1
00:00:01,712 --> 00:00:02,739
(mechanical gears whirring)

2
00:00:02,739 --> 00:00:05,331
(skull cracking)

3
00:00:05,331 --> 00:00:08,309
Get the human.

4
00:00:08,309 --> 00:00:11,662
(lasers zapping)

5
00:00:11,662 --> 00:00:14,495
(Enlai screaming)

6
00:00:19,124 --> 00:00:20,190
Thanks to Hollywood,

7
00:00:20,190 --> 00:00:23,617
this is our image of artificial intelligence.

8
00:00:28,556 --> 00:00:31,556
(suspenseful music)

9
00:00:39,618 --> 00:00:41,300
(Enlai sighing)

10
00:00:41,300 --> 00:00:42,900
Surrender, human.

11
00:00:43,804 --> 00:00:45,643
(suspenseful music)

12
00:00:45,643 --> 00:00:48,040
We imagine AI taking over the planet,

13
00:00:48,040 --> 00:00:49,880
dominating mankind.

14
00:00:49,880 --> 00:00:51,467
But what is the real picture?

15
00:00:51,467 --> 00:00:55,003
What is artificial intelligence?

16
00:00:56,761 --> 00:00:59,401
What?

17
00:00:59,401 --> 00:01:02,409
(mechanical gears whirring)

18
00:01:02,409 --> 00:01:05,322
{\an8}(lights zapping)

19
00:01:05,322 --> 00:01:08,032
{\an8}(hydraulics whirring)

20
00:01:08,032 --> 00:01:11,282
{\an8}(haunting synth music)

21
00:01:13,563 --> 00:01:14,896
{\an8}Get the human.

22
00:01:18,339 --> 00:01:21,114
{\an8}And nothing but the truth.

23
00:01:21,114 --> 00:01:22,900
{\an8}Honey, would you life some coffee?

24
00:01:22,900 --> 00:01:24,145
{\an8}Yes, please.

25
00:01:24,145 --> 00:01:26,895
{\an8}(drill whirring)

26
00:01:29,347 --> 00:01:30,834
(hard drive booting)

27
00:01:30,834 --> 00:01:33,501
(Enlai jolting)

28
00:01:34,821 --> 00:01:35,988
Is it my turn?

29
00:01:36,974 --> 00:01:39,391
I'm still thinking.

30
00:01:40,800 --> 00:01:41,991
(mechanical gears whirring)

31
00:01:41,991 --> 00:01:44,658
(robot beeping)

32
00:01:45,820 --> 00:01:46,848
{\an8}(students bustling)

33
00:01:46,848 --> 00:01:50,431
{\an8}(curious orchestral music)

34
00:01:52,840 --> 00:01:54,190
Alan Turing is considered

35
00:01:54,190 --> 00:01:55,930
the father of computer science

36
00:01:55,930 --> 00:01:57,840
and artificial intelligence.

37
00:01:57,840 --> 00:01:59,890
In 1950, he wrote that

38
00:01:59,890 --> 00:02:03,490
{\an8}a computer would deserve to be considered intelligent

39
00:02:03,490 --> 00:02:07,633
if it could convince a human being that it was human.

40
00:02:09,898 --> 00:02:13,481
{\an8}(playful orchestral music)

41
00:02:19,730 --> 00:02:22,470
Before I try to understand artificial intelligence,

42
00:02:22,470 --> 00:02:26,810
I first want to better appreciate natural intelligence,

43
00:02:26,810 --> 00:02:29,874
so I'm here at the Singapore Zoo to get face to face

44
00:02:29,874 --> 00:02:33,833
with some of nature's most intelligent creatures.

45
00:02:36,934 --> 00:02:40,350
Orangutans are one of mankind's closest relatives.

46
00:02:40,350 --> 00:02:41,530
I don't know about you,

47
00:02:41,530 --> 00:02:44,370
but when I see some of my own human relatives,

48
00:02:44,370 --> 00:02:47,140
I don't always see intelligence.

49
00:02:47,140 --> 00:02:48,860
But when I look at these guys

50
00:02:48,860 --> 00:02:50,361
and see the spark behind their eyes,

51
00:02:50,361 --> 00:02:51,843
wow.

52
00:02:54,002 --> 00:02:56,958
More intelligent animals should be able to learn more

53
00:02:56,958 --> 00:03:00,213
and to solve more complex problems.

54
00:03:02,730 --> 00:03:03,860
Right now, the orangutans are

55
00:03:03,860 --> 00:03:05,528
enjoying their puzzle feeders.

56
00:03:05,528 --> 00:03:06,627
Based on their intelligence,

57
00:03:06,627 --> 00:03:11,290
{\an8}they have to figure out how to get those nuts to enjoy it.

58
00:03:11,290 --> 00:03:12,750
So it's like a test.

59
00:03:12,750 --> 00:03:14,370
A test on their brains.

60
00:03:14,370 --> 00:03:15,836
In the puzzle is like a maze,

61
00:03:15,836 --> 00:03:19,738
so the orangutans have got to use a stick to push that nut

62
00:03:19,738 --> 00:03:22,723
until he gets to the hole and then it drops down.

63
00:03:23,832 --> 00:03:25,560
But the next layer of that opening

64
00:03:25,560 --> 00:03:29,700
is either smaller or bigger, so the stick can't fit in.

65
00:03:29,700 --> 00:03:32,803
So they got to find another stick or modify the stick.

66
00:03:37,430 --> 00:03:38,680
So some of the orangutans are smart

67
00:03:38,680 --> 00:03:41,020
by bringing two or three sticks of various sizes,

68
00:03:41,020 --> 00:03:43,620
so they don't have to go around looking for it.

69
00:03:43,620 --> 00:03:46,060
When we first installed the puzzle feeder,

70
00:03:46,060 --> 00:03:48,640
some of them took only a few seconds.

71
00:03:48,640 --> 00:03:49,610
They outsmarted you.

72
00:03:49,610 --> 00:03:51,450
Outsmarted us sometimes, so now we are trying to

73
00:03:51,450 --> 00:03:54,250
figure out how to outsmart them again

74
00:03:54,250 --> 00:03:55,740
so they won't do it so fast.

75
00:03:55,740 --> 00:03:58,078
They are pretty amazing, aren't they?

76
00:03:58,078 --> 00:04:01,610
Aping the way living creatures see and learn

77
00:04:01,610 --> 00:04:05,933
is the goal of a whole branch of AI: computer vision.

78
00:04:08,940 --> 00:04:10,809
(mechanical clicking)

79
00:04:10,809 --> 00:04:13,359
(robot whistling)

80
00:04:13,359 --> 00:04:16,300
(playful orchestral music)

81
00:04:16,300 --> 00:04:19,320
While a digital camera can take high quality images,

82
00:04:19,320 --> 00:04:22,800
it doesn't really see or know what it's looking at.

83
00:04:22,800 --> 00:04:25,350
To the machine, it's just a bunch of code and data.

84
00:04:31,120 --> 00:04:31,973
Processing.

85
00:04:32,990 --> 00:04:36,754
A group of zebra standing on top of a lush green forest.

86
00:04:36,754 --> 00:04:38,700
But with computer vision,

87
00:04:38,700 --> 00:04:42,050
machines learn how to recognize what they see,

88
00:04:42,050 --> 00:04:45,330
and because this resembles how we see and learn,

89
00:04:45,330 --> 00:04:48,303
you can kind of call it artificial intelligence.

90
00:04:51,652 --> 00:04:52,540
(phone beeping)

91
00:04:52,540 --> 00:04:54,120
A giraffe eating leaves from a tree.

92
00:04:54,120 --> 00:04:55,280
Yes!

93
00:04:55,280 --> 00:04:56,113
Correct.

94
00:04:58,223 --> 00:04:59,190
(phone beeping)

95
00:04:59,190 --> 00:05:00,023
A herd of elephants

96
00:05:00,023 --> 00:05:01,830
standing next to a body of water.

97
00:05:01,830 --> 00:05:02,663
Oh yes!

98
00:05:04,910 --> 00:05:07,776
32-year-old man with black hair, looking neutral.

99
00:05:07,776 --> 00:05:09,532
(Enlai chuckling)

100
00:05:09,532 --> 00:05:12,350
32 years old, heh.

101
00:05:12,350 --> 00:05:13,353
How accurate!

102
00:05:15,070 --> 00:05:18,520
(groovy rock music)

103
00:05:18,520 --> 00:05:19,360
This is Annabelle.

104
00:05:19,360 --> 00:05:20,193
Hi!

105
00:05:21,033 --> 00:05:22,400
She and the good folks at Neural Bay

106
00:05:22,400 --> 00:05:25,860
have prepared a sexy demonstration for me

107
00:05:25,860 --> 00:05:28,750
to show me how a computer can learn to

108
00:05:28,750 --> 00:05:31,900
see and recognize things.

109
00:05:31,900 --> 00:05:36,097
And today we'll be training a computer to hunt for...

110
00:05:36,097 --> 00:05:37,111
Boobies!

111
00:05:37,111 --> 00:05:38,201
(Enlai laughing)

112
00:05:38,201 --> 00:05:40,550
As you know, to be specific, just a cleavage like this,

113
00:05:40,550 --> 00:05:41,400
{\an8}this line here.

114
00:05:41,400 --> 00:05:43,807
{\an8}Otherwise known as a longkang!

115
00:05:43,807 --> 00:05:46,593
Why on earth are we doing this?

116
00:05:46,593 --> 00:05:50,755
{\an8}For fun, but basically raise up, you know,

117
00:05:50,755 --> 00:05:54,020
{\an8}very serious topics about data ethics,

118
00:05:54,020 --> 00:05:57,483
data collection and privacy, and also implementation.

119
00:06:01,150 --> 00:06:05,440
How did you teach the computer cleavage?

120
00:06:05,440 --> 00:06:08,700
It's a very, very human-intensive task,

121
00:06:08,700 --> 00:06:09,533
like what Yen is doing.

122
00:06:09,533 --> 00:06:12,181
He's drawing boxes around the image,

123
00:06:12,181 --> 00:06:15,970
just the part where the cleavage is meant to be detected.

124
00:06:15,970 --> 00:06:17,228
Okay.

125
00:06:17,228 --> 00:06:19,720
You've been staring at how many?

126
00:06:19,720 --> 00:06:20,553
2,000.

127
00:06:20,553 --> 00:06:23,140
2,000 sets of boobies.

128
00:06:23,140 --> 00:06:25,667
So let me give an analogy

129
00:06:25,667 --> 00:06:28,210
as to why we are actually drawing this labels, right?

130
00:06:28,210 --> 00:06:30,490
So imagine a little girl,

131
00:06:30,490 --> 00:06:31,560
and you want to teach a little girl

132
00:06:31,560 --> 00:06:33,600
how to recognize a dog, right?

