3 principles for creating safer AI | Stuart Russell

140,348 views ใƒป 2017-06-06

TED


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

ืžืชืจื’ื: Igal Opendik ืžื‘ืงืจ: Ido Dekkers
00:12
This is Lee Sedol.
0
12532
1552
ื–ื”ื• ืœื™ ืกื™ื“ื•ืœ.
00:14
Lee Sedol is one of the world's greatest Go players,
1
14108
3997
ื”ื•ื ืื—ื“ ืžื’ื“ื•ืœื™ ืฉื—ืงื ื™ ื’ื• ื‘ืขื•ืœื.
00:18
and he's having what my friends in Silicon Valley call
2
18129
2885
ื›ืืŸ ื”ื•ื ื—ื•ื•ื” ืืช ื”ืจื’ืข ืฉื—ื‘ืจื™ื™ ืžืขืžืง ื”ืกื™ืœื™ืงื•ืŸ ืžื›ื ื™ื
00:21
a "Holy Cow" moment --
3
21038
1510
"ื–ื” ื”ื–ื•ื™!" -
00:22
(Laughter)
4
22572
1073
(ืฆื—ื•ืง)
00:23
a moment where we realize
5
23669
2188
ื”ืจื’ืข ื‘ื• ืื ื• ืžื‘ื™ื ื™ื
00:25
that AI is actually progressing a lot faster than we expected.
6
25881
3296
ืฉื”ืชืคืชื—ื•ืชื” ืฉืœ ื”ื‘''ืž (ื‘ื™ื ื” ืžืœืื›ืชื™ืช) ืžืชืงื“ืžืช ื”ืจื‘ื” ื™ื•ืชืจ ืžื”ืจ ืžืฉืฆื™ืคื™ื ื•.
00:29
So humans have lost on the Go board. What about the real world?
7
29974
3047
ืื– ื‘ื ื™ ื”ืื ื•ืฉ ื”ืคืกื™ื“ื• ื‘ืžืฉื—ืง ื’ื•. ืื‘ืœ ืžื” ืขื ื”ืขื•ืœื ื”ืืžื™ืชื™?
00:33
Well, the real world is much bigger,
8
33045
2100
ื•ื‘ื›ืŸ, ื”ืขื•ืœื ื”ืืžื™ืชื™ ื”ืจื‘ื” ื™ื•ืชืจ ื’ื“ื•ืœ,
00:35
much more complicated than the Go board.
9
35169
2249
ื”ืจื‘ื” ื™ื•ืชืจ ืžื•ืจื›ื‘ ืžืžืฉื—ืง ื’ื•.
00:37
It's a lot less visible,
10
37442
1819
ื–ื” ืคื—ื•ืช ื ื’ืœื” ืœืขื™ืŸ,
00:39
but it's still a decision problem.
11
39285
2038
ืื‘ืœ ื–ื• ืขื“ื™ื™ืŸ ื‘ืขื™ื™ืช ืงื‘ืœืช ื”ื—ืœื˜ื•ืช.
00:42
And if we think about some of the technologies
12
42768
2321
ื•ืื ื—ื•ืฉื‘ื™ื ืขืœ ื›ืžื” ื˜ื›ื ื•ืœื•ื’ื™ื•ืช
00:45
that are coming down the pike ...
13
45113
1749
ืฉืžืชืžืžืฉื•ืช ื›ื ื’ื“ ืขื™ื ื™ื™ื ื•...
00:47
Noriko [Arai] mentioned that reading is not yet happening in machines,
14
47558
4335
ื ื•ืจื™ืงื• [ืืจืื™] ื”ื–ื›ื™ืจื” ืฉืžื›ื•ื ื•ืช ืขื“ื™ื™ืŸ ืœื ื™ื•ื“ืขื•ืช ืœืงืจื•ื,
00:51
at least with understanding.
15
51917
1500
ืœืคื—ื•ืช ืœืงืจื•ื ื•ืœื”ื‘ื™ืŸ.
00:53
But that will happen,
16
53441
1536
ืื‘ืœ, ื–ื” ื™ืงืจื”.
00:55
and when that happens,
17
55001
1771
ื•ื›ืืฉืจ ื–ื” ื›ืŸ ื™ืงืจื”,
00:56
very soon afterwards,
18
56796
1187
ืขื“ ืžื”ืจื”
00:58
machines will have read everything that the human race has ever written.
19
58007
4572
ื”ืŸ ืชืงืจืื ื” ืืช ื›ืœ ืžื” ืฉื”ืื ื•ืฉื•ืช ื›ืชื‘ื” ืื™ ืคืขื.
01:03
And that will enable machines,
20
63670
2030
ื–ื” ื™ืงื ื” ืœืžื›ื•ื ื•ืช ื™ื›ื•ืœืช ื—ื“ืฉื”,
01:05
along with the ability to look further ahead than humans can,
21
65724
2920
ืœืฆื“ ื™ื›ื•ืœืช ื”ื—ื™ื–ื•ื™ ืžืขื‘ืจ ืœืžื” ืฉื‘ื ื™ ื”ืื ื•ืฉ ืžืกื•ื’ืœื™ื ืœื—ื–ื•ืช,
01:08
as we've already seen in Go,
22
68668
1680
ื›ืคื™ ืฉื ื•ื›ื—ื ื• ืœื“ืขืช ื‘ืžืฉื—ืง ื’ื•,
01:10
if they also have access to more information,
23
70372
2164
ืื ืชืงื‘ืœื ื” ื’ื™ืฉื” ืœื™ื•ืชืจ ืžื™ื“ืข,
01:12
they'll be able to make better decisions in the real world than we can.
24
72560
4268
ื”ืŸ ืชื•ื›ืœื ื” ืœืงื‘ืœ ื”ื—ืœื˜ื•ืช ื˜ื•ื‘ื•ืช ื™ื•ืชืจ ืžืื™ืชื ื• ื‘ืขื•ืœื ื”ืืžื™ืชื™.
01:18
So is that a good thing?
25
78612
1606
ื”ืื ื–ื” ื˜ื•ื‘ ืœื ื•?
01:21
Well, I hope so.
26
81718
2232
ื•ื‘ื›ืŸ, ืื ื™ ืžืงื•ื•ื” ืฉื›ืŸ.
01:26
Our entire civilization, everything that we value,
27
86514
3255
ื”ืฆื™ื‘ื™ืœื™ื–ืฆื™ื” ืฉืœื ื• ืขืœ ื›ืœ ืขืจื›ื™ื”,
01:29
is based on our intelligence.
28
89793
2068
ืžื‘ื•ืกืกืช ืขืœ ื”ืชื‘ื•ื ื” ืฉืœื ื•.
01:31
And if we had access to a lot more intelligence,
29
91885
3694
ื•ืœื• ื”ื™ืชื” ืœื ื• ื’ื™ืฉื” ืœืชื‘ื•ื ื” ืจื‘ื” ื™ื•ืชืจ,
01:35
then there's really no limit to what the human race can do.
30
95603
3302
ืื–ื™ ืœื ื™ื”ื™ื” ื’ื‘ื•ืœ ืœืžื” ืฉื”ืื ื•ืฉื•ืช ืชื•ื›ืœ ืœืขืฉื•ืช.
01:40
And I think this could be, as some people have described it,
31
100485
3325
ื•ืื ื™ ืกื‘ื•ืจ ื›ื™ ื–ื” ื”ื™ื” ื™ื›ื•ืœ ืœื”ื™ื•ืช, ื›ืคื™ ืฉืื ืฉื™ื ืžืกื•ื™ืžื™ื ืชื™ืืจื• ื–ืืช,
01:43
the biggest event in human history.
32
103834
2016
ื”ืืจื•ืข ื”ื’ื“ื•ืœ ื‘ืชื•ืœื“ื•ืช ื”ืื ื•ืฉื•ืช.
01:48
So why are people saying things like this,
33
108485
2829
ืื– ืžื“ื•ืข ืื ืฉื™ื ืื•ืžืจื™ื ื“ื‘ืจื™ื ื›ื’ื•ืŸ,
01:51
that AI might spell the end of the human race?
34
111338
2876
ื‘''ืž ืขืœื•ืœื” ืœื’ืจื•ื ืœืกื•ืฃ ื”ืื ื•ืฉื•ืช?
01:55
Is this a new thing?
35
115258
1659
ื”ืื ื–ื” ื—ื“ืฉ ืœื ื•?
01:56
Is it just Elon Musk and Bill Gates and Stephen Hawking?
36
116941
4110
ื”ืื ืืœื” ืจืง ืืœื•ืŸ ืžืืกืง, ื‘ื™ืœ ื’ื™ื™ื˜ืก ื•ืกื˜ื™ื‘ืŸ ื”ื•ืงื™ื ื’?
02:01
Actually, no. This idea has been around for a while.
37
121773
3262
ืœื. ืจืขื™ื•ืŸ ื–ื” ื›ื‘ืจ ืงื™ื™ื ื–ืžืŸ ืžื”.
02:05
Here's a quotation:
38
125059
1962
ื”ืจื™ ื”ืฆื™ื˜ื•ื˜:
02:07
"Even if we could keep the machines in a subservient position,
39
127045
4350
"ืืคื™ืœื• ืื ื”ื™ื™ื ื• ืžืกื•ื’ืœื™ื ืœืฉืœื•ื˜ ื‘ืžื›ื•ื ื•ืช ื›ื‘ืžืฉืจืชื™ื ื‘ืœื‘ื“,
02:11
for instance, by turning off the power at strategic moments" --
40
131419
2984
ืœืžืฉืœ ืขืœ ื™ื“ื™ ื›ื™ื‘ื•ื™ ืืกืคืงืช ื—ืฉืžืœ ื‘ืจื’ืขื™ื ืงืจื™ื˜ื™ื™ื" --
02:14
and I'll come back to that "turning off the power" idea later on --
41
134427
3237
ืื—ื–ื•ืจ ืœื ื•ืฉื "ื›ื™ื‘ื•ื™ ื”ื—ืฉืžืœ" ื‘ื”ืžืฉืš --
02:17
"we should, as a species, feel greatly humbled."
42
137688
2804
"ืื ื—ื ื• ื›ืžื™ืŸ ืฆืจื™ื›ื™ื ืœื”ืจื’ื™ืฉ ืขื ื•ื•ื” ื’ื“ื•ืœื”."
02:21
So who said this? This is Alan Turing in 1951.
43
141997
3448
ืžื™ ืืžืจ ื–ืืช? ืืœืŸ ื˜ื•ืจื™ื ื’ ื‘-1951.
