WEBVTT

00:00:00.899 --> 00:00:04.566
Intelligence -- what is it?

00:00:04.566 --> 00:00:06.857
If we take a look back at the history

00:00:06.857 --> 00:00:09.481
of how intelligence has been viewed,

00:00:09.481 --> 00:00:13.099
one seminal example has been

00:00:13.099 --> 00:00:16.576
Edsger Dijkstra's famous quote that

00:00:16.576 --> 00:00:19.687
"the question of whether a machine can think

00:00:19.687 --> 00:00:20.997
is about as interesting

00:00:20.997 --> 00:00:23.968
as the question of whether a submarine

00:00:23.968 --> 00:00:25.758
can swim."

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

00:00:29.602 --> 00:00:31.656
intended it as a criticism

00:00:31.656 --> 00:00:34.656
of the early pioneers of computer science,

00:00:34.656 --> 00:00:36.403
like Alan Turing.

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

00:00:38.902 --> 00:00:40.867
and think about what have been

00:00:40.867 --> 00:00:42.863
the most empowering innovations

00:00:42.863 --> 00:00:44.742
that enabled us to build

00:00:44.742 --> 00:00:46.976
artificial machines that swim

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

00:00:49.549 --> 00:00:53.096
you find that it was only through understanding

00:00:53.096 --> 00:00:55.704
the underlying physical mechanisms

00:00:55.704 --> 00:00:58.483
of swimming and flight

00:00:58.483 --> 00:01:01.655
that we were able to build these machines.

00:01:01.655 --> 00:01:03.911
And so, several years ago,

00:01:03.911 --> 00:01:07.160
I undertook a program to try to understand

00:01:07.160 --> 00:01:09.794
the fundamental physical mechanisms

00:01:09.794 --> 00:01:12.562
underlying intelligence.

NOTE Paragraph

00:01:12.562 --> 00:01:14.422
Let's take a step back.

00:01:14.422 --> 00:01:17.571
Let's first begin with a thought experiment.

00:01:17.571 --> 00:01:20.425
Pretend that you're an alien race

00:01:20.425 --> 00:01:23.466
that doesn't know anything about Earth biology

00:01:23.466 --> 00:01:26.582
or Earth neuroscience or Earth intelligence,

00:01:26.582 --> 00:01:28.774
but you have amazing telescopes

00:01:28.774 --> 00:01:31.136
and you're able to watch the Earth,

00:01:31.136 --> 00:01:33.468
and you have amazingly long lives,

00:01:33.468 --> 00:01:34.967
so you're able to watch the Earth

00:01:34.967 --> 00:01:38.409
over millions, even billions of years.

00:01:38.409 --> 00:01:41.424
And you observe a really strange effect.

00:01:41.424 --> 00:01:45.736
You observe that, over the course of the millennia,

00:01:45.736 --> 00:01:50.021
Earth is continually bombarded with asteroids

00:01:50.021 --> 00:01:52.108
up until a point,

00:01:52.108 --> 00:01:53.639
and that at some point,

00:01:53.639 --> 00:01:57.831
corresponding roughly to our year, 2000 AD,

00:01:57.831 --> 00:01:59.547
asteroids that are on

00:01:59.547 --> 00:02:01.478
a collision course with the Earth

00:02:01.478 --> 00:02:03.453
that otherwise would have collided

00:02:03.453 --> 00:02:05.868
mysteriously get deflected

00:02:05.868 --> 00:02:08.940
or they detonate before they can hit the Earth.

00:02:08.940 --> 00:02:11.023
Now of course, as earthlings,

00:02:11.023 --> 00:02:12.567
we know the reason would be

00:02:12.567 --> 00:02:14.323
that we're trying to save ourselves.

00:02:14.323 --> 00:02:17.403
We're trying to prevent an impact.

00:02:17.403 --> 00:02:19.114
But if you're an alien race

00:02:19.114 --> 00:02:20.260
who doesn't know any of this,

00:02:20.260 --> 00:02:22.774
doesn't have any concept of Earth intelligence,

00:02:22.774 --> 00:02:24.502
you'd be forced to put together

00:02:24.502 --> 00:02:27.420
a physical theory that explains how,

00:02:27.420 --> 00:02:29.958
up until a certain point in time,

00:02:29.958 --> 00:02:34.407
asteroids that would demolish the surface of a planet

00:02:34.407 --> 00:02:37.638
mysteriously stop doing that.

