0:00:16.256,0:00:19.993
Intelligence, what is it?

0:00:19.993,0:00:24.610
If we take a look back at the history[br]of how intelligence is being viewed,

0:00:24.610,0:00:31.187
one seminal example has been[br]Edsger Dijkstra's famous quote

0:00:31.187,0:00:34.898
that the question of[br]whether a machine can think

0:00:34.898,0:00:37.779
is about as interesting as the question of

0:00:37.779,0:00:40.942
whether a submarine can swim.

0:00:41.412,0:00:46.832
Now, Edsger Dijkstra, when he wrote this,[br]intended it as a criticism

0:00:46.832,0:00:51.519
of early pioneers of computer science[br]like Alan Turing.

0:00:52.679,0:00:55.532
However, if you take a look back

0:00:55.532,0:00:59.406
and think about what have been[br]the most empowering innovations

0:00:59.406,0:01:03.007
that enabled us to build[br]artificial machines that swim

0:01:03.507,0:01:06.278
and artificial machines that [fly],

0:01:06.388,0:01:10.700
you find that it was only through[br]understanding the underlying

0:01:10.810,0:01:15.916
physical mechanisms of swimming[br]and flight that we were able

0:01:15.916,0:01:18.442
to build these machines.

0:01:18.442,0:01:22.019
And so, several years ago,[br]I undertook a program

0:01:22.019,0:01:26.487
to try to understand the fundamental[br]physical mechanisms

0:01:26.487,0:01:29.065
underlying intelligence.

0:01:30.275,0:01:32.338
Let's take a step back.

0:01:32.364,0:01:35.613
Let's first begin[br]with a thought experiment.

0:01:35.613,0:01:38.019
Pretend that you're an alien race

0:01:38.019,0:01:42.854
that doesn't know anything[br]about Earth biology or Earth neuroscience

0:01:42.854,0:01:46.565
or Earth intelligence, but you have[br]amazing telescopes

0:01:46.655,0:01:51.021
and you're able to watch the Earth[br]and you have amazingly long lives

0:01:51.021,0:01:55.943
so you're able to watch the Earth[br]over millions, even billions of years.

0:01:55.943,0:02:00.361
And you observe a really strange effect,

0:02:00.361,0:02:03.623
you observe that over the course[br]of the millennia,

0:02:03.649,0:02:09.798
Earth is continually bombarded[br]with asteroids up until a point

0:02:09.798,0:02:13.201
and that at some point,[br]corresponding roughly

0:02:13.221,0:02:18.980
to our year 2000 AD, asteroids that are[br]on a collision course with the Earth,

0:02:19.241,0:02:23.293
that otherwise would have collided,[br]mysteriously get deflected

0:02:23.893,0:02:26.643
or detonate before they can hit the Earth.

0:02:26.863,0:02:30.364
Now, of course, as Earthlings,[br]we know the reason would be

0:02:30.364,0:02:35.088
that we're trying to save ourselves,[br]we're trying to prevent an impact.

0:02:35.088,0:02:38.087
But if you're an alien race[br]that doesn't know any of this,

0:02:38.087,0:02:40.664
that doesn't have any concept[br]of Earth intelligence,

0:02:40.664,0:02:42.543
you'd be forced to put together

0:02:42.543,0:02:46.945
a physical theory that explains how,[br]up until a certain point in time,

0:02:47.925,0:02:51.508
asteroids thad would demolish[br]the surface of the planet,

0:02:52.258,0:02:55.361
mysteriously stop doing that.

0:02:55.361,0:02:59.610
So, I claim that this is the same question

0:02:59.610,0:03:03.112
as understanding the physical[br]nature of intelligence.

0:03:03.762,0:03:08.863
So, in this program that I undertook[br]years ago, I've looked at a variety

0:03:08.863,0:03:13.766
of different threads in crossed science[br]across a variety of disciplines,

0:03:13.766,0:03:19.280
pointing, I think, towards a single[br]underlying mechanism for intelligence.

