0:00:00.899,0:00:04.566
Intelligence -- what is it?

0:00:04.566,0:00:06.857
If we take a look back at the history

0:00:06.857,0:00:09.481
of how intelligence has been viewed,

0:00:09.481,0:00:13.099
one seminal example has been

0:00:13.099,0:00:16.576
Edsger Dijkstra's famous quote that

0:00:16.576,0:00:19.687
"the question of whether a machine can think

0:00:19.687,0:00:20.997
is about as interesting

0:00:20.997,0:00:23.968
as the question of whether a submarine

0:00:23.968,0:00:25.758
can swim."

0:00:25.758,0:00:29.602
Now, Edsger Dijkstra, when he wrote this,

0:00:29.602,0:00:31.656
intended it as a criticism

0:00:31.656,0:00:34.656
of the early pioneers of computer science,

0:00:34.656,0:00:36.403
like Alan Turing.

0:00:36.403,0:00:38.902
However, if you take a look back

0:00:38.902,0:00:40.867
and think about what have been

0:00:40.867,0:00:42.863
the most empowering innovations

0:00:42.863,0:00:44.742
that enabled us to build

0:00:44.742,0:00:46.976
artificial machines that swim

0:00:46.976,0:00:49.549
and artificial machines that [fly],

0:00:49.549,0:00:53.096
you find that it was only through understanding

0:00:53.096,0:00:55.704
the underlying physical mechanisms

0:00:55.704,0:00:58.483
of swimming and flight

0:00:58.483,0:01:01.655
that we were able to build these machines.

0:01:01.655,0:01:03.911
And so, several years ago,

0:01:03.911,0:01:07.160
I undertook a program to try to understand

0:01:07.160,0:01:09.794
the fundamental physical mechanisms

0:01:09.794,0:01:12.562
underlying intelligence.

0:01:12.562,0:01:14.422
Let's take a step back.

0:01:14.422,0:01:17.571
Let's first begin with a thought experiment.

0:01:17.571,0:01:20.425
Pretend that you're an alien race

0:01:20.425,0:01:23.466
that doesn't know anything about Earth biology

0:01:23.466,0:01:26.582
or Earth neuroscience or Earth intelligence,

0:01:26.582,0:01:28.774
but you have amazing telescopes

0:01:28.774,0:01:31.136
and you're able to watch the Earth,

0:01:31.136,0:01:33.468
and you have amazingly long lives,

0:01:33.468,0:01:34.967
so you're able to watch the Earth

0:01:34.967,0:01:38.409
over millions, even billions of years.

0:01:38.409,0:01:41.424
And you observe a really strange effect.

0:01:41.424,0:01:45.736
You observe that, over the course of the millennia,

0:01:45.736,0:01:50.021
Earth is continually bombarded with asteroids

0:01:50.021,0:01:52.108
up until a point,

0:01:52.108,0:01:53.639
and that at some point,

0:01:53.639,0:01:57.831
corresponding roughly to our year, 2000 AD,

0:01:57.831,0:01:59.547
asteroids that are on

0:01:59.547,0:02:01.478
a collision course with the Earth

0:02:01.478,0:02:03.453
that otherwise would have collided

0:02:03.453,0:02:05.868
mysteriously get deflected

0:02:05.868,0:02:08.940
or they detonate before they can hit the Earth.

0:02:08.940,0:02:11.023
Now of course, as earthlings,

0:02:11.023,0:02:12.567
we know the reason would be

0:02:12.567,0:02:14.323
that we're trying to save ourselves.

0:02:14.323,0:02:17.403
We're trying to prevent an impact.

0:02:17.403,0:02:19.114
But if you're an alien race

0:02:19.114,0:02:20.260
who doesn't know any of this,

0:02:20.260,0:02:22.774
doesn't have any concept of Earth intelligence,

0:02:22.774,0:02:24.502
you'd be forced to put together

0:02:24.502,0:02:27.420
a physical theory that explains how,

0:02:27.420,0:02:29.958
up until a certain point in time,

0:02:29.958,0:02:34.407
asteroids that would demolish the surface of a planet

0:02:34.407,0:02:37.638
mysteriously stop doing that.

0:02:37.638,0:02:41.842
And so I claim that this is the same question

0:02:41.842,0:02:45.840
as understanding the physical nature of intelligence.

0:02:45.840,0:02:49.722
So in this program that I[br]undertook several years ago,

0:02:49.722,0:02:52.487
I looked at a variety of different threads

0:02:52.487,0:02:55.649
across science, across a variety of disciplines,

0:02:55.649,0:02:57.541
that were pointing, I think,

0:02:57.541,0:03:00.089
towards a single, underlying mechanism

0:03:00.089,0:03:01.670
for intelligence.