00:02:37.638 --> 00:02:41.842
And so I claim that this is the same question

00:02:41.842 --> 00:02:45.840
as understanding the physical nature of intelligence.

NOTE Paragraph

00:02:45.840 --> 00:02:49.722
So in this program that I
undertook several years ago,

00:02:49.722 --> 00:02:52.487
I looked at a variety of different threads

00:02:52.487 --> 00:02:55.649
across science, across a variety of disciplines,

00:02:55.649 --> 00:02:57.541
that were pointing, I think,

00:02:57.541 --> 00:03:00.089
towards a single, underlying mechanism

00:03:00.089 --> 00:03:01.670
for intelligence.

00:03:01.670 --> 00:03:04.216
In cosmology, for example,

00:03:04.216 --> 00:03:06.963
there have been a variety of
different threads of evidence

00:03:06.963 --> 00:03:10.370
that our universe appears to be finely tuned

00:03:10.370 --> 00:03:12.523
for the development of intelligence,

00:03:12.523 --> 00:03:14.912
and, in particular, for the development

00:03:14.912 --> 00:03:16.798
of universal states

00:03:16.798 --> 00:03:20.896
that maximize the diversity of possible futures.

00:03:20.896 --> 00:03:23.240
In game play, for example, in Go --

00:03:23.240 --> 00:03:26.265
everyone remembers in 1997

00:03:26.265 --> 00:03:30.216
when IBM's Deep Blue beat 
Garry Kasparov at chess --

00:03:30.216 --> 00:03:31.739
fewer people are aware

00:03:31.739 --> 00:03:33.757
that in the past 10 years or so,

00:03:33.757 --> 00:03:34.955
the game of Go,

00:03:34.955 --> 00:03:36.911
arguably a much more challenging game

00:03:36.911 --> 00:03:39.336
because it has a much higher branching factor,

00:03:39.336 --> 00:03:41.038
has also started to succumb

00:03:41.038 --> 00:03:42.903
to computer game players

00:03:42.903 --> 00:03:44.476
for the same reason:

00:03:44.476 --> 00:03:47.276
the best techniques right now
for computers playing Go

00:03:47.276 --> 00:03:50.972
are techniques that try to maximize future options

00:03:50.972 --> 00:03:52.986
during game play.

00:03:52.986 --> 00:03:56.567
Finally, in robotic motion planning,

00:03:56.567 --> 00:03:58.749
there have been a variety of recent techniques

00:03:58.749 --> 00:04:00.651
that have tried to take advantage

00:04:00.651 --> 00:04:03.797
of abilities of robots to maximize

00:04:03.797 --> 00:04:05.303
future freedom of action

00:04:05.303 --> 00:04:08.400
in order to accomplish complex tasks.

00:04:08.400 --> 00:04:10.755
And so, taking all of these different threads

00:04:10.755 --> 00:04:12.377
and putting them together,

00:04:12.377 --> 00:04:15.017
I asked, starting several years ago,

00:04:15.017 --> 00:04:17.867
is there an underlying mechanism for intelligence

00:04:17.867 --> 00:04:19.540
that we can factor out

00:04:19.540 --> 00:04:21.314
of all of these different threads?

00:04:21.314 --> 00:04:25.907
Is there a single equation for intelligence?

NOTE Paragraph

00:04:25.907 --> 00:04:29.278
And the answer, I believe, is yes.
["F = T ∇ Sτ"]

00:04:29.278 --> 00:04:31.191
What you're seeing is probably

00:04:31.191 --> 00:04:34.485
the closest equivalent to an E = mc²

00:04:34.485 --> 00:04:37.315
for intelligence that I've seen.

00:04:37.315 --> 00:04:39.017
So what you're seeing here

00:04:39.017 --> 00:04:41.686
is a statement of correspondence

00:04:41.686 --> 00:04:46.121
that intelligence is a force, F,

00:04:46.121 --> 00:04:50.771
that acts so as to maximize future freedom of action.

