.613 --> 00:01:38.019
Pretend that you're an alien race

00:01:38.019 --> 00:01:42.854
that doesn't know anything
about Earth biology or Earth neuroscience

00:01:42.854 --> 00:01:46.565
or Earth intelligence, but you have
amazing telescopes

00:01:46.655 --> 00:01:51.021
and you're able to watch the Earth
and you have amazingly long lives

00:01:51.021 --> 00:01:55.943
so you're able to watch the Earth
over millions, even billions of years.

00:01:55.943 --> 00:02:00.361
And you observe a really strange effect,

00:02:00.361 --> 00:02:03.623
you observe that over the course
of the millennia,

00:02:03.649 --> 00:02:09.798
Earth is continually bombarded
with asteroids up until a point

00:02:09.798 --> 00:02:13.201
and that at some point,
corresponding roughly

00:02:13.221 --> 00:02:18.980
to our year 2000 AD, asteroids that are
on a collision course with the Earth,

00:02:19.241 --> 00:02:23.293
that otherwise would have collided,
mysteriously get deflected

00:02:23.893 --> 00:02:26.643
or detonate before they can hit the Earth.

00:02:26.863 --> 00:02:30.364
Now, of course, as Earthlings,
we know the reason would be

00:02:30.364 --> 00:02:35.088
that we're trying to save ourselves,
we're trying to prevent an impact.

00:02:35.088 --> 00:02:38.087
But if you're an alien race
that doesn't know any of this,

00:02:38.087 --> 00:02:40.664
that doesn't have any concept
of Earth intelligence,

00:02:40.664 --> 00:02:42.543
you'd be forced to put together

00:02:42.543 --> 00:02:46.945
a physical theory that explains how,
up until a certain point in time,

00:02:47.925 --> 00:02:51.508
asteroids thad would demolish
the surface of the planet,

00:02:52.258 --> 00:02:55.361
mysteriously stop doing that.

00:02:55.361 --> 00:02:59.610
So, I claim that this is the same question

00:02:59.610 --> 00:03:03.112
as understanding the physical
nature of intelligence.

00:03:03.762 --> 00:03:08.863
So, in this program that I undertook
years ago, I've looked at a variety

00:03:08.863 --> 00:03:13.766
of different threads in crossed science
across a variety of disciplines,

00:03:13.766 --> 00:03:19.280
pointing, I think, towards a single
underlying mechanism for intelligence.

00:03:19.910 --> 00:03:22.024
In cosmology, for example,

00:03:22.314 --> 00:03:24.968
there has been a variety
of different threads of evidence

00:03:24.968 --> 00:03:29.695
that our universe appears to be
finely tuned for the development

00:03:29.695 --> 00:03:33.360
of intelligence, and in particular,
for the development

00:03:33.360 --> 00:03:38.941
of universal states that maximize
the diversity of possible futures.

00:03:38.941 --> 00:03:44.063
In gameplay, for example in Go,
everyone remembers in 1997

00:03:44.403 --> 00:03:48.120
when IBM's Deep Blue beat
Gary Kasparov at chess.

00:03:48.480 --> 00:03:51.934
Fewer people are aware
that in the past ten year or so,

00:03:51.934 --> 00:03:56.137
the game of Go, arguably a much more
challenging game because it has

00:03:56.137 --> 00:04:00.804
a much higher branching factor,
has also started to succumb to computer

00:04:00.804 --> 00:04:03.862
game players for the same reason.

00:04:03.862 --> 00:04:06.529
The best techniques, right now,
for computers playing Go,

00:04:06.529 --> 00:04:11.651
are techniques that try to maximize
future options during gameplay.

00:04:12.091 --> 00:04:15.693
Finally, in robotic motion planning,

00:04:15.693 --> 00:04:17.863
there has been a variety
of recent techniques

00:04:17.863 --> 00:04:22.768
that have tried to take advantage
of abilities of robots to maximize

00:04:23.018 --> 00:04:27.116
future freedom of action in order
to accomplish complex tasks.

00:04:27.496 --> 00:04:31.340
And so, taking all of these different
threads and putting them together,

00:04:31.730 --> 00:04:36.090
I asked, starting several years ago,
is there an underlying mechanism

00:04:36.340 --> 00:04:40.249
for intelligence that we can factor out
of all of these different threads?

00:04:40.509 --> 00:04:45.250
Is there, as it were,
a single equation for intelligence?

00:04:46.990 --> 00:04:50.442
And the answer, I believe, is yes.

00:04:50.468 --> 00:04:57.469
What you're seeing is probably the closest
equivalent to an E=mc2 for intelligence

00:04:57.469 --> 00:05:00.072
that I certainly have ever seen.

00:05:00.098 --> 00:05:02.276
So, what you're seeing here

00:05:02.371 --> 00:05:07.835
is a statement of correspondence
that intelligence is a Force (F)

00:05:08.765 --> 00:05:13.390
that acts so as to maximize
future freedom of action;

00:05:13.590 --> 00:05:17.324
It acts to maximize future freedom
of action or keep options open

00:05:17.324 --> 00:05:19.654
with some strength (T),

00:05:19.904 --> 00:05:24.955
with the amount of the diversity
of possible accessible futures (S),

00:05:24.985 --> 00:05:28.295
up to some future time horizon (Ƭ).

