1
00:00:01,000 --> 00:00:05,466
我想講下
一個好多人都經歷過嘅感官錯覺

2
00:00:05,480 --> 00:00:06,920
當呢個錯覺嚟嗰陣

3
00:00:06,920 --> 00:00:09,360
我哋會唔識得留意危險

4
00:00:09,360 --> 00:00:14,320
我亦都想講下一個我認為駭人聽聞

5
00:00:14,320 --> 00:00:16,850
同時又好有可能會發生嘅情景

6
00:00:16,850 --> 00:00:19,666
呢個情景發生嘅話，唔係一件好事嚟

7
00:00:20,080 --> 00:00:22,536
你哋可能唔覺得我依家講緊嘅嘢恐怖

8
00:00:22,560 --> 00:00:24,640
反而覺得好型

9
00:00:25,200 --> 00:00:29,996
所以我想講下
我哋人類喺人工智能方面取得嘅成就

10
00:00:30,000 --> 00:00:31,776
最终會點樣摧毀我哋

11
00:00:31,800 --> 00:00:33,580
而事實上，我認為好難會見到

12
00:00:33,580 --> 00:00:36,960
佢哋唔會摧毀我哋
或者導致我哋自我毀滅

13
00:00:37,400 --> 00:00:39,256
依家你哋或者同我一樣

14
00:00:39,280 --> 00:00:41,936
覺得諗呢啲嘢好得意

15
00:00:41,960 --> 00:00:43,500
正因為覺得得意

16
00:00:43,500 --> 00:00:45,610
亦都成為咗問題嘅一部份

17
00:00:45,610 --> 00:00:48,010
你哋應該擔心你哋嘅反應至真！

18
00:00:48,010 --> 00:00:51,220
如果我喺呢場演講度話畀你哋聽

19
00:00:51,220 --> 00:00:54,106
因為氣候變化或者大災難嘅原因

20
00:00:54,106 --> 00:00:56,800
我哋會遭遇一場饑荒

21
00:00:56,800 --> 00:01:02,606
而你嘅孫，或者佢哋嘅孫
會好似咁樣生活

22
00:01:02,606 --> 00:01:04,170
你就唔會覺得

23
00:01:04,170 --> 00:01:08,220
「好有趣，我鍾意呢場 TED 演講。」

24
00:01:09,230 --> 00:01:10,840
饑荒一啲都唔有趣

25
00:01:11,650 --> 00:01:14,970
但科幻小說描繪嘅死亡就好有趣

26
00:01:14,980 --> 00:01:18,546
呢一刻，人工智能發展
最令我最困擾嘅係

27
00:01:18,546 --> 00:01:25,020
我哋面對近在眼前嘅危險似乎無動於衷

28
00:01:25,036 --> 00:01:27,006
雖然我喺你哋面前演講

29
00:01:27,006 --> 00:01:30,006
但我同你哋一樣都係冇反應

30
00:01:30,006 --> 00:01:32,670
成件事就好似我哋企喺兩道門前面

31
00:01:32,670 --> 00:01:36,876
喺一號門後面，我哋唔再發展智能機器

32
00:01:36,876 --> 00:01:38,490
因為某啲原因

33
00:01:38,490 --> 00:01:41,040
我哋電腦嘅硬件同軟件都停滯不前

34
00:01:41,040 --> 00:01:44,926
依家嚟諗一下點解呢種情況會發生

35
00:01:44,926 --> 00:01:48,910
即係話，因為智能同自動化好重要

36
00:01:48,910 --> 00:01:52,406
所以我哋會喺許可嘅情況之下
繼續改善科技

37
00:01:53,086 --> 00:01:55,520
咁究竟係乜嘢會阻止我哋？

38
00:01:55,533 --> 00:01:57,620
一個全面嘅核戰爭？

39
00:01:58,960 --> 00:02:00,860
一個全球流行病？

40
00:02:02,160 --> 00:02:03,890
一個小行星撞擊？

41
00:02:05,400 --> 00:02:08,020
Justin Bieber 做咗美國總統？

42
00:02:08,020 --> 00:02:09,936
（笑聲）

43
00:02:12,810 --> 00:02:15,330