133
00:06:33,600 --> 00:06:35,867
You hold a little girl's head and your brain is,

134
00:06:35,867 --> 00:06:36,720
"Go, go.

135
00:06:36,720 --> 00:06:38,190
This is dog one, dog two, dog three.

136
00:06:38,190 --> 00:06:41,979
So you point out another dog, totally different to the girl.

137
00:06:41,979 --> 00:06:44,496
The girl is gonna guess, right?

138
00:06:44,496 --> 00:06:46,580
She'll say, "I think it's a dog."

139
00:06:46,580 --> 00:06:49,960
And that's where it's a guess with maybe 80% certainty

140
00:06:49,960 --> 00:06:51,220
that that is a dog.

141
00:06:51,220 --> 00:06:53,370
So that's what we call a confidence score.

142
00:06:53,370 --> 00:06:56,210
So how do we increase this confidence score?

143
00:06:56,210 --> 00:06:59,400
You show a girl hundred thousand images of dogs, right?

144
00:06:59,400 --> 00:07:02,067
And eventually when she sees a dog, she's going to go like,

145
00:07:02,067 --> 00:07:05,597
"Oh yeah, that is definitely for sure a dog."

146
00:07:09,200 --> 00:07:10,330
For this experiment,

147
00:07:10,330 --> 00:07:12,270
they've trained a computer to judge

148
00:07:12,270 --> 00:07:14,853
when cleavage is socially acceptable.

149
00:07:16,630 --> 00:07:18,143
I retrained this as acceptable,

150
00:07:18,143 --> 00:07:20,680
and all the way down here as unacceptable,

151
00:07:20,680 --> 00:07:23,563
and whatever's in the middle, we let the computer decide.

152
00:07:26,350 --> 00:07:29,500
We have a few clothing items right there

153
00:07:29,500 --> 00:07:32,503
and you can dress Vivian Vicky.

154
00:07:35,160 --> 00:07:38,503
So if she were to be fairly conservative.

155
00:07:40,150 --> 00:07:42,770
Okay, yeah, acceptable, acceptable.

156
00:07:42,770 --> 00:07:44,236
I agree.

157
00:07:44,236 --> 00:07:45,836
Okay, oh.

158
00:07:45,836 --> 00:07:47,196
Oh.

159
00:07:47,196 --> 00:07:48,029
(Annabelle chuckling)

160
00:07:48,029 --> 00:07:48,862
Wow.

161
00:07:48,862 --> 00:07:49,963
Put it down.

162
00:07:49,963 --> 00:07:50,796
Still okay.

163
00:07:50,796 --> 00:07:51,629
Just keep going.

164
00:07:51,629 --> 00:07:52,462
We're getting a gray area right here.

165
00:07:52,462 --> 00:07:53,469
Ah, there we go!

166
00:07:53,469 --> 00:07:54,681
Oh!

167
00:07:54,681 --> 00:07:58,420
I wouldn't want my kids in the future to kind of like

168
00:07:58,420 --> 00:08:00,793
see somebody walking around looking like that.

169
00:08:03,772 --> 00:08:05,295
Oh, now it's acceptable.

170
00:08:05,295 --> 00:08:06,224
Okay, yes.

171
00:08:06,224 --> 00:08:07,307
Yes, it works, it works.

172
00:08:07,307 --> 00:08:08,140
Yes!

173
00:08:08,140 --> 00:08:09,843
Yay!

174
00:08:11,680 --> 00:08:14,689
Well, I guess it's time to test it on a human being.

175
00:08:14,689 --> 00:08:15,870
That's you?

176
00:08:15,870 --> 00:08:16,703
Mm-hmm.

177
00:08:16,703 --> 00:08:17,536
Okay.

178
00:08:17,536 --> 00:08:20,008
Because I don't believe my acting is like a mannequin.

179
00:08:20,008 --> 00:08:21,060
(Annabelle laughing)

180
00:08:21,060 --> 00:08:22,053
First button.

181
00:08:24,140 --> 00:08:24,973
Ha ha.

182
00:08:24,973 --> 00:08:25,806
Ta da!

183
00:08:28,568 --> 00:08:30,000
Acceptable or unacceptable?

184
00:08:30,000 --> 00:08:31,020
Unacceptable.

185
00:08:31,020 --> 00:08:33,686
Oh finally, thank goodness.

186
00:08:33,686 --> 00:08:34,519
(Annabelle laughing)

187
00:08:34,519 --> 00:08:35,352
Yep!

188
00:08:35,352 --> 00:08:36,185
I was getting a bit worried.

189
00:08:36,185 --> 00:08:37,563
I thought I'm going to have to take off my shirt.

190
00:08:37,563 --> 00:08:39,670
If you walk out of your house like that,

191
00:08:39,670 --> 00:08:44,670
I'm pretty sure a lot of people will find this unacceptable.

192
00:08:44,683 --> 00:08:45,880
Why?

193
00:08:45,880 --> 00:08:47,340
What's wrong?

194
00:08:47,340 --> 00:08:50,600
From now on, I will be walking out of my house like this.

195
00:08:50,600 --> 00:08:52,133
Hide your kids!

196
00:08:52,133 --> 00:08:54,660
(both laughing)

197
00:08:54,660 --> 00:08:55,630
The computer scientists

198
00:08:55,630 --> 00:08:57,040
at the National University of Singapore

199
00:08:57,040 --> 00:08:59,970
claim that their robot can recognize common objects

200
00:08:59,970 --> 00:09:01,228
that you place in front of it.

201
00:09:01,228 --> 00:09:02,270
(bowl dinging)

202
00:09:02,270 --> 00:09:06,745
So I brought some random things to put it to the test.

203
00:09:06,745 --> 00:09:08,600
(duckie squeaking)

204
00:09:08,600 --> 00:09:12,810
Alexa, ask Jerry to pick up the rubber duck.

205
00:09:12,810 --> 00:09:15,450
Okay, looking for the rubber duck.

206
00:09:15,450 --> 00:09:16,283
Hold on.

207
00:09:20,495 --> 00:09:22,130
(duckie squeaking)

208
00:09:22,130 --> 00:09:23,620
Well, the magic here is that

209
00:09:23,620 --> 00:09:25,960
{\an8}even though Jerry has never seen

210
00:09:25,960 --> 00:09:27,930
{\an8}any of this particular object,

211
00:09:27,930 --> 00:09:30,490
he has seen thousands of images of all sorts of

212
00:09:30,490 --> 00:09:32,010
different kinds of objects,

213
00:09:32,010 --> 00:09:33,470
just like humans as babies.

214
00:09:33,470 --> 00:09:36,699
After we see enough example of them, we start to understand

215
00:09:36,699 --> 00:09:40,940
what animals are like and what toys are like.

216
00:09:40,940 --> 00:09:44,400
Okay, looking for the red truck, hold on.

217
00:09:44,400 --> 00:09:46,503
By looking at thousands and thousands of images,

218
00:09:46,503 --> 00:09:49,010
it will be able to extract some kind of

219
00:09:49,010 --> 00:09:53,610
distinguishing characteristics that identifies the image.

220
00:09:53,610 --> 00:09:57,680
So how has Jerry absorbed all this information?

221
00:09:57,680 --> 00:10:01,040
It used to be that professionals, engineers,

222
00:10:01,040 --> 00:10:03,540
that would handcraft these features,

223
00:10:03,540 --> 00:10:06,840
points, edges or corners.

224
00:10:06,840 --> 00:10:09,030
And it turns out humans are

225
00:10:09,030 --> 00:10:11,910
not nearly as good as the machines

226
00:10:11,910 --> 00:10:14,760
to automatically discover these features.

227
00:10:14,760 --> 00:10:17,090
That's the latest advances in machine learning

228
00:10:17,090 --> 00:10:19,120
and deep learning technologies.

229
00:10:19,120 --> 00:10:21,010
Advances in machine learning and vision

230
00:10:21,010 --> 00:10:24,000
have given computers human-like perception.

231
00:10:24,000 --> 00:10:25,720
They can spot a face in the crowd.

232
00:10:25,720 --> 00:10:26,553
Hello?

233
00:10:27,510 --> 00:10:28,573
I know you.

234
00:10:28,573 --> 00:10:31,313
They can see and anticipate oncoming traffic.

235
00:10:33,830 --> 00:10:35,593
It can even spot diseases.

236
00:10:36,437 --> 00:10:37,974
(bubbles popping)

237
00:10:37,974 --> 00:10:41,639
(bright orchestral music)

238
00:10:41,639 --> 00:10:43,660
At the National University of Singapore,

239
00:10:43,660 --> 00:10:45,740
researchers have trained a computer to

240
00:10:45,740 --> 00:10:49,313
differentiate cancerous cells from healthy cells.

241
00:10:50,930 --> 00:10:53,505
{\an8}It turns out that many diseases that we have

242
00:10:53,505 --> 00:10:57,958
all have a signature in the way DNA is packed.

243
00:10:57,958 --> 00:11:00,811
And inside this cancer cell nucleae,

244
00:11:00,811 --> 00:11:03,243
the packing of the DNA is very different.

245
00:11:08,480 --> 00:11:10,020
Based on his discovery,

246
00:11:10,020 --> 00:11:13,040
Professor Shivashankar's team is using machine learning

247
00:11:13,040 --> 00:11:14,710
to analyze DNA packing

248
00:11:14,710 --> 00:11:17,603
in thousands of cancerous and healthy cells.

249
00:11:19,200 --> 00:11:23,010
With enough data and training, AI could enable pathologists

250
00:11:23,010 --> 00:11:25,890
to diagnose cancer in its early stages

251
00:11:25,890 --> 00:11:28,603
when it's easier to treat, but harder to detect.

252
00:11:30,110 --> 00:11:31,185
If you want to do early detection,

253
00:11:31,185 --> 00:11:35,610
you need some AI-based approaches,

254
00:11:35,610 --> 00:11:39,690
because it's very hard to detect visually, manually.

255
00:11:39,690 --> 00:11:42,010
Just like finding a needle in a haystack.

256
00:11:42,010 --> 00:11:45,721
But for early stages, it's a very subjective decision

257
00:11:45,721 --> 00:11:49,160
because I've met with a number of pathologists.

258
00:11:49,160 --> 00:11:51,410
They have told me that, "My brain is a machine.

259
00:11:51,410 --> 00:11:53,130
It has already learned."

260
00:11:53,130 --> 00:11:55,960
But that depends on the experience of the pathologist.

261
00:11:55,960 --> 00:11:58,703
But we need to eliminate the subjectivity.

