The danger of AI is weirder than you think | Janelle Shane

2,816,115 views ใƒป 2019-11-13

TED


ืื ื ืœื—ืฅ ืคืขืžื™ื™ื ืขืœ ื”ื›ืชื•ื‘ื™ื•ืช ื‘ืื ื’ืœื™ืช ืœืžื˜ื” ื›ื“ื™ ืœื”ืคืขื™ืœ ืืช ื”ืกืจื˜ื•ืŸ.

ืชืจื’ื•ื: Shlomo Adam ืขืจื™ื›ื”: Sigal Tifferet
00:01
So, artificial intelligence
0
1765
3000
ื”ื‘ื™ื ื” ื”ืžืœืื›ื•ืชื™ืช, ื‘"ืž,
00:04
is known for disrupting all kinds of industries.
1
4789
3529
ื™ื“ื•ืขื” ื‘ื›ืš ืฉื”ื™ื ืžืฉื‘ืฉืช ืชืขืฉื™ื•ืช ืฉื•ื ื•ืช.
00:08
What about ice cream?
2
8961
2043
ืžื” ืขื ื’ืœื™ื“ื”?
00:11
What kind of mind-blowing new flavors could we generate
3
11879
3639
ืžื”ื ื”ื˜ืขืžื™ื ื”ื—ื“ืฉื™ื ื•ื”ืžื“ื”ื™ืžื™ื ืฉื ื•ื›ืœ ืœื”ืžืฆื™ื
00:15
with the power of an advanced artificial intelligence?
4
15542
2976
ืื ื ื™ืขื–ืจ ื‘ื›ื•ื—ื” ืฉืœ ื‘"ืž ืžืชืงื“ืžืช?
00:19
So I teamed up with a group of coders from Kealing Middle School
5
19011
4161
ืื™ื—ื“ืชื™ ื›ื•ื—ื•ืช ืขื ืงื‘ื•ืฆืช ืžืชื›ื ืชื™ื ืžื—ื˜ื™ื‘ืช ื”ื‘ื™ื ื™ื™ื ืฉืœ ื‘ื™ื”"ืก ืงื™ืœื™ื ื’
00:23
to find out the answer to this question.
6
23196
2241
ื›ื“ื™ ืœืžืฆื•ื ืืช ื”ืชืฉื•ื‘ื” ืœืฉืืœื” ื–ื•.
00:25
They collected over 1,600 existing ice cream flavors,
7
25461
5081
ื”ื ืืกืคื• ืžืขืœ 1,600 ื˜ืขืžื™ื ืงื™ื™ืžื™ื ืฉืœ ื’ืœื™ื“ื”
00:30
and together, we fed them to an algorithm to see what it would generate.
8
30566
5522
ื•ื‘ื™ื—ื“ ื”ื–ื ื• ืื•ืชื ืœืืœื’ื•ืจื™ืชื ื›ื“ื™ ืœืจืื•ืช ืžื” ื”ื•ื ื™ืฆืœื™ื— ืœื™ืฆื•ืจ.
00:36
And here are some of the flavors that the AI came up with.
9
36112
3753
ื”ื ื” ื›ืžื” ืžื”ื˜ืขืžื™ื ืฉื”ื‘"ืž ื”ืžืฆื™ืื”.
00:40
[Pumpkin Trash Break]
10
40444
1471
[ื”ืคืกืงืช ื–ื‘ืœ ื“ืœืขืช]
00:41
(Laughter)
11
41939
1402
(ืฆื—ื•ืง)
00:43
[Peanut Butter Slime]
12
43365
2469
[ืกืœื™ื™ื ื—ืžืืช ื‘ื•ื˜ื ื™ื]
00:46
[Strawberry Cream Disease]
13
46822
1343
[ืžื—ืœืช ืงืฆืคืช ืชื•ืช]
00:48
(Laughter)
14
48189
2126
(ืฆื—ื•ืง)
00:50
These flavors are not delicious, as we might have hoped they would be.
15
50339
4597
ืืœื” ืื™ื ื ื˜ืขืžื™ื ื˜ืขื™ืžื™ื, ื›ืคื™ ืฉืงื™ื•ื•ื™ื ื• ืฉื™ื”ื™ื•.
00:54
So the question is: What happened?
16
54960
1864
ืื– ื”ืฉืืœื” ื”ื™ื: ืžื” ืงืจื”? ืžื” ื”ืฉืชื‘ืฉ?
00:56
What went wrong?
17
56848
1394
00:58
Is the AI trying to kill us?
18
58266
1959
ื”ืื ื”ื‘"ืž ืžื ืกื” ืœื”ืจื•ื’ ืื•ืชื ื•?
01:01
Or is it trying to do what we asked, and there was a problem?
19
61027
4310
ื•ืื•ืœื™ ื”ื™ื ืžื ืกื” ืœืขืฉื•ืช ืžื” ืฉื‘ื™ืงืฉื ื•, ื•ื”ื™ืชื” ื›ืืŸ ื‘ืขื™ื”?
01:06
In movies, when something goes wrong with AI,
20
66567
2464
ื‘ืกืจื˜ื™ื, ื›ืฉืžืฉื”ื• ืžืฉืชื‘ืฉ ืขื ื‘"ืž,
01:09
it's usually because the AI has decided
21
69055
2712
ื–ื” ื‘ื“ืจืš ื›ืœืœ ื›ื™ ื”ื‘"ืž ื”ื—ืœื™ื˜ื”
01:11
that it doesn't want to obey the humans anymore,
22
71791
2272
ืฉื”ื™ื ื›ื‘ืจ ืœื ืจื•ืฆื” ืœืฆื™ื™ืช ืœื‘ื ื™ ื”ืื“ื
ื•ื™ืฉ ืœื” ืžื˜ืจื•ืช ืžืฉืœื”, ืชื•ื“ื” ืจื‘ื” ืœื›ื.
01:14
and it's got its own goals, thank you very much.