0:03:19.910,0:03:22.024
In cosmology, for example,

0:03:22.314,0:03:24.968
there has been a variety[br]of different threads of evidence

0:03:24.968,0:03:29.695
that our universe appears to be[br]finely tuned for the development

0:03:29.695,0:03:33.360
of intelligence, and in particular,[br]for the development

0:03:33.360,0:03:38.941
of universal states that maximize[br]the diversity of possible futures.

0:03:38.941,0:03:44.063
In gameplay, for example in Go,[br]everyone remembers in 1997

0:03:44.403,0:03:48.120
when IBM's Deep Blue beat[br]Gary Kasparov at chess.

0:03:48.480,0:03:51.934
Fewer people are aware[br]that in the past ten year or so,

0:03:51.934,0:03:56.137
the game of Go, arguably a much more[br]challenging game because it has

0:03:56.137,0:04:00.804
a much higher branching factor,[br]has also started to succumb to computer

0:04:00.804,0:04:03.862
game players for the same reason.

0:04:03.862,0:04:06.529
The best techniques, right now,[br]for computers playing Go,

0:04:06.529,0:04:11.651
are techniques that try to maximize[br]future options during gameplay.

0:04:12.091,0:04:15.693
Finally, in robotic motion planning,

0:04:15.693,0:04:17.863
there has been a variety[br]of recent techniques

0:04:17.863,0:04:22.768
that have tried to take advantage[br]of abilities of robots to maximize

0:04:23.018,0:04:27.116
future freedom of action in order[br]to accomplish complex tasks.

0:04:27.496,0:04:31.340
And so, taking all of these different[br]threads and putting them together,

0:04:31.730,0:04:36.090
I asked, starting several years ago,[br]is there an underlying mechanism

0:04:36.340,0:04:40.249
for intelligence that we can factor out[br]of all of these different threads?

0:04:40.509,0:04:45.250
Is there, as it were,[br]a single equation for intelligence?

0:04:46.990,0:04:50.442
And the answer, I believe, is yes.

0:04:50.468,0:04:57.469
What you're seeing is probably the closest[br]equivalent to an E=mc2 for intelligence

0:04:57.469,0:05:00.072
that I certainly have ever seen.

0:05:00.098,0:05:02.276
So, what you're seeing here

0:05:02.371,0:05:07.835
is a statement of correspondence[br]that intelligence is a Force (F)

0:05:08.765,0:05:13.390
that acts so as to maximize[br]future freedom of action;

0:05:13.590,0:05:17.324
It acts to maximize future freedom[br]of action or keep options open

0:05:17.324,0:05:19.654
with some strength (T),

0:05:19.904,0:05:24.955
with the amount of the diversity[br]of possible accessible futures (S),

0:05:24.985,0:05:28.295
up to some future time horizon (Ƭ).

0:05:28.321,0:05:30.613
In short, intelligence doesn't like

0:05:30.613,0:05:34.498
to get trapped, intelligence tries[br]to maximize future freedom of action

0:05:34.498,0:05:39.526
and keep options open.[br]And so, given this one equation

0:05:39.526,0:05:42.445
it's natural to ask:[br]So, what can you do with this?

0:05:42.445,0:05:45.856
How predictive is it? Does it predict[br]human-level intelligence?

0:05:45.856,0:05:48.609
Does it predict artificial intelligence?

0:05:48.609,0:05:53.726
So, I'm going to show you now a video[br]that will, I think, demonstrate

0:05:54.006,0:05:58.294
some of the amazing applications[br]of just this single equation.

0:06:00.004,0:06:03.357
(Video) Recent research in cosmology[br]has suggested that universes

0:06:03.357,0:06:07.531
that produce more disorder or "entropy"[br]over their lifetimes should tend

0:06:07.531,0:06:11.269
to have more favorable conditions[br]for the existence of intelligent beings

0:06:11.589,0:06:13.445
such as ourselves.

0:06:13.445,0:06:15.763
But what if that tentative[br]cosmological connection

0:06:15.763,0:06:19.449
between entropy and intelligence[br]hints at a deeper relationship?

0:06:19.449,0:06:22.012
What if intelligent behavior[br]doesn't just correlate

0:06:22.012,0:06:26.226
with the production of long-term entropy,[br]but actually emerges directly from it?