之但係，如我哋所知

44
00:02:15,330 --> 00:02:17,480
有一啲嘢會摧毀文明

45
00:02:17,480 --> 00:02:19,360
你必須要想像

46
00:02:19,360 --> 00:02:25,876
如果我哋一代又一代人
永遠改善唔到科技

47
00:02:25,876 --> 00:02:28,086
情況會有幾嚴重

48
00:02:28,086 --> 00:02:29,610
幾乎可以確定嘅係

49
00:02:29,610 --> 00:02:32,186
呢個係人類史上最壞嘅事

50
00:02:32,186 --> 00:02:36,130
所以唯一嘅選擇
就係二號門後嘅做法

51
00:02:36,130 --> 00:02:41,066
我哋繼續年復一年升級改造智能機器

52
00:02:41,066 --> 00:02:42,440
到咗某個地步

53
00:02:42,440 --> 00:02:45,950
我哋就會整出比我哋仲要聰明嘅機器

54
00:02:45,950 --> 00:02:48,440
一旦我哋有咗比我哋自己
仲聰明嘅機器

55
00:02:48,440 --> 00:02:50,646
佢哋就會自我改良

56
00:02:50,646 --> 00:02:55,006
到時我哋就會面臨數學家 IJ Good 
講嘅「智能爆炸」危機

57
00:02:55,006 --> 00:02:57,966
即係話，改良過程唔再需要人類

58
00:02:57,966 --> 00:03:01,130
依家，經常會有人學呢張諷刺漫畫咁

59
00:03:01,130 --> 00:03:05,170
描繪叛變嘅機器人會攻擊我哋

60
00:03:05,170 --> 00:03:08,636
但係呢個唔係最有可能發生嘅情景

61
00:03:08,636 --> 00:03:12,926
我哋嘅機器唔會自動變惡

62
00:03:12,926 --> 00:03:18,076
所以問題在於我哋製造出
比我哋更加做到嘢嘅機器嘅時候

63
00:03:18,086 --> 00:03:22,436
佢哋目標上同我哋嘅細微分歧
會置我哋於死地

64
00:03:24,030 --> 00:03:26,690
就諗下我哋同螞蟻之間嘅關係︰

65
00:03:26,690 --> 00:03:28,080
我哋唔討厭佢哋

66
00:03:28,080 --> 00:03:30,674
我哋唔會傷害佢哋

67
00:03:30,674 --> 00:03:32,780
甚至我哋為咗唔傷害佢哋
而會受一啲苦

68
00:03:32,780 --> 00:03:34,514
例如我哋會為咗唔踩到佢哋
而跨過佢哋

69
00:03:34,514 --> 00:03:39,240
但係一旦佢哋嘅存在
同我哋嘅其中一個目標有嚴重衝突

70
00:03:39,240 --> 00:03:41,707
譬如話要起一棟咁樣嘅樓

71
00:03:41,707 --> 00:03:44,927
我哋諗都唔諗就殺死佢哋

72
00:03:44,927 --> 00:03:48,101
問題係，我哋終有一日整出嘅機器——

73
00:03:48,101 --> 00:03:50,176
無論佢哋自己有冇意識都好

74
00:03:50,176 --> 00:03:52,280
同樣會冷漠咁對待我哋

75
00:03:53,250 --> 00:03:57,030
依家，我估對於你哋大部份人嚟講
呢件情景都係遙不可及嘅

76
00:03:57,430 --> 00:04:03,240
我賭你哋當中有人質疑
超級智能嘅可能性

77
00:04:03,240 --> 00:04:05,440
更加唔好講
人類要避免超級智能

78
00:04:05,440 --> 00:04:08,990
但係你哋肯定會喺下面嘅假設當中
搵到一啲謬誤

79
00:04:08,990 --> 00:04:10,694
呢度一共有三個假設

80
00:04:11,944 --> 00:04:16,820
喺物理系統裏面，智能等如訊息處理

81
00:04:17,460 --> 00:04:19,370
但係，呢個超過咗假設

82
00:04:19,370 --> 00:04:23,449
因為我哋已經喺我哋嘅機器裏面
植入咗弱人工智能

83
00:04:23,449 --> 00:04:28,846
而且呢啲機器好多
已經處於一個超人類智能水平

84
00:04:28,846 --> 00:04:33,940