23
74087
2623
01:17
In real life, though, the AI that we actually have
24
77266
3216
ืื‘ืœ ื‘ืžืฆื™ืื•ืช, ื”ื‘"ืž ืฉื™ืฉ ืœื ื•
01:20
is not nearly smart enough for that.
25
80506
1863
ืžืžืฉ ืœื ื›ืœ-ื›ืš ื—ื›ืžื”.
01:22
It has the approximate computing power
26
82781
2982
ื™ืฉ ืœื” ื›ื•ื— ืžื™ื—ืฉื•ื‘ ื‘ืขืจืš ื›ืžื• ืฉืœ ืฉืœืฉื•ืœ,
01:25
of an earthworm,
27
85787
1276
ื•ืœื›ืœ ื”ื™ื•ืชืจ - ืฉืœ ื“ื‘ื•ืจืช ื“ื‘ืฉ ื™ื—ื™ื“ื”,
01:27
or maybe at most a single honeybee,
28
87087
3403
01:30
and actually, probably maybe less.
29
90514
2215
ื•ื›ื ืจืื” ืคื—ื•ืช ืžื›ืš.
01:32
Like, we're constantly learning new things about brains
30
92753
2594
ื”ืจื™ ืื ื• ื›ืœ ื”ื–ืžืŸ ืœื•ืžื“ื™ื ื“ื‘ืจื™ื ื—ื“ืฉื™ื ืื•ื“ื•ืช ื”ืžื•ื—
01:35
that make it clear how much our AIs don't measure up to real brains.
31
95371
4360
ืฉืžื‘ื”ื™ืจื™ื ืขื“ ื›ืžื” ื”ื‘"ืž ืื™ื ื” ื‘ืช-ื”ืฉื•ื•ืื” ืขื ืžื•ื— ืืžื™ืชื™.
01:39
So today's AI can do a task like identify a pedestrian in a picture,
32
99755
5663
ื”ื‘"ืž ืฉืœ ื”ื™ื•ื ืžืกื•ื’ืœืช ืœืขืฉื•ืช ืœื–ื”ื•ืช ื‘ืชืžื•ื ื” ื”ื•ืœืš-ืจื’ืœ,
01:45
but it doesn't have a concept of what the pedestrian is
33
105442
2983
ืื‘ืœ ืื™ืŸ ืœื” ืžื•ืฉื’ ืžื”ื• ื”ื•ืœืš-ืจื’ืœ
01:48
beyond that it's a collection of lines and textures and things.
34
108449
4824
ืžืขื‘ืจ ืœื”ื™ื•ืชื• ืื•ืกืฃ ืงื•ื•ื™ื, ืžืจืงืžื™ื ื•ื›ื“ื•ืžื”.
01:53
It doesn't know what a human actually is.
35
113792
2521
ื”ื™ื ื‘ื›ืœืœ ืœื ื™ื•ื“ืขืช ืžื”ื• ืื“ื.
01:56
So will today's AI do what we ask it to do?
36
116822
3282
ื”ืื ื”ื‘"ืž ืฉืœ ื™ืžื™ื ื• ืชืขืฉื” ืžื” ืฉื ื‘ืงืฉ ืžืžื ื”?
02:00
It will if it can,
37
120128
1594
ื›ืŸ, ืื ื”ื™ื ืชื•ื›ืœ,
02:01
but it might not do what we actually want.
38
121746
2726
ืื‘ืœ ืื•ืœื™ ืœื ื‘ื“ื™ื•ืง ืžื” ืฉื ืจืฆื”.
02:04
So let's say that you were trying to get an AI
39
124496
2415
ื ื ื™ื— ืฉืืชื ืจื•ืฆื™ื ืฉื”ื‘"ืž
02:06
to take this collection of robot parts
40
126935
2619
ืชื™ืงื— ืืช ืื•ืกืฃ ื—ืœืงื™ ื”ืจื•ื‘ื•ื˜ ื”ื–ื”
02:09
and assemble them into some kind of robot to get from Point A to Point B.
41
129578
4197
ื•ืชืจื›ื™ื‘ ืžื”ืŸ ืจื•ื‘ื•ื˜ ืฉื™ื’ื™ืข ืžื ืงื•ื“ื” ื' ืœื‘'.
02:13
Now, if you were going to try and solve this problem
42
133799
2481
ืื™ืœื• ื ื™ืกื™ืชื ืœืคืชื•ืจ ื‘ืขื™ื” ื–ื• ืข"ื™ ื›ืชื™ื‘ืช ืชื•ื›ื ืช ืžื—ืฉื‘ ืจื’ื™ืœื”,
02:16
by writing a traditional-style computer program,
43
136304
2351
02:18
you would give the program step-by-step instructions
44
138679
3417
ื”ื™ื™ืชื ื ื•ืชื ื™ื ืœืชื•ื›ื ื” ื”ื•ืจืื•ืช ืžืคื•ืจื˜ื•ืช
02:22
on how to take these parts,
45
142120
1329
ืื™ืš ืœืงื—ืช ืืช ื”ื—ืœืงื™ื ื”ืืœื”,
02:23
how to assemble them into a robot with legs
46
143473
2407
ืœื”ืจื›ื™ื‘ ืžื”ื ืจื•ื‘ื•ื˜ ืขื ืจื’ืœื™ื™ื
02:25
and then how to use those legs to walk to Point B.
47
145904
2942
ื•ืœื”ืฉืชืžืฉ ื‘ืจื’ืœื™ื™ื ื”ืืœื” ื›ื“ื™ ืœืœื›ืช ืขื“ ื ืงื•ื“ื” ื‘'.
02:29
But when you're using AI to solve the problem,
48
149441
2340
ืื‘ืœ ื›ืฉืžืฉืชืžืฉื™ื ื‘ื‘"ืž ื›ื“ื™ ืœืคืชื•ืจ ื‘ืขื™ื” ื–ื•,
02:31
it goes differently.
49
151805
1174
ื–ื” ืฉื•ื ื”.