00:05:28.321 --> 00:05:30.613
In short, intelligence doesn't like

00:05:30.613 --> 00:05:34.498
to get trapped, intelligence tries
to maximize future freedom of action

00:05:34.498 --> 00:05:39.526
and keep options open.
And so, given this one equation

00:05:39.526 --> 00:05:42.445
it's natural to ask:
So, what can you do with this?

00:05:42.445 --> 00:05:45.856
How predictive is it? Does it predict
human-level intelligence?

00:05:45.856 --> 00:05:48.609
Does it predict artificial intelligence?

00:05:48.609 --> 00:05:53.726
So, I'm going to show you now a video
that will, I think, demonstrate

00:05:54.006 --> 00:05:58.294
some of the amazing applications
of just this single equation.

00:06:00.004 --> 00:06:03.357
Recent research in cosmology
has suggested that universes

00:06:03.357 --> 00:06:07.531
that produce more disorder or "entropy"
over their lifetimes should tend

00:06:07.531 --> 00:06:11.269
to have more favorable conditions
for the existence of intelligent beings

00:06:11.589 --> 00:06:13.445
such as ourselves.

00:06:13.445 --> 00:06:15.763
But what if that tentative
cosmological connection

00:06:15.763 --> 00:06:19.449
between entropy and intelligence
hints at a deeper relationship?

00:06:19.449 --> 00:06:22.012
What if intelligent behavior
doesn't just correlate

00:06:22.012 --> 00:06:26.226
with the production of long-term entropy,
but actually emerges directly from it?

00:06:26.576 --> 00:06:30.114
To find out, we developed
a software engine called ENTROPICA

00:06:30.114 --> 00:06:34.184
designed to maximize the production
of long-term entropy of any system

00:06:34.184 --> 00:06:36.001
that it finds itself in.

00:06:36.001 --> 00:06:40.645
Amazingly, ENTROPICA was able to pass
multiple animal intelligence tests,

00:06:40.645 --> 00:06:43.766
play human games
and even earn money trading stocks;

00:06:43.766 --> 00:06:46.157
all without being instructed to do so.

00:06:46.157 --> 00:06:48.610
Here are some examples
of ENTROPICA in action:

00:06:48.610 --> 00:06:52.164
just like a human standing upright
without falling over, here we see

00:06:52.254 --> 00:06:56.225
ENTROPICA automatically
balancing a pole using a cart.

00:06:56.225 --> 00:07:00.342
This behavior is remarkable, in part,
because we never gave ENTROPICA a goal,

00:07:00.342 --> 00:07:03.754
it simply decided on its own
to balance the pole.

00:07:03.754 --> 00:07:06.997
This balancing ability would have
applications for humanoid robotics

00:07:06.997 --> 00:07:09.277
and human assistive technologies.

00:07:09.625 --> 00:07:12.679
Just as some animals can use
objects in their environments

00:07:12.679 --> 00:07:15.056
as tools to reach into narrow spaces,

00:07:15.056 --> 00:07:18.967
here we see that ENTROPICA,
again on its own initiative,

00:07:18.967 --> 00:07:22.192
was able to move a large disk,
representing an animal,

00:07:22.192 --> 00:07:25.450
around so as to cause a small disk,
representing a tool,

00:07:25.450 --> 00:07:28.346
to reach into a confined space
holding a third disk

00:07:28.346 --> 00:07:31.953
and release the third disk
from its initially fixed position.

00:07:31.953 --> 00:07:36.658
This tool usability would have application
for smart manufacturing and agriculture.

00:07:37.338 --> 00:07:40.295
In addition, just as some other animals
are able to cooperate

00:07:40.295 --> 00:07:44.043
by pulling opposite ends of a rope
at the same time to release food,

00:07:44.043 --> 00:07:46.740
here we see that ENTROPICA
is able to accomplish

00:07:46.740 --> 00:07:48.497
a model version of that task.

00:07:48.497 --> 00:07:52.136
This cooperative ability has interesting
implications for economic planning

00:07:52.136 --> 00:07:55.450
and a variety of other fields.

00:07:55.450 --> 00:07:59.288
ENTROPICA is broadly applicable
to a variety of domains.

00:07:59.288 --> 00:08:03.661
For example, here we see it successfully
playing a game of pong against itself

00:08:04.371 --> 00:08:06.347
illustrating its potential for gaming.

00:08:08.103 --> 00:08:09.794
Here, we see ENTROPICA orchestrating

00:08:09.794 --> 00:08:13.289
new connections on a social network
where friends are constantly

00:08:13.289 --> 00:08:17.341
falling out of touch and successfully
keeping the network well connected.

00:08:17.671 --> 00:08:22.252
This same network orchestration ability
also has applications in health care,

00:08:22.252 --> 00:08:25.404
energy and intelligence.