262
00:12:01,800 --> 00:12:03,580
The researchers say their technique

263
00:12:03,580 --> 00:12:05,178
is being refined and validated

264
00:12:05,178 --> 00:12:08,623
and could be used in clinics in two to three years.

265
00:12:10,540 --> 00:12:12,640
We'll see you after this commercial break.

266
00:12:13,910 --> 00:12:18,570
Another way AI is becoming human is through language.

267
00:12:18,570 --> 00:12:20,890
How can I help you today?

268
00:12:20,890 --> 00:12:24,600
But can AI give machines a sense of humor?

269
00:12:24,600 --> 00:12:28,102
These jokes are written by a computer.

270
00:12:28,102 --> 00:12:29,600
(comedian chuckling)

271
00:12:29,600 --> 00:12:31,534
Stupid computer.

272
00:12:31,534 --> 00:12:35,379
{\an8}(mechanical gears whirring)

273
00:12:35,379 --> 00:12:38,962
(hard drive booting)

274
00:12:38,962 --> 00:12:42,456
(hard drive booting)

275
00:12:42,456 --> 00:12:46,338
(mechanical gears whirring)

276
00:12:46,338 --> 00:12:48,921
{\an8}(men chatting)

277
00:12:49,882 --> 00:12:52,664
(Enlai giggling)

278
00:12:52,664 --> 00:12:55,300
Another important aspect of intelligence

279
00:12:55,300 --> 00:12:57,110
is our grasp of language.

280
00:12:57,110 --> 00:12:59,330
An ape may be able to learn a few hundred words,

281
00:12:59,330 --> 00:13:03,363
{\an8}but the average human knows over 20,000 words.

282
00:13:04,280 --> 00:13:05,113
Apparently.

283
00:13:07,800 --> 00:13:12,053
There's a huge gulf between zookeeper and zoo exhibit.

284
00:13:14,056 --> 00:13:15,260
(playful childish music)

285
00:13:15,260 --> 00:13:17,570
From birth till age five,

286
00:13:17,570 --> 00:13:20,650
children learn languages at a rapid pace,

287
00:13:20,650 --> 00:13:22,790
and they don't need formal lessons.

288
00:13:22,790 --> 00:13:26,307
Human babies are like language-learning machines.

289
00:13:26,307 --> 00:13:30,200
Hello, I'm Terry the Tiger.

290
00:13:30,200 --> 00:13:33,190
Are you Ollie the Ostrich?

291
00:13:33,190 --> 00:13:34,603
Whoa, you're a heavy boy.

292
00:13:35,527 --> 00:13:36,360
It's pointless to just

293
00:13:36,360 --> 00:13:38,680
sit the child down and teach a new language.

294
00:13:38,680 --> 00:13:40,340
It's not done that way.

295
00:13:40,340 --> 00:13:41,620
It's more like emotion.

296
00:13:41,620 --> 00:13:44,370
{\an8}So if you're washing your hands with your little girl,

297
00:13:44,370 --> 00:13:45,532
{\an8}you say,

298
00:13:45,532 --> 00:13:50,303
{\an8}(Patricia speaking foreign language)

299
00:13:50,303 --> 00:13:52,570
This is green, it's real paint.

300
00:13:52,570 --> 00:13:54,850
There is this language acquisition device

301
00:13:54,850 --> 00:13:56,050
that's in there,

302
00:13:56,050 --> 00:13:59,743
and it's just able to listen to the sounds,

303
00:13:59,743 --> 00:14:02,182
deciphers it, makes sense of it,

304
00:14:02,182 --> 00:14:04,920
and makes use of it when it comes.

305
00:14:04,920 --> 00:14:09,648
So I would say language is often caught, not taught.

306
00:14:09,648 --> 00:14:10,481
(parents clapping)

307
00:14:10,481 --> 00:14:11,314
♪ Baby shark do do do do do ♪

308
00:14:11,314 --> 00:14:13,316
♪ Mommy shark do do do do do ♪

309
00:14:13,316 --> 00:14:14,840
♪ Daddy shark do do do do do ♪

310
00:14:14,840 --> 00:14:16,100
While learning languages

311
00:14:16,100 --> 00:14:17,726
comes naturally to most children,

312
00:14:17,726 --> 00:14:20,583
it's a much harder task for a machine.

313
00:14:21,560 --> 00:14:22,610
Goodbye, everybody!

314
00:14:23,786 --> 00:14:25,355
Bye!

315
00:14:25,355 --> 00:14:27,480
(mechanical whirring)

316
00:14:27,480 --> 00:14:30,313
(robot whistling)

317
00:14:31,490 --> 00:14:35,940
Professor Francis Bond is a linguist who has, for decades,

318
00:14:35,940 --> 00:14:40,080
tried to teach computers to understand language.

319
00:14:40,080 --> 00:14:41,600
Well, a lot of understanding language

320
00:14:41,600 --> 00:14:42,700
is common sense

321
00:14:42,700 --> 00:14:44,917
and we have a lot of trouble getting computers to do that.

322
00:14:44,917 --> 00:14:46,780
{\an8}Out of the many possible meanings,

323
00:14:46,780 --> 00:14:49,510
{\an8}they'll invariably choose not quite the right one.

324
00:14:49,510 --> 00:14:51,840
Okay, do you have any examples of that?

325
00:14:51,840 --> 00:14:53,440
So here's an example of a sentence.

326
00:14:53,440 --> 00:14:55,117
You saw this, you read this.

327
00:14:55,117 --> 00:14:56,950
"I saw a kid with a cat."

328
00:14:56,950 --> 00:14:59,120
And normally you build an image in your mind.

329
00:14:59,120 --> 00:15:00,767
But the computer says,

330
00:15:00,767 --> 00:15:05,000
"Ah, I can combine these words in many different ways."

331
00:15:05,000 --> 00:15:09,037
So here we have "I saw a kid with a cat."

332
00:15:09,037 --> 00:15:09,973
(Enlai laughing)

333
00:15:09,973 --> 00:15:12,107
It's like animal cruelty.

334
00:15:12,107 --> 00:15:15,070
Oh, and child abuse let's not forget that.

335
00:15:15,070 --> 00:15:17,946
So here, "I saw a kid with a cat"

336
00:15:17,946 --> 00:15:19,646
is a possible interpretation.

337
00:15:19,646 --> 00:15:22,588
Using the cat as a telescope.

338
00:15:22,588 --> 00:15:26,000
To overcome computers' lack of common sense,

339
00:15:26,000 --> 00:15:28,064
scientists program them to follow rules

340
00:15:28,064 --> 00:15:31,943
in line with how we interpret language.

341
00:15:32,850 --> 00:15:34,500
To get a computer to do it,

342
00:15:34,500 --> 00:15:37,700
we have to build the framework that exists in your brain.

343
00:15:37,700 --> 00:15:40,360
But as you can see, just for a two-word sentence,

344
00:15:40,360 --> 00:15:42,520
we're already covering the screen.

345
00:15:42,520 --> 00:15:44,170
Yeah I know right?

346
00:15:44,170 --> 00:15:46,623
In a 10-word sentence, you might have 10,000 features.

347
00:15:49,220 --> 00:15:50,820
{\an8}Natural language processing

348
00:15:50,820 --> 00:15:53,520
{\an8}is one of AI's hardest challenges,

349
00:15:53,520 --> 00:15:56,993
but companies like Microsoft have made breakthroughs.

350
00:15:57,940 --> 00:16:01,270
Microsoft say their AI can transcribe speech

351
00:16:01,270 --> 00:16:03,183
better than a human can.

352
00:16:04,120 --> 00:16:05,410
For the first time in history,

353
00:16:05,410 --> 00:16:08,360
we exceeded human's performance.

354
00:16:08,360 --> 00:16:10,610
{\an8}Many people have worked on speech recognition

355
00:16:10,610 --> 00:16:12,880
{\an8}on the so-called switchboarder task.

356
00:16:12,880 --> 00:16:14,646
Secretary of war.

357
00:16:14,646 --> 00:16:16,470
We have two people

358
00:16:16,470 --> 00:16:18,120
having conversation naturally,

359
00:16:18,120 --> 00:16:21,220
and we need to transcribe that conversation

360
00:16:21,220 --> 00:16:23,130
as accurately as possible.

361
00:16:23,130 --> 00:16:26,785
If you look at the speech recognition on the switchboard,

362
00:16:26,785 --> 00:16:29,950
the error rate has been above 10%

363
00:16:29,950 --> 00:16:32,650
for almost like 25 years.

364
00:16:32,650 --> 00:16:35,130
We're able to really make that error come down

365
00:16:35,130 --> 00:16:39,130
from over 10% all the way to 5%.

366
00:16:39,130 --> 00:16:41,500
That is the same level of you and me

367
00:16:41,500 --> 00:16:44,530
to transcribe a natural conversation on the same task.

368
00:16:44,530 --> 00:16:47,940
Transcribing speech at human levels.

369
00:16:47,940 --> 00:16:49,570
How did that happen?

370
00:16:49,570 --> 00:16:51,020
It's through the deep learning.

371
00:16:51,020 --> 00:16:54,160
We're able to really learn from this huge amount of data,

372
00:16:54,160 --> 00:16:55,450
running many experiments,

373
00:16:55,450 --> 00:16:58,040
using the most advanced computing infrastructure.

374
00:16:58,040 --> 00:17:00,880
Those three things really form the core

375
00:17:00,880 --> 00:17:02,443
of speech breakthrough.

376
00:17:04,520 --> 00:17:06,660
This prototype uses AI

377
00:17:06,660 --> 00:17:08,472
to identify who's speaking

378
00:17:08,472 --> 00:17:11,853
and transcribes what they're saying in real time.

379
00:17:14,404 --> 00:17:16,160
You don't need to be worried about, you know,

380
00:17:16,160 --> 00:17:17,640
taking notes anymore.

381
00:17:17,640 --> 00:17:19,943
The poor person who has to write down the minutes,

382
00:17:19,943 --> 00:17:22,460
oh well, you don't have to do that anymore.

383
00:17:22,460 --> 00:17:23,570
How convenient.

384
00:17:23,570 --> 00:17:27,480
Accurate transcription is key to accurate translation.

385
00:17:27,480 --> 00:17:29,050
Our remote teammate in China

386
00:17:29,050 --> 00:17:32,220
can see and hear a translation of this meeting.

387
00:17:34,000 --> 00:17:35,990
On the sentence about center spaces

388
00:17:35,990 --> 00:17:39,793
for the first time, Microsoft translation achieved

389
00:17:39,793 --> 00:17:43,110
the same level or better performance

390
00:17:43,110 --> 00:17:45,540
in comparison to professional people.

391
00:17:45,540 --> 00:17:47,840
This provided us with a framework for us

392
00:17:47,840 --> 00:17:51,650
to approach human level of performance for translation.

