3 principles for creating safer AI | Stuart Russell

139,486 views ・ 2017-06-06

TED


請雙擊下方英文字幕播放視頻。

譯者: 易帆 余 審譯者: Wilde Luo
00:12
This is Lee Sedol.
0
12532
1552
這是李世石。
00:14
Lee Sedol is one of the world's greatest Go players,
1
14108
3997
李世石是全世界 頂尖圍棋高手之一,
00:18
and he's having what my friends in Silicon Valley call
2
18129
2885
此時,他正在經歷的是 我的矽谷朋友們稱之為
00:21
a "Holy Cow" moment --
3
21038
1510
「我的媽呀!」的時刻......
00:22
(Laughter)
4
22572
1073
(笑聲)
00:23
a moment where we realize
5
23669
2188
在這一刻讓我們意識到,
00:25
that AI is actually progressing a lot faster than we expected.
6
25881
3296
原來人工智慧發展的進程 比我們預期的要快得多。
00:29
So humans have lost on the Go board. What about the real world?
7
29974
3047
人類已在圍棋博弈中落敗, 那現實世界中情況又如何?
00:33
Well, the real world is much bigger,
8
33045
2100
當然啦,現實世界要比棋盤 廣闊、複雜得多,
00:35
much more complicated than the Go board.
9
35169
2249
00:37
It's a lot less visible,
10
37442
1819
它也遠不如棋盤上那麽黑白分明,
00:39
but it's still a decision problem.
11
39285
2038
但仍然是個判定問題 (Decision Problem)。
00:42
And if we think about some of the technologies
12
42768
2321
如果我們思考一些 即將問世的新科技……
00:45
that are coming down the pike ...
13
45113
1749
00:47
Noriko [Arai] mentioned that reading is not yet happening in machines,
14
47558
4335
新井紀子提到機器仍無法 「閱讀」,
00:51
at least with understanding.
15
51917
1500
至少無法真正理解文本含義。
00:53
But that will happen,
16
53441
1536
但這項能力最終會被機器掌握,
00:55
and when that happens,
17
55001
1771
而當這一切發生時,
00:56
very soon afterwards,
18
56796
1187
不久之後,
00:58
machines will have read everything that the human race has ever written.
19
58007
4572
機器就能讀遍所有人類寫下的東西。
01:03
And that will enable machines,
20
63670
2030
這會讓機器擁有比人類 更深刻的遠見和洞察力。
01:05
along with the ability to look further ahead than humans can,
21
65724
2920
01:08
as we've already seen in Go,
22
68668
1680
就如我們在這場圍棋博弈中所見,
01:10
if they also have access to more information,
23
70372
2164
如果機器能接觸到比人類更多的信息,
01:12
they'll be able to make better decisions in the real world than we can.
24
72560
4268
那機器將能夠在現實世界中 做出比人類更好的決策。
01:18
So is that a good thing?
25
78612
1606
那這會是一件好事嗎?
01:21
Well, I hope so.
26
81718
2232
我當然希望如此。
01:26
Our entire civilization, everything that we value,
27
86514
3255
人類的全部文明, 我們所珍視的一切,
01:29
is based on our intelligence.
28
89793
2068
都是基於我們的智慧。
01:31
And if we had access to a lot more intelligence,
29
91885
3694
如果我們能獲得更強大的智慧,
01:35
then there's really no limit to what the human race can do.
30
95603
3302
那人類將無所不能了。
01:40
And I think this could be, as some people have described it,
31
100485
3325
我在想,到時後就像 一些人所描述的那樣,
01:43
the biggest event in human history.
32
103834
2016
這會是人類歷史上最重要的事件。
01:48
So why are people saying things like this,
33
108485
2829
那為什麽有的人會說出 以下的言論呢?
01:51
that AI might spell the end of the human race?
34
111338
2876
說人工智慧將是人類的末日呢?
01:55
Is this a new thing?
35
115258
1659
這是新鮮事嗎?
01:56
Is it just Elon Musk and Bill Gates and Stephen Hawking?
36
116941
4110
這僅僅只是伊隆馬斯克、比爾蓋茲、 史蒂芬霍金的新發明嗎?
02:01
Actually, no. This idea has been around for a while.
37
121773
3262
實際上不是,這個概念 已經存在很長的時間了。
02:05
Here's a quotation:
38
125059
1962
請看這段話:
02:07
"Even if we could keep the machines in a subservient position,
39
127045
4350
「即便我們能讓機器屈從於我們,
02:11
for instance, by turning off the power at strategic moments" --
40
131419
2984
比如說,在重要時刻關掉它。」
02:14
and I'll come back to that "turning off the power" idea later on --
41
134427
3237
我等會兒會再來討論 「關機」這一話題。
02:17
"we should, as a species, feel greatly humbled."
42
137688
2804
「我們作為人類,仍應懷着謙卑......」
02:21
So who said this? This is Alan Turing in 1951.