00:08:25.404 --> 00:08:28.816
Here we see ENTROPICA directing
the paths of a fleet of ships

00:08:28.816 --> 00:08:33.260
successfully discovering and utilizing
the Panama Canal to globally extend

00:08:33.260 --> 00:08:35.951
its reach from the Atlantic
to the Pacific.

00:08:35.951 --> 00:08:39.253
By the same token, ENTROPICA
is broadly applicable to problems

00:08:39.253 --> 00:08:43.496
in autonomous defense,
logistics and transportation.

00:08:44.566 --> 00:08:49.370
Finally, here we see ENTROPICA
spontaneously discovering and executing

00:08:49.370 --> 00:08:53.843
a buy low, sell high strategy
on a simulated range traded stock

00:08:53.843 --> 00:08:57.369
successfully growing assets
under management exponentially.

00:08:57.369 --> 00:09:00.513
This risk management ability
would have broad applications

00:09:00.513 --> 00:09:02.911
in finance and insurance.

00:09:08.475 --> 00:09:12.067
So, what you've just seen
is that a variety

00:09:12.114 --> 00:09:16.178
of signature human
intelligent cognitive behavior

00:09:16.204 --> 00:09:18.895
such us tool use and walking upright

00:09:19.490 --> 00:09:24.005
and social cooperation, all follow
from a single equation

00:09:24.255 --> 00:09:29.452
which drives a system to maximize
its future freedom of action.

00:09:30.242 --> 00:09:33.237
Now, there's a profound irony here.

00:09:33.237 --> 00:09:37.663
Going back to the beginning
of the usage of the term robot,

00:09:38.503 --> 00:09:41.499
the play RUR,

00:09:41.499 --> 00:09:46.515
there was always a concept
that if we develop machine, intelligence,

00:09:47.345 --> 00:09:52.622
there will be a cybernetic revolt,
that machines would rise up against us.

00:09:53.452 --> 00:09:58.731
One major consequence of this work
is that maybe all of these decades

00:09:58.731 --> 00:10:02.872
we've had the whole concept
of cybernetic revolt in reverse.

00:10:03.772 --> 00:10:06.918
It's not that machines
first become intelligent

00:10:06.918 --> 00:10:11.283
and then megalomaniacal,
and try to take over the world.

00:10:11.283 --> 00:10:15.621
It's quite the opposite:
that the urge to take control

00:10:15.621 --> 00:10:19.701
of all possible futures
is a more fundamental principle

00:10:20.071 --> 00:10:23.949
than that of intelligence;
that general intelligence may, in fact,

00:10:23.949 --> 00:10:28.456
emerge directly from this sort
of control grabbing,

00:10:28.456 --> 00:10:31.209
rather than vice versa.

00:10:32.589 --> 00:10:36.312
Another important consequence
is goal seeking.

00:10:36.652 --> 00:10:42.443
I'm often asked how does the ability
to seek goals follow from this framework

00:10:42.643 --> 00:10:43.747
and the answer is:

00:10:43.747 --> 00:10:48.203
the ability to seek goals, for example
if you're playing the game of chess,

00:10:48.543 --> 00:10:53.252
to try to win that game of chess
in order to accomplish worldly goods

00:10:53.252 --> 00:10:55.599
and accomplishments outside of that game,

00:10:55.809 --> 00:10:59.124
will follow directly from this
in the following sense:

00:10:59.554 --> 00:11:03.855
Just like you would travel
through a tunnel, a bottleneck,

00:11:03.855 --> 00:11:07.050
in your future path space
in order to achieve many other

00:11:07.050 --> 00:11:11.178
diverse objectives later on
or just like you would invest

00:11:11.178 --> 00:11:15.350
in a financial security reducing
your short term liquidity

00:11:15.350 --> 00:11:17.825
in order to increase your wealth
over the long term,

00:11:17.825 --> 00:11:21.613
goal seeking emerges directly
from a long term drive

00:11:21.613 --> 00:11:25.571
to increase future freedom of action.

00:11:25.571 --> 00:11:29.881
Finally, the famous physicist
Richard Feynman once wrote

00:11:30.361 --> 00:11:34.703
that if human civilization were destroyed
and you could pass only a single concept

00:11:34.703 --> 00:11:38.164
on to our descendents
to help them rebuild civilization,

00:11:38.524 --> 00:11:41.620
that concept should be
that all matter around us

00:11:42.240 --> 00:11:45.506
is made out of tiny elements
that attract each other

00:11:45.506 --> 00:11:48.101
when they're far apart,
but repel each other

00:11:48.341 --> 00:11:50.096
when they're close together.

00:11:50.126 --> 00:11:53.152
My equivalent to that statement
to pass on to descendents

00:11:53.472 --> 00:11:55.915
to help them build
artificial intelligence,

00:11:55.915 --> 00:11:59.988
or to help them to understand
human intelligence, is the following:

00:12:00.108 --> 00:12:03.541
Intelligence should be viewed
as a physical process

00:12:03.541 --> 00:12:06.492
that tries to maximize
future freedom of action

00:12:06.492 --> 00:12:09.624
and avoid constraints in its own future.

00:12:10.194 --> 00:12:11.452
Thank you very much.

00:12:11.478 --> 00:12:14.478
(Applause)