1
00:00:00,512 --> 00:00:03,068
I'd like to tell you about two games of chess.

2
00:00:03,068 --> 00:00:06,932
The first happened in 1997, in which Garry Kasparov,

3
00:00:06,932 --> 00:00:10,648
a human, lost to Deep Blue, a machine.

4
00:00:10,648 --> 00:00:12,888
To many, this was the dawn of a new era,

5
00:00:12,888 --> 00:00:15,667
one where man would be dominated by machine.

6
00:00:15,667 --> 00:00:19,001
But here we are, 20 years on, and the greatest change

7
00:00:19,001 --> 00:00:21,691
in how we relate to computers is the iPad,

8
00:00:21,691 --> 00:00:23,736
not HAL.

9
00:00:23,736 --> 00:00:26,384
The second game was a freestyle chess tournament

10
00:00:26,384 --> 00:00:29,353
in 2005, in which man and machine could enter together

11
00:00:29,353 --> 00:00:34,019
as partners, rather than adversaries, if they so chose.

12
00:00:34,019 --> 00:00:35,870
At first, the results were predictable.

13
00:00:35,870 --> 00:00:38,367
Even a supercomputer was beaten by a grandmaster

14
00:00:38,367 --> 00:00:40,679
with a relatively weak laptop.

15
00:00:40,679 --> 00:00:43,664
The surprise came at the end. Who won?

16
00:00:43,664 --> 00:00:46,440
Not a grandmaster with a supercomputer,

17
00:00:46,440 --> 00:00:47,933
but actually two American amateurs

18
00:00:47,933 --> 00:00:51,755
using three relatively weak laptops.

19
00:00:51,755 --> 00:00:54,351
Their ability to coach and manipulate their computers

20
00:00:54,351 --> 00:00:56,786
to deeply explore specific positions

21
00:00:56,786 --> 00:00:59,176
effectively counteracted the superior chess knowledge

22
00:00:59,176 --> 00:01:01,785
of the grandmasters and the superior computational power

23
00:01:01,785 --> 00:01:03,694
of other adversaries.

24
00:01:03,694 --> 00:01:06,599
This is an astonishing result: average men,

25
00:01:06,599 --> 00:01:10,680
average machines beating the best man, the best machine.

26
00:01:10,680 --> 00:01:13,879
And anyways, isn't it supposed to be man versus machine?

27
00:01:13,879 --> 00:01:18,031
Instead, it's about cooperation, and the right type of cooperation.

28
00:01:18,031 --> 00:01:20,888
We've been paying a lot of attention to Marvin Minsky's

29
00:01:20,888 --> 00:01:24,130
vision for artificial intelligence over the last 50 years.

30
00:01:24,130 --> 00:01:26,392
It's a sexy vision, for sure. Many have embraced it.

31
00:01:26,392 --> 00:01:29,145
It's become the dominant school of thought in computer science.

32
00:01:29,145 --> 00:01:32,217
But as we enter the era of big data, of network systems,

33
00:01:32,217 --> 00:01:34,915
of open platforms, and embedded technology,

34
00:01:34,915 --> 00:01:38,307
I'd like to suggest it's time to reevaluate an alternative vision

35
00:01:38,307 --> 00:01:41,377
that was actually developed around the same time.

36
00:01:41,377 --> 00:01:44,709
I'm talking about J.C.R. Licklider's human-computer symbiosis,

37
00:01:44,709 --> 00:01:48,517
perhaps better termed "intelligence augmentation," I.A.

38
00:01:48,517 --> 00:01:51,157
Licklider was a computer science titan who had a profound

39
00:01:51,157 --> 00:01:54,163
effect on the development of technology and the Internet.

40
00:01:54,163 --> 00:01:57,031
His vision was to enable man and machine to cooperate

41
00:01:57,031 --> 00:02:00,621
in making decisions, controlling complex situations

42
00:02:00,621 --> 00:02:02,391
without the inflexible dependence

43
00:02:02,391 --> 00:02:04,924
on predetermined programs.

44
00:02:04,924 --> 00:02:07,422
Note that word "cooperate."

