TEDxBeaconStreet
Alex Wissner-Gross: A new equation for intelligence
Filmed:
Readability: 5.9
2,098,891 views
Is there an equation for intelligence? Yes. It's F = T ∇ Sτ. In a fascinating and informative talk, physicist and computer scientist Alex Wissner-Gross explains what in the world that means.
Alex Wissner-Gross - Scientist, entrepreneur, inventor
Alex Wissner-Gross applies science and engineering principles to big (and diverse) questions, like: "What is the equation for intelligence?" and "What's the best way to raise awareness about climate change?" Full bio
Alex Wissner-Gross applies science and engineering principles to big (and diverse) questions, like: "What is the equation for intelligence?" and "What's the best way to raise awareness about climate change?" Full bio
Double-click the English transcript below to play the video.
00:12
Intelligence -- what is it?
0
899
3667
00:16
If we take a look back at the history
1
4566
2291
00:18
of how intelligence has been viewed,
2
6857
2624
00:21
one seminal example has been
3
9481
3618
00:25
Edsger Dijkstra's famous quote that
4
13099
3477
00:28
"the question of whether a machine can think
5
16576
3111
00:31
is about as interesting
6
19687
1310
00:32
as the question of whether a submarine
7
20997
2971
00:35
can swim."
8
23968
1790
00:37
Now, Edsger Dijkstra, when he wrote this,
9
25758
3844
00:41
intended it as a criticism
10
29602
2054
00:43
of the early pioneers of computer science,
11
31656
3000
00:46
like Alan Turing.
12
34656
1747
00:48
However, if you take a look back
13
36403
2499
00:50
and think about what have been
14
38902
1965
00:52
the most empowering innovations
15
40867
1996
00:54
that enabled us to build
16
42863
1879
00:56
artificial machines that swim
17
44742
2234
00:58
and artificial machines that [fly],
18
46976
2573
01:01
you find that it was only through understanding
19
49549
3547
01:05
the underlying physical mechanisms
20
53096
2608
01:07
of swimming and flight
21
55704
2779
01:10
that we were able to build these machines.
22
58483
3172
01:13
And so, several years ago,
23
61655
2256
01:15
I undertook a program to try to understand
24
63911
3249
01:19
the fundamental physical mechanisms
25
67160
2634
01:21
underlying intelligence.
26
69794
2768
01:24
Let's take a step back.
27
72562
1860
01:26
Let's first begin with a thought experiment.
28
74422
3149
01:29
Pretend that you're an alien race
29
77571
2854
01:32
that doesn't know anything about Earth biology
30
80425
3041
01:35
or Earth neuroscience or Earth intelligence,
31
83466
3116
01:38
but you have amazing telescopes
32
86582
2192
01:40
and you're able to watch the Earth,
33
88774
2362
01:43
and you have amazingly long lives,
34
91136
2332
01:45
so you're able to watch the Earth
35
93468
1499
01:46
over millions, even billions of years.
36
94967
3442
01:50
And you observe a really strange effect.
37
98409
3015
01:53
You observe that, over the course of the millennia,
38
101424
4312
01:57
Earth is continually bombarded with asteroids
39
105736
4285
02:02
up until a point,
40
110021
2087
02:04
and that at some point,
41
112108
1531
02:05
corresponding roughly to our year, 2000 AD,
42
113639
4192
02:09
asteroids that are on
43
117831
1716
02:11
a collision course with the Earth
44
119547
1931
02:13
that otherwise would have collided
45
121478
1975
02:15
mysteriously get deflected
46
123453
2415
02:17
or they detonate before they can hit the Earth.
47
125868
3072
02:20
Now of course, as earthlings,
48
128940
2083
02:23
we know the reason would be
49
131023
1544
02:24
that we're trying to save ourselves.
50
132567
1756
02:26
We're trying to prevent an impact.
51
134323
3080
02:29
But if you're an alien race
52
137403
1711
02:31
who doesn't know any of this,
53
139114
1146
02:32
doesn't have any concept of Earth intelligence,
54
140260
2514
02:34
you'd be forced to put together
55
142774
1728
02:36
a physical theory that explains how,
56
144502
2918
02:39
up until a certain point in time,
57
147420
2538
02:41
asteroids that would demolish the surface of a planet
58
149958
4449
02:46
mysteriously stop doing that.
59
154407
3231
02:49
And so I claim that this is the same question
60
157638
4204
02:53
as understanding the physical nature of intelligence.