0:03:01.670,0:03:04.216
In cosmology, for example,

0:03:04.216,0:03:06.963
there have been a variety of[br]different threads of evidence

0:03:06.963,0:03:10.370
that our universe appears to be finely tuned

0:03:10.370,0:03:12.523
for the development of intelligence,

0:03:12.523,0:03:14.912
and, in particular, for the development

0:03:14.912,0:03:16.798
of universal states

0:03:16.798,0:03:20.896
that maximize the diversity of possible futures.

0:03:20.896,0:03:23.240
In game play, for example, in Go --

0:03:23.240,0:03:26.265
everyone remembers in 1997

0:03:26.265,0:03:30.216
when IBM's Deep Blue beat [br]Garry Kasparov at chess --

0:03:30.216,0:03:31.739
fewer people are aware

0:03:31.739,0:03:33.757
that in the past 10 years or so,

0:03:33.757,0:03:34.955
the game of Go,

0:03:34.955,0:03:36.911
arguably a much more challenging game

0:03:36.911,0:03:39.336
because it has a much higher branching factor,

0:03:39.336,0:03:41.038
has also started to succumb

0:03:41.038,0:03:42.903
to computer game players

0:03:42.903,0:03:44.476
for the same reason:

0:03:44.476,0:03:47.276
the best techniques right now[br]for computers playing Go

0:03:47.276,0:03:50.972
are techniques that try to maximize future options

0:03:50.972,0:03:52.986
during game play.

0:03:52.986,0:03:56.567
Finally, in robotic motion planning,

0:03:56.567,0:03:58.749
there have been a variety of recent techniques

0:03:58.749,0:04:00.651
that have tried to take advantage

0:04:00.651,0:04:03.797
of abilities of robots to maximize

0:04:03.797,0:04:05.303
future freedom of action

0:04:05.303,0:04:08.400
in order to accomplish complex tasks.

0:04:08.400,0:04:10.755
And so, taking all of these different threads

0:04:10.755,0:04:12.377
and putting them together,

0:04:12.377,0:04:15.017
I asked, starting several years ago,

0:04:15.017,0:04:17.867
is there an underlying mechanism for intelligence

0:04:17.867,0:04:19.540
that we can factor out

0:04:19.540,0:04:21.314
of all of these different threads?

0:04:21.314,0:04:25.907
Is there a single equation for intelligence?

0:04:25.907,0:04:29.278
And the answer, I believe, is yes.[br]["F = T ∇ Sτ"]

0:04:29.278,0:04:31.191
What you're seeing is probably

0:04:31.191,0:04:34.485
the closest equivalent to an E = mc²

0:04:34.485,0:04:37.315
for intelligence that I've seen.

0:04:37.315,0:04:39.017
So what you're seeing here

0:04:39.017,0:04:41.686
is a statement of correspondence

0:04:41.686,0:04:46.121
that intelligence is a force, F,

0:04:46.121,0:04:50.771
that acts so as to maximize future freedom of action.

0:04:50.771,0:04:53.146
It acts to maximize future freedom of action,

0:04:53.146,0:04:54.774
or keep options open,

0:04:54.774,0:04:56.999
with some strength T,

0:04:56.999,0:05:01.776
with the diversity of possible accessible futures, S,

0:05:01.776,0:05:04.326
up to some future time horizon, tau.

0:05:04.326,0:05:07.535
In short, intelligence doesn't like to get trapped.

0:05:07.535,0:05:10.590
Intelligence tries to maximize[br]future freedom of action

0:05:10.590,0:05:13.263
and keep options open.

0:05:13.263,0:05:15.696
And so, given this one equation,

0:05:15.696,0:05:18.228
it's natural to ask, so what can you do with this?

0:05:18.228,0:05:19.579
How predictive is it?

0:05:19.579,0:05:21.714
Does it predict human-level intelligence?

0:05:21.714,0:05:24.532
Does it predict artificial intelligence?

0:05:24.532,0:05:26.574
So I'm going to show you now a video

0:05:26.574,0:05:29.994
that will, I think, demonstrate

0:05:29.994,0:05:32.282
some of the amazing applications

0:05:32.282,0:05:34.601
of just this single equation.

0:05:34.601,0:05:36.580
(Video) Narrator: Recent research in cosmology

0:05:36.580,0:05:38.627
has suggested that universes that produce

0:05:38.627,0:05:42.108
more disorder, or "entropy," over their lifetimes

0:05:42.108,0:05:44.586
should tend to have more favorable conditions

0:05:44.586,0:05:47.602
for the existence of intelligent[br]beings such as ourselves.

0:05:47.602,0:05:50.176
But what if that tentative cosmological connection

0:05:50.176,0:05:52.019
between entropy and intelligence

0:05:52.019,0:05:53.790
hints at a deeper relationship?