45
00:02:07,422 --> 00:02:10,169
Licklider encourages us not to take a toaster

46
00:02:10,169 --> 00:02:12,453
and make it Data from "Star Trek,"

47
00:02:12,453 --> 00:02:15,988
but to take a human and make her more capable.

48
00:02:15,988 --> 00:02:17,899
Humans are so amazing -- how we think,

49
00:02:17,899 --> 00:02:20,517
our non-linear approaches, our creativity,

50
00:02:20,517 --> 00:02:22,648
iterative hypotheses, all very difficult if possible at all

51
00:02:22,648 --> 00:02:23,993
for computers to do.

52
00:02:23,993 --> 00:02:26,445
Licklider intuitively realized this, contemplating humans

53
00:02:26,445 --> 00:02:28,772
setting the goals, formulating the hypotheses,

54
00:02:28,772 --> 00:02:32,148
determining the criteria, and performing the evaluation.

55
00:02:32,148 --> 00:02:33,923
Of course, in other ways, humans are so limited.

56
00:02:33,923 --> 00:02:37,158
We're terrible at scale, computation and volume.

57
00:02:37,158 --> 00:02:38,994
We require high-end talent management

58
00:02:38,994 --> 00:02:41,058
to keep the rock band together and playing.

59
00:02:41,058 --> 00:02:43,262
Licklider foresaw computers doing all the routinizable work

60
00:02:43,262 --> 00:02:46,538
that was required to prepare the way for insights and decision making.

61
00:02:46,538 --> 00:02:48,762
Silently, without much fanfare,

62
00:02:48,762 --> 00:02:52,116
this approach has been compiling victories beyond chess.

63
00:02:52,116 --> 00:02:55,472
Protein folding, a topic that shares the incredible expansiveness of chess —

64
00:02:55,472 --> 00:02:58,514
there are more ways of folding a protein than there are atoms in the universe.

65
00:02:58,514 --> 00:03:00,867
This is a world-changing problem with huge implications

66
00:03:00,867 --> 00:03:03,175
for our ability to understand and treat disease.

67
00:03:03,175 --> 00:03:07,423
And for this task, supercomputer field brute force simply isn't enough.

68
00:03:07,423 --> 00:03:09,807
Foldit, a game created by computer scientists,

69
00:03:09,807 --> 00:03:12,309
illustrates the value of the approach.

70
00:03:12,309 --> 00:03:15,350
Non-technical, non-biologist amateurs play a video game

71
00:03:15,350 --> 00:03:18,423
in which they visually rearrange the structure of the protein,

72
00:03:18,423 --> 00:03:19,922
allowing the computer to manage the atomic forces

73
00:03:19,922 --> 00:03:22,879
and interactions and identify structural issues.

74
00:03:22,879 --> 00:03:25,902
This approach beat supercomputers 50 percent of the time

75
00:03:25,902 --> 00:03:28,486
and tied 30 percent of the time.

76
00:03:28,486 --> 00:03:31,623
Foldit recently made a notable and major scientific discovery

77
00:03:31,623 --> 00:03:34,783
by deciphering the structure of the Mason-Pfizer monkey virus.

78
00:03:34,783 --> 00:03:37,798
A protease that had eluded determination for over 10 years

79
00:03:37,798 --> 00:03:40,424
was solved was by three players in a matter of days,

80
00:03:40,424 --> 00:03:42,449
perhaps the first major scientific advance

81
00:03:42,449 --> 00:03:44,772
to come from playing a video game.

82
00:03:44,772 --> 00:03:46,953
Last year, on the site of the Twin Towers,

83
00:03:46,953 --> 00:03:48,426
the 9/11 memorial opened.

84
00:03:48,426 --> 00:03:51,147
It displays the names of the thousands of victims

85
00:03:51,147 --> 00:03:54,210
using a beautiful concept called "meaningful adjacency."

86
00:03:54,210 --> 00:03:56,376
It places the names next to each other based on their

87
00:03:56,376 --> 00:03:58,589
relationships to one another: friends, families, coworkers.

88
00:03:58,589 --> 00:04:01,617
When you put it all together, it's quite a computational

89
00:04:01,617 --> 00:04:05,840
challenge: 3,500 victims, 1,800 adjacency requests,

90
00:04:05,840 --> 00:04:08,932
the importance of the overall physical specifications

91
00:04:08,932 --> 00:04:11,069
and the final aesthetics.