43
141997
3448
這段話是誰說的呢? 是艾倫 · 圖靈在 1951 年說的。
02:26
Alan Turing, as you know, is the father of computer science
44
146120
2763
眾所皆知艾倫 · 圖靈是計算機科學之父,
02:28
and in many ways, the father of AI as well.
45
148907
3048
並且從很多方面來講, 他也是人工智慧之父。
02:33
So if we think about this problem,
46
153059
1882
所以,當我們在思考「創造出 比自己更聰明的物種」這個問題時,
02:34
the problem of creating something more intelligent than your own species,
47
154965
3787
02:38
we might call this "the gorilla problem,"
48
158776
2622
我們不妨將它稱為「大猩猩問題」。
02:42
because gorillas' ancestors did this a few million years ago,
49
162165
3750
因為大猩猩的祖先們 在幾百萬年前就親歷此境,
02:45
and now we can ask the gorillas:
50
165939
1745
我們可以去問大猩猩們:
02:48
Was this a good idea?
51
168572
1160
「這是不是一個好主意?」
02:49
So here they are having a meeting to discuss whether it was a good idea,
52
169756
3530
圖片中,牠們正在開會討論 那麽做是不是一個好主意,
02:53
and after a little while, they conclude, no,
53
173310
3346
過了一會兒,牠們總結出:「不。」
02:56
this was a terrible idea.
54
176680
1345
這是個很爛的主意──
02:58
Our species is in dire straits.
55
178049
1782
作為靈長類的我們正岌岌可危。
03:00
In fact, you can see the existential sadness in their eyes.
56
180358
4263
你可以從牠們的眼神中 看到存亡攸關的憂傷。
03:04
(Laughter)
57
184645
1640
(笑聲)
03:06
So this queasy feeling that making something smarter than your own species
58
186309
4840
「創造出比你自己更聰明的物種 並不是什麽妙計」
03:11
is maybe not a good idea --
59
191173
2365
這種感覺很倒胃口。
03:14
what can we do about that?
60
194308
1491
那我們能做些什麽呢?
03:15
Well, really nothing, except stop doing AI,
61
195823
4767
其實,除非停止人工智慧的研究, 否則束手無策。
03:20
and because of all the benefits that I mentioned
62
200614
2510
因為我所提到的人工智慧的各種裨益,
03:23
and because I'm an AI researcher,
63
203148
1716
也因為我是人工智慧的研究人員,
03:24
I'm not having that.
64
204888
1791
我可不同意就此止步。
03:27
I actually want to be able to keep doing AI.
65
207103
2468
實際上,我想一直研究人工智慧。
03:30
So we actually need to nail down the problem a bit more.
66
210435
2678
所以我們需要更加明確問題所在。
03:33
What exactly is the problem?
67
213137
1371
這個問題到底是什麽呢?
03:34
Why is better AI possibly a catastrophe?
68
214532
3246
為什麽更強大的人工智慧 可能會是個災難呢?
03:39
So here's another quotation:
69
219218
1498
還有一句名言:
03:41
"We had better be quite sure that the purpose put into the machine
70
221755
3335
「我們最好確保我們向機器發出的指令 與我們的真正目的相吻合。」
03:45
is the purpose which we really desire."
71
225114
2298
03:48
This was said by Norbert Wiener in 1960,
72
228102
3498
這句話是諾伯特 · 維納在 1960 年說的,
03:51
shortly after he watched one of the very early learning systems
73
231624
4002
就在他看完一個早期的學習系統 (Learning System)之後。
03:55
learn to play checkers better than its creator.
74
235650
2583
這個系統在學習如何能把 西洋棋下得比發明它的人更好。
04:00
But this could equally have been said
75
240422
2683
但如出一轍的一句話,
04:03
by King Midas.
76
243129
1167
邁達斯國王也說過。
04:04
King Midas said, "I want everything I touch to turn to gold,"
77
244903
3134
他說:「我希望我觸碰的 所有東西都變成金子。」
04:08
and he got exactly what he asked for.
78
248061
2473
結果他真的獲得了點石成金的能力。
04:10
That was the purpose that he put into the machine,
79
250558
2751
可以說,這就是他給機器下的指令。
04:13
so to speak,
80
253333
1450
04:14
and then his food and his drink and his relatives turned to gold
81
254807
3444
結果他的食物、飲料 和家人都變成了金子,
04:18
and he died in misery and starvation.
82
258275
2281
最後他死於痛苦與饑餓當中。
04:22
So we'll call this "the King Midas problem"
83
262264
2341
所以我們把這類問題叫做 「邁達斯國王問題」,
04:24
of stating an objective which is not, in fact,
84
264629
3305
這個比喻是要說明這種 不符合實際需求的 「目的」。
04:27
truly aligned with what we want.
85
267958
2413
04:30
In modern terms, we call this "the value alignment problem."