ืื™ื ื›ื ืื•ืžืจื™ื ืœื” ืื™ืš ืœืคืชื•ืจ ืืช ื”ื‘ืขื™ื”,
02:33
You don't tell it how to solve the problem,
50
153003
2382
02:35
you just give it the goal,
51
155409
1479
ืืœื ืจืง ืื•ืžืจื™ื ืœื” ืžื”ื™ ื”ืžื˜ืจื”,
02:36
and it has to figure out for itself via trial and error
52
156912
3262
ื•ื”ื™ื ืฆืจื™ื›ื” ืœื”ื‘ื™ืŸ ืœื‘ื“ ื“ืจืš ื ื™ืกื•ื™ ื•ื˜ืขื™ื™ื”
02:40
how to reach that goal.
53
160198
1484
ืื™ืš ืœื”ื’ื™ืข ืœืžื˜ืจื” ื–ื•.
02:42
And it turns out that the way AI tends to solve this particular problem
54
162254
4102
ื•ืžืกืชื‘ืจ ืฉื”ื‘"ืž ื ื•ื˜ื” ืœืคืชื•ืจ ื‘ืขื™ื” ืžืกื•ื™ืžืช ื–ื• ื‘ืื•ืคืŸ ื”ื‘ื:
02:46
is by doing this:
55
166380
1484
02:47
it assembles itself into a tower and then falls over
56
167888
3367
ื”ื™ื ืžืจื›ื™ื‘ื” ืžื”ื—ืœืงื™ื ืžื’ื“ืœ ืฉื ื•ืคืœ ื•ื ื•ื—ืช ืขืœ ื ืงื•ื“ื” ื‘'.
02:51
and lands at Point B.
57
171279
1827
ื•ื˜ื›ื ื™ืช, ื–ื” ืคื•ืชืจ ืืช ื”ื‘ืขื™ื”.
02:53
And technically, this solves the problem.
58
173130
2829
02:55
Technically, it got to Point B.
59
175983
1639
ื˜ื›ื ื™ืช, ื”ื™ื ื”ื’ื™ืขื” ืœื ืงื•ื“ื” ื‘'.
02:57
The danger of AI is not that it's going to rebel against us,
60
177646
4265
ื”ืกื›ื ื” ืื™ื ื ื” ืฉื”ื‘"ืž ืชืชืžืจื“ ื ื’ื“ื ื•,
03:01
it's that it's going to do exactly what we ask it to do.
61
181935
4274
ืืœื ืฉืชืขืฉื” ื‘ื“ื™ื•ืง ืžื” ืฉืื ื• ื“ื•ืจืฉื™ื ืžืžื ื”.
03:06
So then the trick of working with AI becomes:
62
186876
2498
ืœื›ืŸ, ื”ื—ื•ื›ืžื” ื‘ืขื‘ื•ื“ื” ืขื ื‘"ืž ื”ื™ื,
03:09
How do we set up the problem so that it actually does what we want?
63
189398
3828
ืื™ืš ืœื”ื’ื“ื™ืจ ืืช ื”ื‘ืขื™ื” ื›ืš ืฉื”ื™ื ืชืขืฉื” ื‘ื“ื™ื•ืง ืืช ืžื” ืฉื‘ื™ืงืฉื ื•?
03:14
So this little robot here is being controlled by an AI.
64
194726
3306
ื”ืจื•ื‘ื•ื˜ ื”ืงื˜ืŸ ื”ื–ื” ื ืฉืœื˜ ืข"ื™ ื‘"ืž.
ื”ื‘"ืž ืขื™ืฆื‘ื” ืืช ืจื’ืœื™ ื”ืจื•ื‘ื•ื˜
03:18
The AI came up with a design for the robot legs
65
198056
2814
03:20
and then figured out how to use them to get past all these obstacles.
66
200894
4078
ื•ืžืฆืื” ื“ืจืš ืœื”ืฉืชืžืฉ ื‘ื”ืŸ ื›ื“ื™ ืœืขื‘ื•ืจ ืืช ื›ืœ ื”ืžื›ืฉื•ืœื™ื ื”ืืœื”.
03:24
But when David Ha set up this experiment,
67
204996
2741
ืื‘ืœ ื›ืฉื“ื™ื™ื•ื™ื“ ื”ื ืขืจืš ืืช ื”ื ื™ืกื•ื™ ื”ื–ื”,
03:27
he had to set it up with very, very strict limits
68
207761
2856
ื”ื™ื” ืขืœื™ื• ืœื”ื’ื“ื™ืจ ืžื’ื‘ืœื•ืช ื ื•ืงืฉื•ืช ื‘ื™ื•ืชืจ
03:30
on how big the AI was allowed to make the legs,
69
210641
3292
ืœื’ื‘ื™ ื’ื•ื“ืœ ื”ืจื’ืœื™ื™ื ืฉื”ื‘"ืž ืจืฉืื™ืช ืœืงื‘ื•ืข,
03:33
because otherwise ...
70
213957
1550
ื›ื™ ืื—ืจืช...
03:43
(Laughter)
71
223058
3931
(ืฆื—ื•ืง)
03:48
And technically, it got to the end of that obstacle course.
72
228563
3745
ื•ื˜ื›ื ื™ืช, ื”ื™ื ืขื‘ืจื” ืืช ืžืกืœื•ืœ ื”ืžื›ืฉื•ืœื™ื.
03:52
So you see how hard it is to get AI to do something as simple as just walk.
73
232332
4942
ืื– ืืชื ืžื‘ื™ื ื™ื ื›ืžื” ืงืฉื” ืœื’ืจื•ื ืœื‘"ืž ืœืขืฉื•ืช ืžืฉื”ื• ืคืฉื•ื˜ ื›ืžื• ื”ืœื™ื›ื”.
03:57
So seeing the AI do this, you may say, OK, no fair,
74
237298
3820
ืื– ื›ืฉืืชื ืจื•ืื™ื ืืช ื”ื‘"ืž ืขื•ืฉื” ื›ืš, ื•ื“ืื™ ืชื’ื™ื“ื•: "ื–ื” ืœื ื”ื•ื’ืŸ,
04:01
you can't just be a tall tower and fall over,
75
241142
2580
"ืืกื•ืจ ืœืš ืœื‘ื ื•ืช ืกืชื ืžื’ื“ืœ ืฉื ื•ืคืœ.