00:04:50.771 --> 00:04:53.146
It acts to maximize future freedom of action,

00:04:53.146 --> 00:04:54.774
or keep options open,

00:04:54.774 --> 00:04:56.999
with some strength T,

00:04:56.999 --> 00:05:01.776
with the diversity of possible accessible futures, S,

00:05:01.776 --> 00:05:04.326
up to some future time horizon, tau.

00:05:04.326 --> 00:05:07.535
In short, intelligence doesn't like to get trapped.

00:05:07.535 --> 00:05:10.590
Intelligence tries to maximize
future freedom of action

00:05:10.590 --> 00:05:13.263
and keep options open.

00:05:13.263 --> 00:05:15.696
And so, given this one equation,

00:05:15.696 --> 00:05:18.228
it's natural to ask, so what can you do with this?

00:05:18.228 --> 00:05:19.579
How predictive is it?

00:05:19.579 --> 00:05:21.714
Does it predict human-level intelligence?

00:05:21.714 --> 00:05:24.532
Does it predict artificial intelligence?

00:05:24.532 --> 00:05:26.574
So I'm going to show you now a video

00:05:26.574 --> 00:05:29.994
that will, I think, demonstrate

00:05:29.994 --> 00:05:32.282
some of the amazing applications

00:05:32.282 --> 00:05:34.601
of just this single equation.

NOTE Paragraph

00:05:34.601 --> 00:05:36.580
(Video) Narrator: Recent research in cosmology

00:05:36.580 --> 00:05:38.627
has suggested that universes that produce

00:05:38.627 --> 00:05:42.108
more disorder, or "entropy," over their lifetimes

00:05:42.108 --> 00:05:44.586
should tend to have more favorable conditions

00:05:44.586 --> 00:05:47.602
for the existence of intelligent
beings such as ourselves.

00:05:47.602 --> 00:05:50.176
But what if that tentative cosmological connection

00:05:50.176 --> 00:05:52.019
between entropy and intelligence

00:05:52.019 --> 00:05:53.790
hints at a deeper relationship?

00:05:53.790 --> 00:05:56.354
What if intelligent behavior doesn't just correlate

00:05:56.354 --> 00:05:58.198
with the production of long-term entropy,

00:05:58.198 --> 00:06:00.516
but actually emerges directly from it?

00:06:00.516 --> 00:06:02.922
To find out, we developed a software engine

00:06:02.922 --> 00:06:05.425
called Entropica, designed to maximize

00:06:05.425 --> 00:06:07.193
the production of long-term entropy

00:06:07.193 --> 00:06:09.769
of any system that it finds itself in.

00:06:09.769 --> 00:06:11.924
Amazingly, Entropica was able to pass

00:06:11.924 --> 00:06:15.380
multiple animal intelligence
tests, play human games,

00:06:15.380 --> 00:06:17.526
and even earn money trading stocks,

00:06:17.526 --> 00:06:19.637
all without being instructed to do so.

00:06:19.637 --> 00:06:22.155
Here are some examples of Entropica in action.

NOTE Paragraph

00:06:22.155 --> 00:06:25.360
Just like a human standing
upright without falling over,

00:06:25.360 --> 00:06:26.590
here we see Entropica

00:06:26.590 --> 00:06:29.475
automatically balancing a pole using a cart.

00:06:29.475 --> 00:06:31.487
This behavior is remarkable in part

00:06:31.487 --> 00:06:33.818
because we never gave Entropica a goal.

00:06:33.818 --> 00:06:36.975
It simply decided on its own to balance the pole.

00:06:36.975 --> 00:06:39.107
This balancing ability will have appliactions

00:06:39.107 --> 00:06:40.504
for humanoid robotics

00:06:40.504 --> 00:06:43.019
and human assistive technologies.

00:06:43.019 --> 00:06:45.020
Just as some animals can use objects

00:06:45.020 --> 00:06:46.462
in their environments as tools

00:06:46.462 --> 00:06:48.449
to reach into narrow spaces,

00:06:48.449 --> 00:06:50.331
here we see that Entropica,

00:06:50.331 --> 00:06:52.169
again on its own initiative,

00:06:52.169 --> 00:06:55.079
was able to move a large
disk representing an animal

00:06:55.079 --> 00:06:57.424
around so as to cause a small disk,

00:06:57.424 --> 00:07:00.195
representing a tool, to reach into a confined space

00:07:00.195 --> 00:07:01.732
holding a third disk

00:07:01.732 --> 00:07:04.704
and release the third disk
from its initially fixed position.