02:26
Alan Turing, as you know, is the father of computer science
44
146120
2763
ื›ื™ื“ื•ืข ืœื›ื, ืืœืŸ ื˜ื™ื•ืจื™ื ื’ ื”ื•ื ืื‘ื™ ืžื“ืข ื”ืžื—ืฉื‘
02:28
and in many ways, the father of AI as well.
45
148907
3048
ื•ื‘ืžื•ื‘ื ื™ื ืจื‘ื™ื, ื’ื ืื‘ื™ ื”ื‘''ืž.
02:33
So if we think about this problem,
46
153059
1882
ืื ื—ื•ืฉื‘ื™ื ืขืœ ื”ื‘ืขื™ื™ืชื™ื•ืช ืฉื‘ื™ืฆื™ืจืช
02:34
the problem of creating something more intelligent than your own species,
47
154965
3787
ืžืฉื”ื• ืฉื—ื›ื ื™ื•ืชืจ ืžืžื™ื ืš ืฉืœืš,
02:38
we might call this "the gorilla problem,"
48
158776
2622
ื ื™ืชืŸ ืœื›ื ื•ืช ื–ืืช "ื‘ืขื™ื™ืช ื”ื’ื•ืจื™ืœื”",
02:42
because gorillas' ancestors did this a few million years ago,
49
162165
3750
ื”ื™ื•ืช ื•ืื‘ื•ืช ืื‘ื•ืชื™ื”ืŸ ืฉืœ ื”ื’ื•ืจื™ืœื•ืช ืขืฉื• ื–ืืช ืœืคื ื™ ืžื™ืœื™ื•ื ื™ ืฉื ื”,
02:45
and now we can ask the gorillas:
50
165939
1745
ื•ื›ื™ื•ื ื ื•ื›ืœ ืœืฉืื•ืœ ืืช ื”ื’ื•ืจื™ืœื•ืช:
02:48
Was this a good idea?
51
168572
1160
ื”ืื ื–ื” ื”ื™ื” ืจืขื™ื•ืŸ ื˜ื•ื‘?
02:49
So here they are having a meeting to discuss whether it was a good idea,
52
169756
3530
ื”ื™ื ื” ื”ื ืขื•ืจื›ื™ื ื“ื™ื•ืŸ ืขืœ ื˜ื™ื‘ ื”ื“ื‘ืจ
02:53
and after a little while, they conclude, no,
53
173310
3346
ื•ืื—ืจื™ ื–ืžืŸ ืžื” ื”ื ืงื•ื‘ืขื™ื ืฉื–ื”- ืœื!
02:56
this was a terrible idea.
54
176680
1345
ื–ื” ื”ื™ื” ืจืขื™ื•ืŸ ื ื•ืจืื™.
02:58
Our species is in dire straits.
55
178049
1782
ืžื™ืŸ ื”ื’ื•ืจื™ืœื•ืช ื‘ืžืฆื•ืงื” ืงืฉื”.
03:00
In fact, you can see the existential sadness in their eyes.
56
180358
4263
ืœืžืขืฉื”, ืืชื ื™ื›ื•ืœื™ื ืœืจืื•ืช ืืช ื”ืขืฆื‘ ื”ืงื™ื•ืžื™ ื‘ืขื™ื ื™ื”ืŸ.
03:04
(Laughter)
57
184645
1640
(ืฆื—ื•ืง)
03:06
So this queasy feeling that making something smarter than your own species
58
186309
4840
ืœื’ื‘ื™ ื”ื”ืจื’ืฉื” ื”ืžื‘ื—ื™ืœื” ื”ื–ืืช ืฉื™ืฆื™ืจืช ืžืฉื”ื• ืฉืขื•ืœื” ืขืœื™ืš ื‘ื—ื•ื›ืžืชื•
03:11
is maybe not a good idea --
59
191173
2365
ืื™ื ื ื• ืจืขื™ื•ืŸ ื˜ื•ื‘ --
03:14
what can we do about that?
60
194308
1491
ืžื” ืขื•ืฉื™ื ืื™ืชื”?
03:15
Well, really nothing, except stop doing AI,
61
195823
4767
ื”ืืžืช ื”ื™ื ืฉื›ืœื•ื, ื—ื•ืฅ ืžืขืฆื™ืจืช ืคื™ืชื•ื— ื‘''ืž.
03:20
and because of all the benefits that I mentioned
62
200614
2510
ื‘ื’ืœืœ ื›ืœ ื”ื™ืชืจื•ื ื•ืช ืฉื”ื–ื›ืจืชื™
03:23
and because I'm an AI researcher,
63
203148
1716
ื•ื‘ื’ืœืœ ืฉืื ื™ ื—ื•ืงืจ ื‘''ืž,
03:24
I'm not having that.
64
204888
1791
ืื ื™ ืœื ืžืงื‘ืœ ืืช ืืคืฉืจื•ืช ื”ืขืฆื™ืจื”.
03:27
I actually want to be able to keep doing AI.
65
207103
2468
ืื ื™ ื›ืŸ ืžืขื•ื ื™ื™ืŸ ืœื”ืžืฉื™ืš ื‘ืคื™ืชื•ื— ื‘''ืž.
03:30
So we actually need to nail down the problem a bit more.
66
210435
2678
ืœื›ืŸ, ืื ื—ื ื• ื—ื™ื™ื‘ื™ื ืœื”ื’ื“ื™ืจ ืืช ื”ื‘ืขื™ื” ื˜ื•ื‘ ื™ื•ืชืจ.
03:33
What exactly is the problem?
67
213137
1371
ืžื”ื™ ืžื”ื•ืช ื”ื‘ืขื™ื” ื‘ื“ื™ื•ืง?
03:34
Why is better AI possibly a catastrophe?
68
214532
3246
ืœืžื” ื‘''ืž ื˜ื•ื‘ื” ื™ื•ืชืจ ืขืœื•ืœื” ืœื”ื•ื•ืช ืืกื•ืŸ?
03:39
So here's another quotation:
69
219218
1498
ืื– ื”ื™ื ื” ืœื›ื ืขื•ื“ ืฆื™ื˜ื•ื˜:
03:41
"We had better be quite sure that the purpose put into the machine
70
221755
3335
"ืขื“ื™ืฃ ืฉื ื”ื™ื” ื‘ื˜ื•ื—ื™ื ืฉื”ื™ืขื•ื“ ืฉืื ื—ื ื• ืžื˜ืžื™ืขื™ื ื‘ืžื›ื•ื ื”
03:45
is the purpose which we really desire."
71
225114
2298
ื–ื” ืื•ืชื• ื”ื™ืขื•ื“ ืฉืื ื—ื ื• ื—ืคืฆื™ื ื‘ื• ื‘ืืžืช ".
03:48
This was said by Norbert Wiener in 1960,
72
228102
3498
ื–ื” ื ืืžืจ ืขืœ ื™ื“ื™ ื ื•ืจื‘ืจื˜ ื•ื•ื™ื™ื ืจ ื‘-1960
03:51
shortly after he watched one of the very early learning systems
73
231624
4002
ื–ืžืŸ ืงืฆืจ ืื—ืจื™ ืฉืฆืคื” ื‘ืื—ืช ืžืžืขืจื›ื•ืช ื”ืœืžื™ื“ื” ื”ืžื•ืงื“ืžื•ืช
03:55
learn to play checkers better than its creator.
74
235650
2583
ืœื•ืžื“ืช ืœืฉื—ืง ื“ืžืงื” ื˜ื•ื‘ ืžื™ื•ืฆืจื”.
04:00
But this could equally have been said
75
240422
2683
ืื™ืžืจื” ื–ื• ื”ื™ืชื” ื’ื ื™ื›ื•ืœื” ืœื”ื™ืืžืจ
04:03
by King Midas.
76
243129
1167
ืขืœ ื™ื“ื™ ื”ืžืœืš ืžื™ื“ืืก.
04:04
King Midas said, "I want everything I touch to turn to gold,"
77
244903
3134
ื”ืžืœืš ืžื™ื“ืืก ืืžืจ, "ืื ื™ ืจื•ืฆื” ืฉื›ืœ ื“ื‘ืจ ืฉืื’ืข ื‘ื• ื™ื”ืคื•ืš ืœื–ื”ื‘".
04:08
and he got exactly what he asked for.
78
248061
2473
ื•ื”ื•ื ืงื™ื‘ืœ ืืช ืžื‘ื•ืงืฉื• ื‘ื“ื™ื•ืง.
04:10
That was the purpose that he put into the machine,
79
250558
2751
ื–ื” ื”ื™ืขื•ื“ ืื•ืชื• ื”ื•ื ื”ื˜ืžื™ืข ื‘ืžื›ื•ื ื”,
04:13
so to speak,
80
253333
1450
ื›ื‘ื™ื›ื•ืœ,
04:14
and then his food and his drink and his relatives turned to gold
81
254807
3444
ื•ื›ืœ ื”ืžื–ื•ืŸ ื•ื”ืžืฉืงื” ืฉืœื• ื•ื›ืœ ืงืจื•ื‘ื™ื• ื”ืคื›ื• ืœื–ื”ื‘
04:18
and he died in misery and starvation.
82
258275
2281
ื•ื”ื•ื ืžืช ืื•ืžืœืœ ื•ืžื•ืจืขื‘.
04:22
So we'll call this "the King Midas problem"
83
262264
2341
ื ืงืจื ืœื›ืš "ื‘ืขื™ืช ืžืœืš ืžื™ื“ืืก".
04:24
of stating an objective which is not, in fact,
84
264629
3305
ืžืชืŸ ืžืฉื™ืžื” ืืฉืจ ืœื ืขื•ืœื”
04:27
truly aligned with what we want.
85
267958
2413
ื‘ืงื ื” ืื—ื“ ืขื ืžื” ืฉืื ื• ืจื•ืฆื™ื.
04:30
In modern terms, we call this "the value alignment problem."
86
270395
3253
ื‘ืžื•ืฉื’ื™ื ืžื•ื“ืจื ื™ื™ื ืื ื• ืงื•ืจืื™ื ืœื–ื” "ื‘ืขื™ื™ืช ืชืื•ื ืขืจื›ื™ื".
04:36
Putting in the wrong objective is not the only part of the problem.
87
276867
3485
ืžืชืŸ ืžื˜ืจื” ืฉื’ื•ื™ื” ืื™ื ื” ื”ืžืจื›ื™ื‘ ื”ื™ื—ื™ื“ ืฉืœ ื”ื‘ืขื™ื”.