61
161842
3998
02:57
So in this program that I
undertook several years ago,
undertook several years ago,
62
165840
3882
03:01
I looked at a variety of different threads
63
169722
2765
03:04
across science, across a variety of disciplines,
64
172487
3162
03:07
that were pointing, I think,
65
175649
1892
03:09
towards a single, underlying mechanism
66
177541
2548
03:12
for intelligence.
67
180089
1581
03:13
In cosmology, for example,
68
181670
2546
03:16
there have been a variety of
different threads of evidence
different threads of evidence
69
184216
2747
03:18
that our universe appears to be finely tuned
70
186963
3407
03:22
for the development of intelligence,
71
190370
2153
03:24
and, in particular, for the development
72
192523
2389
03:26
of universal states
73
194912
1886
03:28
that maximize the diversity of possible futures.
74
196798
4098
03:32
In game play, for example, in Go --
75
200896
2344
03:35
everyone remembers in 1997
76
203240
3025
03:38
when IBM's Deep Blue beat
Garry Kasparov at chess --
Garry Kasparov at chess --
77
206265
3951
03:42
fewer people are aware
78
210216
1523
03:43
that in the past 10 years or so,
79
211739
2018
03:45
the game of Go,
80
213757
1198
03:46
arguably a much more challenging game
81
214955
1956
03:48
because it has a much higher branching factor,
82
216911
2425
03:51
has also started to succumb
83
219336
1702
03:53
to computer game players
84
221038
1865
03:54
for the same reason:
85
222903
1573
03:56
the best techniques right now
for computers playing Go
for computers playing Go
86
224476
2800
03:59
are techniques that try to maximize future options
87
227276
3696
04:02
during game play.
88
230972
2014
04:04
Finally, in robotic motion planning,
89
232986
3581
04:08
there have been a variety of recent techniques
90
236567
2182
04:10
that have tried to take advantage
91
238749
1902
04:12
of abilities of robots to maximize
92
240651
3146
04:15
future freedom of action
93
243797
1506
04:17
in order to accomplish complex tasks.
94
245303
3097
04:20
And so, taking all of these different threads
95
248400
2355
04:22
and putting them together,
96
250755
1622
04:24
I asked, starting several years ago,
97
252377
2640
04:27
is there an underlying mechanism for intelligence
98
255017
2850
04:29
that we can factor out
99
257867
1673
04:31
of all of these different threads?
100
259540
1774
04:33
Is there a single equation for intelligence?
101
261314
4593
04:37
And the answer, I believe, is yes.
["F = T ∇ Sτ"]
["F = T ∇ Sτ"]
102
265907
3371
04:41
What you're seeing is probably
103
269278
1913
04:43
the closest equivalent to an E = mc²
104
271191
3294
04:46
for intelligence that I've seen.
105
274485
2830
04:49
So what you're seeing here
106
277315
1702
04:51
is a statement of correspondence
107
279017
2669
04:53
that intelligence is a force, F,
108
281686
4435
04:58
that acts so as to maximize future freedom of action.
109
286121
4650
05:02
It acts to maximize future freedom of action,
110
290771
2375
05:05
or keep options open,
111
293146
1628
05:06
with some strength T,
112
294774
2225
05:08
with the diversity of possible accessible futures, S,
113
296999
4777
05:13
up to some future time horizon, tau.
114
301776
2550
05:16
In short, intelligence doesn't like to get trapped.
115
304326
3209
05:19
Intelligence tries to maximize
future freedom of action
future freedom of action
116
307535
3055
05:22
and keep options open.
117
310590
2673
05:25
And so, given this one equation,
118
313263
2433
05:27
it's natural to ask, so what can you do with this?
119
315696
2532
05:30
How predictive is it?
120
318228
1351
05:31
Does it predict human-level intelligence?
121
319579
2135
05:33
Does it predict artificial intelligence?
122
321714
2818
05:36
So I'm going to show you now a video
123
324532
2042
05:38
that will, I think, demonstrate
124
326574
3420
05:41
some of the amazing applications
125
329994
2288
05:44
of just this single equation.
126
332282
2319
05:46
(Video) Narrator: Recent research in cosmology
127
334601
1979
05:48
has suggested that universes that produce
128
336580
2047
05:50
more disorder, or "entropy," over their lifetimes
129
338627
3481
05:54
should tend to have more favorable conditions
130
342108
2478
05:56
for the existence of intelligent
beings such as ourselves.
beings such as ourselves.
131
344586
3016
05:59
But what if that tentative cosmological connection
132
347602
2574
06:02
between entropy and intelligence
133
350176
1843
06:04
hints at a deeper relationship?
134
352019
1771
06:05
What if intelligent behavior doesn't just correlate
135
353790
2564
06:08
with the production of long-term entropy,
136
356354
1844
06:10
but actually emerges directly from it?