393
00:17:51,650 --> 00:17:53,972
This translation app actually

394
00:17:53,972 --> 00:17:58,972
somewhat eradicates all language barriers, right?

395
00:17:59,094 --> 00:18:02,980
(hip hop electronica music)

396
00:18:02,980 --> 00:18:04,600
{\an8}I'm in Chongqing's Hongyadong

397
00:18:04,600 --> 00:18:07,330
{\an8}to try out Microsoft's translator.

398
00:18:07,330 --> 00:18:09,480
My Mandarin is infamously bad,

399
00:18:09,480 --> 00:18:12,423
so if it can help me, it can help anybody.

400
00:18:13,570 --> 00:18:14,403
Right?

401
00:18:21,533 --> 00:18:22,616
Can I try it?

402
00:18:24,796 --> 00:18:27,420
(AI translator speaking Mandarin)

403
00:18:27,420 --> 00:18:29,860
{\an8}(vendor speaking Mandarin)

404
00:18:29,860 --> 00:18:30,693
Okay.

405
00:18:34,195 --> 00:18:35,528
It's quite nice.

406
00:18:38,441 --> 00:18:40,683
Which one is spicy?

407
00:18:41,533 --> 00:18:43,009
(AI translator speaking Mandarin)

408
00:18:43,009 --> 00:18:44,430
{\an8}(vendor speaking Mandarin)

409
00:18:44,430 --> 00:18:45,740
Mala?

410
00:18:45,740 --> 00:18:46,790
Yeah.

411
00:18:46,790 --> 00:18:47,623
{\an8}Let's try it.

412
00:18:47,623 --> 00:18:48,832
{\an8}(vendor speaking Mandarin)

413
00:18:48,832 --> 00:18:50,320
Oh, it's good.

414
00:18:50,320 --> 00:18:51,733
How much is it?

415
00:18:55,106 --> 00:18:56,772
(AI translator speaking Mandarin)

416
00:18:56,772 --> 00:19:00,120
{\an8}(vendor speaking Mandarin)

417
00:19:00,120 --> 00:19:01,200
It's doing-

418
00:19:01,200 --> 00:19:03,773
9,481 pounds?

419
00:19:05,580 --> 00:19:08,770
I'm sorry auntie, but that's too expensive for me.

420
00:19:08,770 --> 00:19:09,971
I want one bag.

421
00:19:09,971 --> 00:19:11,713
(AI translator speaking Mandarin)

422
00:19:11,713 --> 00:19:15,796
{\an8}(vendor speaking Mandarin)

423
00:19:15,796 --> 00:19:16,963
Okay, xie xie.

424
00:19:19,800 --> 00:19:22,529
I managed to get some shopping done with the app.

425
00:19:22,529 --> 00:19:23,716
Was it perfect?

426
00:19:23,716 --> 00:19:25,500
No, not really.

427
00:19:25,500 --> 00:19:27,820
I think it only works with very simple statements,

428
00:19:27,820 --> 00:19:30,645
but it's better than nothing, right?

429
00:19:30,645 --> 00:19:34,050
(vendor hollering)

430
00:19:34,050 --> 00:19:34,900
I wonder what he said.

431
00:19:34,900 --> 00:19:36,640
Could you say that again?

432
00:19:36,640 --> 00:19:38,283
I haven't got my app out yet.

433
00:19:40,440 --> 00:19:41,823
Are you ignoring me now?

434
00:19:43,320 --> 00:19:44,153
People nowadays.

435
00:19:46,471 --> 00:19:47,392
(phone ringing)

436
00:19:47,392 --> 00:19:48,990
Hello, how can I help you?

437
00:19:48,990 --> 00:19:50,950
Hi, I'm calling to book a women's haircut

438
00:19:50,950 --> 00:19:52,196
for a client.

439
00:19:52,196 --> 00:19:53,587
I'm looking for something on May 3rd.

440
00:19:53,587 --> 00:19:54,939
Natural language processing

441
00:19:54,939 --> 00:19:57,950
has enabled computers to talk like us, debate with us-

442
00:19:57,950 --> 00:19:59,080
Allow me to respond to

443
00:19:59,080 --> 00:20:01,660
some of my opponent's most recent claims.

444
00:20:01,660 --> 00:20:03,402
And even compose poetry.

445
00:20:03,402 --> 00:20:06,303
AI wrote this book of poems.

446
00:20:07,747 --> 00:20:10,380
"I have failed to love my own life.

447
00:20:10,380 --> 00:20:12,807
Yet I touched the spirit in your eyes."

448
00:20:13,750 --> 00:20:15,730
Not bad, right?

449
00:20:15,730 --> 00:20:18,760
But there's a part of language that AI struggles with:

450
00:20:18,760 --> 00:20:20,983
humor and sarcasm.

451
00:20:21,934 --> 00:20:25,040
(playful blues music)

452
00:20:25,040 --> 00:20:27,720
The science of humor is very underdeveloped.

453
00:20:27,720 --> 00:20:30,980
{\an8}Why is one person funny and someone else not funny?

454
00:20:30,980 --> 00:20:32,620
{\an8}We still really don't know.

455
00:20:32,620 --> 00:20:34,970
Maybe I can challenge the computer to try and

456
00:20:36,100 --> 00:20:38,377
write some jokes for me.

457
00:20:38,377 --> 00:20:41,300
Yes, we can, we'll give you what we've got.

458
00:20:41,300 --> 00:20:44,473
And if you can make them funny, anyone can make them funny.

459
00:20:50,726 --> 00:20:53,726
{\an8}(upbeat jazz music)

460
00:20:57,320 --> 00:20:59,390
In a few moments, I'm going to be going up on stage

461
00:20:59,390 --> 00:21:01,171
just before Kuma,

462
00:21:01,171 --> 00:21:03,587
everyone's expecting Kuma.

463
00:21:03,587 --> 00:21:06,320
(Enlai groaning)

464
00:21:06,320 --> 00:21:07,153
I'm really nervous.

465
00:21:07,153 --> 00:21:12,153
I have them written here, all our computer-generated jokes.

466
00:21:12,880 --> 00:21:13,813
Wish me luck!

467
00:21:16,340 --> 00:21:18,460
Hello, hello, hello ladies and gentlemen!

468
00:21:18,460 --> 00:21:20,510
Welcome, welcome, welcome!

469
00:21:20,510 --> 00:21:21,810
(audience cheering)

470
00:21:21,810 --> 00:21:23,110
Don't worry, I'm not Kuma.

471
00:21:24,176 --> 00:21:25,009
(audience chuckling)

472
00:21:25,009 --> 00:21:26,923
I think you can tell the difference.

473
00:21:26,923 --> 00:21:27,756
(audience chuckling)

474
00:21:27,756 --> 00:21:28,793
I wear more makeup than Kuma.

475
00:21:29,651 --> 00:21:30,755
(audience chuckling)

476
00:21:30,755 --> 00:21:32,257
Okay, this is the first joke, guys.

477
00:21:32,257 --> 00:21:37,257
"What do you get when you cross a frog with a road?"

478
00:21:37,891 --> 00:21:39,807
(audience chattering)

479
00:21:39,807 --> 00:21:41,839
"A main toad!"

480
00:21:41,839 --> 00:21:45,307
(audience jeering)

481
00:21:45,307 --> 00:21:46,860
Nay?

482
00:21:46,860 --> 00:21:48,570
Nay?

483
00:21:48,570 --> 00:21:49,850
I would throw this card away,

484
00:21:49,850 --> 00:21:52,070
but there are two jokes on one card, so I cannot.

485
00:21:52,070 --> 00:21:53,527
So the next one, one second.

486
00:21:53,527 --> 00:21:57,486
"When is a job not a job?"

487
00:21:57,486 --> 00:21:58,821
(audience laughing)

488
00:21:58,821 --> 00:22:01,136
[Woman In Audience] When it's a (bleep) job?

489
00:22:01,136 --> 00:22:01,969
"When it's a (bleep) job?"

490
00:22:01,969 --> 00:22:03,143
You sure that's not a job?

491
00:22:03,143 --> 00:22:05,967
(audience laughing)

492
00:22:05,967 --> 00:22:07,907
"When it is a nose job."

493
00:22:08,959 --> 00:22:10,317
(audience groaning)

494
00:22:10,317 --> 00:22:12,015
I like that answer, though.

495
00:22:12,015 --> 00:22:13,887
(audience cheering)

496
00:22:13,887 --> 00:22:16,137
Ladies and gentlemen, Kuma!

497
00:22:19,900 --> 00:22:21,150
Thanks Enlai.

498
00:22:22,671 --> 00:22:24,387
(Kuma chuckling)

499
00:22:24,387 --> 00:22:27,983
Stupid computer cannot fight human being, la, not ever.

500
00:22:28,968 --> 00:22:30,430
(audience laughing)

501
00:22:30,430 --> 00:22:33,690
But what did you think of the jokes that I did just now?

502
00:22:33,690 --> 00:22:35,600
{\an8}It was really, really bad.

503
00:22:35,600 --> 00:22:37,630
{\an8}It's not even about knock-knock jokes,

504
00:22:37,630 --> 00:22:39,270
{\an8}or, you know, why did the chicken cross the road?

505
00:22:39,270 --> 00:22:41,287
{\an8}It was so bad.

506
00:22:41,287 --> 00:22:43,340
It was so bad, it was like, "Whoa."

507
00:22:43,340 --> 00:22:44,888
I actually thought it was gonna be like that

508
00:22:44,888 --> 00:22:47,752
because you know, it needs a human to relate to people.

509
00:22:47,752 --> 00:22:49,530
You know, people want to relate with the humor.

510
00:22:49,530 --> 00:22:52,910
They want to relate with things about life, you know?

511
00:22:52,910 --> 00:22:54,400
And a computer can't see that.

512
00:22:54,400 --> 00:22:57,130
You were funnier before you even started telling jokes

513
00:22:57,130 --> 00:22:57,963
from the, you know?

514
00:22:57,963 --> 00:23:00,616
Okay, so slowly as it learns things like

515
00:23:00,616 --> 00:23:02,750
wordplay and puns,

516
00:23:02,750 --> 00:23:03,763
which you also do.

517
00:23:03,763 --> 00:23:05,710
Yes, yes, of course, of course.

518
00:23:05,710 --> 00:23:07,150
Do you think it will-?

519
00:23:07,150 --> 00:23:08,010
Never.

520
00:23:08,010 --> 00:23:08,843
Confirm, never?

521
00:23:08,843 --> 00:23:10,920
Never, ever, ever, ever.

522
00:23:10,920 --> 00:23:12,523
Never.