04:40
There's another part.
88
280376
1152
ื™ืฉ ืžืจื›ื™ื‘ ื ื•ืกืฃ.
04:41
If you put an objective into a machine,
89
281980
1943
ืื ืืชื ืžื˜ืžื™ืขื™ื ืžืฉื™ืžื” ื‘ืžื›ื•ื ื”
04:43
even something as simple as, "Fetch the coffee,"
90
283947
2448
ืืคื™ืœื• ืžืฉื”ื• ืคืฉื•ื˜ ื›ืžื• "ืชื‘ื™ืื™ ืœื™ ืงืคื”",
04:47
the machine says to itself,
91
287728
1841
ื”ืžื›ื•ื ื” ืื•ืžืจืช ืœืขืฆืžื”,
04:50
"Well, how might I fail to fetch the coffee?
92
290553
2623
"ืžื” ืขืœื•ืœ ืœืžื ื•ืข ืžืžื ื™ ืœื”ื‘ื™ื ืืช ื”ืงืคื”?
04:53
Someone might switch me off.
93
293200
1580
ืžื™ืฉื”ื• ืขืœื•ืœ ืœื ืชืง ืื•ืชื™ ืžื”ื—ืฉืžืœ.
04:55
OK, I have to take steps to prevent that.
94
295465
2387
ืœื›ืŸ ืื ื™ ื—ื™ื™ื‘ืช ืœื ืงื•ื˜ ื‘ืฆืขื“ื™ื ืฉื™ืžื ืขื• ื–ืืช.
04:57
I will disable my 'off' switch.
95
297876
1906
ืื ื˜ืจืœ ืืช ืžืชื’ ื”ื›ื™ื‘ื•ื™ ืฉืœื™.
05:00
I will do anything to defend myself against interference
96
300354
2959
ืืขืฉื” ื”ื›ืœ ื›ื“ื™ ืœื”ื’ืŸ ืขืœ ืขืฆืžื™ ื›ื ื’ื“ ืžื” ืฉื™ืžื ืข ืžืžื ื™
05:03
with this objective that I have been given."
97
303337
2629
ืœื‘ืฆืข ืืช ื”ืžืฉื™ืžื” ืฉื ื™ืชื ื” ืœื™".
05:05
So this single-minded pursuit
98
305990
2012
ืœื›ืŸ, ื—ืชื™ืจื” ื—ื“-ื›ื™ื•ื•ื ื™ืช ืœืžืฉื™ืžื”
05:09
in a very defensive mode of an objective that is, in fact,
99
309033
2945
ื‘ืชืฆื•ืจื” ื”ื’ื ืชื™ืช ื”ืžื‘ื˜ื™ื—ื” ืืช ื‘ื™ืฆื•ืข ื”ืžืฉื™ืžื”, ืืฉืจ
05:12
not aligned with the true objectives of the human race --
100
312002
2814
ืื™ื ื ื” ืชื•ืืžืช ืžืฉื™ืžื•ืช ืืžื™ืชื™ื•ืช ืฉืœ ื”ืื ื•ืฉื•ืช --
05:15
that's the problem that we face.
101
315942
1862
ื–ื• ื”ื‘ืขื™ื” ืืชื” ืื ื• ืžืชืžื•ื“ื“ื™ื.
05:18
And in fact, that's the high-value takeaway from this talk.
102
318827
4767
ื–ื• ื”ืชื•ื‘ื ื” ื‘ืขืœืช ืขืจืš ืื•ืชื” ืื ื• ื ืงื— ืžื”ืจืฆืื” ื–ื•.
05:23
If you want to remember one thing,
103
323618
2055
ืื ืชืจืฆื• ืœื–ื›ื•ืจ ื“ื‘ืจ ืื—ื“ --
05:25
it's that you can't fetch the coffee if you're dead.
104
325697
2675
-- ืœื ื ื™ืชืŸ ืœื”ื‘ื™ื ืืช ื”ืงืคื” ืื ืืชื” ืžืช.
05:28
(Laughter)
105
328396
1061
(ืฆื—ื•ืง)
05:29
It's very simple. Just remember that. Repeat it to yourself three times a day.
106
329481
3829
ื–ื” ืคืฉื•ื˜ ืžืื•ื“. ืจืง ืชื–ื›ืจื• ืืช ื–ื”. ืฉื ื ื• ื–ืืช ืœืขืฆืžื›ื ืฉืœื•ืฉ ืคืขืžื™ื ื‘ื™ื•ื.
05:33
(Laughter)
107
333334
1821
(ืฆื—ื•ืง)
05:35
And in fact, this is exactly the plot
108
335179
2754
ื–ื•ื”ื™ ื”ืขืœื™ืœื” ื”ืžื“ื•ื™ืงืช ืฉืœ
05:37
of "2001: [A Space Odyssey]"
109
337957
2648
"2001: ืื•ื“ื™ืกืื” ื‘ื—ืœืœ"
05:41
HAL has an objective, a mission,
110
341046
2090
ืœ-HAL ื™ืฉ ืžืฉื™ืžื”,
05:43
which is not aligned with the objectives of the humans,
111
343160
3732
ืืฉืจ ืื™ื ื ื” ืžืชื•ืืžืช ืขื ืžืฉื™ืžื•ืช ืฉืœ ื‘ื ื™ ื”ืื ื•ืฉ,
05:46
and that leads to this conflict.
112
346916
1810
ื•ื–ื” ืžื•ื‘ื™ืœ ืœืงื•ื ืคืœื™ืงื˜.
05:49
Now fortunately, HAL is not superintelligent.
113
349314
2969
ืœืžืจื‘ื” ื”ืžื–ืœ HAL ืื™ื ื ื• ื‘ืขืœ ืชื‘ื•ื ืช-ืขืœ.
05:52
He's pretty smart, but eventually Dave outwits him
114
352307
3587
ื”ื•ื ื“ื™ ื—ื›ื, ืื‘ืœ ืœื‘ืกื•ืฃ ื“ื™ื™ื‘ ืžืฆืœื™ื— ืœื”ืขืจื™ื ืขืœื™ื•
05:55
and manages to switch him off.
115
355918
1849
ื•ืžืฆืœื™ื— ืœื›ื‘ื•ืช ืื•ืชื•.
06:01
But we might not be so lucky.
116
361648
1619
ืื‘ืœ ืื ื—ื ื• ืขืœื•ืœื™ื ืœื ืœื”ื™ื•ืช ื‘ืจื™ ืžื–ืœ ื‘ืื•ืชื” ืžื™ื“ื”.
06:08
So what are we going to do?
117
368013
1592
ืื– ืžื” ื ืขืฉื”?
06:12
I'm trying to redefine AI
118
372191
2601
ืื ื™ ืžื ืกื” ืœื”ื’ื“ื™ืจ ืืช ื”ื‘''ืž ื‘ืžื•ืฉื’ื™ื ืื—ืจื™ื
06:14
to get away from this classical notion
119
374816
2061
ื›ื“ื™ ืœื”ืชื ืชืง ืžืŸ ื”ื”ืขืงืจื•ืŸ ื”ืงืœืืกื™
06:16
of machines that intelligently pursue objectives.
120
376901
4567
ืœืคื™ื• ืžื›ื•ื ื•ืช ืฉื•ืืคื•ืช ืœื‘ืฆืข ืžืฉื™ืžื•ืช ื‘ืฆื•ืจื” ืชื‘ื•ื ืชื™ืช.
06:22
There are three principles involved.
121
382532
1798
ื™ืฉื ื ืฉืœื•ืฉื” ืขืงืจื•ื ื•ืช ื‘ื‘ืกื™ืก ื”ืขื ื™ื™ืŸ.
06:24
The first one is a principle of altruism, if you like,
122
384354
3289
ื”ืจืืฉื•ืŸ ื”ื•ื, ืื ืชืจืฆื•, ืขืงืจื•ืŸ ื”ืืœื˜ืจื•ื™ื–ื --
06:27
that the robot's only objective
123
387667
3262
-- ื”ืžืฉื™ืžื” ื”ื™ื—ื™ื“ื” ืฉืœ ื”ืจื•ื‘ื•ื˜ื™ื ื”ื™ื
06:30
is to maximize the realization of human objectives,
124
390953
4246
ืžื™ืงืกื•ื ื”ื’ืฉืžืชืŸ ืฉืœ ื”ืžื˜ืจื•ืช ืฉืœ ื‘ื ื™ ืื ื•ืฉ,
06:35
of human values.
125
395223
1390
ืฉืœ ืขืจื›ื™ ื‘ื ื™ ืื ื•ืฉ.
06:36
And by values here I don't mean touchy-feely, goody-goody values.
126
396637
3330
ื‘ืื•ืžืจื™ "ืขืจื›ื™ื" ืื™ื ื ื™ ืžืชื›ื•ื•ืŸ ื›ืืŸ ืœืขืจื›ื™ื ื ืฉื’ื‘ื™ื ืฉืœ ื™ืคื™ ื ืคืฉ.
06:39
I just mean whatever it is that the human would prefer
127
399991
3787
ืื ื™ ืคืฉื•ื˜ ืžืชื›ื•ื•ืŸ ืœื›ืœ ืžื” ืฉื‘ื ื™ ืื ื•ืฉ
06:43
their life to be like.
128
403802
1343
ืžืขื“ื™ืคื™ื ื‘ื—ื™ื™ื”ื.
06:47
And so this actually violates Asimov's law
129
407184
2309
ื–ื” ืœืžืขืฉื” ื ื•ื’ื“ ืœื—ื•ืง ืืกื™ืžื•ื‘
06:49
that the robot has to protect its own existence.
130
409517
2329
ืœืคื™ื• ืจื•ื‘ื•ื˜ ื—ื™ื™ื‘ ืœื”ื’ืŸ ืขืœ ืงื™ื•ืžื•.
06:51
It has no interest in preserving its existence whatsoever.
131
411870
3723
ืื™ืŸ ืœื• ืขื ื™ื™ืŸ ื‘ืฉื™ืžื•ืจ ืงื™ื•ืžื• ื‘ื›ืœืœ.
06:57
The second law is a law of humility, if you like.
132
417240
3768
ื”ืขืงืจื•ืŸ ื”ืฉื ื™, ืื ืชืจืฆื•, ื”ื•ื ืขืงืจื•ืŸ ื”ืขื ื•ื•ื”.
07:01
And this turns out to be really important to make robots safe.