04:03
you have to actually, like, use legs to walk.
76
243746
3435
"ืขืœื™ื™ืš ืœื”ืฉืชืžืฉ ื‘ืจื’ืœื™ื™ื ื›ื“ื™ ืœืœื›ืช."
04:07
And it turns out, that doesn't always work, either.
77
247205
2759
ื•ืžืกืชื‘ืจ ืฉื’ื ื–ื” ืœื ืชืžื™ื“ ืขื•ื‘ื“.
04:09
This AI's job was to move fast.
78
249988
2759
ื”ืžื˜ืœื” ืฉืœ ื”ื‘"ืž ื”ื–ืืช ื”ื™ืชื” ืœื ื•ืข ืžื”ืจ.
04:13
They didn't tell it that it had to run facing forward
79
253115
3593
ืœื ื ืืžืจ ืœื” ืฉืฆืจื™ืš ืœื ื•ืข ืงื“ื™ืžื”
04:16
or that it couldn't use its arms.
80
256732
2258
ืื• ืฉืืกื•ืจ ืœื” ืœื”ืฉืชืžืฉ ื‘ื–ืจื•ืขื•ืช.
04:19
So this is what you get when you train AI to move fast,
81
259487
4618
ื–ื” ืžื” ืฉืžืงื‘ืœื™ื ื›ืฉืžืืžื ื™ื ืืช ื”ื‘"ืž ืœืชื ื•ืขื” ืžื”ื™ืจื”:
04:24
you get things like somersaulting and silly walks.
82
264129
3534
ืกืœื˜ื•ืช ื•ื”ืœื™ื›ื” ืžืฉื•ื ื”.
04:27
It's really common.
83
267687
1400
ื–ื” ื ืคื•ืฅ ืžืื“.
04:29
So is twitching along the floor in a heap.
84
269667
3179
ื›ืš ื’ื ื”ืชืงื“ืžื•ืช ืฉืœ ืขืจื™ืžื” ื‘ืขื•ื•ื™ืชื•ืช.
04:32
(Laughter)
85
272870
1150
(ืฆื—ื•ืง)
04:35
So in my opinion, you know what should have been a whole lot weirder
86
275241
3254
ืื– ืœื“ืขืชื™, ืžื” ืฉื”ื™ื” ืฆืจื™ืš ืœื”ื™ื•ืช ื”ืจื‘ื” ื™ื•ืชืจ ืžื•ื–ืจ
04:38
is the "Terminator" robots.
87
278519
1396
ื”ื ื”ืจื•ื‘ื•ื˜ื™ื ืฉืœ "ืฉืœื™ื—ื•ืช ืงื˜ืœื ื™ืช".
04:40
Hacking "The Matrix" is another thing that AI will do if you give it a chance.
88
280256
3755
ื’ื ืคื™ืฆื•ื— ื”"ืžื˜ืจื™ืงืก" ื”ื•ื ืžืฉื”ื• ืฉื‘"ืž ืชื•ื›ืœ ืœืขืฉื•ืช ืื ื™ืชื ื• ืœื” ื”ื–ื“ืžื ื•ืช.
04:44
So if you train an AI in a simulation,
89
284035
2517
ืื– ืื ืชืืžื ื• ื‘"ืž ื‘ื”ื“ืžื™ื™ื”,
04:46
it will learn how to do things like hack into the simulation's math errors
90
286576
4113
ื”ื™ื ืชืœืžื“ ืœืžืฉืœ ืœืคืฆื— ืืช ื”ืฉื’ื™ืื•ืช ื”ืžืชืžื˜ื™ื•ืช ืฉืœ ื”ื”ื“ืžื™ื™ื”
04:50
and harvest them for energy.
91
290713
2207
ื•ืœื”ืคื™ืง ืžื”ืŸ ืื ืจื’ื™ื”.
04:52
Or it will figure out how to move faster by glitching repeatedly into the floor.
92
292944
5475
ืื• ืชืžืฆื ืื™ืš ืœื ื•ืข ืžื”ืจ ื™ื•ืชืจ ืข"ื™ ื’ืœื™ืฉื” ืžืชืžื“ืช ืขืœ ื”ืจืฆืคื”.
04:58
When you're working with AI,
93
298443
1585
ื›ืฉืขื•ื‘ื“ื™ื ืขื ื‘"ืž,
05:00
it's less like working with another human
94
300052
2389
ื–ื” ื“ื•ืžื” ืคื—ื•ืช ืœืขื‘ื•ื“ื” ืขื ื™ืฆื•ืจ ืื ื•ืฉื™ ืื—ืจ,
05:02
and a lot more like working with some kind of weird force of nature.
95
302465
3629
ื•ื“ื•ืžื” ื™ื•ืชืจ ืœืขื‘ื•ื“ื” ืขื ืื™ื–ื” ื›ื•ื—-ื˜ื‘ืข ืžื•ื–ืจ.
05:06
And it's really easy to accidentally give AI the wrong problem to solve,
96
306562
4623
ื•ืœืžืขืฉื”, ืงืœ ืžืื“ ืœืชืช ืœื‘"ืž ื‘ื˜ืขื•ืช ืืช ื”ื‘ืขื™ื” ื”ืœื-ื ื›ื•ื ื”,
05:11
and often we don't realize that until something has actually gone wrong.
97
311209
4538
ื•ืขืœ ืคื™ ืจื•ื‘ ืื™ื ื ื• ืžื‘ื™ื ื™ื ื–ืืช ืขื“ ืฉืžืฉื”ื• ืžืฉืชื‘ืฉ.