同時我哋知道僅僅係物質
就可以產生所謂嘅「一般智能」

85
00:04:33,940 --> 00:04:37,406
一種可以喺唔同領域之間
靈活思考嘅能力

86
00:04:37,406 --> 00:04:40,576
咁係因為我哋嘅大腦
已經可以做到，係唔係？

87
00:04:40,576 --> 00:04:44,776
我嘅意思係，大腦凈係得原子

88
00:04:44,776 --> 00:04:49,436
只要我哋繼續加設原子系統

89
00:04:49,436 --> 00:04:52,036
機器就可以有更加多智能行為

90
00:04:52,036 --> 00:04:55,026
除非進度有咩停頓

91
00:04:55,026 --> 00:04:59,136
否則我哋最終會喺機器裏面
建構出一般智能

92
00:04:59,146 --> 00:05:02,556
明白進度嘅快慢並唔影響係好重要

93
00:05:02,556 --> 00:05:06,206
因為任何過程都足以令我哋返唔到轉頭

94
00:05:06,206 --> 00:05:08,106
我哋唔需要按照摩爾定律進行

95
00:05:08,106 --> 00:05:10,126
我哋唔需要指數式增長

96
00:05:10,126 --> 00:05:11,686
我哋只需要繼續做

97
00:05:13,280 --> 00:05:16,630
第二個假設就係我哋會繼續做

98
00:05:16,630 --> 00:05:20,030
我哋會繼續改造我哋嘅智能機器

99
00:05:21,440 --> 00:05:25,090
而考慮到智能嘅價值…

100
00:05:25,090 --> 00:05:29,240
我係話，因為有智能
我哋至會珍重事物

101
00:05:29,240 --> 00:05:31,916
或者我哋需要智能
去保護我哋珍重嘅一切

102
00:05:31,916 --> 00:05:34,116
智能係我哋最有寶貴嘅資源

103
00:05:34,116 --> 00:05:36,200
所以我哋想繼續發展智能

104
00:05:36,200 --> 00:05:38,876
我哋有極需解決嘅問題

105
00:05:38,876 --> 00:05:42,546
例如我哋想治療類似阿茲海默症
同癌症嘅疾病

106
00:05:42,546 --> 00:05:44,610
我哋想認識經濟系統

107
00:05:44,610 --> 00:05:46,780
我哋想改善我哋嘅氣候科學

108
00:05:46,780 --> 00:05:49,366
所以如果可以做到嘅話
我哋會繼續發展智能

109
00:05:49,366 --> 00:05:52,656
件事亦都可以比喻為︰
列車已經開出，但冇刹車掣可以踩

110
00:05:53,836 --> 00:06:01,086
最終，我哋唔會去到
智能嘅頂峰或者高智能水平

111
00:06:01,726 --> 00:06:03,576
而呢個就係非常重要嘅觀察結果

112
00:06:03,576 --> 00:06:06,116
就係呢個結果
將我哋置於岌岌可危嘅境地

113
00:06:06,116 --> 00:06:10,926
亦令到我哋對於危險嘅觸覺唔可靠

114
00:06:10,926 --> 00:06:13,900
依家，就諗下史上最聰明嘅人

115
00:06:13,900 --> 00:06:18,310
幾乎喺每個人嘅名單上面
都會有 John von Neumann

116
00:06:18,310 --> 00:06:20,916
我嘅意思係 John von Neumann 
畀佢周圍嘅人嘅印象

117
00:06:20,916 --> 00:06:25,326
包括佢畀嗰個時代最犀利嘅數學家
同物理學家嘅印象

118
00:06:25,326 --> 00:06:27,300
都係有紀錄低嘅

119
00:06:27,300 --> 00:06:31,266
如果一半關於佢嘅故事有一半係真嘅

120
00:06:31,266 --> 00:06:32,356
咁毫無疑問

121
00:06:32,356 --> 00:06:34,366
佢係有史以來其中一個最聰明嘅人

122
00:06:34,366 --> 00:06:37,646
所以當我哋畫一幅比較智力嘅圖

123
00:06:37,646 --> 00:06:41,460
喺右邊高分嘅位置
我哋有 John von Neumann