86
270395
3253
用現代的術語來說,我們把它稱為 「價值取向不一致問題」。
04:36
Putting in the wrong objective is not the only part of the problem.
87
276867
3485
「設錯了目標」不是唯一的問題,
04:40
There's another part.
88
280376
1152
還有其他的。
04:41
If you put an objective into a machine,
89
281980
1943
如果你給機器人設了個目標,
04:43
even something as simple as, "Fetch the coffee,"
90
283947
2448
即使簡單如「去把咖啡端來。」
04:47
the machine says to itself,
91
287728
1841
那機器人會對自己說:
04:50
"Well, how might I fail to fetch the coffee?
92
290553
2623
「什麼會讓我無法去拿咖啡?
04:53
Someone might switch me off.
93
293200
1580
說不定有人會把我關機;
04:55
OK, I have to take steps to prevent that.
94
295465
2387
好,那我要想辦法阻止,
04:57
I will disable my 'off' switch.
95
297876
1906
我得讓我的「關機」開關失效。
05:00
I will do anything to defend myself against interference
96
300354
2959
我得盡一切可能防衛自己,
免得別人干涉我去達成 所被賦予的任務。」
05:03
with this objective that I have been given."
97
303337
2629
05:05
So this single-minded pursuit
98
305990
2012
這種專注的行事,以一種 極端自我保護的模式在執行,
05:09
in a very defensive mode of an objective that is, in fact,
99
309033
2945
05:12
not aligned with the true objectives of the human race --
100
312002
2814
實際上與我們人類 想要的目標並不一致。
05:15
that's the problem that we face.
101
315942
1862
這就是我們面臨的問題。
05:18
And in fact, that's the high-value takeaway from this talk.
102
318827
4767
而這就是這場演講的 核心想法,也是價值所在。
05:23
If you want to remember one thing,
103
323618
2055
如果你想從這場演講中汲取什麽,
05:25
it's that you can't fetch the coffee if you're dead.
104
325697
2675
那你只要記得: 如果死了,就不能端咖啡了。
05:28
(Laughter)
105
328396
1061
(笑聲)
05:29
It's very simple. Just remember that. Repeat it to yourself three times a day.
106
329481
3829
這很簡單,記住就行了, 每天早晚覆誦三遍。
05:33
(Laughter)
107
333334
1821
(笑聲)
05:35
And in fact, this is exactly the plot
108
335179
2754
實際上,這正是電影 《2001太空漫步》的劇情。
05:37
of "2001: [A Space Odyssey]"
109
337957
2648
05:41
HAL has an objective, a mission,
110
341046
2090
HAL 有一個目標,一個任務,
05:43
which is not aligned with the objectives of the humans,
111
343160
3732
但這個目標與人類的目標不一致,
05:46
and that leads to this conflict.
112
346916
1810
最後導致了衝突。
05:49
Now fortunately, HAL is not superintelligent.
113
349314
2969
幸運的是, HAL 並沒有超級智慧,
05:52
He's pretty smart, but eventually Dave outwits him
114
352307
3587
它挺聰明的, 但還是比不過人類戴夫,
05:55
and manages to switch him off.
115
355918
1849
戴夫可以把 HAL 關掉。
06:01
But we might not be so lucky.
116
361648
1619
但我們可能就沒有這麽幸運了。
06:08
So what are we going to do?
117
368013
1592
那我們應該怎麽辦呢?
06:12
I'm trying to redefine AI
118
372191
2601
我想要重新定義人工智慧,
06:14
to get away from this classical notion
119
374816
2061
不再囿於傳統的概念:
06:16
of machines that intelligently pursue objectives.
120
376901
4567
能明智地達成目標的機器。
06:22
There are three principles involved.
121
382532
1798
新的定義涉及三條原則。
06:24
The first one is a principle of altruism, if you like,
122
384354
3289
第一個原則是利他主義原則,
06:27
that the robot's only objective
123
387667
3262
也就是說,機器的唯一目標
06:30
is to maximize the realization of human objectives,
124
390953
4246
就是要最大化地實現 人類的目標、人類的價值。
06:35
of human values.
125
395223
1390
06:36
And by values here I don't mean touchy-feely, goody-goody values.
126
396637
3330
這種價值不是指多愁善感 或者假裝乖巧,
06:39
I just mean whatever it is that the human would prefer
127
399991
3787
而是指人類所嚮往、追求的生活, 無論現狀如何。
06:43
their life to be like.
128
403802
1343
06:47
And so this actually violates Asimov's law
129
407184
2309
事實上,這樣就違反了艾西莫夫定律,
06:49
that the robot has to protect its own existence.
130
409517
2329
定律裡的機器人必須維護自己的生存。
06:51
It has no interest in preserving its existence whatsoever.
131
411870
3723
而在這條原則裡 機器對自身生存與否毫不關心。
06:57
The second law is a law of humility, if you like.
132
417240
3768
第二個原則,不妨稱之為謙遜原則。
07:01
And this turns out to be really important to make robots safe.