05:16
So here's an experiment I did,
98
316242
2080
ื”ื ื” ื ื™ืกื•ื™ ืฉืขืจื›ืชื™,
05:18
where I wanted the AI to copy paint colors,
99
318346
3182
ื•ื‘ื• ืจืฆื™ืชื™ ืฉื”ื‘"ืž ืชืขืชื™ืง ื’ื•ื•ื ื™ ืฆื‘ืข
05:21
to invent new paint colors,
100
321552
1746
ื›ื“ื™ ืœื”ืžืฆื™ื ื’ื•ื•ื ื™ื ื—ื“ืฉื™ื,
05:23
given the list like the ones here on the left.
101
323322
2987
ื‘ื”ื™ื ืชืŸ ืจืฉื™ืžื” ื›ืžื• ื–ื• ืฉืžืฉืžืืœ.
05:26
And here's what the AI actually came up with.
102
326798
3004
ื•ื”ื ื” ื”ืจืฉื™ืžื” ืฉื”ืคื™ืงื” ื”ื‘"ืž.
05:29
[Sindis Poop, Turdly, Suffer, Gray Pubic]
103
329826
3143
[ืงืงื™ ืกื™ื ื“ื™ืก, ื’ื•ื•ืŸ ื’ืœืœ, ืกื‘ืœ, ืืคื•ืจ ืขืจื•ื•ื”]
05:32
(Laughter)
104
332993
4230
(ืฆื—ื•ืง)
05:39
So technically,
105
339177
1886
ืื– ื˜ื›ื ื™ืช,
05:41
it did what I asked it to.
106
341087
1864
ื”ื‘"ืž ืขืฉืชื” ืžื” ืฉื‘ื™ืงืฉืชื™.
05:42
I thought I was asking it for, like, nice paint color names,
107
342975
3308
ื—ืฉื‘ืชื™ ืฉืื ื™ ืžื‘ืงืฉืช ืžืžื ื” ืฉืžื•ืช ื—ื“ืฉื™ื ื•ื ื—ืžื“ื™ื ืฉืœ ื’ื•ื•ื ื™ื,
05:46
but what I was actually asking it to do
108
346307
2307
ืื‘ืœ ืžื” ืฉื‘ื™ืงืฉืชื™ ืžืžื ื” ื‘ืคื•ืขืœ
05:48
was just imitate the kinds of letter combinations
109
348638
3086
ื”ื™ื” ืจืง ืœื—ืงื•ืช ืืช ืฆื™ืจื•ืคื™ ื”ืื•ืชื™ื•ืช
05:51
that it had seen in the original.
110
351748
1905
ืฉื”ื™ื ืจืืชื” ื‘ืžืงื•ืจ.
05:53
And I didn't tell it anything about what words mean,
111
353677
3098
ื•ืœื ืืžืจืชื™ ืœื” ื“ื‘ืจ ืขืœ ืžืฉืžืขื•ืช ื”ืžื™ืœื™ื,
05:56
or that there are maybe some words
112
356799
2560
ืื• ืฉืื•ืœื™ ื™ืฉ ืžื™ืœื™ื
05:59
that it should avoid using in these paint colors.
113
359383
2889
ืฉื›ื“ืื™ ืœื”ื™ืžื ืข ืžื”ืŸ ื‘ืฉืžื•ืช ืฉืœ ื’ื•ื•ื ื™ ืฆื‘ืข.
06:03
So its entire world is the data that I gave it.
114
363141
3494
ื›ืœ ืขื•ืœืžื” ื”ื™ื” ื”ื ืชื•ื ื™ื ืฉื ืชืชื™ ืœื”.
06:06
Like with the ice cream flavors, it doesn't know about anything else.
115
366659
4028
ื›ืžื• ืขื ื˜ืขืžื™ ื”ื’ืœื™ื“ื”, ื”ื™ื ืœื ืžื›ื™ืจื” ืฉื•ื ื“ื‘ืจ ื ื•ืกืฃ.
06:12
So it is through the data
116
372491
1638
ืžืฉืžืข ืฉื“ืจืš ื”ื ืชื•ื ื™ื
06:14
that we often accidentally tell AI to do the wrong thing.
117
374153
4044
ืื ื• ืžืจื‘ื™ื ืœื˜ืขื•ืช ื•ืœื•ืžืจ ืœื‘"ืž ืœืขืฉื•ืช ืืช ื”ื“ื‘ืจ ื”ืœื-ื ื›ื•ืŸ.
06:18
This is a fish called a tench.
118
378694
3032
ื–ื”ื• ื“ื’ ื‘ืฉื ื˜ื™ื ืงื”.
06:21
And there was a group of researchers
119
381750
1815
ื•ื”ื™ืชื” ืงื‘ื•ืฆืช ื—ื•ืงืจื™ื
06:23
who trained an AI to identify this tench in pictures.
120
383589
3874
ืฉืื™ืžื ื” ื‘"ืž ืœื–ื”ื•ืช ื˜ื™ื ืงื” ื‘ืชืžื•ื ื•ืช.
06:27
But then when they asked it
121
387487
1296
ืื‘ืœ ื›ืฉื”ื ืฉืืœื• ืืช ื”ื‘"ืž
06:28
what part of the picture it was actually using to identify the fish,
122
388807
3426
ื‘ืื™ื–ื” ื—ืœืง ืžื”ืชืžื•ื ื” ื”ื™ื ื”ืฉืชืžืฉื” ื›ื“ื™ ืœื–ื”ื•ืช ืืช ื”ื“ื’,
06:32
here's what it highlighted.
123
392257
1358
ื–ื” ืžื” ืฉื”ื™ื ื”ื“ื’ื™ืฉื”.
06:35
Yes, those are human fingers.
124
395203
2189
ื ื›ื•ืŸ, ืืฆื‘ืขื•ืช ืื ื•ืฉื™ื•ืช.
06:37
Why would it be looking for human fingers
125
397416
2059
ืžื“ื•ืข ื”ื™ื ืžื—ืคืฉืช ืืฆื‘ืขื•ืช ืื ื•ืฉื™ื•ืช ื›ืฉื”ื™ื ืžื ืกื” ืœื–ื”ื•ืช ื“ื’?
06:39
if it's trying to identify a fish?