124
00:06:41,460 --> 00:06:43,239
喺中間有你同我

125
00:06:44,269 --> 00:06:45,640
去到最左邊，我哋有雞仔

126
00:06:45,640 --> 00:06:46,816
（笑聲）

127
00:06:46,816 --> 00:06:48,906
係吖，就係一隻雞仔

128
00:06:48,906 --> 00:06:49,796
（笑聲）

129
00:06:49,796 --> 00:06:52,916
我冇理由將呢個演講搞到咁灰㗎

130
00:06:52,916 --> 00:06:56,560
（笑聲）

131
00:06:56,560 --> 00:07:03,550
但好有可能智力分佈
遠比我哋目前認知嘅廣

132
00:07:03,550 --> 00:07:06,750
如果我哋建造出
比我哋擁有更高智慧嘅機器

133
00:07:06,750 --> 00:07:11,566
佢哋嘅智力好有可能會
超越我哋認知嘅最高智力

134
00:07:11,566 --> 00:07:14,490
同埋以無法想像嘅方式超越我哋

135
00:07:15,140 --> 00:07:19,170
同樣重要嘅係
單憑運算速度就可以超越我哋

136
00:07:19,170 --> 00:07:21,246
啱唔啱？諗下如果我哋整咗一個

137
00:07:21,246 --> 00:07:29,886
冇哈佛或者麻省理工研究人員
咁聰明嘅超級人工智能

138
00:07:29,886 --> 00:07:34,326
但電路運行速度大概
比生化電路快一百萬倍

139
00:07:34,326 --> 00:07:39,446
所以呢個機器嘅思考速度應該會
比佢嘅創造者快大概一百萬倍

140
00:07:39,446 --> 00:07:41,296
所以如果佢運行一個星期

141
00:07:41,296 --> 00:07:48,376
佢就可以完成人類要兩萬年
先至完成得到嘅工作

142
00:07:49,410 --> 00:07:50,724
而我哋又點會明白

143
00:07:50,724 --> 00:07:55,360
人工智能係點樣完成咁龐大嘅運算呢？

144
00:07:56,610 --> 00:08:01,730
另一個令人擔憂嘅事，老實講

145
00:08:01,730 --> 00:08:04,486
就係…不如想像一下最好嘅情形

146
00:08:04,486 --> 00:08:09,560
想像一下我哋設計咗一個
冇安全問題嘅超級人工智能

147
00:08:09,560 --> 00:08:12,946
我哋第一次擁有完美嘅設計

148
00:08:12,946 --> 00:08:17,216
就好似我哋摞住
一個按照預期發展嘅神諭

149
00:08:18,200 --> 00:08:21,580
呢個機器仲會變成完美嘅慳力設備

150
00:08:21,580 --> 00:08:23,880
事關機器可以生產另一款機器出嚟

151
00:08:23,880 --> 00:08:25,683
做任何體力勞動

152
00:08:25,683 --> 00:08:27,396
兼由太陽能驅動

153
00:08:27,396 --> 00:08:29,826
成本仲同買原材料差唔多

154
00:08:29,826 --> 00:08:33,306
所以，我哋唔單止講緊咕哩勞力嘅終結

155
00:08:33,306 --> 00:08:36,616
我哋同時講緊大部份用腦工作嘅終結

156
00:08:37,196 --> 00:08:40,260
咁我哋人類面對工作削減
應該何去何從？

157
00:08:40,260 --> 00:08:44,946
我哋會好自由咁去掟飛盤 、同人按摩

158
00:08:44,946 --> 00:08:48,690
服食一啲 LSD 精神藥
同埋著上怪異服飾

159
00:08:48,690 --> 00:08:50,896
於是成個世界都會變成火人節嘅人咁

160
00:08:50,896 --> 00:08:53,026
（笑聲）

161
00:08:54,596 --> 00:08:56,700
頭先講到嘅嘢聽起上嚟好似好好咁

162
00:08:57,220 --> 00:08:58,550
但係撫心自問

163
00:08:58,550 --> 00:09:02,186
面對目前嘅經濟政治秩序
乜嘢會發生呢？

164