0:06:26.576,0:06:30.114
To find out, we developed[br]a software engine called ENTROPICA

0:06:30.114,0:06:34.184
designed to maximize the production[br]of long-term entropy of any system

0:06:34.184,0:06:36.001
that it finds itself in.

0:06:36.001,0:06:40.645
Amazingly, ENTROPICA was able to pass[br]multiple animal intelligence tests,

0:06:40.645,0:06:43.766
play human games[br]and even earn money trading stocks;

0:06:43.766,0:06:46.157
all without being instructed to do so.

0:06:46.157,0:06:48.610
Here are some examples[br]of ENTROPICA in action:

0:06:48.610,0:06:52.164
just like a human standing upright[br]without falling over, here we see

0:06:52.254,0:06:56.225
ENTROPICA automatically[br]balancing a pole using a cart.

0:06:56.225,0:07:00.342
This behavior is remarkable, in part,[br]because we never gave ENTROPICA a goal,

0:07:00.342,0:07:03.754
it simply decided on its own[br]to balance the pole.

0:07:03.754,0:07:06.997
This balancing ability would have[br]applications for humanoid robotics

0:07:06.997,0:07:09.277
and human assistive technologies.

0:07:09.625,0:07:12.679
Just as some animals can use[br]objects in their environments

0:07:12.679,0:07:15.056
as tools to reach into narrow spaces,

0:07:15.056,0:07:18.967
here we see that ENTROPICA,[br]again on its own initiative,

0:07:18.967,0:07:22.192
was able to move a large disk,[br]representing an animal,

0:07:22.192,0:07:25.450
around so as to cause a small disk,[br]representing a tool,

0:07:25.450,0:07:28.346
to reach into a confined space[br]holding a third disk

0:07:28.346,0:07:31.953
and release the third disk[br]from its initially fixed position.

0:07:31.953,0:07:36.658
This tool usability would have application[br]for smart manufacturing and agriculture.

0:07:37.338,0:07:40.295
In addition, just as some other animals[br]are able to cooperate

0:07:40.295,0:07:44.043
by pulling opposite ends of a rope[br]at the same time to release food,

0:07:44.043,0:07:46.740
here we see that ENTROPICA[br]is able to accomplish

0:07:46.740,0:07:48.497
a model version of that task.

0:07:48.497,0:07:52.136
This cooperative ability has interesting[br]implications for economic planning

0:07:52.136,0:07:55.450
and a variety of other fields.

0:07:55.450,0:07:59.288
ENTROPICA is broadly applicable[br]to a variety of domains.

0:07:59.288,0:08:03.661
For example, here we see it successfully[br]playing a game of pong against itself

0:08:04.371,0:08:06.347
illustrating its potential for gaming.

0:08:08.103,0:08:09.794
Here, we see ENTROPICA orchestrating

0:08:09.794,0:08:13.289
new connections on a social network[br]where friends are constantly

0:08:13.289,0:08:17.341
falling out of touch and successfully[br]keeping the network well connected.

0:08:17.671,0:08:22.252
This same network orchestration ability[br]also has applications in health care,

0:08:22.252,0:08:25.404
energy and intelligence.

0:08:25.404,0:08:28.816
Here we see ENTROPICA directing[br]the paths of a fleet of ships

0:08:28.816,0:08:33.260
successfully discovering and utilizing[br]the Panama Canal to globally extend

0:08:33.260,0:08:35.951
its reach from the Atlantic[br]to the Pacific.

0:08:35.951,0:08:39.253
By the same token, ENTROPICA[br]is broadly applicable to problems

0:08:39.253,0:08:43.496
in autonomous defense,[br]logistics and transportation.

0:08:44.566,0:08:49.370
Finally, here we see ENTROPICA[br]spontaneously discovering and executing

0:08:49.370,0:08:53.843
a buy low, sell high strategy[br]on a simulated range traded stock

0:08:53.843,0:08:57.369
successfully growing assets[br]under management exponentially.

0:08:57.369,0:09:00.513
This risk management ability[br]would have broad applications

0:09:00.513,0:09:02.911
in finance and insurance.