523
00:23:12,523 --> 00:23:14,820
Okay, but then what if the computer just writes the script

524
00:23:14,820 --> 00:23:16,880
and you still perform it.

525
00:23:16,880 --> 00:23:18,511
No.

526
00:23:18,511 --> 00:23:19,380
If you, if Kuma-

527
00:23:19,380 --> 00:23:21,524
I'd rather bleed to death.

528
00:23:21,524 --> 00:23:22,809
(Kuma laughing)

529
00:23:22,809 --> 00:23:25,350
(Enlai laughing)

530
00:23:25,350 --> 00:23:28,050
So AI won't be a comedian soon,

531
00:23:28,050 --> 00:23:31,630
but it could be a doctor or a scientist.

532
00:23:31,630 --> 00:23:35,649
She'll become as good as Einsteins or Newtons of science.

533
00:23:35,649 --> 00:23:38,399
(piano tinkling)

534
00:23:39,410 --> 00:23:41,683
Actually, it's surprising to me it's composed by

535
00:23:41,683 --> 00:23:43,725
artificial intelligence.

536
00:23:43,725 --> 00:23:47,840
AI could even turn me into an artist.

537
00:23:47,840 --> 00:23:49,470
So I can hang this up in a gallery?

538
00:23:49,470 --> 00:23:50,500
You can try.

539
00:23:50,500 --> 00:23:54,167
(mechanical gears whirring)

540
00:23:57,317 --> 00:23:59,056
(mechanical gears whirring)

541
00:23:59,056 --> 00:24:00,501
(hard drive booting)

542
00:24:00,501 --> 00:24:02,870
(mechanical gears whirring)

543
00:24:02,870 --> 00:24:05,537
(robot beeping)

544
00:24:07,770 --> 00:24:11,320
{\an8}As an actor, I'm constantly confronted by new technology

545
00:24:11,320 --> 00:24:13,290
in theater and the arts.

546
00:24:13,290 --> 00:24:14,690
Let's play with our horns.

547
00:24:15,540 --> 00:24:17,283
That's interactivity, isn't it?

548
00:24:18,250 --> 00:24:21,983
Everything is so vivid, and we all get doubled.

549
00:24:23,810 --> 00:24:25,770
{\an8}They're really so close to you.

550
00:24:25,770 --> 00:24:26,750
{\an8}It's like:

551
00:24:29,641 --> 00:24:30,474
Like this.

552
00:24:32,330 --> 00:24:35,610
This experimental performance is run by Steve Dixon,

553
00:24:35,610 --> 00:24:39,060
the president of LaSalle College of the Arts.

554
00:24:39,060 --> 00:24:41,400
Because you're Enlai!

555
00:24:41,400 --> 00:24:43,363
Help, there's other people!

556
00:24:44,640 --> 00:24:46,970
When he isn't acting like a maniac.

557
00:24:46,970 --> 00:24:50,519
Steve wrote the book on digital performance.

558
00:24:50,519 --> 00:24:53,860
Artists always will look towards technology

559
00:24:53,860 --> 00:24:58,010
to see if it can really increase the spectacle and so on.

560
00:24:58,010 --> 00:24:59,750
Artificial intelligence is being used

561
00:24:59,750 --> 00:25:01,571
in lots of interesting ways in theater now.

562
00:25:01,571 --> 00:25:03,930
{\an8}I think one early example,

563
00:25:03,930 --> 00:25:06,210
{\an8}and I can show you on video here,

564
00:25:06,210 --> 00:25:08,840
an artificially intelligent performer

565
00:25:08,840 --> 00:25:11,570
in the form of a large head called Jeremiah.

566
00:25:11,570 --> 00:25:13,940
So it has an emotional intelligence engine,

567
00:25:13,940 --> 00:25:16,020
but it is entirely AI.

568
00:25:16,020 --> 00:25:19,534
It isn't being operated by someone offstage.

569
00:25:19,534 --> 00:25:22,670
We will see more synth thespians.

570
00:25:22,670 --> 00:25:25,680
So either projected synthetic performers

571
00:25:25,680 --> 00:25:29,800
or robotic, android type of performers.

572
00:25:29,800 --> 00:25:32,400
But I don't think the actor is ever going to be replaced.

573
00:25:32,400 --> 00:25:34,250
I think that that human being

574
00:25:34,250 --> 00:25:36,340
will still remain center stage.

575
00:25:36,340 --> 00:25:39,470
So your job is secure in that.

576
00:25:39,470 --> 00:25:42,800
Phew, thank goodness!

577
00:25:42,800 --> 00:25:46,180
Then again, I could be a robot.

578
00:25:46,180 --> 00:25:48,760
(Steve chuckling)

579
00:25:48,760 --> 00:25:50,808
My job may be safe for now,

580
00:25:50,808 --> 00:25:54,781
but other artists may be in for a shock.

581
00:25:54,781 --> 00:25:57,907
You will be the third doctor!

582
00:25:59,903 --> 00:26:02,154
(playful orchestral music)

583
00:26:02,154 --> 00:26:05,814
{\an8}In England, at Cambridge Consultants,

584
00:26:05,814 --> 00:26:09,710
researchers have trained a computer to create art.

585
00:26:09,710 --> 00:26:12,520
They call it Vincent.

586
00:26:12,520 --> 00:26:14,167
Since it's called Vincent,

587
00:26:14,167 --> 00:26:17,840
I want to try and do a Vincent Van Gogh painting.

588
00:26:17,840 --> 00:26:18,673
Go for it.

589
00:26:19,720 --> 00:26:21,633
I'm going to try drawing, anyway.

590
00:26:22,600 --> 00:26:24,620
So you'll see what it's doing here is it's trying to

591
00:26:24,620 --> 00:26:28,000
interpret your sketches, the lines you add on,

592
00:26:28,000 --> 00:26:29,653
and as you're drawing.

593
00:26:31,580 --> 00:26:34,140
So we trained on 8,000 images

594
00:26:34,140 --> 00:26:36,340
{\an8}from Renaissance art right through to 20th century.

595
00:26:36,340 --> 00:26:39,212
{\an8}So it's got people like Jackson Pollack and Rothko in there

596
00:26:39,212 --> 00:26:43,093
but all of their opinions and their styles and influences

597
00:26:43,093 --> 00:26:45,038
will be being used together

598
00:26:45,038 --> 00:26:47,010
to try and interpret the lines you make

599
00:26:47,010 --> 00:26:48,100
and create some art.

600
00:26:48,100 --> 00:26:49,260
Oh, okay.

601
00:26:49,260 --> 00:26:54,260
Well, I'm going to try and do something by Rothko,

602
00:26:54,500 --> 00:26:56,433
which is basically just sort of like.

603
00:26:57,740 --> 00:26:59,170
It's ultra abstract, isn't it?

604
00:26:59,170 --> 00:27:00,610
Like that.

605
00:27:00,610 --> 00:27:03,013
{\an8}This looks like kueh lapis.

606
00:27:07,790 --> 00:27:09,130
Very good.

607
00:27:09,130 --> 00:27:11,380
Tim was quite dismissive of my art.

608
00:27:11,380 --> 00:27:14,250
I think Vincent was more forgiving.

609
00:27:14,250 --> 00:27:16,997
I thought my drawings were quite good, weren't they?

610
00:27:16,997 --> 00:27:18,609
And I would certainly say that

611
00:27:18,609 --> 00:27:21,890
the more professional or the more talented you are,

612
00:27:21,890 --> 00:27:24,425
I think the better results that you can achieve.

613
00:27:24,425 --> 00:27:25,309
Okay.

614
00:27:25,309 --> 00:27:27,350
So with this experiment that you have here,

615
00:27:27,350 --> 00:27:29,810
do you think this is going to change the future of art?

616
00:27:29,810 --> 00:27:31,960
I think that you could say that this machine

617
00:27:31,960 --> 00:27:34,041
has some level of creativity,

618
00:27:34,041 --> 00:27:36,980
but it understands only a very small world.

619
00:27:36,980 --> 00:27:39,300
This machine only knows those 8,000 images.

620
00:27:39,300 --> 00:27:42,190
So its inspiration is quite limited, you could say.

621
00:27:42,190 --> 00:27:44,640
It might change the future of my art, maybe,

622
00:27:44,640 --> 00:27:47,333
but for somebody who's actually a good artist, I think not.

623
00:27:48,310 --> 00:27:50,970
Simple forms of this kind of AI

624
00:27:50,970 --> 00:27:55,143
{\an8}let us make our selfies resemble artistic masterpieces.

625
00:28:02,040 --> 00:28:05,060
But I wonder what the art establishment thinks of

626
00:28:05,060 --> 00:28:08,290
AI emulating the masters.

627
00:28:08,290 --> 00:28:10,120
Actually artists draw from

628
00:28:10,120 --> 00:28:12,340
diverse cultural sources.

629
00:28:12,340 --> 00:28:15,300
The curator says Singapore's National Gallery

630
00:28:15,300 --> 00:28:17,060
houses a style of paintings

631
00:28:17,060 --> 00:28:20,183
that combines Eastern and Western influences.

632
00:28:22,400 --> 00:28:24,270
This is a very good example of what we call

633
00:28:24,270 --> 00:28:25,150
the Nan Yung style.

634
00:28:25,150 --> 00:28:27,040
This is a painting by Lee Man Fong

635
00:28:27,040 --> 00:28:28,840
{\an8}and it's titled "Balinese Life."

636
00:28:28,840 --> 00:28:31,550
{\an8}And here we see an example of his use of

637
00:28:31,550 --> 00:28:34,497
Chinese ink painting brush techniques,

638
00:28:34,497 --> 00:28:38,200
but bringing in kind of the Western atmospheric perspective

639
00:28:38,200 --> 00:28:40,950
to give a sense of distance to the painting.

640
00:28:40,950 --> 00:28:42,960
If you look at this face,

641
00:28:42,960 --> 00:28:44,820
it looks like it's from a Chinese painting.

642
00:28:44,820 --> 00:28:45,653
Exactly, yeah.

643
00:28:45,653 --> 00:28:47,179
Right?

644
00:28:47,179 --> 00:28:48,221
Yes.

645
00:28:48,221 --> 00:28:50,410
But then the way these ladies are reclined.

646
00:28:50,410 --> 00:28:51,522
Exactly.

647
00:28:51,522 --> 00:28:52,750
It's taken from the Western pictorial traditions,

648
00:28:52,750 --> 00:28:53,763
the reclining nude.

649
00:28:57,990 --> 00:29:02,410
So I have used a combination of styles,

650
00:29:02,410 --> 00:29:07,230
something very primitive, which is orangutans-

651
00:29:07,230 --> 00:29:08,063
Wow.

652
00:29:08,063 --> 00:29:09,307
Uh huh.