92
00:04:11,069 --> 00:04:13,684
When first reported by the media, full credit for such a feat

93
00:04:13,684 --> 00:04:15,576
was given to an algorithm from the New York City

94
00:04:15,576 --> 00:04:19,577
design firm Local Projects. The truth is a bit more nuanced.

95
00:04:19,577 --> 00:04:22,448
While an algorithm was used to develop the underlying framework,

96
00:04:22,448 --> 00:04:25,456
humans used that framework to design the final result.

97
00:04:25,456 --> 00:04:27,681
So in this case, a computer had evaluated millions

98
00:04:27,681 --> 00:04:31,016
of possible layouts, managed a complex relational system,

99
00:04:31,016 --> 00:04:33,430
and kept track of a very large set of measurements

100
00:04:33,430 --> 00:04:35,840
and variables, allowing the humans to focus

101
00:04:35,840 --> 00:04:38,642
on design and compositional choices.

102
00:04:38,642 --> 00:04:39,678
So the more you look around you,

103
00:04:39,678 --> 00:04:41,640
the more you see Licklider's vision everywhere.

104
00:04:41,640 --> 00:04:44,944
Whether it's augmented reality in your iPhone or GPS in your car,

105
00:04:44,944 --> 00:04:47,914
human-computer symbiosis is making us more capable.

106
00:04:47,914 --> 00:04:49,569
So if you want to improve human-computer symbiosis,

107
00:04:49,569 --> 00:04:50,998
what can you do?

108
00:04:50,998 --> 00:04:53,450
You can start by designing the human into the process.

109
00:04:53,450 --> 00:04:55,654
Instead of thinking about what a computer will do to solve the problem,

110
00:04:55,654 --> 00:04:59,523
design the solution around what the human will do as well.

111
00:04:59,523 --> 00:05:01,460
When you do this, you'll quickly realize that you spent

112
00:05:01,460 --> 00:05:04,339
all of your time on the interface between man and machine,

113
00:05:04,339 --> 00:05:07,438
specifically on designing away the friction in the interaction.

114
00:05:07,438 --> 00:05:10,204
In fact, this friction is more important than the power

115
00:05:10,204 --> 00:05:12,256
of the man or the power of the machine

116
00:05:12,256 --> 00:05:14,187
in determining overall capability.

117
00:05:14,187 --> 00:05:16,164
That's why two amateurs with a few laptops

118
00:05:16,164 --> 00:05:18,620
handily beat a supercomputer and a grandmaster.

119
00:05:18,620 --> 00:05:21,625
What Kasparov calls process is a byproduct of friction.

120
00:05:21,625 --> 00:05:24,026
The better the process, the less the friction.

121
00:05:24,026 --> 00:05:28,282
And minimizing friction turns out to be the decisive variable.

122
00:05:28,282 --> 00:05:30,525
Or take another example: big data.

123
00:05:30,525 --> 00:05:32,431
Every interaction we have in the world is recorded

124
00:05:32,431 --> 00:05:35,490
by an ever growing array of sensors: your phone,

125
00:05:35,490 --> 00:05:37,863
your credit card, your computer. The result is big data,

126
00:05:37,863 --> 00:05:39,605
and it actually presents us with an opportunity

127
00:05:39,605 --> 00:05:42,267
to more deeply understand the human condition.

128
00:05:42,267 --> 00:05:44,572
The major emphasis of most approaches to big data

129
00:05:44,572 --> 00:05:46,787
focus on, "How do I store this data? How do I search

130
00:05:46,787 --> 00:05:49,063
this data? How do I process this data?"

131
00:05:49,063 --> 00:05:51,267
These are necessary but insufficient questions.

132
00:05:51,267 --> 00:05:53,738
The imperative is not to figure out how to compute,

133
00:05:53,738 --> 00:05:55,922
but what to compute. How do you impose human intuition

134
00:05:55,922 --> 00:05:57,713
on data at this scale?

135
00:05:57,713 --> 00:06:01,212
Again, we start by designing the human into the process.