0:09:08.475,0:09:12.067
AWG: So, what you've just seen[br]is that a variety

0:09:12.114,0:09:16.178
of signature human[br]intelligent cognitive behavior

0:09:16.204,0:09:18.895
such us tool use and walking upright

0:09:19.490,0:09:24.005
and social cooperation, all follow[br]from a single equation

0:09:24.255,0:09:29.452
which drives a system to maximize[br]its future freedom of action.

0:09:30.242,0:09:33.237
Now, there's a profound irony here.

0:09:33.237,0:09:37.663
Going back to the beginning[br]of the usage of the term robot,

0:09:38.503,0:09:41.499
the play RUR,

0:09:41.499,0:09:46.515
there was always a concept[br]that if we develop machine, intelligence,

0:09:47.345,0:09:52.622
there will be a cybernetic revolt,[br]that machines would rise up against us.

0:09:53.452,0:09:58.731
One major consequence of this work[br]is that maybe all of these decades

0:09:58.731,0:10:02.872
we've had the whole concept[br]of cybernetic revolt in reverse.

0:10:03.772,0:10:06.918
It's not that machines[br]first become intelligent

0:10:06.918,0:10:11.283
and then megalomaniacal,[br]and try to take over the world.

0:10:11.283,0:10:15.621
It's quite the opposite:[br]that the urge to take control

0:10:15.621,0:10:19.701
of all possible futures[br]is a more fundamental principle

0:10:20.071,0:10:23.949
than that of intelligence;[br]that general intelligence may, in fact,

0:10:23.949,0:10:28.456
emerge directly from this sort[br]of control grabbing,

0:10:28.456,0:10:31.209
rather than vice versa.

0:10:32.589,0:10:36.312
Another important consequence[br]is goal seeking.

0:10:36.652,0:10:42.443
I'm often asked how does the ability[br]to seek goals follow from this framework

0:10:42.643,0:10:43.747
and the answer is:

0:10:43.747,0:10:48.203
the ability to seek goals, for example[br]if you're playing the game of chess,

0:10:48.543,0:10:53.252
to try to win that game of chess[br]in order to accomplish worldly goods

0:10:53.252,0:10:55.599
and accomplishments outside of that game,

0:10:55.809,0:10:59.124
will follow directly from this[br]in the following sense:

0:10:59.554,0:11:03.855
Just like you would travel[br]through a tunnel, a bottleneck,

0:11:03.855,0:11:07.050
in your future path space[br]in order to achieve many other

0:11:07.050,0:11:11.178
diverse objectives later on[br]or just like you would invest

0:11:11.178,0:11:15.350
in a financial security reducing[br]your short term liquidity

0:11:15.350,0:11:17.825
in order to increase your wealth[br]over the long term,

0:11:17.825,0:11:21.613
goal seeking emerges directly[br]from a long term drive

0:11:21.613,0:11:25.571
to increase future freedom of action.

0:11:25.571,0:11:29.881
Finally, the famous physicist[br]Richard Feynman once wrote

0:11:30.361,0:11:34.703
that if human civilization were destroyed[br]and you could pass only a single concept

0:11:34.703,0:11:38.164
on to our descendents[br]to help them rebuild civilization,

0:11:38.524,0:11:41.620
that concept should be[br]that all matter around us

0:11:42.240,0:11:45.506
is made out of tiny elements[br]that attract each other

0:11:45.506,0:11:48.101
when they're far apart,[br]but repel each other

0:11:48.341,0:11:50.096
when they're close together.

0:11:50.126,0:11:53.152
My equivalent to that statement[br]to pass on to descendents

0:11:53.472,0:11:55.915
to help them build[br]artificial intelligence,

0:11:55.915,0:11:59.988
or to help them to understand[br]human intelligence, is the following:

0:12:00.108,0:12:03.541
Intelligence should be viewed[br]as a physical process

0:12:03.541,0:12:06.492
that tries to maximize[br]future freedom of action

0:12:06.492,0:12:09.624
and avoid constraints in its own future.

0:12:10.194,0:12:11.452
Thank you very much.

0:12:11.478,0:12:14.478
(Applause)