00:09:02,606 --> 00:09:04,816
似乎我哋會目睹

165
00:09:04,816 --> 00:09:10,336
我哋從未見過咁嚴重嘅
貧富懸殊同失業率

166
00:09:10,336 --> 00:09:15,446
如果呢筆新財富唔即時用嚟服務全人類

167
00:09:15,446 --> 00:09:19,170
就算一啲億萬富翁使好多錢
㨘靚商業雜誌嘅封面

168
00:09:19,170 --> 00:09:21,896
世界上其他人都要挨餓

169
00:09:21,896 --> 00:09:23,920
咁如果俄羅斯人或者中國人

170
00:09:23,920 --> 00:09:29,170
聽到矽谷嘅一啲公司
打算使用一個超級人工智能

171
00:09:29,170 --> 00:09:30,330
佢哋會點諗？


172
00:09:30,330 --> 00:09:35,216
呢個機器有能力用未見過嘅力度
發動地面或者網絡戰爭

173
00:09:35,216 --> 00:09:37,160
呢個係「勝者全取」嘅情況

174
00:09:38,120 --> 00:09:43,006
喺呢場人工智能較量中有六個月嘅優勢

175
00:09:43,006 --> 00:09:47,366
就係至少要做多人類五十萬年做到嘅嘢

176
00:09:47,366 --> 00:09:54,166
甚至只係關於人工智能突破嘅謠言
就可以令到人類亂起上嚟

177
00:09:54,920 --> 00:10:00,000
依家最驚人嘅一件事，我覺得

178
00:10:00,000 --> 00:10:05,986
就係人工智能研究人員
安定人心時講嘅說話

179
00:10:06,856 --> 00:10:10,250
佢哋成日話，因為我哋有時間
所以我哋唔需要擔心

180
00:10:10,250 --> 00:10:12,636
「乜你唔知有排咩？

181
00:10:12,636 --> 00:10:15,000
仲有五十年或者一百年先到。」

182
00:10:15,590 --> 00:10:17,010
一位研究人員曾經咁講︰

183
00:10:17,010 --> 00:10:21,000
「擔心人工智能嘅安全就好似
擔心火星人口爆棚一樣。」

184
00:10:22,038 --> 00:10:23,984
呢句嘢等如矽谷同你講︰

185
00:10:23,984 --> 00:10:26,194
「你十八廿二就杞人憂天！」

186
00:10:26,194 --> 00:10:27,150
（笑聲）

187
00:10:27,150 --> 00:10:33,586
冇人意識到
攞時間嚟到講完全係無稽之談

188
00:10:34,086 --> 00:10:37,736
如果智能凈係同處理訊息有關

189
00:10:37,736 --> 00:10:40,626
同埋我哋繼續改良我哋嘅機器嘅話

190
00:10:40,626 --> 00:10:44,116
我哋最終會生產到超級智能

191
00:10:44,116 --> 00:10:46,760
但我哋唔知道要用幾長時間

192
00:10:46,760 --> 00:10:50,896
先可以生產安全嘅超級智能

193
00:10:52,256 --> 00:10:53,420
等我再講多一次

194
00:10:53,420 --> 00:10:56,516
我哋唔知道要用幾長時間

195
00:10:56,516 --> 00:10:59,496
先可以生產安全嘅超級智能

196
00:11:01,176 --> 00:11:02,290
如果你仲未意識到

197
00:11:02,290 --> 00:11:04,700
五十年嘅概念已經唔同咗喇

198
00:11:04,700 --> 00:11:06,984
呢幅圖顯示咗以月份計嘅五十年

199
00:11:06,984 --> 00:11:09,370
先係 iPhone 面世至今嘅時間

200
00:11:09,370 --> 00:11:12,480
再係阿森一族出現係電視上嘅時間

201
00:11:12,480 --> 00:11:18,390
五十年不足以畀人類應對最大挑戰

202
00:11:19,560 --> 00:11:23,240
再一次，我哋對於有理由發生嘅事

203
00:11:23,240 --> 00:11:26,246
未有採取適當嘅情緒反應

204
00:11:26,246 --> 00:11:30,366