133
421794
3743
ืžืชื‘ืจืจ ืฉื”ื•ื ื—ืฉื•ื‘ ืžืื•ื“ ืœื™ืฆื™ืจืช ืจื•ื‘ื•ื˜ื™ื ื‘ื˜ื•ื—ื™ื.
07:05
It says that the robot does not know
134
425561
3142
ืœืคื™ื• ืจื•ื‘ื•ื˜ ืื™ื ื ื• ื™ื•ื“ืข
07:08
what those human values are,
135
428727
2028
ืžื”ื ืขืจื›ื™ ื‘ื ื™ ื”ืื ื•ืฉ.
07:10
so it has to maximize them, but it doesn't know what they are.
136
430779
3178
ื”ื•ื ื—ื™ื™ื‘ ืœืžืžืฉ ืื•ืชื ืขืœ ื”ืฆื“ ื”ื˜ื•ื‘ ื‘ื™ื•ืชืจ, ืื‘ืœ ื”ื•ื ืœื ื™ื•ื“ืข ืžื” ื”ื.
07:15
And that avoids this problem of single-minded pursuit
137
435074
2626
ื–ื” ืขื•ืงืฃ ืืช ื‘ืขื™ื™ืช ื—ืชื™ืจื” ื—ื“-ื›ื™ื•ื•ื ื™ืช
07:17
of an objective.
138
437724
1212
ืœืžืฉื™ืžื”.
07:18
This uncertainty turns out to be crucial.
139
438960
2172
ืื™ ื”ื•ื•ื“ืื•ืช ื”ื–ืืช ื”ื•ืคื›ืช ืœืขื ื™ื™ืŸ ืžื›ืจื™ืข.
07:21
Now, in order to be useful to us,
140
441546
1639
ื‘ื›ื“ื™ ืฉืจื•ื‘ื•ื˜ ื™ื”ื™ื” ืžื•ืขื™ืœ ืœื ื•
07:23
it has to have some idea of what we want.
141
443209
2731
ืขืœื™ื• ืœื“ืขืช ืžื” ืื ื• ืจื•ืฆื™ื, ื‘ืจืžื” ื›ืœืฉื”ื™.
07:27
It obtains that information primarily by observation of human choices,
142
447043
5427
ื”ื•ื ืžืงื‘ืœ ืžื™ื“ืข ื–ื” ื‘ืขื™ืงืจ ืžืฆืคื™ื” ื‘ื‘ื—ื™ืจื•ืช ืื ื•ืฉื™ื•ืช,
07:32
so our own choices reveal information
143
452494
2801
ื›ืš ืฉื”ื‘ื—ื™ืจื•ืช ืฉืœื ื• ืžืœืžื“ื•ืช ืžื™ื“ืข
07:35
about what it is that we prefer our lives to be like.
144
455319
3300
ืขืœ ื›ื™ืฆื“ ืื ื• ืžืขื“ื™ืคื™ื ืฉื—ื™ื™ื ื• ื™ื”ื™ื•.
07:40
So those are the three principles.
145
460452
1683
ืืœื” ื”ื ืฉืœื•ืฉืช ื”ืขืงืจื•ื ื•ืช.
07:42
Let's see how that applies to this question of:
146
462159
2318
ื”ื‘ื” ื ืจืื” ื›ื™ืฆื“ ื”ื ืžื™ื•ืฉืžื™ื ื‘ื‘ืขื™ื” ืฉืœ:
07:44
"Can you switch the machine off?" as Turing suggested.
147
464501
2789
"ื”ืื ืืชื” ื™ื›ื•ืœ ืœื›ื‘ื•ืช ืืช ื”ืžื›ื•ื ื”?" ื›ืคื™ ืฉื”ืฆื™ืข ื˜ื™ื•ืจื™ื ื’.
07:48
So here's a PR2 robot.
148
468893
2120
ื”ื™ื ื” ืจื•ื‘ื•ื˜ ืžื“ื’ื PR2 --
07:51
This is one that we have in our lab,
149
471037
1821
-- ืื—ื“ ืžืืœื” ืฉื™ืฉ ืœื ื• ื‘ืžืขื‘ื“ื”
07:52
and it has a big red "off" switch right on the back.
150
472882
2903
ื•ื™ืฉ ืœื• ืขืœ ื”ื’ื‘ ื›ืคืชื•ืจ ื›ื™ื‘ื•ื™ ื’ื“ื•ืœ ื•ืื“ื•ื.
07:56
The question is: Is it going to let you switch it off?
151
476361
2615
ื”ืฉืืœื” ื”ื™ื: ื”ืื ื”ื•ื ื™ืจืฉื” ืœืš ืœื›ื‘ื•ืช ืื•ืชื•?
07:59
If we do it the classical way,
152
479000
1465
ืื ื ืœืš ืœืคื™ ื”ืžื•ื“ืœ ื”ืงืœืืกื™,
08:00
we give it the objective of, "Fetch the coffee, I must fetch the coffee,
153
480489
3482
ื›ืš ื”ื•ื ื™ื‘ืฆืข ืืช ืžืฉื™ืžืช "ืชื‘ื™ื ืืช ื”ืงืคื”": "ืื ื™ ื—ื™ื™ื‘ ืœื”ื‘ื™ื ืืช ื”ืงืคื”,
08:03
I can't fetch the coffee if I'm dead,"
154
483995
2580
ืื ื™ ืœื ื™ื›ื•ืœ ืœื”ื‘ื™ื ืืช ื”ืงืคื” ืื ืื ื™ ืžืช",
08:06
so obviously the PR2 has been listening to my talk,
155
486599
3341
ื•ื”ื™ื•ืช ื• - PR2 ื›ืžื•ื‘ืŸ ื”ืงืฉื™ื‘ ืœื”ืจืฆืืชื™
08:09
and so it says, therefore, "I must disable my 'off' switch,
156
489964
3753
ื”ื•ื ืื•ืžืจ ืื™ืคื•ื, "ืื ื™ ื—ื™ื™ื‘ ืœื ื˜ืจืœ ืืช ืžืชื’ ื”ื›ื™ื‘ื•ื™ ืฉืœื™
08:14
and probably taser all the other people in Starbucks
157
494796
2694
ื•ืื•ืœื™ ื’ื ืœื—ืฉืžืœ ืืช ื›ืœ ื”ืื ืฉื™ื ื‘ืกื˜ืืจื‘ืงืก
08:17
who might interfere with me."
158
497514
1560
ืืฉืจ ืขืœื•ืœื™ื ืœื”ืคืจื™ืข ืœื™".
08:19
(Laughter)
159
499098
2062
(ืฆื—ื•ืง)
08:21
So this seems to be inevitable, right?
160
501184
2153
ื ืจืื” ื‘ืœืชื™ ื ืžื ืข, ื ื›ื•ืŸ?
08:23
This kind of failure mode seems to be inevitable,
161
503361
2398
ืื•ืคืŸ ืคืขื•ืœื” ื–ื” ื ืจืื” ื‘ืœืชื™ ื ืžื ืข
08:25
and it follows from having a concrete, definite objective.
162
505783
3543
ื‘ื’ืœืœ ืฉื”ื•ื ื ื•ื‘ืข ืžืžืชืŸ ืžืฉื™ืžื” ืžื•ื’ื“ืจืช ื•ืกื•ืคื™ืช.
08:30
So what happens if the machine is uncertain about the objective?
163
510632
3144
ืื‘ืœ ืžื” ื™ืงืจื” ืื ื”ืžื›ื•ื ื” ืœื ื‘ื˜ื•ื—ื” ืœื’ื‘ื™ ื”ืžืฉื™ืžื”?
08:33
Well, it reasons in a different way.
164
513800
2127
ื‘ืžืงืจื” ื–ื” ื”ื™ื ื—ื•ืฉื‘ืช ื‘ืฆื•ืจื” ืฉื•ื ื”.
08:35
It says, "OK, the human might switch me off,
165
515951
2424
ื”ื™ื ืื•ืžืจืช. "ื˜ื•ื‘, ื‘ืŸ ืื ื•ืฉ ืขืœื•ืœ ืœื›ื‘ื•ืช ืื•ืชื™,
08:38
but only if I'm doing something wrong.
166
518964
1866
ืื‘ืœ ื–ื” ืจืง ืื ืืขืฉื” ืžืฉื”ื• ืœื ื ื›ื•ืŸ.
08:41
Well, I don't really know what wrong is,
167
521567
2475
ืื™ื ื ื™ ื™ื•ื“ืขืช ืžื”ื• ืœื ื ื›ื•ืŸ,
08:44
but I know that I don't want to do it."
168
524066
2044
ืื‘ืœ ืื ื™ ื™ื•ื“ืขืช ืฉืื™ื ื ื™ ืจื•ืฆื” ืœืขืฉื•ืช ืื•ืชื•."
08:46
So that's the first and second principles right there.
169
526134
3010
ื›ืš, ื™ืฉ ืœื ื• ืคื” ื”ืขืงืจื•ืŸ ื”ืจืืฉื•ืŸ ื•ื”ืฉื ื™ ื’ื ื™ื—ื“.
08:49
"So I should let the human switch me off."
170
529168
3359
"ืœื›ืŸ, ืื ื™ ืฆืจื™ื›ื” ืœืืคืฉืจ ืœื‘ื ื™ ืื ื•ืฉ ืœื›ื‘ื•ืช ืื•ืชื™".
08:53
And in fact you can calculate the incentive that the robot has
171
533541
3956
ื ื™ืชืŸ ืœื”ื˜ืžื™ืข ืืช ืฉื™ืงื•ืœ ื”ืจื•ื•ื— ืฉื™ื”ื™ื” ืœืจื•ื‘ื•ื˜
08:57
to allow the human to switch it off,
172
537521
2493
ื›ื“ื™ ืฉื™ืืคืฉืจ ืœื ื• ืœื›ื‘ื•ืช ืื•ืชื•.
09:00
and it's directly tied to the degree
173
540038
1914
ื–ื” ื™ืฉื™ืจื•ืช ืงืฉื•ืจ ื‘ืจืžืช
09:01
of uncertainty about the underlying objective.
174
541976
2746
ืื™-ื”ื•ื•ื“ืื•ืช ืœื’ื‘ื™ ืžื”ื•ืช ื”ืžืฉื™ืžื”.
09:05
And then when the machine is switched off,
175
545797
2949
ื•ื›ืืฉืจ ื”ืžื›ื•ื ื” ื›ื‘ืจ ืžื›ื•ื‘ื”,
09:08
that third principle comes into play.
176
548770
1805
ื”ืขืงืจื•ืŸ ื”ืฉืœื™ืฉื™ ื ื›ื ืก ืœืžืฉื—ืง.