653
00:29:09,307 --> 00:29:12,157
Something very contemporary, which is a selfie of myself,

654
00:29:13,550 --> 00:29:16,810
And a style of painting

655
00:29:16,810 --> 00:29:19,747
which you could kind of describe as cubism,

656
00:29:19,747 --> 00:29:23,710
using artificial intelligence technology

657
00:29:23,710 --> 00:29:26,900
to come up with this.

658
00:29:26,900 --> 00:29:29,140
Right, well I think it's interesting.

659
00:29:29,140 --> 00:29:30,270
The composition, you know,

660
00:29:30,270 --> 00:29:32,481
having your self portrait at the center

661
00:29:32,481 --> 00:29:34,300
draws our attention to it.

662
00:29:34,300 --> 00:29:37,070
And as you mentioned, it's also a kind of cubistic

663
00:29:37,070 --> 00:29:38,850
or a bit of surrealistic elements.

664
00:29:38,850 --> 00:29:40,500
So I can hang this up in a gallery?

665
00:29:40,500 --> 00:29:41,575
You can try.

666
00:29:41,575 --> 00:29:43,095
(Seng Fu laughing)

667
00:29:43,095 --> 00:29:44,370
You may not succeed.

668
00:29:44,370 --> 00:29:45,995
Okay, how can I improve myself?

669
00:29:45,995 --> 00:29:48,070
When you make digital art, of course,

670
00:29:48,070 --> 00:29:50,820
it's not just using an application,

671
00:29:50,820 --> 00:29:54,550
because that's just kind of a very formulaic way of working.

672
00:29:54,550 --> 00:29:58,190
I think for important art or art that's kind of creative,

673
00:29:58,190 --> 00:30:02,130
we have to push out of these boundaries.

674
00:30:02,130 --> 00:30:07,130
Will artificial intelligence aid artists in the future?

675
00:30:07,290 --> 00:30:10,060
Or will they replace artists?

676
00:30:10,060 --> 00:30:11,250
Somewhere in between,

677
00:30:11,250 --> 00:30:14,660
whereby technology or augmented intelligence

678
00:30:14,660 --> 00:30:18,400
is used to collaborate with artists to make art.

679
00:30:18,400 --> 00:30:19,233
Cool.

680
00:30:19,233 --> 00:30:20,867
All they need is a platform.

681
00:30:20,867 --> 00:30:24,869
And all this needs is a space.

682
00:30:24,869 --> 00:30:28,303
I'm going to put it right here, okay?

683
00:30:31,180 --> 00:30:32,110
Right here.

684
00:30:32,110 --> 00:30:34,930
Yeah, let it stay there, it looks great.

685
00:30:34,930 --> 00:30:35,766
Okay?

686
00:30:35,766 --> 00:30:36,793
Okay.

687
00:30:38,466 --> 00:30:40,253
Enjoy, guys, enjoy.

688
00:30:47,520 --> 00:30:50,603
{\an8}AI is going beyond visual art.

689
00:30:52,000 --> 00:30:53,970
This piece of music was created

690
00:30:53,970 --> 00:30:57,193
with the help of an AI composer called AIVA.

691
00:30:58,623 --> 00:31:02,290
(dramatic orchestral music)

692
00:31:09,920 --> 00:31:13,830
Wow, what a stirring soundtrack.

693
00:31:13,830 --> 00:31:17,060
You know, it's very emotional.

694
00:31:17,060 --> 00:31:19,100
There are highs and there are lows.

695
00:31:19,100 --> 00:31:22,443
Strangely, it's all written by AI.

696
00:31:28,431 --> 00:31:29,700
{\an8}AIVA is an artificial intelligence

697
00:31:29,700 --> 00:31:32,980
{\an8}that has been taught music composition

698
00:31:32,980 --> 00:31:35,500
{\an8}by reading essentially 30,000 scores

699
00:31:35,500 --> 00:31:37,040
of the great composers of history.

700
00:31:37,040 --> 00:31:39,930
So, you know, Mozart, Bach, Beethoven.

701
00:31:39,930 --> 00:31:42,150
What AIVA looks in those scores is for patterns

702
00:31:42,150 --> 00:31:45,110
in the melody, in the rhythm and the harmony.

703
00:31:45,110 --> 00:31:46,550
And based on those patterns,

704
00:31:46,550 --> 00:31:48,480
it will be able to create a model of

705
00:31:48,480 --> 00:31:50,432
what music is supposed to sound like.

706
00:31:50,432 --> 00:31:54,099
(uplifting classical music)

707
00:31:55,810 --> 00:31:57,950
On the screen because the representation of

708
00:31:57,950 --> 00:31:59,763
how AIVA understands music.

709
00:32:01,340 --> 00:32:03,641
This looks like a colorblindness test.

710
00:32:03,641 --> 00:32:04,474
{\an8}(Denis chuckling)

711
00:32:04,474 --> 00:32:06,480
{\an8}Aell, AIVA definitely isn't colorblind

712
00:32:06,480 --> 00:32:08,650
{\an8}when it comes to musical style.

713
00:32:08,650 --> 00:32:10,250
Each one of these points

714
00:32:10,250 --> 00:32:13,311
is a particular piece in our database

715
00:32:13,311 --> 00:32:16,353
labeled here by, you know, the composer.

716
00:32:18,820 --> 00:32:23,820
I'd like AIVA to compose a piece called "Becoming Human."

717
00:32:26,060 --> 00:32:30,310
I would like something that can tug at the heartstrings.

718
00:32:30,310 --> 00:32:31,143
Okay.

719
00:32:31,143 --> 00:32:32,610
Something emotional.

720
00:32:32,610 --> 00:32:33,443
You got it.

721
00:32:33,443 --> 00:32:36,861
Yeah, I need one tear by the first 15 seconds.

722
00:32:36,861 --> 00:32:39,090
That's, um-

723
00:32:39,090 --> 00:32:40,730
We'll try that.

724
00:32:40,730 --> 00:32:42,690
We'll ask AIVA to do her best,

725
00:32:42,690 --> 00:32:44,693
but you know, that'll be subjective.

726
00:32:45,801 --> 00:32:48,270
AIVA composes original music

727
00:32:48,270 --> 00:32:50,760
in less than three days.

728
00:32:50,760 --> 00:32:54,070
AIVA's registered with a French composer society

729
00:32:54,070 --> 00:32:56,476
and has a built-in plagiarism checker.

730
00:32:56,476 --> 00:32:59,320
She's written music for the City of Dubai,

731
00:32:59,320 --> 00:33:01,817
global corporations, and now, me.

732
00:33:05,036 --> 00:33:08,286
(dramatic piano music)

733
00:33:15,870 --> 00:33:19,510
{\an8}AIVA's latest composition, "Becoming Human," is presented

734
00:33:19,510 --> 00:33:22,773
before a panel of music students and their teachers.

735
00:33:30,890 --> 00:33:32,379
{\an8}Kind of soothing,

736
00:33:32,379 --> 00:33:33,953
{\an8}but also quite dramatic at the same time.

737
00:33:34,880 --> 00:33:37,110
{\an8}Seems to me as if I would have heard it before.

738
00:33:37,110 --> 00:33:38,910
And there were a lot of quotes

739
00:33:38,910 --> 00:33:40,773
from music that we had heard before.

740
00:33:42,706 --> 00:33:44,579
{\an8}It's like, strange, a bit.

741
00:33:44,579 --> 00:33:45,629
{\an8}A little bit strange.

742
00:33:48,821 --> 00:33:51,970
{\an8}I like this piece, and it's actually surprising to me

743
00:33:51,970 --> 00:33:55,033
it's composed by artificial intelligence.

744
00:34:03,270 --> 00:34:04,270
A bit of pain,

745
00:34:04,270 --> 00:34:06,123
but also at the same time, some hope.

746
00:34:07,590 --> 00:34:11,180
The ending sort of leaves you wanting more of it.

747
00:34:11,180 --> 00:34:12,700
{\an8}And I think that's a good sign of

748
00:34:12,700 --> 00:34:14,593
{\an8}a well-crafted piece of music.

749
00:34:30,800 --> 00:34:35,170
It did tug on my heartstrings a bit, I guess,

750
00:34:35,170 --> 00:34:37,910
because it was kind of like melodramatic,

751
00:34:37,910 --> 00:34:39,830
but this piece of music, "Becoming Human,"

752
00:34:39,830 --> 00:34:43,510
may not be an appropriate soundtrack for this show.

753
00:34:43,510 --> 00:34:47,367
It's probably better for like "Titanic II."

754
00:34:50,510 --> 00:34:53,100
{\an8}AI isn't just mastering the arts.

755
00:34:53,100 --> 00:34:55,663
It's taking on science and medicine.

756
00:34:56,820 --> 00:34:58,250
Welcome to Babylon.

757
00:34:58,250 --> 00:35:00,130
How can I help you today?

758
00:35:00,130 --> 00:35:01,800
{\an8}I've got a really bad headache.

759
00:35:01,800 --> 00:35:03,200
Do you get any warning signs

760
00:35:03,200 --> 00:35:04,880
before your main symptoms start?

761
00:35:04,880 --> 00:35:07,470
Babylon Healthsave have taught their AI chat bot

762
00:35:07,470 --> 00:35:09,133
to think like a doctor.

763
00:35:10,015 --> 00:35:12,470
Have you felt dizzy, unsteady, lightheaded,

764
00:35:12,470 --> 00:35:14,921
or like the room is spinning, recently?

765
00:35:14,921 --> 00:35:16,160
No.

766
00:35:16,160 --> 00:35:17,850
The most likely cause of your symptoms

767
00:35:17,850 --> 00:35:19,070
is a migraine.

768
00:35:19,070 --> 00:35:22,220
For more information on possible causes and treatments,

769
00:35:22,220 --> 00:35:24,250
I've sent you a full report in your Alexa app.

770
00:35:24,250 --> 00:35:25,980
The text-based Babylon chat bot

771
00:35:25,980 --> 00:35:28,290
is already serving the National Health Service

772
00:35:28,290 --> 00:35:31,340
and 32,000 people in central London.

773
00:35:31,340 --> 00:35:33,750
Once Babylon diagnosis an illness,

774
00:35:33,750 --> 00:35:35,413
it helps you get medical treatment

775
00:35:35,413 --> 00:35:37,830
or a doctor's appointment.

776
00:35:37,830 --> 00:35:39,070
The beauty of the app is

777
00:35:39,070 --> 00:35:41,460
{\an8}if you click on book an appointment,

778
00:35:41,460 --> 00:35:44,753
{\an8}we can see when the next video consultation is available.