137
358198
2318
06:12
To find out, we developed a software engine
138
360516
2406
06:14
called Entropica, designed to maximize
139
362922
2503
06:17
the production of long-term entropy
140
365425
1768
06:19
of any system that it finds itself in.
141
367193
2576
06:21
Amazingly, Entropica was able to pass
142
369769
2155
06:23
multiple animal intelligence
tests, play human games,
tests, play human games,
143
371924
3456
06:27
and even earn money trading stocks,
144
375380
2146
06:29
all without being instructed to do so.
145
377526
2111
06:31
Here are some examples of Entropica in action.
146
379637
2518
06:34
Just like a human standing
upright without falling over,
upright without falling over,
147
382155
3205
06:37
here we see Entropica
148
385360
1230
06:38
automatically balancing a pole using a cart.
149
386590
2885
06:41
This behavior is remarkable in part
150
389475
2012
06:43
because we never gave Entropica a goal.
151
391487
2331
06:45
It simply decided on its own to balance the pole.
152
393818
3157
06:48
This balancing ability will have appliactions
153
396975
2132
06:51
for humanoid robotics
154
399107
1397
06:52
and human assistive technologies.
155
400504
2515
06:55
Just as some animals can use objects
156
403019
2001
06:57
in their environments as tools
157
405020
1442
06:58
to reach into narrow spaces,
158
406462
1987
07:00
here we see that Entropica,
159
408449
1882
07:02
again on its own initiative,
160
410331
1838
07:04
was able to move a large
disk representing an animal
disk representing an animal
161
412169
2910
07:07
around so as to cause a small disk,
162
415079
2345
07:09
representing a tool, to reach into a confined space
163
417424
2771
07:12
holding a third disk
164
420195
1537
07:13
and release the third disk
from its initially fixed position.
from its initially fixed position.
165
421732
2972
07:16
This tool use ability will have applications
166
424704
2189
07:18
for smart manufacturing and agriculture.
167
426893
2359
07:21
In addition, just as some other animals
168
429252
1944
07:23
are able to cooperate by pulling
opposite ends of a rope
opposite ends of a rope
169
431196
2696
07:25
at the same time to release food,
170
433892
2053
07:27
here we see that Entropica is able to accomplish
171
435945
2295
07:30
a model version of that task.
172
438240
1988
07:32
This cooperative ability has interesting implications
173
440228
2522
07:34
for economic planning and a variety of other fields.
174
442750
3435
07:38
Entropica is broadly applicable
175
446185
2071
07:40
to a variety of domains.
176
448256
1943
07:42
For example, here we see it successfully
177
450199
2442
07:44
playing a game of pong against itself,
178
452641
2559
07:47
illustrating its potential for gaming.
179
455200
2343
07:49
Here we see Entropica orchestrating
180
457543
1919
07:51
new connections on a social network
181
459462
1839
07:53
where friends are constantly falling out of touch
182
461301
2760
07:56
and successfully keeping
the network well connected.
the network well connected.
183
464061
2856
07:58
This same network orchestration ability
184
466917
2298
08:01
also has applications in health care,
185
469215
2328
08:03
energy, and intelligence.
186
471543
3232
08:06
Here we see Entropica directing the paths
187
474775
2085
08:08
of a fleet of ships,
188
476860
1486
08:10
successfully discovering and
utilizing the Panama Canal
utilizing the Panama Canal
189
478346
3175
08:13
to globally extend its reach from the Atlantic
190
481521
2458
08:15
to the Pacific.
191
483979
1529
08:17
By the same token, Entropica
192
485508
1727
08:19
is broadly applicable to problems
193
487235
1620
08:20
in autonomous defense, logistics and transportation.
194
488855
5302
08:26
Finally, here we see Entropica
195
494173
2030
08:28
spontaneously discovering and executing
196
496203
2723
08:30
a buy-low, sell-high strategy
197
498926
2067
08:32
on a simulated range traded stock,
198
500993
2178
08:35
successfully growing assets under management
199
503171
2331
08:37
exponentially.
200
505502
1424
08:38
This risk management ability
201
506926
1308
08:40
will have broad applications in finance
202
508234
2487
08:42
and insurance.
203
510721
3328
08:46
Alex Wissner-Gross: So what you've just seen
204
514049
2091
08:48
is that a variety of signature human intelligent
205
516140
4392
08:52
cognitive behaviors
206
520532
1757
08:54
such as tool use and walking upright
207
522289
2831
08:57
and social cooperation
208
525120
2029
08:59
all follow from a single equation,
209
527149
2972
09:02
which drives a system
210
530121
1932
09:04
to maximize its future freedom of action.