136
00:06:01,212 --> 00:06:04,024
When PayPal was first starting as a business, their biggest

137
00:06:04,024 --> 00:06:06,828
challenge was not, "How do I send money back and forth online?"

138
00:06:06,828 --> 00:06:10,700
It was, "How do I do that without being defrauded by organized crime?"

139
00:06:10,700 --> 00:06:12,788
Why so challenging? Because while computers can learn

140
00:06:12,788 --> 00:06:15,932
to detect and identify fraud based on patterns,

141
00:06:15,932 --> 00:06:17,411
they can't learn to do that based on patterns

142
00:06:17,411 --> 00:06:19,527
they've never seen before, and organized crime

143
00:06:19,527 --> 00:06:22,236
has a lot in common with this audience: brilliant people,

144
00:06:22,236 --> 00:06:25,876
relentlessly resourceful, entrepreneurial spirit — (Laughter) —

145
00:06:25,876 --> 00:06:28,588
and one huge and important difference: purpose.

146
00:06:28,588 --> 00:06:31,420
And so while computers alone can catch all but the cleverest

147
00:06:31,420 --> 00:06:33,673
fraudsters, catching the cleverest is the difference

148
00:06:33,673 --> 00:06:36,218
between success and failure.

149
00:06:36,218 --> 00:06:38,439
There's a whole class of problems like this, ones with

150
00:06:38,439 --> 00:06:41,014
adaptive adversaries. They rarely if ever present with a

151
00:06:41,014 --> 00:06:43,750
repeatable pattern that's discernable to computers.

152
00:06:43,750 --> 00:06:47,743
Instead, there's some inherent component of innovation or disruption,

153
00:06:47,743 --> 00:06:50,478
and increasingly these problems are buried in big data.

154
00:06:50,478 --> 00:06:52,978
For example, terrorism. Terrorists are always adapting

155
00:06:52,978 --> 00:06:55,030
in minor and major ways to new circumstances, and despite

156
00:06:55,030 --> 00:06:58,124
what you might see on TV, these adaptations,

157
00:06:58,124 --> 00:07:00,417
and the detection of them, are fundamentally human.

158
00:07:00,417 --> 00:07:03,534
Computers don't detect novel patterns and new behaviors,

159
00:07:03,534 --> 00:07:06,769
but humans do. Humans, using technology, testing hypotheses,

160
00:07:06,769 --> 00:07:11,389
searching for insight by asking machines to do things for them.

161
00:07:11,389 --> 00:07:13,709
Osama bin Laden was not caught by artificial intelligence.

162
00:07:13,709 --> 00:07:16,262
He was caught by dedicated, resourceful, brilliant people

163
00:07:16,262 --> 00:07:20,531
in partnerships with various technologies.

164
00:07:20,531 --> 00:07:23,349
As appealing as it might sound, you cannot algorithmically

165
00:07:23,349 --> 00:07:24,950
data mine your way to the answer.

166
00:07:24,950 --> 00:07:27,805
There is no "Find Terrorist" button, and the more data

167
00:07:27,805 --> 00:07:30,107
we integrate from a vast variety of sources

168
00:07:30,107 --> 00:07:32,240
across a wide variety of data formats from very

169
00:07:32,240 --> 00:07:35,549
disparate systems, the less effective data mining can be.

170
00:07:35,549 --> 00:07:37,573
Instead, people will have to look at data

171
00:07:37,573 --> 00:07:41,029
and search for insight, and as Licklider foresaw long ago,

172
00:07:41,029 --> 00:07:43,714
the key to great results here is the right type of cooperation,

173
00:07:43,714 --> 00:07:45,238
and as Kasparov realized,

174
00:07:45,238 --> 00:07:48,269
that means minimizing friction at the interface.

175
00:07:48,269 --> 00:07:51,027
Now this approach makes possible things like combing

176
00:07:51,027 --> 00:07:54,413
through all available data from very different sources,

177
00:07:54,413 --> 00:07:57,205
identifying key relationships and putting them in one place,

178
00:07:57,205 --> 00:08:00,133
something that's been nearly impossible to do before.