0:05:53.790,0:05:56.354
What if intelligent behavior doesn't just correlate

0:05:56.354,0:05:58.198
with the production of long-term entropy,

0:05:58.198,0:06:00.516
but actually emerges directly from it?

0:06:00.516,0:06:02.922
To find out, we developed a software engine

0:06:02.922,0:06:05.425
called Entropica, designed to maximize

0:06:05.425,0:06:07.193
the production of long-term entropy

0:06:07.193,0:06:09.769
of any system that it finds itself in.

0:06:09.769,0:06:11.924
Amazingly, Entropica was able to pass

0:06:11.924,0:06:15.380
multiple animal intelligence[br]tests, play human games,

0:06:15.380,0:06:17.526
and even earn money trading stocks,

0:06:17.526,0:06:19.637
all without being instructed to do so.

0:06:19.637,0:06:22.155
Here are some examples of Entropica in action.

0:06:22.155,0:06:25.360
Just like a human standing[br]upright without falling over,

0:06:25.360,0:06:26.590
here we see Entropica

0:06:26.590,0:06:29.475
automatically balancing a pole using a cart.

0:06:29.475,0:06:31.487
This behavior is remarkable in part

0:06:31.487,0:06:33.818
because we never gave Entropica a goal.

0:06:33.818,0:06:36.975
It simply decided on its own to balance the pole.

0:06:36.975,0:06:39.107
This balancing ability will have appliactions

0:06:39.107,0:06:40.504
for humanoid robotics

0:06:40.504,0:06:43.019
and human assistive technologies.

0:06:43.019,0:06:45.020
Just as some animals can use objects

0:06:45.020,0:06:46.462
in their environments as tools

0:06:46.462,0:06:48.449
to reach into narrow spaces,

0:06:48.449,0:06:50.331
here we see that Entropica,

0:06:50.331,0:06:52.169
again on its own initiative,

0:06:52.169,0:06:55.079
was able to move a large[br]disk representing an animal

0:06:55.079,0:06:57.424
around so as to cause a small disk,

0:06:57.424,0:07:00.195
representing a tool, to reach into a confined space

0:07:00.195,0:07:01.732
holding a third disk

0:07:01.732,0:07:04.704
and release the third disk[br]from its initially fixed position.

0:07:04.704,0:07:06.893
This tool use ability will have applications

0:07:06.893,0:07:09.252
for smart manufacturing and agriculture.

0:07:09.252,0:07:11.196
In addition, just as some other animals

0:07:11.196,0:07:13.892
are able to cooperate by pulling[br]opposite ends of a rope

0:07:13.892,0:07:15.945
at the same time to release food,

0:07:15.945,0:07:18.240
here we see that Entropica is able to accomplish

0:07:18.240,0:07:20.228
a model version of that task.

0:07:20.228,0:07:22.750
This cooperative ability has interesting implications

0:07:22.750,0:07:26.185
for economic planning and a variety of other fields.

0:07:26.185,0:07:28.256
Entropica is broadly applicable

0:07:28.256,0:07:30.199
to a variety of domains.

0:07:30.199,0:07:32.641
For example, here we see it successfully

0:07:32.641,0:07:35.200
playing a game of pong against itself,

0:07:35.200,0:07:37.543
illustrating its potential for gaming.

0:07:37.543,0:07:39.462
Here we see Entropica orchestrating

0:07:39.462,0:07:41.301
new connections on a social network

0:07:41.301,0:07:44.061
where friends are constantly falling out of touch

0:07:44.061,0:07:46.917
and successfully keeping[br]the network well connected.

0:07:46.917,0:07:49.215
This same network orchestration ability

0:07:49.215,0:07:51.543
also has applications in health care,

0:07:51.543,0:07:54.775
energy, and intelligence.

0:07:54.775,0:07:56.860
Here we see Entropica directing the paths

0:07:56.860,0:07:58.346
of a fleet of ships,

0:07:58.346,0:08:01.521
successfully discovering and[br]utilizing the Panama Canal

0:08:01.521,0:08:03.979
to globally extend its reach from the Atlantic

0:08:03.979,0:08:05.508
to the Pacific.

0:08:05.508,0:08:07.235
By the same token, Entropica

0:08:07.235,0:08:08.855
is broadly applicable to problems

0:08:08.855,0:08:14.157
in autonomous defense, logistics and transportation.

0:08:14.173,0:08:16.203
Finally, here we see Entropica

0:08:16.203,0:08:18.926
spontaneously discovering and executing

0:08:18.926,0:08:20.993
a buy-low, sell-high strategy

0:08:20.993,0:08:23.171
on a simulated range traded stock,

0:08:23.171,0:08:25.502
successfully growing assets under management

0:08:25.502,0:08:26.926
exponentially.