133
421794
3743
這一條對製造出安全的機器人十分重要。
07:05
It says that the robot does not know
134
425561
3142
它是指機器人不知道人類的價值是什麽,
07:08
what those human values are,
135
428727
2028
07:10
so it has to maximize them, but it doesn't know what they are.
136
430779
3178
它只知道將該價值最大化, 但卻不知道該價值究竟是什麽。
07:15
And that avoids this problem of single-minded pursuit
137
435074
2626
這就避免了「追求單一目的 而不知變通」的現象。
07:17
of an objective.
138
437724
1212
07:18
This uncertainty turns out to be crucial.
139
438960
2172
這種不確定性就變得很重要了。
07:21
Now, in order to be useful to us,
140
441546
1639
為了對我們有益,
07:23
it has to have some idea of what we want.
141
443209
2731
機械就得大概明白我們想要什麽。
07:27
It obtains that information primarily by observation of human choices,
142
447043
5427
它要獲取這類信息,主要是 透過觀察人類的決策,
07:32
so our own choices reveal information
143
452494
2801
所以我們的決策會揭露 我們生活的意願,
07:35
about what it is that we prefer our lives to be like.
144
455319
3300
07:40
So those are the three principles.
145
460452
1683
所以,這三條原則,
07:42
Let's see how that applies to this question of:
146
462159
2318
讓我們來看看要如何 應用到圖靈所說的問題:
07:44
"Can you switch the machine off?" as Turing suggested.
147
464501
2789
「你能不能將機器關掉?」
07:48
So here's a PR2 robot.
148
468893
2120
這是 PR2 機器人,
07:51
This is one that we have in our lab,
149
471037
1821
這是我們實驗室裡的其中一台,
07:52
and it has a big red "off" switch right on the back.
150
472882
2903
它的背面有一個大大的紅色開關。
07:56
The question is: Is it going to let you switch it off?
151
476361
2615
那問題來了:它會讓你把它關掉嗎?
07:59
If we do it the classical way,
152
479000
1465
如果我們用傳統的定義製造它,
08:00
we give it the objective of, "Fetch the coffee, I must fetch the coffee,
153
480489
3482
我們給它一個「去拿咖啡」的目標, 它會想:「我必須去拿咖啡,
08:03
I can't fetch the coffee if I'm dead,"
154
483995
2580
但如果我死了,就不能拿咖啡了。」
08:06
so obviously the PR2 has been listening to my talk,
155
486599
3341
看來, PR2 聽過我的演講了,
08:09
and so it says, therefore, "I must disable my 'off' switch,
156
489964
3753
因此它說:「我必須讓自己的開關失靈,
08:14
and probably taser all the other people in Starbucks
157
494796
2694
可能還要通過電擊把那些在 星巴克裡干擾我的人都擊暈。」
08:17
who might interfere with me."
158
497514
1560
08:19
(Laughter)
159
499098
2062
(笑聲)
08:21
So this seems to be inevitable, right?
160
501184
2153
這無法避免,對吧?
08:23
This kind of failure mode seems to be inevitable,
161
503361
2398
這種失敗看起來是必然的,
08:25
and it follows from having a concrete, definite objective.
162
505783
3543
因為機器人會遵循一個 十分明確的目標。
08:30
So what happens if the machine is uncertain about the objective?
163
510632
3144
那如果機器對目標 不那麽確定會發生什麽呢?
08:33
Well, it reasons in a different way.
164
513800
2127
那它的思路就不一樣了。
08:35
It says, "OK, the human might switch me off,
165
515951
2424
它會說:「好的,人類可能會把我關掉,
08:38
but only if I'm doing something wrong.
166
518964
1866
但只有我做錯事了,才會把我關掉。
08:41
Well, I don't really know what wrong is,
167
521567
2475
沒錯,我真的不知道什麽才是錯,
08:44
but I know that I don't want to do it."
168
524066
2044
但我知道我不該做錯的事。」
08:46
So that's the first and second principles right there.
169
526134
3010
這就是第一和第二原則。
08:49
"So I should let the human switch me off."
170
529168
3359
「所以我應該讓人類把我關掉。」
08:53
And in fact you can calculate the incentive that the robot has
171
533541
3956
事實上你可以推斷出機器人為了 允許讓人類關掉它所包含的動機,
08:57
to allow the human to switch it off,
172
537521
2493
09:00
and it's directly tied to the degree
173
540038
1914
而且這與根本目標的 不確定性程度直接相關。
09:01
of uncertainty about the underlying objective.
174
541976
2746
09:05
And then when the machine is switched off,
175
545797
2949
當機器被關閉後,
09:08
that third principle comes into play.
176
548770
1805
第三條原則就起作用了。
09:10
It learns something about the objectives it should be pursuing,
177
550599
3062
機器開始學習它應追求的目標,
09:13
because it learns that what it did wasn't right.