126
399499
1921
06:42
Well, it turns out that the tench is a trophy fish,
127
402126
3164
ืžืกืชื‘ืจ ืฉืืช ื”ื˜ื™ื ืงื” ื“ื’ื™ื ื‘ืขื™ืงืจ ื›ืฉืœืœ ืฆื™ื“,
06:45
and so in a lot of pictures that the AI had seen of this fish
128
405314
3811
ื•ืœื›ืŸ ื‘ืชืžื•ื ื•ืช ืจื‘ื•ืช ืฉืœ ื”ื“ื’ ื”ื–ื” ืฉื”ื‘"ืž ืจืืชื” ื‘ืื™ืžื•ืŸ ืฉืœื”,
06:49
during training,
129
409149
1151
06:50
the fish looked like this.
130
410324
1490
ื”ื“ื’ ื ืจืื” ื›ื›ื”.
06:51
(Laughter)
131
411838
1635
(ืฆื—ื•ืง)
06:53
And it didn't know that the fingers aren't part of the fish.
132
413497
3330
ื”ื™ื ืœื ื™ื“ืขื” ืฉื”ืืฆื‘ืขื•ืช ืื™ื ืŸ ืื™ื‘ืจื™ื ืฉืœ ื”ื“ื’.
06:58
So you see why it is so hard to design an AI
133
418808
4120
ืื– ืืชื ืžื‘ื™ื ื™ื ืœืžื” ืงืฉื” ื›ืœ-ื›ืš ืœืชื›ื ืŸ ื‘"ืž
07:02
that actually can understand what it's looking at.
134
422952
3319
ืฉืชื•ื›ืœ ืœื”ื‘ื™ืŸ ืžื” ื”ื™ื ืจื•ืื”.
07:06
And this is why designing the image recognition
135
426295
2862
ื•ืœื›ืŸ ืงืฉื” ื›ืš-ื›ืš ืœืชื›ื ืŸ ื–ื™ื”ื•ื™ ืชืžื•ื ื”
07:09
in self-driving cars is so hard,
136
429181
2067
ื‘ืžื›ื•ื ื™ื•ืช ืื•ื˜ื•ื ื•ืžื™ื•ืช,
07:11
and why so many self-driving car failures
137
431272
2205
ื•ืžื“ื•ืข ื›ืฉืœื™ื ืจื‘ื™ื ื›ืœ-ื›ืš ื‘ืžื›ื•ื ื™ื•ืช ืื•ื˜ื•ื ื•ืžื™ื•ืช
07:13
are because the AI got confused.
138
433501
2885
ื ื•ื‘ืขื™ื ืžื‘ืœื‘ื•ืœ ืฉืœ ื”ื‘"ืž.
07:16
I want to talk about an example from 2016.
139
436410
4008
ืื ื™ ืจื•ืฆื” ืœื“ื‘ืจ ืขืœ ื“ื•ื’ืžื” ืž-2016.
07:20
There was a fatal accident when somebody was using Tesla's autopilot AI,
140
440442
4455
ื”ื™ืชื” ืชืื•ื ื” ืงื˜ืœื ื™ืช ื›ืฉืžื™ืฉื”ื• ื”ืฉืชืžืฉ ื‘ื‘"ืž ืฉืœ ืจื›ื‘ ืื•ื˜ื•ื ื•ืžื™ ืฉืœ "ื˜ืกืœื”",
07:24
but instead of using it on the highway like it was designed for,
141
444921
3414
ืื‘ืœ ื‘ืžืงื•ื ืœื”ืฉืชืžืฉ ื‘ื” ื‘ื›ื‘ื™ืฉ ื”ืžื”ื™ืจ, ื›ืคื™ ืฉืชื•ื›ื ื ื”,
07:28
they used it on city streets.
142
448359
2205
ื”ื•ื ื”ืฉืชืžืฉ ื‘ื” ื‘ืจื—ื•ื‘ื•ืช ื”ืขื™ืจ.
07:31
And what happened was,
143
451239
1175
ื•ืžื” ืฉืงืจื” ื”ื•ื,
07:32
a truck drove out in front of the car and the car failed to brake.
144
452438
3396
ืฉืžืฉืื™ืช ื—ืฆืชื” ืœืคื ื™ ื”ืจื›ื‘ ื•ื”ืจื›ื‘ ืœื ื‘ืœื.
07:36
Now, the AI definitely was trained to recognize trucks in pictures.
145
456507
4762
ื”ื‘"ืž ื”ื•ื›ืฉืจื” ื‘ื”ื—ืœื˜ ืœื–ื”ื•ืช ืžืฉืื™ื•ืช ื‘ืชืžื•ื ื•ืช.
07:41
But what it looks like happened is
146
461293
2145
ืื‘ืœ ืžื” ืฉื›ื ืจืื” ืงืจื” ื”ื•ื,
07:43
the AI was trained to recognize trucks on highway driving,
147
463462
2931
ืฉื”ื‘"ืž ืื•ืžื ื” ืœื–ื”ื•ืช ืžืฉืื™ื•ืช ื‘ื›ื‘ื™ืฉ ื”ืžื”ื™ืจ,
07:46
where you would expect to see trucks from behind.
148
466417
2899
ื•ืฉื ื”ืฆื™ืคื™ื” ื”ื™ื ืœืจืื•ืช ืžืฉืื™ื•ืช ืžืื—ื•ืจ.
07:49
Trucks on the side is not supposed to happen on a highway,
149
469340
3420
ืžืฉืื™ื•ืช ื‘ืžื‘ื˜ ืฆื™ื“ื™ ืœื ืืžื•ืจื•ืช ืœื”ื™ืจืื•ืช ื‘ื›ื‘ื™ืฉ ืžื”ื™ืจ,
07:52
and so when the AI saw this truck,
150
472784
3455
ืื– ื›ืฉื‘"ืž ืจืืชื” ืืช ื”ืžืฉืื™ืช,
07:56
it looks like the AI recognized it as most likely to be a road sign
151
476263
4827
ื”ื™ื ื›ื ืจืื” ื–ื™ื”ืชื” ืื•ืชื” ื›ืฉืœื˜ ื“ืจื›ื™ื,
08:01
and therefore, safe to drive underneath.