09:10
It learns something about the objectives it should be pursuing,
177
550599
3062
ื”ืžื›ื•ื ื” ืœื•ืžื“ืช ืžืฉื”ื• ืื•ื“ื•ืช ื”ืžืฉื™ืžื•ืช ืฉืขืœื™ื” ืœื‘ืฆืข,
09:13
because it learns that what it did wasn't right.
178
553685
2533
ื›ื™ ื”ื™ื ืœื•ืžื“ืช ืฉืžื” ืฉืขืฉืชื” ื”ื™ื” ืœื ื ื›ื•ืŸ.
09:16
In fact, we can, with suitable use of Greek symbols,
179
556242
3570
ื‘ืืžืฆืขื•ืช ืฉื™ืžื•ืฉ ืจืื•ื™ ื‘ืื•ืชื™ื•ืช ื™ื•ื•ื ื™ื•ืช,
09:19
as mathematicians usually do,
180
559836
2131
ื›ืคื™ ืฉื ื”ื•ื’ ืืฆืœ ื”ืžืชืžื˜ื™ืงืื™ื,
09:21
we can actually prove a theorem
181
561991
1984
ื ื™ืชืŸ ืœื”ื•ื›ื™ื— ื”ื”ื ื—ื”
09:23
that says that such a robot is provably beneficial to the human.
182
563999
3553
ืฉืื•ืžืจืช ืฉืจื•ื‘ื•ื˜ ื›ื–ื” ื”ื™ื ื• ื‘ื”ื—ืœื˜ ืžื•ืขื™ืœ ืœืื ื•ืฉื•ืช.
09:27
You are provably better off with a machine that's designed in this way
183
567576
3803
ืขื“ื™ืฃ ืœื›ื ืฉื”ืžื›ื•ื ื” ืชื”ื™ื” ื›ื–ืืช
09:31
than without it.
184
571403
1246
ื•ืœื ืื—ืจืช.
09:33
So this is a very simple example, but this is the first step
185
573057
2906
ื–ื•ื”ื™ ื“ื•ื’ืžื ืžืื•ื“ ืคืฉื•ื˜ื”, ืื‘ืœ ื–ื” ื”ืฆืขื“ ื”ืจืืฉื•ืŸ
09:35
in what we're trying to do with human-compatible AI.
186
575987
3903
ืœืงืจืืช ืžื” ืฉืื ื• ืžื ืกื™ื ืœืขืฉื•ืช ืขื ื‘''ืž ืžื•ืชืืžืช ืื ื•ืฉื•ืช.
09:42
Now, this third principle,
187
582477
3257
ื”ืขืงืจื•ืŸ ื”ืฉืœื™ืฉื™ ื”ื•ื ื–ื”
09:45
I think is the one that you're probably scratching your head over.
188
585758
3112
ืฉืœื’ื‘ื™ื• ืืชื ืžื”ืจื”ืจื™ื ืœื“ืขืชื™.
09:48
You're probably thinking, "Well, you know, I behave badly.
189
588894
3239
ืืชื ื‘ื•ื•ื“ืื™ ื—ื•ืฉื‘ื™ื, "ืื ื™ ืžืชื ื”ื’ ืœื ื›ืจืื•ื™.
09:52
I don't want my robot to behave like me.
190
592157
2929
ื•ืื™ื ื ื™ ืจื•ืฆื” ืฉื”ืจื•ื‘ื•ื˜ ืฉืœื™ ื™ืชื ื”ื’ ื›ืžื•ื ื™.
09:55
I sneak down in the middle of the night and take stuff from the fridge.
191
595110
3434
ืื ื™ ืžืชื’ื ื‘ ื‘ืืžืฆืข ื”ืœื™ืœื” ื•ืœื•ืงื— ื“ื‘ืจื™ื ืžืŸ ื”ืžืงืจืจ.
09:58
I do this and that."
192
598568
1168
ืื ื™ ืขื•ืฉื” ืืช ื–ื” ื•ืืช ื–ื”."
09:59
There's all kinds of things you don't want the robot doing.
193
599760
2797
ื™ืฉ ื›ืœ ืžื™ื ื™ ื“ื‘ืจื™ื ืฉืœื ืชืจืฆื• ืฉื”ืจื•ื‘ื•ื˜ ืฉืœื›ื ื™ืขืฉื”.
10:02
But in fact, it doesn't quite work that way.
194
602581
2071
ืื‘ืœ, ื”ื“ื‘ืจื™ื ืœื ื‘ื“ื™ื•ืง ืขื•ื‘ื“ื™ื ื›ื›ื”.
10:04
Just because you behave badly
195
604676
2155
ืจืง ื‘ื’ืœืœ ืฉืืชื ืžืชื ื”ื’ื™ื ืœื ื›ืจืื•ื™
10:06
doesn't mean the robot is going to copy your behavior.
196
606855
2623
ื”ืจื•ื‘ื•ื˜ ืฉืœื›ื ืœื ื‘ื”ื›ืจื— ื™ืขืชื™ืง ืืช ื”ืชื ื”ื’ื•ืชื›ื.
10:09
It's going to understand your motivations and maybe help you resist them,
197
609502
3910
ื”ื•ื ื™ื‘ื™ืŸ ืืช ื”ืžื ื™ืขื™ื ืฉืœื›ื ื•ืื•ืœื™ ื™ืขื–ื•ืจ ืœื›ื ืœื”ืชื ื’ื“ ืœื”ื,
10:13
if appropriate.
198
613436
1320
ื‘ืžื™ื“ื” ื•ื–ื” ื™ื”ื™ื” ืจืื•ื™.
10:16
But it's still difficult.
199
616026
1464
ืื‘ืœ ืขื“ื™ื™ืŸ, ื”ืขื ื™ื™ืŸ ืžืกื•ื‘ืš.
10:18
What we're trying to do, in fact,
200
618122
2545
ืžื” ืฉืื ื• ืžื ืกื™ื ืœืขืฉื•ืช,
10:20
is to allow machines to predict for any person and for any possible life
201
620691
5796
ื–ื” ืœืืคืฉืจ ืœืžื›ื•ื ื•ืช ืœื—ื–ื•ืช ื‘ืขื‘ื•ืจ ื›ืœ ืื“ื ื•ืขื‘ื•ืจ ื›ืœ ืžืกืœื•ืœื™ ื”ื—ื™ื™ื
10:26
that they could live,
202
626511
1161
ืฉื”ื•ื ืขืฉื•ื™ ืœื—ื™ื•ืช,
10:27
and the lives of everybody else:
203
627696
1597
ื•ื›ืŸ ื—ื™ื™ื”ื ืฉืœ ื›ืœ ื”ืฉืืจ:
10:29
Which would they prefer?
204
629317
2517
ืžื” ื™ืขื“ื™ืคื•?
10:33
And there are many, many difficulties involved in doing this;
205
633881
2954
ื•ื™ืฉื ื ื”ืจื‘ื” ืžืื•ื“ ืงืฉื™ื™ื ื‘ื“ืจืš ืœื‘ื™ืฆื•ืข,
10:36
I don't expect that this is going to get solved very quickly.
206
636859
2932
ืื™ื ื ื™ ืžืฆืคื” ืฉื”ืขื ื™ื™ืŸ ื™ื™ืคืชืจ ื‘ืžื”ืจื”.
10:39
The real difficulties, in fact, are us.
207
639815
2643
ื”ืงืฉื™ื™ื ื”ืืžื™ืชื™ื™ื ื”ื ืœืžืขืฉื” ืื ื—ื ื•.
10:43
As I have already mentioned, we behave badly.
208
643969
3117
ื›ืคื™ ืฉื›ื‘ืจ ืฆื™ื™ื ืชื™, ืื ื—ื ื• ืžืชื ื”ื’ื™ื ืœื ื›ืจืื•ื™.
10:47
In fact, some of us are downright nasty.
209
647110
2321
ืœืžืขืฉื”, ื—ืœืง ืžืื™ืชื ื• ืคืฉื•ื˜ ื ื•ืจืื™ื.
10:50
Now the robot, as I said, doesn't have to copy the behavior.
210
650251
3052
ื›ืคื™ ืฉืืžืจืชื™ ื”ืจื•ื‘ื•ื˜ ืœื ื—ื™ื™ื‘ ืœื”ืขืชื™ืง ืืช ื”ืชื ื”ื’ื•ืชื™ื ื•.
10:53
The robot does not have any objective of its own.
211
653327
2791
ืœืจื•ื‘ื•ื˜ ืื™ืŸ ื›ืœ ืžืฉื™ืžื” ืžืฉืœื•.
10:56
It's purely altruistic.
212
656142
1737
ื”ื•ื ืืœื˜ืจื•ื™ืกื˜ ื˜ื”ื•ืจ.
10:59
And it's not designed just to satisfy the desires of one person, the user,
213
659113
5221
ื•ื”ื•ื ืœื ืขื•ืฆื‘ ืœืฉื ืžื™ืžื•ืฉ ืจืฆื•ื ื•ืช ืฉืœ ืื“ื ืื—ื“ ื‘ืœื‘ื“,ื”ืžืฉืชืžืฉ,
11:04
but in fact it has to respect the preferences of everybody.
214
664358
3138
ืืœื ืœืงื—ืช ื‘ื—ืฉื‘ื•ืŸ ืืช ื”ืขื“ืคื•ืช ืฉืœ ื›ื•ืœื.
11:09
So it can deal with a certain amount of nastiness,
215
669083
2570
ื›ืš, ื”ื•ื ื™ื›ื•ืœ ืœื”ืชืžื•ื“ื“ ืขื ืจืžื” ืžืกื•ื™ืžืช ืฉืœ ืจื•ืข,
11:11
and it can even understand that your nastiness, for example,
216
671677
3701
ื•ื”ื•ื ืืคื™ืœื• ื™ื›ื•ืœ ืœื”ื‘ื™ืŸ ืืช ื”ื”ืชื ื”ื’ื•ืช ื”ืจืขื” ืฉืœืš.
11:15
you may take bribes as a passport official
217
675402
2671
ืœืžืฉืœ, ืืชื” ืื•ืœื™ ืœื•ืงื— ืฉื•ื—ื“ ื‘ืชื•ืจ ืคืงื™ื“ ื“ืจื›ื•ื ื™ื,
11:18
because you need to feed your family and send your kids to school.