00:07:04.704 --> 00:07:06.893
This tool use ability will have applications

00:07:06.893 --> 00:07:09.252
for smart manufacturing and agriculture.

00:07:09.252 --> 00:07:11.196
In addition, just as some other animals

00:07:11.196 --> 00:07:13.892
are able to cooperate by pulling
opposite ends of a rope

00:07:13.892 --> 00:07:15.945
at the same time to release food,

00:07:15.945 --> 00:07:18.240
here we see that Entropica is able to accomplish

00:07:18.240 --> 00:07:20.228
a model version of that task.

00:07:20.228 --> 00:07:22.750
This cooperative ability has interesting implications

00:07:22.750 --> 00:07:26.185
for economic planning and a variety of other fields.

NOTE Paragraph

00:07:26.185 --> 00:07:28.256
Entropica is broadly applicable

00:07:28.256 --> 00:07:30.199
to a variety of domains.

00:07:30.199 --> 00:07:32.641
For example, here we see it successfully

00:07:32.641 --> 00:07:35.200
playing a game of pong against itself,

00:07:35.200 --> 00:07:37.543
illustrating its potential for gaming.

00:07:37.543 --> 00:07:39.462
Here we see Entropica orchestrating

00:07:39.462 --> 00:07:41.301
new connections on a social network

00:07:41.301 --> 00:07:44.061
where friends are constantly falling out of touch

00:07:44.061 --> 00:07:46.917
and successfully keeping
the network well connected.

00:07:46.917 --> 00:07:49.215
This same network orchestration ability

00:07:49.215 --> 00:07:51.543
also has applications in health care,

00:07:51.543 --> 00:07:54.775
energy, and intelligence.

00:07:54.775 --> 00:07:56.860
Here we see Entropica directing the paths

00:07:56.860 --> 00:07:58.346
of a fleet of ships,

00:07:58.346 --> 00:08:01.521
successfully discovering and
utilizing the Panama Canal

00:08:01.521 --> 00:08:03.979
to globally extend its reach from the Atlantic

00:08:03.979 --> 00:08:05.508
to the Pacific.

00:08:05.508 --> 00:08:07.235
By the same token, Entropica

00:08:07.235 --> 00:08:08.855
is broadly applicable to problems

00:08:08.855 --> 00:08:14.157
in autonomous defense, logistics and transportation.

NOTE Paragraph

00:08:14.173 --> 00:08:16.203
Finally, here we see Entropica

00:08:16.203 --> 00:08:18.926
spontaneously discovering and executing

00:08:18.926 --> 00:08:20.993
a buy-low, sell-high strategy

00:08:20.993 --> 00:08:23.171
on a simulated range traded stock,

00:08:23.171 --> 00:08:25.502
successfully growing assets under management

00:08:25.502 --> 00:08:26.926
exponentially.

00:08:26.926 --> 00:08:28.234
This risk management ability

00:08:28.234 --> 00:08:30.721
will have broad applications in finance

00:08:30.721 --> 00:08:34.049
and insurance.

NOTE Paragraph

00:08:34.049 --> 00:08:36.140
Alex Wissner-Gross: So what you've just seen

00:08:36.140 --> 00:08:40.532
is that a variety of signature human intelligent

00:08:40.532 --> 00:08:42.289
cognitive behaviors

00:08:42.289 --> 00:08:45.120
such as tool use and walking upright

00:08:45.120 --> 00:08:47.149
and social cooperation

00:08:47.149 --> 00:08:50.121
all follow from a single equation,

00:08:50.121 --> 00:08:52.053
which drives a system

00:08:52.053 --> 00:08:55.964
to maximize its future freedom of action.

NOTE Paragraph

00:08:55.964 --> 00:08:58.971
Now, there's a profound irony here.

00:08:58.971 --> 00:09:00.995
Going back to the beginning

00:09:00.995 --> 00:09:04.268
of the usage of the term robot,

00:09:04.268 --> 00:09:07.171
the play "RUR,"

00:09:07.171 --> 00:09:09.406
there was always a concept

00:09:09.406 --> 00:09:12.632
that if we developed machine intelligence,

00:09:12.632 --> 00:09:15.659
there would be a cybernetic revolt.