178
553685
2533
因為它知道它剛才做的事是不對的。
09:16
In fact, we can, with suitable use of Greek symbols,
179
556242
3570
實際上,我們可以適當地 使用些希臘字母,
09:19
as mathematicians usually do,
180
559836
2131
就像數學家們經常做的那樣,
09:21
we can actually prove a theorem
181
561991
1984
直接證明這一個理論:這樣的 機器人對人類是絕對有利的。
09:23
that says that such a robot is provably beneficial to the human.
182
563999
3553
09:27
You are provably better off with a machine that's designed in this way
183
567576
3803
可以證明如此設計出來的機器人, 對我們的生活是是有益的。
09:31
than without it.
184
571403
1246
09:33
So this is a very simple example, but this is the first step
185
573057
2906
這個例子很簡單,
但它是我們嘗試實現 能與人類和諧共處的 AI 的第一步。
09:35
in what we're trying to do with human-compatible AI.
186
575987
3903
09:42
Now, this third principle,
187
582477
3257
現在來看第三個原則,
09:45
I think is the one that you're probably scratching your head over.
188
585758
3112
我知道各位可能還在為 這一個原則傷腦筋。
09:48
You're probably thinking, "Well, you know, I behave badly.
189
588894
3239
你可能會想:「你懂的, 我行為舉止比較差勁。
09:52
I don't want my robot to behave like me.
190
592157
2929
我的機器人可不能被我帶壞。
09:55
I sneak down in the middle of the night and take stuff from the fridge.
191
595110
3434
我有時後會大半夜偷偷摸摸地 從冰箱裡找東西吃,
09:58
I do this and that."
192
598568
1168
東瞅瞅,西摸摸。」
09:59
There's all kinds of things you don't want the robot doing.
193
599760
2797
有各種各樣的事 你是不希望機器人去做的。
10:02
But in fact, it doesn't quite work that way.
194
602581
2071
但實際上不是那樣。
10:04
Just because you behave badly
195
604676
2155
你行為不檢,
10:06
doesn't mean the robot is going to copy your behavior.
196
606855
2623
不代表機器人就得有樣學樣。
10:09
It's going to understand your motivations and maybe help you resist them,
197
609502
3910
它會去嘗試理解你做事的動機,
而且可能會在合適的情況下 幫助你、制止你。
10:13
if appropriate.
198
613436
1320
但這仍然十分困難。
10:16
But it's still difficult.
199
616026
1464
10:18
What we're trying to do, in fact,
200
618122
2545
實際上,我們是要讓機器
10:20
is to allow machines to predict for any person and for any possible life
201
620691
5796
為任何人、任何一種 可能的生活去預測:
10:26
that they could live,
202
626511
1161
他們更想怎樣?更想要什麽?
10:27
and the lives of everybody else:
203
627696
1597
10:29
Which would they prefer?
204
629317
2517
10:33
And there are many, many difficulties involved in doing this;
205
633881
2954
這涉及到諸多困難,
10:36
I don't expect that this is going to get solved very quickly.
206
636859
2932
我不認為這會很快地就被解決。
10:39
The real difficulties, in fact, are us.
207
639815
2643
實際上,真正的困難是我們自己。
10:43
As I have already mentioned, we behave badly.
208
643969
3117
就像我剛說的那樣, 我們做事不守規矩。
10:47
In fact, some of us are downright nasty.
209
647110
2321
我們當中就有人是非常惡劣的。
10:50
Now the robot, as I said, doesn't have to copy the behavior.
210
650251
3052
如前所說,機器人 未必得要複製那些行為。
10:53
The robot does not have any objective of its own.
211
653327
2791
機器人沒有自己的目標,
10:56
It's purely altruistic.
212
656142
1737
它是完全利他的。
10:59
And it's not designed just to satisfy the desires of one person, the user,
213
659113
5221
它的誕生不僅僅是為了去滿足 某一個人、某一個用戶的欲望,
11:04
but in fact it has to respect the preferences of everybody.
214
664358
3138
而是去尊重所有人的意願。
11:09
So it can deal with a certain amount of nastiness,
215
669083
2570
所以它懂得抵制一些惡劣的行為,
11:11
and it can even understand that your nastiness, for example,
216
671677
3701
它甚至能理解你為什麼惡劣,比如說,
11:15
you may take bribes as a passport official
217
675402
2671
如果你是一個邊境護照官員, 你可能會收取賄賂,
11:18
because you need to feed your family and send your kids to school.
218
678097
3812
因為你得養家、供孩子們上學。
11:21
It can understand that; it doesn't mean it's going to steal.
219
681933
2906
機器人能理解這一點, 但不代表它也會學你偷錢,
11:24
In fact, it'll just help you send your kids to school.
220
684863
2679
它反而會幫助你去供孩子們上學。
11:28
We are also computationally limited.
221
688796
3012
我們的計算能力也是有限的。
11:31
Lee Sedol is a brilliant Go player,
222
691832
2505
李世石是一個傑出的圍棋大師,
11:34
but he still lost.