218
678097
3812
ื‘ื’ืœืœ ืฉืืชื” ืฆืจื™ืš ืœื”ืื›ื™ืœ ืืช ืžืฉืคื—ืชืš ื•ืœืฉืœื•ื— ืืช ื™ืœื“ื™ืš ืœื‘ื™ื”''ืก.
11:21
It can understand that; it doesn't mean it's going to steal.
219
681933
2906
ื”ื•ื ื™ื›ื•ืœ ืœื”ื‘ื™ืŸ ื–ืืช. ืื™ืŸ ื–ื” ืื•ืžืจ ืฉื”ื•ื ื‘ืขืฆืžื• ื™ื’ื ื•ื‘.
11:24
In fact, it'll just help you send your kids to school.
220
684863
2679
ื”ื”ื‘ื ื” ืฉืœื• ืจืง ืชืขื–ื•ืจ ืœืš ืœืฉืœื•ื— ืืช ื”ื™ืœื“ื™ื ืœื‘''ืก.
11:28
We are also computationally limited.
221
688796
3012
ืื ื—ื ื• ื’ื ืžื•ื’ื‘ืœื™ื ื‘ืชื—ื•ื ื”ื—ื™ืฉื•ื‘ื™ื.
11:31
Lee Sedol is a brilliant Go player,
222
691832
2505
ืœื™ ืกื“ื•ืœ ื”ื•ื ืฉื—ืงืŸ ื’ื• ืžื‘ืจื™ืง,
11:34
but he still lost.
223
694361
1325
ืื‘ืœ ื”ื•ื ืขื“ื™ื™ืŸ ื”ืคืกื™ื“.
11:35
So if we look at his actions, he took an action that lost the game.
224
695710
4239
ืื– ืื ื ืชื‘ื•ื ืŸ ื‘ืฆืขื“ื™ื•, ื”ื•ื ืขืฉื” ืฆืขื“ ืฉื”ื•ื‘ื™ืœ ืœื”ืคืกื“.
11:39
That doesn't mean he wanted to lose.
225
699973
2161
ื–ื” ืœื ืื•ืžืจ ืฉื”ื•ื ืจืฆื” ืœื”ืคืกื™ื“.
11:43
So to understand his behavior,
226
703160
2040
ืœื›ืŸ, ืขืœ ืžื ืช ืœื”ื‘ื™ืŸ ืืช ื”ืชื ื”ื’ื•ืชื•,
11:45
we actually have to invert through a model of human cognition
227
705224
3644
ืขืœื™ื ื• ืœืจื“ืช ืœืคืจื˜ื™ ื”ืžื•ื“ืœ ื”ืงื•ื’ื ื™ื˜ื™ื‘ื™ ื”ืื ื•ืฉื™
11:48
that includes our computational limitations -- a very complicated model.
228
708892
4977
ืฉื›ื•ืœืœ ืืช ื”ืžื’ื‘ืœื•ืช ื”ื—ื™ืฉื•ื‘ื™ื•ืช ืฉืœื ื• - ื•ื–ื” ืžื•ื“ืœ ืžืกื•ื‘ืš ืžืื•ื“.
11:53
But it's still something that we can work on understanding.
229
713893
2993
ืื‘ืœ, ื–ื” ืขื“ื™ื™ืŸ ืžืฉื”ื• ืฉืื ื—ื ื• ื™ื›ื•ืœื™ื ืœื ืกื•ืช ืœื”ื‘ื™ืŸ.
11:57
Probably the most difficult part, from my point of view as an AI researcher,
230
717696
4320
ืื•ืœื™ ื”ื—ืœืง ื”ืงืฉื” ื‘ื™ื•ืชืจ ืขื‘ื•ืจื™ ื›ื—ื•ืงืจ ื‘''ืž
12:02
is the fact that there are lots of us,
231
722040
2575
ื”ื•ื ื”ืขื•ื‘ื“ื” ืฉืื ื—ื ื• ืจื‘ื™ื,
12:06
and so the machine has to somehow trade off, weigh up the preferences
232
726114
3581
ื•ืœื›ืŸ ื”ืžื›ื•ื ื” ื—ื™ื™ื‘ืช ืื™ื›ืฉื”ื• ืœื ืชื‘ ื•ืœืงื—ืช ื‘ื—ืฉื‘ื•ืŸ ืืช ื”ืขื“ืคื•ืช
12:09
of many different people,
233
729719
2225
ืฉืœ ื”ืžื•ืŸ ืื ืฉื™ื ืฉื•ื ื™ื.
12:11
and there are different ways to do that.
234
731968
1906
ื•ื™ืฉื ืŸ ื“ืจื›ื™ื ืฉื•ื ื•ืช ืœืขืฉื•ืช ื–ืืช.
12:13
Economists, sociologists, moral philosophers have understood that,
235
733898
3689
ื›ืœื›ืœื ื™ื, ืกื•ืฆื™ื•ืœื•ื’ื™ื, ืคื™ืœื•ืกื•ืคื™ื ื—ื•ืงืจื™ ืžื•ืกืจ ื”ื‘ื™ื ื• ื–ืืช
12:17
and we are actively looking for collaboration.
236
737611
2455
ื•ืื ื—ื ื• ืžื—ืคืฉื™ื ืืช ืฉื™ืชื•ืฃ ื”ืคืขื•ืœื” ื‘ืื•ืคืŸ ืคืขื™ืœ.
12:20
Let's have a look and see what happens when you get that wrong.
237
740090
3251
ื‘ื•ื ื ืจืื” ืžื” ืงื•ืจื” ื›ืืฉืจ ืœื ืžื‘ื™ื ื™ื ืืช ื–ื” ื ื›ื•ื ื”.
12:23
So you can have a conversation, for example,
238
743365
2133
ื ื’ื™ื“ ืฉืืชื ืžื ื”ืœื™ื ืฉื™ื—ื”
12:25
with your intelligent personal assistant
239
745522
1944
ืขื ื”ืขื•ื–ืจ ื”ืื™ืฉื™ ื”ืชื‘ื•ื ื™ ืฉืœื›ื,
12:27
that might be available in a few years' time.
240
747490
2285
ืืฉืจ ื™ื›ื•ืœ ืœื”ื™ื•ืช ื–ืžื™ืŸ ื‘ืขื•ื“ ื›ืžื” ืฉื ื™ื.
12:29
Think of a Siri on steroids.
241
749799
2524
ืชื—ืฉื‘ื• ืขืœ ืกื™ืจื™ ืขืœ ืกื˜ืจื•ืื™ื“ื™ื.
12:33
So Siri says, "Your wife called to remind you about dinner tonight."
242
753447
4322
ืื– ืกื™ืจื™ ืื•ืžืจืช, "ืื™ืฉืชืš ื”ืชืงืฉืจื” ืœื”ื–ื›ื™ืจ ืขืœ ืืจื•ื—ืช ื”ืขืจื‘".
12:38
And of course, you've forgotten. "What? What dinner?
243
758436
2508
ื•ื›ืžื•ื‘ืŸ ืืชื ืฉื›ื—ืชื. "ืžื”? ืื™ื–ื” ืืจื•ื—ื”?
12:40
What are you talking about?"
244
760968
1425
ืขืœ ืžื” ืืชื” ืžื“ื‘ืจ?"
12:42
"Uh, your 20th anniversary at 7pm."
245
762417
3746
"ืื”, ื™ื•ื ื”ื ื™ืฉื•ืื™ืŸ ื”-20 ืฉืœื›ื ื‘- 19:00"
12:48
"I can't do that. I'm meeting with the secretary-general at 7:30.
246
768735
3719
"ืื ื™ ืœื ื™ื›ื•ืœ. ื™ืฉ ืœื™ ืคื’ื™ืฉื” ืขื ื”ืžื ื›''ืœ ื‘- 19:30.
12:52
How could this have happened?"
247
772478
1692
ืื™ืš ื–ื” ื”ื™ื” ื™ื›ื•ืœ ืœืงืจื•ืช?"
12:54
"Well, I did warn you, but you overrode my recommendation."
248
774194
4660
"ื”ื–ื›ืจืชื™ ืœืš, ืื‘ืœ ืืชื” ื”ืชืขืœืžืช ืžื”ื”ืžืœืฆื” ืฉืœื™".
12:59
"Well, what am I going to do? I can't just tell him I'm too busy."
249
779966
3328
"ื˜ื•ื‘, ืื–ื” ืžื” ืื ื™ ืืขืฉื”? ืื ื™ ืœื ื™ื›ื•ืœ ืคืฉื•ื˜ ืœื•ืžืจ ืœื• ืฉืื ื™ ืขืกื•ืง."
13:04
"Don't worry. I arranged for his plane to be delayed."
250
784310
3281
"ืืœ ืชื“ืื’. ื“ืื’ืชื™ ืฉื™ื”ื™ื” ืœื• ืขื™ื›ื•ื‘ ื‘ื˜ื™ืกื”".
13:07
(Laughter)
251
787615
1682
(ืฆื—ื•ืง)
13:10
"Some kind of computer malfunction."
252
790069
2101
"ืกื•ื’ ืฉืœ ืชืงืœืช ืžื—ืฉื‘".
13:12
(Laughter)
253
792194
1212
(ืฆื—ื•ืง)
13:13
"Really? You can do that?"
254
793430
1617
"ื‘ืืžืช? ืืชื” ื™ื›ื•ืœ ืœืขืฉื•ืช ืืช ื–ื”?"
13:16
"He sends his profound apologies
255
796220
2179
"ื”ื•ื ืฉื•ืœื— ืืช ื”ืชื ืฆืœื•ืชื• ื”ืขืžื•ืงื”
13:18
and looks forward to meeting you for lunch tomorrow."
256
798423
2555
ื•ืžืฆืคื” ืœืคื’ื•ืฉ ืื•ืชืš ืžื—ืจ ืœืืจื•ื—ืช ื”ืฆื”ืจื™ื".
13:21
(Laughter)
257
801002
1299
(ืฆื—ื•ืง)
13:22
So the values here -- there's a slight mistake going on.
258
802325
4403
ืื– ืžื‘ื—ื™ื ืช ื”ืขืจื›ื™ื ืคื” -- ื™ืฉื ื• ืฉื™ื‘ื•ืฉ ืงืœ.
13:26
This is clearly following my wife's values
259
806752
3009
ื–ื” ื‘ื‘ื™ืจื•ืจ ืชื•ืื ืœืขืจื›ื™ื” ืฉืœ ืื™ืฉืชื™ -
13:29
which is "Happy wife, happy life."
260
809785
2069
"ืื™ืฉื” ืžืื•ืฉืจืช- ื—ื™ื™ื ืžืื•ืฉืจื™ื".