152
481114
2273
ื›ื–ื” ืฉื‘ื˜ื•ื— ืœืขื‘ื•ืจ ืžืชื—ืชื™ื•.
08:04
Here's an AI misstep from a different field.
153
484114
2580
ื”ื ื” ื˜ืขื•ืช ืฉืœ ื‘"ืž ืžืชื—ื•ื ืฉื•ื ื”.
08:06
Amazon recently had to give up on a rรฉsumรฉ-sorting algorithm
154
486718
3460
ืœืื—ืจื•ื ื” "ืืžื–ื•ืŸ" ื ืืœืฆื” ืœื ื˜ื•ืฉ ืืœื’ื•ืจื™ืชื ืœืžื™ื•ืŸ ืงื•ืจื•ืช-ื—ื™ื™ื
08:10
that they were working on
155
490202
1220
ืฉืขืœื™ื• ื”ื ืขื‘ื“ื•,
08:11
when they discovered that the algorithm had learned to discriminate against women.
156
491446
3908
ื›ืฉื’ื™ืœื• ืฉื”ืืœื’ื•ืจื™ืชื ืœืžื“ ืœื”ืคืœื•ืช ื ื’ื“ ื ืฉื™ื.
08:15
What happened is they had trained it on example rรฉsumรฉs
157
495378
2716
ืžื” ืฉืงืจื” ื”ื•ื ืฉื”ื ืื™ืžื ื• ืื•ืชื• ืขื ื“ื’ื™ืžื•ืช ืฉืœ ืงื•ืจื•ืช-ื—ื™ื™ื
08:18
of people who they had hired in the past.
158
498118
2242
ืฉืœ ืื ืฉื™ื ืฉื”ื ื”ืขืกื™ืงื• ื‘ืขื‘ืจ.
08:20
And from these examples, the AI learned to avoid the rรฉsumรฉs of people
159
500384
4023
ื•ืžื”ื“ื•ื’ืžืื•ืช, ื”ื‘"ืž ืœืžื“ื” ืœื”ื™ืžื ืข ืžืงื•ืจื•ืช-ื—ื™ื™ื ืฉืœ ืื ืฉื™ื
08:24
who had gone to women's colleges
160
504431
2026
ืฉื”ืœื›ื• ืœืžื›ืœืœื•ืช ืฉืœ ื ืฉื™ื
08:26
or who had the word "women" somewhere in their resume,
161
506481
2806
ืื• ืฉื”ืžื™ืœื” "ื ืฉื™ื" ื”ื•ืคื™ืขื” ื‘ืงื•ืจื•ืช ื”ื—ื™ื™ื,
08:29
as in, "women's soccer team" or "Society of Women Engineers."
162
509311
4576
ื›ืžื• "ื ื‘ื—ืจืช ื›ื“ื•ืจื’ืœ ื ืฉื™ื" ืื• "ืื’ื•ื“ืช ื”ื ืฉื™ื ื”ืžื”ื ื“ืกื•ืช".
08:33
The AI didn't know that it wasn't supposed to copy this particular thing
163
513911
3974
ื”ื‘"ืž ืœื ื™ื“ืขื” ืฉื”ื™ื ืœื ืืžื•ืจื” ืœื”ืขืชื™ืง ืืช ื”ื“ื‘ืจ ื”ืžืกื•ื™ื ื”ื–ื”
08:37
that it had seen the humans do.
164
517909
1978
ืžื‘ื ื™ ื”ืื“ื.
08:39
And technically, it did what they asked it to do.
165
519911
3177
ื•ื˜ื›ื ื™ืช, ื”ื™ื ืขืฉืชื” ื›ืคื™ ืฉื”ืชื‘ืงืฉื”.
08:43
They just accidentally asked it to do the wrong thing.
166
523112
2797
ืืœื ืฉื‘ื™ืงืฉื• ืžืžื ื” ื‘ื˜ืขื•ืช ืœืขืฉื•ืช ืืช ื”ื“ื‘ืจ ื”ืœื-ื ื›ื•ืŸ.
08:46
And this happens all the time with AI.
167
526653
2895
ื•ื–ื” ืงื•ืจื” ื›ืœ ื”ื–ืžืŸ ืขื ื‘"ืž.
08:50
AI can be really destructive and not know it.
168
530120
3591
ื‘"ืž ื™ื›ื•ืœื” ืœื”ื™ื•ืช ื”ืจืกื ื™ืช ืžืื“ ืžื‘ืœื™ ืœื“ืขืช ื–ืืช.
08:53
So the AIs that recommend new content in Facebook, in YouTube,
169
533735
5078
ืื– ื”ื‘"ืž ืฉืžืžืœื™ืฆื•ืช ืขืœ ืชื›ื ื™ื ื‘"ืคื™ื™ืกื‘ื•ืง" ื•"ื™ื•-ื˜ื™ื•ื‘",
08:58
they're optimized to increase the number of clicks and views.
170
538837
3539
ืžืžื•ื˜ื‘ื•ืช ืœื”ื’ื“ื™ืœ ืืช ืžืกืคืจ ื”ืงืœื™ืงื™ื ื•ื”ืฆืคื™ื•ืช.
09:02
And unfortunately, one way that they have found of doing this
171
542400
3436
ืœืžืจื‘ื” ื”ืฆืขืจ, ืื—ืช ื”ื“ืจื›ื™ื ืฉื”ืŸ ืžืฆืื• ืœืขืฉื•ืช ื–ืืช
09:05
is to recommend the content of conspiracy theories or bigotry.
172
545860
4503
ื”ื™ื ืœื”ืžืœื™ืฅ ืขืœ ืชื›ื ื™ ืชื™ืื•ืจื™ื•ืช-ืงืฉืจ ืื• ื’ื–ืขื ื•ืช.