179
00:08:00,133 --> 00:08:02,075
To some, this has terrifying privacy and civil liberties

180
00:08:02,075 --> 00:08:05,485
implications. To others it foretells of an era of greater

181
00:08:05,485 --> 00:08:07,394
privacy and civil liberties protections,

182
00:08:07,394 --> 00:08:10,330
but privacy and civil liberties are of fundamental importance.

183
00:08:10,330 --> 00:08:12,523
That must be acknowledged, and they can't be swept aside,

184
00:08:12,523 --> 00:08:15,053
even with the best of intents.

185
00:08:15,053 --> 00:08:17,571
So let's explore, through a couple of examples, the impact

186
00:08:17,571 --> 00:08:19,977
that technologies built to drive human-computer symbiosis

187
00:08:19,977 --> 00:08:22,896
have had in recent time.

188
00:08:22,896 --> 00:08:26,312
In October, 2007, U.S. and coalition forces raided

189
00:08:26,312 --> 00:08:28,728
an al Qaeda safe house in the city of Sinjar

190
00:08:28,728 --> 00:08:30,662
on the Syrian border of Iraq.

191
00:08:30,662 --> 00:08:33,038
They found a treasure trove of documents:

192
00:08:33,038 --> 00:08:35,373
700 biographical sketches of foreign fighters.

193
00:08:35,373 --> 00:08:37,957
These foreign fighters had left their families in the Gulf,

194
00:08:37,957 --> 00:08:41,103
the Levant and North Africa to join al Qaeda in Iraq.

195
00:08:41,103 --> 00:08:42,719
These records were human resource forms.

196
00:08:42,719 --> 00:08:45,574
The foreign fighters filled them out as they joined the organization.

197
00:08:45,574 --> 00:08:46,785
It turns out that al Qaeda, too,

198
00:08:46,785 --> 00:08:49,382
is not without its bureaucracy. (Laughter)

199
00:08:49,382 --> 00:08:51,480
They answered questions like, "Who recruited you?"

200
00:08:51,480 --> 00:08:54,334
"What's your hometown?" "What occupation do you seek?"

201
00:08:54,334 --> 00:08:57,503
In that last question, a surprising insight was revealed.

202
00:08:57,503 --> 00:08:59,903
The vast majority of foreign fighters

203
00:08:59,903 --> 00:09:02,303
were seeking to become suicide bombers for martyrdom --

204
00:09:02,303 --> 00:09:06,641
hugely important, since between 2003 and 2007, Iraq

205
00:09:06,641 --> 00:09:10,885
had 1,382 suicide bombings, a major source of instability.

206
00:09:10,885 --> 00:09:12,943
Analyzing this data was hard. The originals were sheets

207
00:09:12,943 --> 00:09:15,685
of paper in Arabic that had to be scanned and translated.

208
00:09:15,685 --> 00:09:17,877
The friction in the process did not allow for meaningful

209
00:09:17,877 --> 00:09:21,227
results in an operational time frame using humans, PDFs

210
00:09:21,227 --> 00:09:23,445
and tenacity alone.

211
00:09:23,445 --> 00:09:25,398
The researchers had to lever up their human minds

212
00:09:25,398 --> 00:09:27,743
with technology to dive deeper, to explore non-obvious

213
00:09:27,743 --> 00:09:30,961
hypotheses, and in fact, insights emerged.

214
00:09:30,961 --> 00:09:33,605
Twenty percent of the foreign fighters were from Libya,

215
00:09:33,605 --> 00:09:36,573
50 percent of those from a single town in Libya,

216
00:09:36,573 --> 00:09:39,023
hugely important since prior statistics put that figure at

217
00:09:39,023 --> 00:09:41,406
three percent. It also helped to hone in on a figure

218
00:09:41,406 --> 00:09:44,383
of rising importance in al Qaeda, Abu Yahya al-Libi,

219
00:09:44,383 --> 00:09:47,014
a senior cleric in the Libyan Islamic fighting group.

220
00:09:47,014 --> 00:09:49,678
In March of 2007, he gave a speech, after which there was

221
00:09:49,678 --> 00:09:53,144
a surge in participation amongst Libyan foreign fighters.