00:09:15.659 --> 00:09:19.210
The machines would rise up against us.

00:09:19.210 --> 00:09:21.529
One major consequence of this work

00:09:21.529 --> 00:09:24.298
is that maybe all of these decades,

00:09:24.298 --> 00:09:27.274
we've had the whole concept of cybernetic revolt

00:09:27.274 --> 00:09:29.285
in reverse.

00:09:29.285 --> 00:09:32.564
It's not that machines first become intelligent

00:09:32.564 --> 00:09:34.579
and then megalomaniacal

00:09:34.579 --> 00:09:36.803
and try to take over the world.

00:09:36.803 --> 00:09:38.237
It's quite the opposite,

00:09:38.237 --> 00:09:41.143
that the urge to take control

00:09:41.143 --> 00:09:43.404
of all possible futures

00:09:43.404 --> 00:09:45.522
is a more fundamental principle

00:09:45.522 --> 00:09:46.885
than that of intelligence,

00:09:46.885 --> 00:09:50.585
that general intelligence may in fact emerge

00:09:50.585 --> 00:09:54.144
directly from this sort of control-grabbing,

00:09:54.144 --> 00:09:58.329
rather than vice versa.

NOTE Paragraph

00:09:58.329 --> 00:10:02.098
Another important consequence is goal seeking.

00:10:02.098 --> 00:10:06.458
I'm often asked, how does the ability to seek goals

00:10:06.458 --> 00:10:08.078
follow from this sort of framework?

00:10:08.078 --> 00:10:11.106
And the answer is, the ability to seek goals

00:10:11.106 --> 00:10:12.988
will follow directly from this

00:10:12.988 --> 00:10:14.822
in the following sense:

00:10:14.822 --> 00:10:17.687
just like you would travel through a tunnel,

00:10:17.687 --> 00:10:20.192
a bottleneck in your future path space,

00:10:20.192 --> 00:10:22.063
in order to achieve many other

00:10:22.063 --> 00:10:24.084
diverse objectives later on,

00:10:24.084 --> 00:10:26.456
or just like you would invest

00:10:26.456 --> 00:10:28.243
in a financial security,

00:10:28.243 --> 00:10:30.480
reducing your short-term liquidity

00:10:30.480 --> 00:10:32.880
in order to increase your wealth over the long term,

00:10:32.880 --> 00:10:35.217
goal seeking emerges directly

00:10:35.217 --> 00:10:36.946
from a long-term drive

00:10:36.946 --> 00:10:40.983
to increase future freedom of action.

NOTE Paragraph

00:10:40.983 --> 00:10:44.511
Finally, Richard Feynman, famous physicist,

00:10:44.511 --> 00:10:48.183
once wrote that if human civilization were destroyed

00:10:48.183 --> 00:10:50.076
and you could pass only a single concept

00:10:50.076 --> 00:10:51.447
on to our descendants

00:10:51.447 --> 00:10:53.754
to help them rebuild civilization,

00:10:53.754 --> 00:10:55.440
that concept should be

00:10:55.440 --> 00:10:57.292
that all matter around us

00:10:57.292 --> 00:10:59.615
is made out of tiny elements

00:10:59.615 --> 00:11:02.123
that attract each other when they're far apart

00:11:02.123 --> 00:11:05.453
but repel each other when they're close together.

00:11:05.453 --> 00:11:07.234
My equivalent of that statement

00:11:07.234 --> 00:11:08.502
to pass on to descendants

00:11:08.502 --> 00:11:11.214
to help them build artificial intelligences

00:11:11.214 --> 00:11:14.163
or to help them understand human intelligence,

00:11:14.163 --> 00:11:15.430
is the following:

00:11:15.430 --> 00:11:17.483
Intelligence should be viewed

00:11:17.483 --> 00:11:18.896
as a physical process

00:11:18.896 --> 00:11:21.861
that tries to maximize future freedom of action

00:11:21.861 --> 00:11:25.477
and avoid constraints in its own future.

NOTE Paragraph

00:11:25.477 --> 00:11:26.835
Thank you very much.

NOTE Paragraph

00:11:26.835 --> 00:11:30.835
(Applause)