13:31
(Laughter)
261
811878
1583
(ืฆื—ื•ืง)
13:33
It could go the other way.
262
813485
1444
ื–ื” ื”ื™ื” ื™ื›ื•ืœ ืœืœื›ืช ื’ื ืœื›ื™ื•ื•ืŸ ืื—ืจ.
13:35
You could come home after a hard day's work,
263
815641
2201
ื ื ื™ื— ืฉื—ื–ืจืชื ื”ื‘ื™ืชื” ืื—ืจื™ ื™ื•ื ืขื‘ื•ื“ื” ืงืฉื”
13:37
and the computer says, "Long day?"
264
817866
2195
ื•ื”ืžื—ืฉื‘ ืฉื•ืืœ "ื™ื•ื ืืจื•ืš?"
13:40
"Yes, I didn't even have time for lunch."
265
820085
2288
"ื›ืŸ, ืืคื™ืœื• ืœื ื”ื™ื” ืœื™ ื–ืžืŸ ืœืื›ื•ืœ ืฆื”ืจื™ื".
13:42
"You must be very hungry."
266
822397
1282
"ืืชื” ื‘ื˜ื— ืžืื•ื“ ืจืขื‘".
13:43
"Starving, yeah. Could you make some dinner?"
267
823703
2646
ื›ืŸ, ื’ื•ื•ืข ืžืจืขื‘. ืชื•ื›ืœ ืœื‘ืฉืœ ืืจื•ื—ืช ืขืจื‘?"
13:47
"There's something I need to tell you."
268
827890
2090
"ื™ืฉ ืžืฉื”ื• ืฉืื ื™ ื—ื™ื™ื‘ ืœืกืคืจ ืœืš"
13:50
(Laughter)
269
830004
1155
(ืฆื—ื•ืง)
13:52
"There are humans in South Sudan who are in more urgent need than you."
270
832013
4905
"ื™ืฉ ื‘ื ื™ ืื ื•ืฉ ื‘ื“ืจื•ื ืกื•ื“ืŸ ืฉื ื–ืงืงื™ื ื”ืจื‘ื” ื™ื•ืชืจ ืžืžืš."
13:56
(Laughter)
271
836942
1104
(ืฆื—ื•ืง)
13:58
"So I'm leaving. Make your own dinner."
272
838070
2075
"ืื– ืื ื™ ืขื•ื–ื‘. ืชื›ื™ืŸ ืœืš ืื•ื›ืœ ื‘ืขืฆืžืš."
14:00
(Laughter)
273
840169
2000
(ืฆื—ื•ืง)
14:02
So we have to solve these problems,
274
842643
1739
ืื– ืขืœื™ื ื• ืœืคืชื•ืจ ืืช ื”ื‘ืขื™ื•ืช ื”ืœืœื•
14:04
and I'm looking forward to working on them.
275
844406
2515
ื•ืื ื™ ืžืื•ื“ ืจื•ืฆื” ืœืขื‘ื•ื“ ืขืœ ื–ื”.
14:06
There are reasons for optimism.
276
846945
1843
ื™ืฉ ืกื™ื‘ื•ืช ืœืื•ืคื˜ื™ืžื™ื•ืช.
14:08
One reason is,
277
848812
1159
ืกื™ื‘ื” ืจืืฉื•ื ื” ื”ื™ื
14:09
there is a massive amount of data.
278
849995
1868
ืฉืงื™ื™ืžืช ื›ืžื•ืช ืื“ื™ืจื” ืฉืœ ืžื™ื“ืข.
14:11
Because remember -- I said they're going to read everything
279
851887
2794
ื–ื•ื›ืจื™ื -- ืืžืจืชื™ ืฉื”ื ื™ืงืจืื• ืืช ื›ืœ
14:14
the human race has ever written.
280
854705
1546
ืžื” ืฉื”ืื ื•ืฉื•ืช ื›ืชื‘ื” ืื™ ืคืขื?
14:16
Most of what we write about is human beings doing things
281
856275
2724
ืœืจื•ื‘ ืื ื—ื ื• ื›ื•ืชื‘ื™ื ืื•ื“ื•ืช ืžืขืฉื™ ืื ืฉื™ื
14:19
and other people getting upset about it.
282
859023
1914
ื•ืื•ื“ื•ืช ืื ืฉื™ื ืื—ืจื™ื ืฉืžืชื•ืกื›ืœื™ื ืžื›ืš.
14:20
So there's a massive amount of data to learn from.
283
860961
2398
ืœื›ืŸ, ื™ืฉื ื• ื ืคื— ืขื ืง ืฉืœ ืžื™ื“ืข ืฉืืคืฉืจ ืœืœืžื•ื“ ืžืžื ื•.
14:23
There's also a very strong economic incentive
284
863383
2236
ื™ืฉื ื• ื’ื ืžื ื™ืข ื›ืœื›ืœื™ ื—ื–ืง ืžืื•ื“
14:27
to get this right.
285
867151
1186
ืœืขืฉื•ืช ื–ืืช ื ื›ื•ืŸ.
14:28
So imagine your domestic robot's at home.
286
868361
2001
ื“ืžื™ื™ื ื• ืืช ื”ืจื•ื‘ื•ื˜ ื”ื‘ื™ืชื™ ืฉืœื›ื.
14:30
You're late from work again and the robot has to feed the kids,
287
870386
3067
ืฉื•ื‘ ื—ื–ืจืชื ืžืื•ื—ืจ ืžื”ืขื‘ื•ื“ื” ื•ื”ืจื•ื‘ื•ื˜ ื—ื™ื™ื‘ ืœื”ืื›ื™ืœ ืืช ื™ืœื“ื™ื›ื
14:33
and the kids are hungry and there's nothing in the fridge.
288
873477
2823
ื”ื™ืœื“ื™ื ืจืขื‘ื™ื ื•ืื™ืŸ ื›ืœื•ื ื‘ืžืงืจืจ.
14:36
And the robot sees the cat.
289
876324
2605
ื•ืื– ื”ืจื•ื‘ื•ื˜ ืžื‘ื—ื™ืŸ ื‘ื—ืชื•ืœ.
14:38
(Laughter)
290
878953
1692
(ืฆื—ื•ืง)
14:40
And the robot hasn't quite learned the human value function properly,
291
880669
4190
ื”ืจื•ื‘ื•ื˜ ืขื“ื™ื™ืŸ ืœื ืœื’ืžืจื™ ืœืžื“ ื›ื™ืฆื“ ื”ืขืจื›ื™ื ื”ืื ื•ืฉื™ื™ื ืขื•ื‘ื“ื™ื,
14:44
so it doesn't understand
292
884883
1251
ืœื›ืŸ ืื™ื ื ื• ืžื‘ื™ืŸ
14:46
the sentimental value of the cat outweighs the nutritional value of the cat.
293
886158
4844
ืฉื”ืขืจืš ื”ืจื’ืฉื™ ืฉืœ ื”ื—ืชื•ืœ ืจื‘ ืขืœ ืขืจื›ื• ื”ืชื–ื•ื ืชื™.
14:51
(Laughter)
294
891026
1095
(ืฆื—ื•ืง)
14:52
So then what happens?
295
892145
1748
ืื– ืžื” ืงื•ืจื” ืื—ืจื™ ื–ื”?
14:53
Well, it happens like this:
296
893917
3297
ืžื” ืฉืงื•ืจื” ื–ื” ื›ื›ื”:
14:57
"Deranged robot cooks kitty for family dinner."
297
897238
2964
"ื”ืจื•ื‘ื•ื˜ ื”ืžื˜ื•ืจืฃ ืžื‘ืฉืœ ืืช ื”ื—ืชื•ืœ ืœืืจื•ื—ืช ืขืจื‘ ืžืฉืคื—ืชื™ืช".
15:00
That one incident would be the end of the domestic robot industry.
298
900226
4523
ืืจื•ืข ืื—ื“ ืฉื›ื–ื” ื™ื”ื™ื” ืกื•ืคื” ืฉืœ ืชืขืฉื™ื™ืช ืจื•ื‘ื•ื˜ื™ื ื‘ื™ืชื™ื™ื.
15:04
So there's a huge incentive to get this right
299
904773
3372
ืœื›ืŸ, ื™ืฉื ื” ืกื™ื‘ื” ืžืžืฉ ื˜ื•ื‘ื” ืœืขืฉื•ืช ื”ื›ืœ ื ื›ื•ืŸ.
15:08
long before we reach superintelligent machines.
300
908169
2715
ื”ืจื‘ื” ืœืคื ื™ ืฉื ื’ื™ืข ืœืฉืœื‘ ืžื›ื•ื ื•ืช ืขื ืชื‘ื•ื ืช-ืขืœ.
15:11
So to summarize:
301
911948
1535
ืœืกื™ื›ื•ื:
15:13
I'm actually trying to change the definition of AI
302
913507
2881
ืื ื™ ืžื ืกื” ืœืฉื ื•ืช ืืช ื”ื’ื“ืจืช ื‘''ืž,
15:16
so that we have provably beneficial machines.
303
916412
2993
ื›ืš ืฉื ื™ื™ืฆืจ ืžื›ื•ื ื•ืช ืฉืชื”ื™ื™ื ื” ื‘ื”ื—ืœื˜ ื˜ื•ื‘ื•ืช ืœื ื•.
15:19
And the principles are:
304
919429
1222
ื•ื”ืจื™ ื”ืขืงืจื•ื ื•ืช:
15:20
machines that are altruistic,
305
920675
1398
ืžื›ื•ื ื•ืช ืฉื”ืŸ ืืœื˜ืจื•ื™ืกื˜ื™ื•ืช,
15:22
that want to achieve only our objectives,
306
922097
2804
ืฉืจื•ืฆื•ืช ืœืžืžืฉ ืืช ื”ืžืฉื™ืžื•ืช ืฉืœื ื• ื‘ืœื‘ื“,
15:24
but that are uncertain about what those objectives are,
307
924925
3116
ืื‘ืœ ืื™ื ืŸ ื‘ื˜ื•ื—ื•ืช ืžื” ื”ืŸ ื”ืžืฉื™ืžื•ืช ื”ืœืœื•
15:28
and will watch all of us
308
928065
1998
ื•ืฉืชืชื‘ื•ื ื ื” ื‘ื›ื•ืœื ื•
15:30
to learn more about what it is that we really want.
309
930087
3203
ื›ื“ื™ ืœืœืžื•ื“ ื™ื•ืชืจ ืขืœ ืžื” ืฉืื ื—ื ื• ืจื•ืฆื™ื ื‘ืืžืช.