779
00:35:47,020 --> 00:35:48,470
Behind Babylon as a team of

780
00:35:48,470 --> 00:35:51,190
scientists, engineers, and medical specialists

781
00:35:51,190 --> 00:35:52,777
who train the AI with medical knowledge

782
00:35:52,777 --> 00:35:55,640
from textbooks and research papers.

783
00:35:55,640 --> 00:35:58,000
Their chief scientist says their AI

784
00:35:58,000 --> 00:36:01,530
can match human doctors at certain tasks.

785
00:36:01,530 --> 00:36:03,870
{\an8}When we compare the average accuracy

786
00:36:03,870 --> 00:36:05,350
{\an8}in terms of a diagnosis

787
00:36:05,350 --> 00:36:09,507
{\an8}of the AI against a set of human doctors,

788
00:36:09,507 --> 00:36:11,170
we're very close.

789
00:36:11,170 --> 00:36:13,760
But then when we consider triaging ability,

790
00:36:13,760 --> 00:36:15,264
that's the ability to tell the patient

791
00:36:15,264 --> 00:36:17,137
where to go to seek advice,

792
00:36:17,137 --> 00:36:20,803
stay at home, pharmacy, G.P., emergency department,

793
00:36:20,803 --> 00:36:24,030
our AI is already safer than a human doctor.

794
00:36:24,030 --> 00:36:26,410
Could it one day surpass the human doctor?

795
00:36:26,410 --> 00:36:31,410
Our aim is not to surpass or to replace human doctors.

796
00:36:31,750 --> 00:36:33,040
Our aim is always

797
00:36:33,040 --> 00:36:35,640
to augment the intelligence of human doctors

798
00:36:35,640 --> 00:36:38,540
with artificial intelligence.

799
00:36:38,540 --> 00:36:40,495
Since the AI is able to be kept up to date

800
00:36:40,495 --> 00:36:42,410
with the latest medical knowledge,

801
00:36:42,410 --> 00:36:44,409
and because we're able to see, you know,

802
00:36:44,409 --> 00:36:48,350
hundreds of thousands of patients potentially per day,

803
00:36:48,350 --> 00:36:51,880
and so because of the application of machine learning,

804
00:36:51,880 --> 00:36:53,803
we expect that over time,

805
00:36:53,803 --> 00:36:56,330
we should start to get close to

806
00:36:56,330 --> 00:36:58,513
the level of intelligence of human doctor.

807
00:37:00,670 --> 00:37:03,700
{\an8}If an AI doctor is on the horizon,

808
00:37:03,700 --> 00:37:06,250
{\an8}how about a scientist?

809
00:37:06,250 --> 00:37:09,100
Computer science pioneer Alan Turing

810
00:37:09,100 --> 00:37:13,880
theorized about AI at the University of Manchester.

811
00:37:13,880 --> 00:37:17,552
His successors at the university's biotechnology institute

812
00:37:17,552 --> 00:37:20,677
have created a robot scientist called Eve.

813
00:37:23,010 --> 00:37:26,380
She's designed to automate

814
00:37:26,380 --> 00:37:27,960
simple forms of scientific research,

815
00:37:27,960 --> 00:37:30,682
not just the experiments you see physically happening here,

816
00:37:30,682 --> 00:37:34,910
{\an8}but also the thinking and reasoning involved in science.

817
00:37:34,910 --> 00:37:37,100
{\an8}She forms her own hypotheses,

818
00:37:37,100 --> 00:37:39,700
she thinks of experiments to test the hypotheses

819
00:37:39,700 --> 00:37:43,010
and continues the cycle of scientific research.

820
00:37:43,010 --> 00:37:44,930
Scientist Eve discovered that

821
00:37:44,930 --> 00:37:47,450
an antibacterial chemical in toothpaste

822
00:37:47,450 --> 00:37:50,963
also kills the organism that causes malaria.

823
00:37:52,310 --> 00:37:53,143
Eve doesn't do anything

824
00:37:53,143 --> 00:37:55,710
which human scientists couldn't potentially do,

825
00:37:55,710 --> 00:37:58,170
but human scientists didn't discover it, you know,

826
00:37:58,170 --> 00:37:59,570
it was her discovery.

827
00:37:59,570 --> 00:38:02,220
How does Eve's intelligence

828
00:38:02,220 --> 00:38:04,490
match up with human intelligence?

829
00:38:04,490 --> 00:38:07,457
Because we regard human scientists as, you know,

830
00:38:07,457 --> 00:38:10,850
being very smart, very knowledgeable.

831
00:38:10,850 --> 00:38:14,030
Humans are not very good at reasoning about science,

832
00:38:14,030 --> 00:38:16,718
whereas it actually plays to the strengths of

833
00:38:16,718 --> 00:38:19,020
computers and AI systems.

834
00:38:19,020 --> 00:38:22,126
At the moment, Eve can only do very simple forms of science,

835
00:38:22,126 --> 00:38:25,836
but I think over time with better software and hardware,

836
00:38:25,836 --> 00:38:29,272
she'll become as good as average scientists

837
00:38:29,272 --> 00:38:31,320
and eventually over time

838
00:38:31,320 --> 00:38:34,830
as good as Einsteins or Newtons at science.

839
00:38:34,830 --> 00:38:36,910
That's my prediction.

840
00:38:36,910 --> 00:38:37,798
Really?

841
00:38:37,798 --> 00:38:39,110
(urgent electronica music)

842
00:38:39,110 --> 00:38:42,340
AI as smart as Einstein?

843
00:38:42,340 --> 00:38:47,340
If AI outsmarts us, machines could take over our world.

844
00:38:47,539 --> 00:38:50,260
(lasers zapping)

845
00:38:50,260 --> 00:38:52,880
We could totally lose control over them.

846
00:38:52,880 --> 00:38:55,210
The next generation of AI could come from

847
00:38:55,210 --> 00:38:57,610
unlocking secrets in our brains

848
00:38:57,610 --> 00:39:00,590
and solving childhood mysteries.

849
00:39:00,590 --> 00:39:02,120
I can see a lot of promise for

850
00:39:02,120 --> 00:39:05,210
artificial intelligence that learn like infants.

851
00:39:05,210 --> 00:39:08,770
I just hope we can avoid a nightmare future.

852
00:39:08,770 --> 00:39:10,710
Help me!

853
00:39:10,710 --> 00:39:14,377
{\an8}(mechanical gears whirring)

854
00:39:17,385 --> 00:39:21,052
{\an8}(mechanical gears whirring)

855
00:39:22,228 --> 00:39:24,895
(robot beeping)

856
00:39:26,080 --> 00:39:27,830
Dear Professor Tegmark,

857
00:39:27,830 --> 00:39:32,061
I read your book, "Life 3.0," and had a bad dream.

858
00:39:32,061 --> 00:39:37,061
I dreamt I was kept in a zoo that was run by robots and AI.

859
00:39:37,550 --> 00:39:38,383
Help me!

860
00:39:42,691 --> 00:39:44,114
(electronics beeping)

861
00:39:44,114 --> 00:39:46,623
(Enlai gasping)

862
00:39:46,623 --> 00:39:50,338
Oh food, food, feeding time!

863
00:39:50,338 --> 00:39:51,171
Feed me!

864
00:39:53,429 --> 00:39:55,380
Do you think this could really happen?

865
00:39:55,380 --> 00:39:57,103
Yours sincerely, Enlai.

866
00:39:58,320 --> 00:39:59,960
Thanks for your message, Enlai.

867
00:39:59,960 --> 00:40:01,290
{\an8}I think this could totally happen,

868
00:40:01,290 --> 00:40:03,510
{\an8}because the reason that we humans

869
00:40:03,510 --> 00:40:07,068
{\an8}have more power on this planet than the lions in the zoo

870
00:40:07,068 --> 00:40:10,990
is not because we're stronger, but because we're smarter.

871
00:40:10,990 --> 00:40:13,530
So if we build machines which are smarter than us,

872
00:40:13,530 --> 00:40:15,450
which I think we eventually will,

873
00:40:15,450 --> 00:40:18,870
then we could totally lose control over them.

874
00:40:18,870 --> 00:40:20,520
But it doesn't have to be a bad thing.

875
00:40:20,520 --> 00:40:24,060
We were all in the presence of smarter entities already,

876
00:40:24,060 --> 00:40:25,760
mommy and daddy, when we were little,

877
00:40:25,760 --> 00:40:27,100
and it worked out well, you know,

878
00:40:27,100 --> 00:40:29,639
because their goals were aligned with ours.

879
00:40:29,639 --> 00:40:33,070
If AI grows up to rule the world,

880
00:40:33,070 --> 00:40:36,133
I hope it learns from the best in humanity.

881
00:40:36,133 --> 00:40:38,350
We have a lot to offer.

882
00:40:38,350 --> 00:40:40,263
Just look how cute we are.

883
00:40:42,727 --> 00:40:44,602
(mechanical whirring)

884
00:40:44,602 --> 00:40:46,240
(robot whistling)

885
00:40:46,240 --> 00:40:48,740
{\an8}Nature inspires science.

886
00:40:48,740 --> 00:40:51,712
Just as bird swings inspired airplane wings,

887
00:40:51,712 --> 00:40:54,893
brains inspire artificial intelligence.

888
00:40:55,741 --> 00:40:59,260
{\an8}MIT, The Massachusetts Institute of Technology,

889
00:40:59,260 --> 00:41:02,590
{\an8}is home to the Center for Brains, Minds, and Machines.

890
00:41:02,590 --> 00:41:05,300
They draw on psychology and neuroscience

891
00:41:05,300 --> 00:41:07,413
to build intelligent machines.

892
00:41:08,360 --> 00:41:09,547
I want to understand the brain,

893
00:41:09,547 --> 00:41:13,480
and if intelligent machines comes as a consequence of that,

894
00:41:13,480 --> 00:41:15,663
good, but I would be happy either way.

895
00:41:17,920 --> 00:41:19,040
The center's director is

896
00:41:19,040 --> 00:41:21,614
a neuroscientist slash computer scientist,

897
00:41:21,614 --> 00:41:24,343
Dr. Tomaso A. Poggio.

898
00:41:25,750 --> 00:41:28,170
{\an8}The founder of computer science,

899
00:41:28,170 --> 00:41:31,020
{\an8}you can argue it was Alan Turing.

900
00:41:31,020 --> 00:41:33,110
He wanted to build a brain.

901
00:41:33,110 --> 00:41:35,390
He ended up building a computer.

902
00:41:35,390 --> 00:41:39,040
A good way to understand, to build an intelligent machine

903
00:41:39,040 --> 00:41:40,310
is to understand first

904
00:41:40,310 --> 00:41:42,470
the only example of intelligence we have,

905
00:41:42,470 --> 00:41:44,640
which is the human brain.