211
532053
3911
09:07
Now, there's a profound irony here.
212
535964
3007
09:10
Going back to the beginning
213
538971
2024
09:12
of the usage of the term robot,
214
540995
3273
09:16
the play "RUR,"
215
544268
2903
09:19
there was always a concept
216
547171
2235
09:21
that if we developed machine intelligence,
217
549406
3226
09:24
there would be a cybernetic revolt.
218
552632
3027
09:27
The machines would rise up against us.
219
555659
3551
09:31
One major consequence of this work
220
559210
2319
09:33
is that maybe all of these decades,
221
561529
2769
09:36
we've had the whole concept of cybernetic revolt
222
564298
2976
09:39
in reverse.
223
567274
2011
09:41
It's not that machines first become intelligent
224
569285
3279
09:44
and then megalomaniacal
225
572564
2015
09:46
and try to take over the world.
226
574579
2224
09:48
It's quite the opposite,
227
576803
1434
09:50
that the urge to take control
228
578237
2906
09:53
of all possible futures
229
581143
2261
09:55
is a more fundamental principle
230
583404
2118
09:57
than that of intelligence,
231
585522
1363
09:58
that general intelligence may in fact emerge
232
586885
3700
10:02
directly from this sort of control-grabbing,
233
590585
3559
10:06
rather than vice versa.
234
594144
4185
10:10
Another important consequence is goal seeking.
235
598329
3769
10:14
I'm often asked, how does the ability to seek goals
236
602098
4360
10:18
follow from this sort of framework?
237
606458
1620
10:20
And the answer is, the ability to seek goals
238
608078
3028
10:23
will follow directly from this
239
611106
1882
10:24
in the following sense:
240
612988
1834
10:26
just like you would travel through a tunnel,
241
614822
2865
10:29
a bottleneck in your future path space,
242
617687
2505
10:32
in order to achieve many other
243
620192
1871
10:34
diverse objectives later on,
244
622063
2021
10:36
or just like you would invest
245
624084
2372
10:38
in a financial security,
246
626456
1787
10:40
reducing your short-term liquidity
247
628243
2237
10:42
in order to increase your wealth over the long term,
248
630480
2400
10:44
goal seeking emerges directly
249
632880
2337
10:47
from a long-term drive
250
635217
1729
10:48
to increase future freedom of action.
251
636946
4037
10:52
Finally, Richard Feynman, famous physicist,
252
640983
3528
10:56
once wrote that if human civilization were destroyed
253
644511
3672
11:00
and you could pass only a single concept
254
648183
1893
11:02
on to our descendants
255
650076
1371
11:03
to help them rebuild civilization,
256
651447
2307
11:05
that concept should be
257
653754
1686
11:07
that all matter around us
258
655440
1852
11:09
is made out of tiny elements
259
657292
2323
11:11
that attract each other when they're far apart
260
659615
2508
11:14
but repel each other when they're close together.
261
662123
3330
11:17
My equivalent of that statement
262
665453
1781
11:19
to pass on to descendants
263
667234
1268
11:20
to help them build artificial intelligences
264
668502
2712
11:23
or to help them understand human intelligence,
265
671214
2949
11:26
is the following:
266
674163
1267
11:27
Intelligence should be viewed
267
675430
2053
11:29
as a physical process
268
677483
1413
11:30
that tries to maximize future freedom of action
269
678896
2965
11:33
and avoid constraints in its own future.
270
681861
3616
11:37
Thank you very much.
271
685477
1358
11:38
(Applause)
272
686835
4000
ABOUT THE SPEAKER
Alex Wissner-Gross - Scientist, entrepreneur, inventorAlex Wissner-Gross applies science and engineering principles to big (and diverse) questions, like: "What is the equation for intelligence?" and "What's the best way to raise awareness about climate change?"
Why you should listen
Alex Wissner-Gross is a serial big-picture thinker. He applies physics and computer science principles to a wide variety of topics, like human intelligence, climate change and financial trading.
Lately Wissner-Gross started wondering: Why have we searched for so long to understand intelligence? Can it really be this elusive? His latest work posits that intelligence can indeed be defined physically, as a dynamic force, rather than a static property. He explains intelligence in terms of causal entropic forces, ultimately defining it as "a force to maximize future freedom of action."
Wissner-Gross is a fellow at the Harvard Institute for Applied Computational Science and a research affiliate at the MIT Media Lab. He has a Ph.D. in physics from Harvard and bachelor's degrees in physics, electrical science and engineering, and mathematics from MIT.
Alex Wissner-Gross | Speaker | TED.com