222
00:09:53,144 --> 00:09:56,250
Perhaps most clever of all, though, and least obvious,

223
00:09:56,250 --> 00:09:58,323
by flipping the data on its head, the researchers were

224
00:09:58,323 --> 00:10:01,223
able to deeply explore the coordination networks in Syria

225
00:10:01,223 --> 00:10:03,740
that were ultimately responsible for receiving and

226
00:10:03,740 --> 00:10:06,204
transporting the foreign fighters to the border.

227
00:10:06,204 --> 00:10:08,837
These were networks of mercenaries, not ideologues,

228
00:10:08,837 --> 00:10:11,235
who were in the coordination business for profit.

229
00:10:11,235 --> 00:10:13,139
For example, they charged Saudi foreign fighters

230
00:10:13,139 --> 00:10:15,338
substantially more than Libyans, money that would have

231
00:10:15,338 --> 00:10:17,658
otherwise gone to al Qaeda.

232
00:10:17,658 --> 00:10:19,703
Perhaps the adversary would disrupt their own network

233
00:10:19,703 --> 00:10:22,738
if they knew they cheating would-be jihadists.

234
00:10:22,738 --> 00:10:26,483
In January, 2010, a devastating 7.0 earthquake struck Haiti,

235
00:10:26,483 --> 00:10:29,399
third deadliest earthquake of all time, left one million people,

236
00:10:29,399 --> 00:10:31,983
10 percent of the population, homeless.

237
00:10:31,983 --> 00:10:35,120
One seemingly small aspect of the overall relief effort

238
00:10:35,120 --> 00:10:37,296
became increasingly important as the delivery of food

239
00:10:37,296 --> 00:10:39,456
and water started rolling.

240
00:10:39,456 --> 00:10:40,914
January and February are the dry months in Haiti,

241
00:10:40,914 --> 00:10:43,856
yet many of the camps had developed standing water.

242
00:10:43,856 --> 00:10:45,978
The only institution with detailed knowledge of Haiti's

243
00:10:45,978 --> 00:10:47,275
floodplains had been leveled

244
00:10:47,275 --> 00:10:50,283
in the earthquake, leadership inside.

245
00:10:50,283 --> 00:10:52,858
So the question is, which camps are at risk,

246
00:10:52,858 --> 00:10:54,779
how many people are in these camps, what's the

247
00:10:54,779 --> 00:10:57,090
timeline for flooding, and given very limited resources

248
00:10:57,090 --> 00:11:00,474
and infrastructure, how do we prioritize the relocation?

249
00:11:00,474 --> 00:11:02,818
The data was incredibly disparate. The U.S. Army had

250
00:11:02,818 --> 00:11:05,747
detailed knowledge for only a small section of the country.

251
00:11:05,747 --> 00:11:08,258
There was data online from a 2006 environmental risk

252
00:11:08,258 --> 00:11:10,922
conference, other geospatial data, none of it integrated.

253
00:11:10,922 --> 00:11:13,880
The human goal here was to identify camps for relocation

254
00:11:13,880 --> 00:11:16,275
based on priority need.

255
00:11:16,275 --> 00:11:18,715
The computer had to integrate a vast amount of geospacial

256
00:11:18,715 --> 00:11:21,299
information, social media data and relief organization

257
00:11:21,299 --> 00:11:24,779
information to answer this question.

258
00:11:24,779 --> 00:11:27,194
By implementing a superior process, what was otherwise

259
00:11:27,194 --> 00:11:29,802
a task for 40 people over three months became

260
00:11:29,802 --> 00:11:32,978
a simple job for three people in 40 hours,

261
00:11:32,978 --> 00:11:35,606
all victories for human-computer symbiosis.

262
00:11:35,606 --> 00:11:37,660
We're more than 50 years into Licklider's vision

263
00:11:37,660 --> 00:11:39,902
for the future, and the data suggests that we should be

264
00:11:39,902 --> 00:11:42,932
quite excited about tackling this century's hardest problems,

265
00:11:42,932 --> 00:11:45,879
man and machine in cooperation together.

266
00:11:45,879 --> 00:11:48,076
Thank you. (Applause)

267
00:11:48,076 --> 00:11:50,581
(Applause)