09:10
The AIs themselves don't have any concept of what this content actually is,
173
550902
5302
ืœื‘"ืž ืขืฆืžืŸ ืื™ืŸ ืฉื•ื ืžื•ืฉื’ ืžื”ื ื‘ืขืฆื ื”ืชื›ื ื™ื ื”ืืœื”,
09:16
and they don't have any concept of what the consequences might be
174
556228
3395
ื•ื’ื ืœื ืžื” ืขืœื•ืœื•ืช ืœื”ื™ื•ืช ื”ืชื•ืฆืื•ืช
09:19
of recommending this content.
175
559647
2109
ืฉืœ ื”ื”ืžืœืฆื” ืขืœ ืชื›ื ื™ื ื›ืืœื”.
09:22
So, when we're working with AI,
176
562296
2011
ืื– ื›ืฉืื ื• ืขื•ื‘ื“ื™ื ืขื ื‘"ืž,
09:24
it's up to us to avoid problems.
177
564331
4182
ื‘ืื—ืจื™ื•ืชื ื• ืœืžื ื•ืข ื‘ืขื™ื•ืช.
09:28
And avoiding things going wrong,
178
568537
2323
ื•ืžื ื™ืขื” ืžื“ื‘ืจื™ื ืœื”ืฉืชื‘ืฉ
09:30
that may come down to the age-old problem of communication,
179
570884
4526
ื™ื›ื•ืœื” ืœื”ืกืชื›ื ื‘ื‘ืขื™ื™ื” ืขืชื™ืงืช ื”ื™ื•ืžื™ืŸ ืฉืœ ืงืฆืจื™ื ื‘ืชืงืฉื•ืจืช.
09:35
where we as humans have to learn how to communicate with AI.
180
575434
3745
ื›ืฉืื ื•, ื‘ื ื™ ื”ืื“ื, ืฆืจื™ื›ื™ื ืœืœืžื•ื“ ืื™ืš ืœืชืงืฉืจ ืขื ื‘"ืž,
09:39
We have to learn what AI is capable of doing and what it's not,
181
579203
4039
ืขืœื™ื ื• ืœืœืžื•ื“ ืžื” ื”ื‘"ืž ืžืกื•ื’ืœืช ืœืขืฉื•ืช ื•ืžื” ืœื,
09:43
and to understand that, with its tiny little worm brain,
182
583266
3086
ื•ืœื”ื‘ื™ืŸ ืฉืขื ืžื•ื— ื”ืฉืœืฉื•ืœ ื”ื–ืขื™ืจ ืฉืœื”,
09:46
AI doesn't really understand what we're trying to ask it to do.
183
586376
4013
ื”ื‘"ืž ืœื ืžืžืฉ ืžื‘ื™ื ื” ืžื” ืื ื• ืžื ืกื™ื ืœื‘ืงืฉ ืžืžื ื” ืœืขืฉื•ืช.
09:51
So in other words, we have to be prepared to work with AI
184
591148
3321
ื‘ืžืœื™ื ืื—ืจื•ืช, ืขืœื™ื ื• ืœื”ื™ื•ืช ืžื•ื›ื ื™ื ืœืขื‘ื•ื“ ืขื ื‘"ืž
09:54
that's not the super-competent, all-knowing AI of science fiction.
185
594493
5258
ืฉืื™ื ื ื” ื›ืœ-ื™ื›ื•ืœื” ื•ื™ื•ื“ืขืช-ื›ืœ ื›ืžื• ื‘ืžื“ืข ื”ื‘ื“ื™ื•ื ื™:
09:59
We have to be prepared to work with an AI
186
599775
2862
ืขืœื™ื ื• ืœื”ื™ื•ืช ืžื•ื›ื ื™ื ืœืขื‘ื•ื“ ืขื ื”ื‘"ืž ืฉื™ืฉ ืœื ื• ื‘ื”ื•ื•ื”.
10:02
that's the one that we actually have in the present day.
187
602661
2938
10:05
And present-day AI is plenty weird enough.
188
605623
4205
ื•ื”ื‘"ืž ื‘ื”ื•ื•ื” ื”ื™ื ืžืกืคื™ืง ืžื•ื–ืจื”.
10:09
Thank you.
189
609852
1190
ืชื•ื“ื” ืœื›ื.
10:11
(Applause)
190
611066
5225
(ืžื—ื™ืื•ืช ื›ืคื™ื™ื)
ืขืœ ืืชืจ ื–ื”

ืืชืจ ื–ื” ื™ืฆื™ื’ ื‘ืคื ื™ื›ื ืกืจื˜ื•ื ื™ YouTube ื”ืžื•ืขื™ืœื™ื ืœืœื™ืžื•ื“ ืื ื’ืœื™ืช. ืชื•ื›ืœื• ืœืจืื•ืช ืฉื™ืขื•ืจื™ ืื ื’ืœื™ืช ื”ืžื•ืขื‘ืจื™ื ืขืœ ื™ื“ื™ ืžื•ืจื™ื ืžื”ืฉื•ืจื” ื”ืจืืฉื•ื ื” ืžืจื—ื‘ื™ ื”ืขื•ืœื. ืœื—ืฅ ืคืขืžื™ื™ื ืขืœ ื”ื›ืชื•ื‘ื™ื•ืช ื‘ืื ื’ืœื™ืช ื”ืžื•ืฆื’ื•ืช ื‘ื›ืœ ื“ืฃ ื•ื™ื“ืื• ื›ื“ื™ ืœื”ืคืขื™ืœ ืืช ื”ืกืจื˜ื•ืŸ ืžืฉื. ื”ื›ืชื•ื‘ื™ื•ืช ื’ื•ืœืœื•ืช ื‘ืกื ื›ืจื•ืŸ ืขื ื”ืคืขืœืช ื”ื•ื•ื™ื“ืื•. ืื ื™ืฉ ืœืš ื”ืขืจื•ืช ืื• ื‘ืงืฉื•ืช, ืื ื ืฆื•ืจ ืื™ืชื ื• ืงืฉืจ ื‘ืืžืฆืขื•ืช ื˜ื•ืคืก ื™ืฆื™ืจืช ืงืฉืจ ื–ื”.

https://forms.gle/WvT1wiN1qDtmnspy7