906
00:41:44,640 --> 00:41:48,080
Professor, talking about AI and machine learning today,

907
00:41:48,080 --> 00:41:50,880
how much of it is inspired by

908
00:41:50,880 --> 00:41:54,820
modeling the way our brains and minds work?

909
00:41:54,820 --> 00:41:57,100
Deep neural networks have been

910
00:41:57,100 --> 00:42:02,100
real big advance in the last six years or so,

911
00:42:02,170 --> 00:42:07,076
suddenly making machine learning able to attain

912
00:42:07,076 --> 00:42:09,590
practical use, you know,

913
00:42:09,590 --> 00:42:14,590
in terms of gadgets like Alexa or autonomous driving.

914
00:42:15,297 --> 00:42:19,100
And I think this is really the beginning of a golden age

915
00:42:19,100 --> 00:42:23,260
for artificial intelligence applications.

916
00:42:23,260 --> 00:42:25,380
Recent success stories.

917
00:42:25,380 --> 00:42:27,110
What is inside of this?

918
00:42:27,110 --> 00:42:31,100
Two algorithms: deep learning and reinforcement learning.

919
00:42:31,100 --> 00:42:35,710
And both of them come from neuroscience as an inspiration.

920
00:42:35,710 --> 00:42:39,780
Deep learning comes from a model from neurons

921
00:42:39,780 --> 00:42:41,800
in the visual cortex of monkeys.

922
00:42:41,800 --> 00:42:45,390
That's the basic architecture today of deep networks.

923
00:42:45,390 --> 00:42:48,100
And the other one is reinforcement learning.

924
00:42:48,100 --> 00:42:53,100
It's this a trade off between exploration and verification

925
00:42:53,369 --> 00:42:57,690
that you can use or a mouse uses to solve a maze

926
00:42:57,690 --> 00:42:59,350
and go towards a food.

927
00:42:59,350 --> 00:43:02,190
And so I think if somebody asked me

928
00:43:02,190 --> 00:43:03,980
what is the next breakthrough-

929
00:43:03,980 --> 00:43:05,601
Yes, what is the next breakthrough?

930
00:43:05,601 --> 00:43:06,490
(Enlai chuckling)

931
00:43:06,490 --> 00:43:07,510
Exactly.

932
00:43:07,510 --> 00:43:09,423
I would say, "I don't know."

933
00:43:09,423 --> 00:43:10,770
(both chuckling)

934
00:43:10,770 --> 00:43:15,770
But I think it's a good bet to think that

935
00:43:15,822 --> 00:43:18,663
it will come also from neuroscience.

936
00:43:20,580 --> 00:43:23,420
Neuroscience isn't their only bet.

937
00:43:23,420 --> 00:43:24,340
The center partners

938
00:43:24,340 --> 00:43:26,400
the Harvard Lab for Developmental Studies,

939
00:43:26,400 --> 00:43:29,820
uncovering how children and infants learn

940
00:43:29,820 --> 00:43:32,173
could spark next-generation AI.

941
00:43:33,380 --> 00:43:34,473
I can see a lot of promise

942
00:43:34,473 --> 00:43:36,740
for artificial intelligence

943
00:43:36,740 --> 00:43:39,230
from attempts to build machines that learn like infants

944
00:43:39,230 --> 00:43:43,870
{\an8}because such a machine I think would stand the best chance

945
00:43:43,870 --> 00:43:47,200
of developing something like our common sense.

946
00:43:47,200 --> 00:43:49,630
And they'll be most useful if they can understand us

947
00:43:49,630 --> 00:43:51,840
and share our general picture of the world.

948
00:43:51,840 --> 00:43:54,695
So how does the development of intelligence

949
00:43:54,695 --> 00:43:57,423
in children and infants

950
00:43:57,423 --> 00:44:01,570
inform the development of artificial intelligence?

951
00:44:01,570 --> 00:44:03,360
There's lots of domains in which children

952
00:44:03,360 --> 00:44:05,030
develop intuitive sense, right?

953
00:44:05,030 --> 00:44:06,870
There's the domain of what we like to call

954
00:44:06,870 --> 00:44:08,810
common sense or naive physics,

955
00:44:08,810 --> 00:44:10,360
how to objects behave,

956
00:44:10,360 --> 00:44:12,427
what kind of actions do you need to

957
00:44:12,427 --> 00:44:14,810
perform on a set of objects

958
00:44:14,810 --> 00:44:16,547
to build a tower that won't fall over

959
00:44:16,547 --> 00:44:19,730
or a bridge between two supports?

960
00:44:19,730 --> 00:44:22,013
So there's a whole domain of knowledge there.

961
00:44:22,013 --> 00:44:24,770
It's really over the last decade that I feel that

962
00:44:24,770 --> 00:44:28,400
enough information has been amassed

963
00:44:28,400 --> 00:44:29,905
that we really are in a position to

964
00:44:29,905 --> 00:44:33,660
develop computer models of human infant learning,

965
00:44:33,660 --> 00:44:35,450
which could serve as a basis for

966
00:44:35,450 --> 00:44:38,250
developing smarter, more flexible machines.

967
00:44:38,250 --> 00:44:40,400
How confident are you

968
00:44:40,400 --> 00:44:45,400
that we may be able to create a truly intelligent machine?

969
00:44:45,550 --> 00:44:47,453
I think it's the greatest problem,

970
00:44:48,300 --> 00:44:51,730
besides making a machine, a computer

971
00:44:51,730 --> 00:44:55,168
that is as intelligent as we are,

972
00:44:55,168 --> 00:44:57,523
to be a long time.

973
00:44:58,359 --> 00:45:01,348
(gentle childish music)

974
00:45:01,348 --> 00:45:04,650
How smart AI can get is anyone's guess.

975
00:45:04,650 --> 00:45:08,550
There are many books and theories about AI's future impact.

976
00:45:08,550 --> 00:45:12,955
Here's best-selling author and futurist Max Tegmark.

977
00:45:12,955 --> 00:45:15,030
All the recent progress in AI

978
00:45:15,030 --> 00:45:17,560
makes me wonder how far it's going to go.

979
00:45:17,560 --> 00:45:19,470
And I think about this question in terms of

980
00:45:19,470 --> 00:45:21,805
this abstract landscape of tasks,

981
00:45:21,805 --> 00:45:24,380
where the elevation of each task

982
00:45:24,380 --> 00:45:27,432
represents how hard it is for AI to do at human level

983
00:45:27,432 --> 00:45:31,610
and the water level represents where AI is today.

984
00:45:31,610 --> 00:45:36,062
So that beach over there, for instance, represents driving.

985
00:45:36,062 --> 00:45:38,890
AI is improving, the water level is rising,

986
00:45:38,890 --> 00:45:41,510
{\an8}and it's a really bad idea to try to make a career now

987
00:45:41,510 --> 00:45:42,623
{\an8}out of being a driver.

988
00:45:44,110 --> 00:45:47,729
The tall tree out there might represent investing,

989
00:45:47,729 --> 00:45:51,000
which will take longer to get outperformed by AI.

990
00:45:51,000 --> 00:45:53,610
And then the distance I imagined tall mountains

991
00:45:53,610 --> 00:45:57,400
like art, cinematography, science.

992
00:45:57,400 --> 00:46:00,267
The really big question that keeps me awake at night is

993
00:46:00,267 --> 00:46:03,650
"Will the water eventually flood all land?"

994
00:46:03,650 --> 00:46:05,830
Because if we do, then by definition,

995
00:46:05,830 --> 00:46:08,660
AI can do all jobs, even the job of AI design,

996
00:46:08,660 --> 00:46:10,140
better than us,

997
00:46:10,140 --> 00:46:13,740
rapidly leaving human intelligence far, far behind.

998
00:46:13,740 --> 00:46:14,760
Either it's going to be

999
00:46:14,760 --> 00:46:17,770
the best thing ever to happen to humanity or the worst.

1000
00:46:17,770 --> 00:46:20,516
I want to work hard to make sure it's the former.

1001
00:46:20,516 --> 00:46:23,410
(spectators chattering)

1002
00:46:23,410 --> 00:46:25,056
Well, there you go.

1003
00:46:25,056 --> 00:46:26,723
Va-va-va-voom.

1004
00:46:26,723 --> 00:46:27,881
(Enlai chuckling)

1005
00:46:27,881 --> 00:46:29,070
It's like animal cruelty.

1006
00:46:29,070 --> 00:46:31,780
This journey I've had studying AI

1007
00:46:31,780 --> 00:46:34,206
has been a real eye-opener.

1008
00:46:34,206 --> 00:46:36,395
Whoa, you're a heavy boy.

1009
00:46:36,395 --> 00:46:39,373
AI being able to make art,

1010
00:46:41,332 --> 00:46:42,910
(Enlai laughing)

1011
00:46:42,910 --> 00:46:44,613
AI being able to make music,

1012
00:46:45,610 --> 00:46:48,624
AI being able to understand many languages,

1013
00:46:48,624 --> 00:46:50,842
{\an8}(AI translator speaking Mandarin)

1014
00:46:50,842 --> 00:46:53,229
and I think Turing would be quite thrilled

1015
00:46:53,229 --> 00:46:55,800
by how far AI has come.

1016
00:46:55,800 --> 00:46:56,633
Right?

1017
00:46:57,890 --> 00:47:00,280
But I don't think I will see robots

1018
00:47:00,280 --> 00:47:02,080
that are smarter than me.

1019
00:47:02,080 --> 00:47:02,913
Bingo!

1020
00:47:04,004 --> 00:47:05,590
Not in my lifetime.

1021
00:47:05,590 --> 00:47:09,251
Stupid computer cannot fight human being, la.

1022
00:47:09,251 --> 00:47:13,296
As long as we use these tools to help us

1023
00:47:13,296 --> 00:47:15,832
and not take over us.

1024
00:47:15,832 --> 00:47:17,170
(performers singing)

1025
00:47:17,170 --> 00:47:19,007
I think I'm good with that.

1026
00:47:19,007 --> 00:47:20,150
(robot dog speaking foreign language)

1027
00:47:20,150 --> 00:47:22,269
What's the most pressing when it comes to the use of AI

1028
00:47:22,269 --> 00:47:26,229
is our ethical use of AI.

1029
00:47:26,229 --> 00:47:29,223
What's right, what's wrong.

1030
00:47:30,567 --> 00:47:33,160
That's the question I seek to answer

1031
00:47:33,160 --> 00:47:35,682
in my next AI adventure.

1032
00:47:35,682 --> 00:47:37,260
(gavel clanking)

1033
00:47:37,260 --> 00:47:38,450
Order in the court!

1034
00:47:38,450 --> 00:47:40,698
Court is in session!

1035
00:47:40,698 --> 00:47:44,031
{\an8}(uplifting synth music)