15:34
And hopefully in the process, we will learn to be better people.
310
934193
3559
ื ืงื•ื•ื” ืฉืชื•ืš ื›ื“ื™ ื›ืš ื ืœืžื“ ืœื”ื™ื•ืช ืื ืฉื™ื ื˜ื•ื‘ื™ื ื™ื•ืชืจ.
15:37
Thank you very much.
311
937776
1191
ืชื•ื“ื” ืจื‘ื”.
15:38
(Applause)
312
938991
3709
(ืžื—ื™ืื•ืช ื›ืคื™ื™ื)
15:42
Chris Anderson: So interesting, Stuart.
313
942724
1868
ื›ืจื™ืก ืื ื“ืจืกื•ืŸ (ื›.ื.): ื–ื” ื›ืœ ื›ืš ืžืขื ื™ื™ืŸ, ืกื˜ื•ืืจื˜.
15:44
We're going to stand here a bit because I think they're setting up
314
944616
3170
ื‘ื•ื ืจื’ืข ื ืขืžื•ื“ ืคื” ื›ื™ ืื ื™ ื—ื•ืฉื‘ ืฉื”ืžืืจื’ื ื™ื ืžืชื›ื•ื ื ื™ื
15:47
for our next speaker.
315
947810
1151
ืœื“ื•ื‘ืจ ื”ื‘ื.
15:48
A couple of questions.
316
948985
1538
ืžืกืคืจ ืฉืืœื•ืช.
15:50
So the idea of programming in ignorance seems intuitively really powerful.
317
950547
5453
ืจืขื™ื•ืŸ ื”ื˜ืžืขืช ื‘ื•ืจื•ืช ืžืจื’ื™ืฉ ืžืžืฉ ืขื•ืฆืžืชื™.
15:56
As you get to superintelligence,
318
956024
1594
ืื‘ืœ, ื›ืืฉืจ ื ื’ื™ืข ืœืฉืœื‘ ืชื‘ื•ื ืช-ืขืœ
15:57
what's going to stop a robot
319
957642
2258
ืžื” ื™ืžื ืข ืžืจื•ื‘ื•ื˜
15:59
reading literature and discovering this idea that knowledge
320
959924
2852
ืœืขื™ื™ืŸ ื‘ืกืคืจื•ืช ื•ืœื’ืœื•ืช ืืช ื”ืจืขื™ื•ืŸ ืฉื™ื“ืข
16:02
is actually better than ignorance
321
962800
1572
ืขื“ื™ืฃ ืขืœ ื‘ื•ืจื•ืช,
16:04
and still just shifting its own goals and rewriting that programming?
322
964396
4218
ืœืฉื ื•ืช ืืช ืžื˜ืจื•ืชื™ื• ื•ืœืฉื›ืชื‘ ืืช ื”ืชื•ื›ื ื” ื‘ื”ืชืื?
16:09
Stuart Russell: Yes, so we want it to learn more, as I said,
323
969512
6356
ืก.ืจ.: ื›ืŸ. ืื ื—ื ื• ืจื•ืฆื™ื ืฉื™ืœืžื“ ื™ื•ืชืจ, ื›ืคื™ ืฉืืžืจืชื™,
16:15
about our objectives.
324
975892
1287
ืื•ื“ื•ืช ืžื˜ืจื•ืชื™ื ื•.
16:17
It'll only become more certain as it becomes more correct,
325
977203
5521
ื–ื” ื™ืชื‘ื”ืจ ื™ื•ืชืจ ืจืง ื›ืฉื–ื” ื™ื”ื™ื” ื ื›ื•ืŸ ื™ื•ืชืจ
16:22
so the evidence is there
326
982748
1945
ืื– ื”ืขื•ื‘ื“ื•ืช ื ืžืฆืื•ืช ื‘ืคื ื™ื•
16:24
and it's going to be designed to interpret it correctly.
327
984717
2724
ื•ื”ืจื•ื‘ื•ื˜ ื™ืขื•ืฆื‘ ื›ืš ืฉื™ื•ื›ืœ ืœืคืจืฉ ืื•ืชืŸ ื ื›ื•ืŸ.
16:27
It will understand, for example, that books are very biased
328
987465
3956
ืœืžืฉืœ, ื”ื•ื ื™ื‘ื™ืŸ ืฉืกืคืจื™ื ืžื•ื˜ื™ื ืžืื•ื“
16:31
in the evidence they contain.
329
991445
1483
ื‘ืขื•ื‘ื“ื•ืช ืฉื”ื ืžื›ื™ืœื™ื.
16:32
They only talk about kings and princes
330
992952
2397
ื”ื ืžื“ื‘ืจื™ื ืจืง ืขืœ ื”ืžืœื›ื™ื ื•ื”ื ืกื™ื›ื•ืช
16:35
and elite white male people doing stuff.
331
995373
2800
ื•ืืœื™ื˜ื•ืช ืฉืœ ื’ื‘ืจื™ื ืœื‘ื ื™ื ืฉืขื•ืฉื™ื ื›ืœ ืžื™ื ื™ ื“ื‘ืจื™ื.
16:38
So it's a complicated problem,
332
998197
2096
ืœื›ืŸ, ื–ื•ื”ื™ ื‘ืขื™ื” ืžื•ืจื›ื‘ืช.
16:40
but as it learns more about our objectives
333
1000317
3872
ืื‘ืœ, ื›ื›ืœ ืฉื”ืจื•ื‘ื•ื˜ ื™ืœืžื“ ืืช ืžื˜ืจื•ืชื™ื ื•
16:44
it will become more and more useful to us.
334
1004213
2063
ื”ื•ื ื™ื™ืœืš ื•ื™ื™ืขืฉื” ืžื•ืขื™ืœ ื™ื•ืชืจ ืขื‘ื•ืจื ื•.
16:46
CA: And you couldn't just boil it down to one law,
335
1006300
2526
ื›.ื.: ืื™ ืืคืฉืจ ืœืกื›ื ืืช ื–ื” ืœื›ื“ื™ ื—ื•ืง ืื—ื“,
16:48
you know, hardwired in:
336
1008850
1650
ืืชื” ื™ื•ื“ืข, ืžืฉื”ื• ื—ืฆื•ื‘ ื‘ืกืœืข:
16:50
"if any human ever tries to switch me off,
337
1010524
3293
" ืื ื‘ืŸ ืื ื•ืฉ ื™ื ืกื” ืื™ ืคืขื ืœื›ื‘ื•ืช ืื•ืชื™,
16:53
I comply. I comply."
338
1013841
1935
ืื ื™ ืžืฆื™ื™ืช. ืื ื™ ืžืฆื™ื™ืช."
16:55
SR: Absolutely not.
339
1015800
1182
ืก.ืจ.: ืœื’ืžืจื™ ืœื.
16:57
That would be a terrible idea.
340
1017006
1499
ื–ื” ื™ื”ื™ื” ืจืขื™ื•ืŸ ื ื•ืจืื™.
16:58
So imagine that you have a self-driving car
341
1018529
2689
ื“ืžื™ื™ื ื• ืฉื™ืฉ ืœื›ื ืจื›ื‘ ืื•ื˜ื•ืžื˜ื™
17:01
and you want to send your five-year-old
342
1021242
2433
ื•ืืชื ืจื•ืฆื™ื ืœืฉืœื•ื— ืืช ื™ืœื“ื™ื›ื ื‘ืŸ ื”ื—ืžืฉ
17:03
off to preschool.
343
1023699
1174
ืœื’ืŸ ื™ืœื“ื™ื.
17:04
Do you want your five-year-old to be able to switch off the car
344
1024897
3101
ื”ื™ื™ืชื ืจื•ืฆื™ื ืฉื”ื™ืœื“ ื™ื•ื›ืœ ืœื›ื‘ื•ืช ืืช ื”ืจื›ื‘
17:08
while it's driving along?
345
1028022
1213
ืชื•ืš ื›ื“ื™ ื”ื ืกื™ืขื”?
17:09
Probably not.
346
1029259
1159
ื›ื ืจืื” ืฉืœื.
17:10
So it needs to understand how rational and sensible the person is.
347
1030442
4703
ื™ื•ืฆื ืฉื”ืžื›ื•ื ื™ืช ื—ื™ื™ื‘ืช ืœื”ื—ืœื™ื˜ ืขื“ ื›ืžื” ื”ื ื•ืกืข ื ื‘ื•ืŸ ื•ื”ื’ื™ื•ื ื™.
17:15
The more rational the person,
348
1035169
1676
ื›ื›ืœ ืฉื”ื ื•ืกืข ื”ื’ื™ื•ื ื™ ื™ื•ืชืจ,
17:16
the more willing you are to be switched off.
349
1036869
2103
ื›ืš ื”ืžื›ื•ื ื™ืช ืชื”ื™ื” ืžื•ื›ื ื” ื™ื•ืชืจ ืฉื™ื›ื‘ื• ืื•ืชื”.
17:18
If the person is completely random or even malicious,
350
1038996
2543
ืื ื”ื ื•ืกืข ืคื–ื™ื– ื•ื—ืกืจ ื”ื’ื™ื•ืŸ ืœื—ืœื•ื˜ื™ืŸ ืื• ืืคื™ืœื• ื–ื“ื•ื ื™,
17:21
then you're less willing to be switched off.
351
1041563
2512
ืื– ื”ืžื›ื•ื ื™ืช ืคื—ื•ืช ืชืจืฆื” ืฉื™ื›ื‘ื• ืื•ืชื”.
17:24
CA: All right. Stuart, can I just say,
352
1044099
1866
ืง.ื.: ื‘ืกื“ืจ ื’ืžื•ืจ. ืชืจืฉื” ืœื™ ืจืง ืœื•ืžืจ
17:25
I really, really hope you figure this out for us.
353
1045989
2314
ืฉืื ื™ ืžืžืฉ ืžืงื•ื•ื” ืฉืชืคืชื•ืจ ืืช ื–ื” ืขื‘ื•ืจื™ื ื•.
17:28
Thank you so much for that talk. That was amazing.
354
1048327
2375
ืชื•ื“ื” ืจื‘ื” ืขืœ ื”ื”ืจืฆืื” ื”ื–ืืช. ื–ื” ื”ื™ื” ืžื“ื”ื™ื.
17:30
SR: Thank you.
355
1050726
1167
ืก.ืจ: ืชื•ื“ื”.
17:31
(Applause)
356
1051917
1837
(ื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