223
694361
1325
但他還是輸了。
11:35
So if we look at his actions, he took an action that lost the game.
224
695710
4239
如果我們仔細觀察他的棋路, 他下錯了那幾步以致輸棋,
11:39
That doesn't mean he wanted to lose.
225
699973
2161
但這不意味著他想要輸。
11:43
So to understand his behavior,
226
703160
2040
所以要理解他的行為,
11:45
we actually have to invert through a model of human cognition
227
705224
3644
我們得從人類認知的模型回推過來,
11:48
that includes our computational limitations -- a very complicated model.
228
708892
4977
它包含了我們計算能力上的局限,
是一個很覆雜的模型。
11:53
But it's still something that we can work on understanding.
229
713893
2993
但我們仍然可以嘗試去理解。
11:57
Probably the most difficult part, from my point of view as an AI researcher,
230
717696
4320
可能對於我這樣的 AI 研究人員來說,
12:02
is the fact that there are lots of us,
231
722040
2575
最大的困難是,人有很多種,
12:06
and so the machine has to somehow trade off, weigh up the preferences
232
726114
3581
所以機器必須想辦法去協調、 權衡不同人之間的喜好、需求,
12:09
of many different people,
233
729719
2225
12:11
and there are different ways to do that.
234
731968
1906
而要做到這一點有多種不同的方法。
12:13
Economists, sociologists, moral philosophers have understood that,
235
733898
3689
經濟學家、社會學家、 道德哲學家都理解這一點,
12:17
and we are actively looking for collaboration.
236
737611
2455
我們正積極地尋求合作。
12:20
Let's have a look and see what happens when you get that wrong.
237
740090
3251
讓我們來看看,如果我們把這一步 走錯了會怎麽樣。
12:23
So you can have a conversation, for example,
238
743365
2133
比如說,你可能會與你的 人工智慧助理有這樣的對話,
12:25
with your intelligent personal assistant
239
745522
1944
12:27
that might be available in a few years' time.
240
747490
2285
這樣的人工智慧可能幾年內就會出現。
12:29
Think of a Siri on steroids.
241
749799
2524
可以把它想成是強化版的 Siri 。
12:33
So Siri says, "Your wife called to remind you about dinner tonight."
242
753447
4322
Siri 對你說:「你老婆打電話 提醒你別忘了今天的晚宴。」
12:38
And of course, you've forgotten. "What? What dinner?
243
758436
2508
當然你早就忘了這回事:
「什麽?什麽晚宴?你在說什麽?」
12:40
What are you talking about?"
244
760968
1425
12:42
"Uh, your 20th anniversary at 7pm."
245
762417
3746
「呃.....今晚 7 點 慶祝結婚 20 周年。」
12:48
"I can't do that. I'm meeting with the secretary-general at 7:30.
246
768735
3719
「我可去不了, 我晚上 7 點半要見秘書長。
12:52
How could this have happened?"
247
772478
1692
怎麽會這樣呢?」
12:54
"Well, I did warn you, but you overrode my recommendation."
248
774194
4660
「呃,我可是提醒過你的, 但你沒有理會我的建議。」
12:59
"Well, what am I going to do? I can't just tell him I'm too busy."
249
779966
3328
「我該怎麽辦呢?我可不能跟秘書長說 我有事,沒空見他。」
13:04
"Don't worry. I arranged for his plane to be delayed."
250
784310
3281
「別擔心。我已經安排了, 讓他的航班延誤。」
13:07
(Laughter)
251
787615
1682
(笑聲)
13:10
"Some kind of computer malfunction."
252
790069
2101
「用某種電腦故障。」
13:12
(Laughter)
253
792194
1212
(笑聲)
13:13
"Really? You can do that?"
254
793430
1617
「真的嗎?這個你也能做到?」
13:16
"He sends his profound apologies
255
796220
2179
「秘書長很不好意思,跟你道歉,
13:18
and looks forward to meeting you for lunch tomorrow."
256
798423
2555
並邀請你明天中午吃飯。」
13:21
(Laughter)
257
801002
1299
(笑聲)
13:22
So the values here -- there's a slight mistake going on.
258
802325
4403
所以這裡談的價值觀就有點問題了,
13:26
This is clearly following my wife's values
259
806752
3009
這顯然是在遵循我老婆的價值觀,
13:29
which is "Happy wife, happy life."
260
809785
2069
也就是「老婆開心,生活舒心」。
13:31
(Laughter)
261
811878
1583
(笑聲)
13:33
It could go the other way.
262
813485
1444
它也有可能發展成另一種情況。
13:35
You could come home after a hard day's work,
263
815641
2201
你忙碌一天,回到家裏,
13:37
and the computer says, "Long day?"
264
817866
2195
電腦對你說:「今天很忙喔?」
13:40
"Yes, I didn't even have time for lunch."