0:08:26.926,0:08:28.234
This risk management ability

0:08:28.234,0:08:30.721
will have broad applications in finance

0:08:30.721,0:08:34.049
and insurance.

0:08:34.049,0:08:36.140
Alex Wissner-Gross: So what you've just seen

0:08:36.140,0:08:40.532
is that a variety of signature human intelligent

0:08:40.532,0:08:42.289
cognitive behaviors

0:08:42.289,0:08:45.120
such as tool use and walking upright

0:08:45.120,0:08:47.149
and social cooperation

0:08:47.149,0:08:50.121
all follow from a single equation,

0:08:50.121,0:08:52.053
which drives a system

0:08:52.053,0:08:55.964
to maximize its future freedom of action.

0:08:55.964,0:08:58.971
Now, there's a profound irony here.

0:08:58.971,0:09:00.995
Going back to the beginning

0:09:00.995,0:09:04.268
of the usage of the term robot,

0:09:04.268,0:09:07.171
the play "RUR,"

0:09:07.171,0:09:09.406
there was always a concept

0:09:09.406,0:09:12.632
that if we developed machine intelligence,

0:09:12.632,0:09:15.659
there would be a cybernetic revolt.

0:09:15.659,0:09:19.210
The machines would rise up against us.

0:09:19.210,0:09:21.529
One major consequence of this work

0:09:21.529,0:09:24.298
is that maybe all of these decades,

0:09:24.298,0:09:27.274
we've had the whole concept of cybernetic revolt

0:09:27.274,0:09:29.285
in reverse.

0:09:29.285,0:09:32.564
It's not that machines first become intelligent

0:09:32.564,0:09:34.579
and then megalomaniacal

0:09:34.579,0:09:36.803
and try to take over the world.

0:09:36.803,0:09:38.237
It's quite the opposite,

0:09:38.237,0:09:41.143
that the urge to take control

0:09:41.143,0:09:43.404
of all possible futures

0:09:43.404,0:09:45.522
is a more fundamental principle

0:09:45.522,0:09:46.885
than that of intelligence,

0:09:46.885,0:09:50.585
that general intelligence may in fact emerge

0:09:50.585,0:09:54.144
directly from this sort of control-grabbing,

0:09:54.144,0:09:58.329
rather than vice versa.

0:09:58.329,0:10:02.098
Another important consequence is goal seeking.

0:10:02.098,0:10:06.458
I'm often asked, how does the ability to seek goals

0:10:06.458,0:10:08.078
follow from this sort of framework?

0:10:08.078,0:10:11.106
And the answer is, the ability to seek goals

0:10:11.106,0:10:12.988
will follow directly from this

0:10:12.988,0:10:14.822
in the following sense:

0:10:14.822,0:10:17.687
just like you would travel through a tunnel,

0:10:17.687,0:10:20.192
a bottleneck in your future path space,

0:10:20.192,0:10:22.063
in order to achieve many other

0:10:22.063,0:10:24.084
diverse objectives later on,

0:10:24.084,0:10:26.456
or just like you would invest

0:10:26.456,0:10:28.243
in a financial security,

0:10:28.243,0:10:30.480
reducing your short-term liquidity

0:10:30.480,0:10:32.880
in order to increase your wealth over the long term,

0:10:32.880,0:10:35.217
goal seeking emerges directly

0:10:35.217,0:10:36.946
from a long-term drive

0:10:36.946,0:10:40.983
to increase future freedom of action.

0:10:40.983,0:10:44.511
Finally, Richard Feynman, famous physicist,

0:10:44.511,0:10:48.183
once wrote that if human civilization were destroyed

0:10:48.183,0:10:50.076
and you could pass only a single concept

0:10:50.076,0:10:51.447
on to our descendants

0:10:51.447,0:10:53.754
to help them rebuild civilization,

0:10:53.754,0:10:55.440
that concept should be

0:10:55.440,0:10:57.292
that all matter around us

0:10:57.292,0:10:59.615
is made out of tiny elements

0:10:59.615,0:11:02.123
that attract each other when they're far apart

0:11:02.123,0:11:05.453
but repel each other when they're close together.

0:11:05.453,0:11:07.234
My equivalent of that statement

0:11:07.234,0:11:08.502
to pass on to descendants

0:11:08.502,0:11:11.214
to help them build artificial intelligences

0:11:11.214,0:11:14.163
or to help them understand human intelligence,

0:11:14.163,0:11:15.430
is the following:

0:11:15.430,0:11:17.483
Intelligence should be viewed

0:11:17.483,0:11:18.896
as a physical process

0:11:18.896,0:11:21.861
that tries to maximize future freedom of action

0:11:21.861,0:11:25.477
and avoid constraints in its own future.

0:11:25.477,0:11:26.835
Thank you very much.

0:11:26.835,0:11:30.835
(Applause)