265
820085
2288
「是啊,我連午飯都沒來得及吃。」
13:42
"You must be very hungry."
266
822397
1282
「那你一定很餓了吧。」
13:43
"Starving, yeah. Could you make some dinner?"
267
823703
2646
「快餓暈了。你能做點晚飯嗎?」
13:47
"There's something I need to tell you."
268
827890
2090
「有一件事我得告訴你。」
13:50
(Laughter)
269
830004
1155
(笑聲)
13:52
"There are humans in South Sudan who are in more urgent need than you."
270
832013
4905
「南蘇丹人民的情況 比你更緊急,更需要照顧。」
13:56
(Laughter)
271
836942
1104
(笑聲)
13:58
"So I'm leaving. Make your own dinner."
272
838070
2075
「所以我要走了。你自己做飯去吧。」
14:00
(Laughter)
273
840169
2000
(笑聲)
14:02
So we have to solve these problems,
274
842643
1739
我們得解決這類的問題,
14:04
and I'm looking forward to working on them.
275
844406
2515
我也很期待能解決這樣的問題。
14:06
There are reasons for optimism.
276
846945
1843
我們有理由感到樂觀。
14:08
One reason is,
277
848812
1159
理由之一是,
14:09
there is a massive amount of data.
278
849995
1868
我們有大量的數據資料。
14:11
Because remember -- I said they're going to read everything
279
851887
2794
記住,我說過機器將能夠 閱讀所有人類寫下來的東西。
14:14
the human race has ever written.
280
854705
1546
而我們寫下的文字大都類似於
14:16
Most of what we write about is human beings doing things
281
856275
2724
「人類做了一些事情 導致其他人對此感到沮喪」。
14:19
and other people getting upset about it.
282
859023
1914
14:20
So there's a massive amount of data to learn from.
283
860961
2398
所以機器可以從 大量的數據中去學習。
14:23
There's also a very strong economic incentive
284
863383
2236
同時從經濟的角度, 我們也有足夠的動機去做好這件事。
14:27
to get this right.
285
867151
1186
14:28
So imagine your domestic robot's at home.
286
868361
2001
想像一下,你家裡有個居家機器人。
14:30
You're late from work again and the robot has to feed the kids,
287
870386
3067
而你又得加班, 機器人得給孩子們做飯,
14:33
and the kids are hungry and there's nothing in the fridge.
288
873477
2823
孩子們很餓, 但冰箱裡什麽都沒有。
14:36
And the robot sees the cat.
289
876324
2605
然後機器人看到了家裡的貓。
14:38
(Laughter)
290
878953
1692
(笑聲)
14:40
And the robot hasn't quite learned the human value function properly,
291
880669
4190
機器人還沒學透人類的價值觀。
14:44
so it doesn't understand
292
884883
1251
所以它不知道, 貓的情感價值大於其營養價值。
14:46
the sentimental value of the cat outweighs the nutritional value of the cat.
293
886158
4844
14:51
(Laughter)
294
891026
1095
(笑聲)
14:52
So then what happens?
295
892145
1748
接下來會發生什麽事?
14:53
Well, it happens like this:
296
893917
3297
頭版頭條可能會是這樣:
14:57
"Deranged robot cooks kitty for family dinner."
297
897238
2964
「瘋狂機器人煮了貓咪當晚餐!」
15:00
That one incident would be the end of the domestic robot industry.
298
900226
4523
這場意外就足以結束 整個居家機器人的產業。
15:04
So there's a huge incentive to get this right
299
904773
3372
所以在我們實現超級 AI 之前, 我們有足夠的動機把它做對做好。
15:08
long before we reach superintelligent machines.
300
908169
2715
15:11
So to summarize:
301
911948
1535
總結來說:
15:13
I'm actually trying to change the definition of AI
302
913507
2881
我事實上想要改變人工智慧的定義,
15:16
so that we have provably beneficial machines.
303
916412
2993
這樣我們就可以製造出 對我們有益無害的機器人。
15:19
And the principles are:
304
919429
1222
這三個原則是:
15:20
machines that are altruistic,
305
920675
1398
機器是利他的,
15:22
that want to achieve only our objectives,
306
922097
2804
只想著實現我們的目標,
15:24
but that are uncertain about what those objectives are,
307
924925
3116
但它不確定我們的目標是什麽,
15:28
and will watch all of us
308
928065
1998
並且它會觀察我們,
15:30
to learn more about what it is that we really want.
309
930087
3203
從中學習我們想要的究竟是什麽。
15:34
And hopefully in the process, we will learn to be better people.
310
934193
3559
希望在這個過程中, 我們也能學會成為更好的人。
15:37
Thank you very much.
311
937776
1191
謝謝大家。
15:38
(Applause)
312
938991
3709
(掌聲)
15:42
Chris Anderson: So interesting, Stuart.
313
942724
1868
克里斯安德森:非常有意思,斯圖爾特。
15:44
We're going to stand here a bit because I think they're setting up
314
944616
3170
趁工作人員為下一位講者佈置的時候,
15:47
for our next speaker.
315
947810
1151
我們先站在這裡聊幾句。
15:48
A couple of questions.
316
948985
1538
我有幾個問題。
15:50
So the idea of programming in ignorance seems intuitively really powerful.
317
950547
5453
將「無知」編寫到程式中, 這種思想真的很有衝擊力。
15:56
As you get to superintelligence,
318
956024
1594
當機器人有超級智慧時,
15:57
what's going to stop a robot
319
957642
2258
還有什麽東西能阻檔機器人閱讀書籍,
15:59
reading literature and discovering this idea that knowledge
320
959924
2852
並了解到:博學比無知要好得多,
16:02
is actually better than ignorance
321
962800
1572
16:04
and still just shifting its own goals and rewriting that programming?
322
964396
4218
進而改變它的目標, 重新編寫自己的程式呢?
16:09
Stuart Russell: Yes, so we want it to learn more, as I said,
323
969512
6356
斯圖爾特拉塞爾:是的, 我們想要它去學習,就像我說的,
16:15
about our objectives.
324
975892
1287
讓機器人學習我們的目標,
16:17
It'll only become more certain as it becomes more correct,
325
977203
5521
只有在理解得越正確的時候, 它們才會更明確我們要的東西,
16:22
so the evidence is there
326
982748
1945
佐證擺在那裡,
16:24
and it's going to be designed to interpret it correctly.
327
984717
2724
並且我們使它能夠正確解讀這些目標。
16:27
It will understand, for example, that books are very biased
328
987465
3956
比如說,它能夠從書中的佐證 判斷出那些富含偏見的書,
16:31
in the evidence they contain.
329
991445
1483
16:32
They only talk about kings and princes
330
992952
2397
像是只講述國王、王子, 和男性精英白人之類的書。
16:35
and elite white male people doing stuff.
331
995373
2800
16:38
So it's a complicated problem,
332
998197
2096
所以這是一個複雜的問題,
16:40
but as it learns more about our objectives
333
1000317
3872
但當它更深入地學習我們的目標時,
16:44
it will become more and more useful to us.
334
1004213
2063
它會變得越來越有用。
16:46
CA: And you couldn't just boil it down to one law,
335
1006300
2526
CA:所以它十分複雜, 遠不足以濃縮成一條法則嗎?
16:48
you know, hardwired in:
336
1008850
1650
像是,把這樣的命令燒録進去:
16:50
"if any human ever tries to switch me off,
337
1010524
3293
「如果人類想把我關掉,
16:53
I comply. I comply."
338
1013841
1935
我要服從。我要服從。」
16:55
SR: Absolutely not.
339
1015800
1182
SR:絕對不行。
16:57
That would be a terrible idea.
340
1017006
1499
那將是一個很糟糕的主意。
16:58
So imagine that you have a self-driving car
341
1018529
2689
試想一下,你有一輛無人駕駛汽車,
17:01
and you want to send your five-year-old
342
1021242
2433
你想讓它送你五歲的孩子去幼稚園。
17:03
off to preschool.
343
1023699
1174
17:04
Do you want your five-year-old to be able to switch off the car
344
1024897
3101
你會希望你五歲的孩子 在汽車運行的過程中將它關閉嗎?
17:08
while it's driving along?
345
1028022
1213
應該不會吧。
17:09
Probably not.
346
1029259
1159
所以它得理解
17:10
So it needs to understand how rational and sensible the person is.
347
1030442
4703
下指令的人有多理智、有多講道理。
17:15
The more rational the person,
348
1035169
1676
這個人越理智,
17:16
the more willing you are to be switched off.
349
1036869
2103
它就越願意被你關掉。
17:18
If the person is completely random or even malicious,
350
1038996
2543
如果這個人是完全思緒混亂 或者甚至是有惡意的,
17:21
then you're less willing to be switched off.
351
1041563
2512
那它就不太願意被你關掉了。
17:24
CA: All right. Stuart, can I just say,
352
1044099
1866
CA:好吧。斯圖爾特,我得說,
17:25
I really, really hope you figure this out for us.
353
1045989
2314
我真的希望你為我們所有人, 找到解決的辦法。
17:28
Thank you so much for that talk. That was amazing.
354
1048327
2375
很感謝你的演講。 十分精彩。
SR:謝謝。
CA:謝謝。
17:30
SR: Thank you.
355
1050726
1167
17:31
(Applause)
356
1051917
1837
(掌聲)
關於本網站

本網站將向您介紹對學習英語有用的 YouTube 視頻。 您將看到來自世界各地的一流教師教授的英語課程。 雙擊每個視頻頁面上顯示的英文字幕,從那裡播放視頻。 字幕與視頻播放同步滾動。 如果您有任何意見或要求,請使用此聯繫表與我們聯繫。

https://forms.gle/WvT1wiN1qDtmnspy7