對此，電腦科學家 Stuart Russell 
有一個好嘅比喻

205
00:11:30,366 --> 00:11:35,146
佢話︰想像一下我哋收到
一個來自外星文明嘅信息

206
00:11:35,146 --> 00:11:36,206
上面寫住：

207
00:11:36,206 --> 00:11:38,196
「地球上嘅人類，

208
00:11:38,196 --> 00:11:41,666
我哋五十年之後會到達你哋嘅星球。

209
00:11:41,666 --> 00:11:42,600
請準備好。」

210
00:11:43,490 --> 00:11:46,636
咁我哋依家凈係會倒數外星人來臨？

211
00:11:47,556 --> 00:11:50,606
我哋應該更加緊張至係

212
00:11:52,626 --> 00:11:55,400
另一個我哋被告知唔使擔心嘅原因係

213
00:11:55,400 --> 00:11:57,636
呢啲機器只會識得
將我哋嘅價值觀傳開

214
00:11:57,636 --> 00:11:59,956
因為佢哋係我哋人類嘅附屬嘅一部分

215
00:11:59,956 --> 00:12:02,336
但同時佢哋會被植入我哋嘅大腦

216
00:12:02,336 --> 00:12:04,696
所以我哋會成為佢哋嘅邊緣系統

217
00:12:04,696 --> 00:12:09,360
依家使啲時間諗下
最安全同唯一審慎嘅做法

218
00:12:09,360 --> 00:12:14,130
而推薦嘅做法就係
直接將呢種科技植入我哋嘅大腦

219
00:12:14,680 --> 00:12:17,940
呢種做法可能係最安全同唯一審慎嘅

220
00:12:17,940 --> 00:12:21,246
但係喺你將佢植入你個腦之前

221
00:12:21,246 --> 00:12:24,886
科技嘅安全問題需要解決

222
00:12:24,886 --> 00:12:26,706
（笑聲）

223
00:12:26,706 --> 00:12:28,340
更深一層嘅問題係

224
00:12:28,340 --> 00:12:32,060
人工智能自己整超級人工智能

225
00:12:32,060 --> 00:12:35,876
似乎比整一個可以喺神經科學上

226
00:12:35,876 --> 00:12:39,620
同我哋腦部無縫接合嘅
超級人工智能簡單

227
00:12:41,120 --> 00:12:47,340
考慮到從事研發人工智能嘅公司
同政府好可能會互相競爭

228
00:12:47,340 --> 00:12:50,720
考慮到要贏呢場比賽就要贏成個世界

229
00:12:50,720 --> 00:12:53,574
同埋先假設如果你下一刻
唔會糟塌人工智能嘅成果

230
00:12:53,574 --> 00:12:57,850
咁樣，似乎更加簡單嘅事會完成咗先

231
00:12:58,490 --> 00:13:01,120
但唔好彩嘅係
我除咗叫大家反思呢個問題

232
00:13:01,120 --> 00:13:03,476
我就再冇辦法解決呢個問題

233
00:13:03,476 --> 00:13:05,716
我覺得我哋喺人工智能方面

234
00:13:05,716 --> 00:13:07,906
需要好似「曼哈頓計劃」咁嘅計劃

235
00:13:08,416 --> 00:13:12,076
唔係講點樣整人工智能
因為我認為人工智能終有一日會整到

236
00:13:12,076 --> 00:13:14,416
而係搞清楚點樣避免一場軍備競賽

237
00:13:14,416 --> 00:13:17,690
同埋往符合我哋利益嘅方向
發展人工智能

238
00:13:18,140 --> 00:13:22,626
當你講緊可以自我改造嘅超級人工智能

239
00:13:22,626 --> 00:13:27,126
我哋似乎只有一個機會
令到人工智能發展得安全

240
00:13:27,126 --> 00:13:28,266
就算發展得安全

241
00:13:28,266 --> 00:13:32,946
我哋都要接受
人工智能對經濟同政治產生嘅結果

242
00:13:33,806 --> 00:13:40,360
但係當我哋同意
訊息處理係智能嘅起步點

243
00:13:40,360 --> 00:13:46,160
同意一啲適當嘅計算系統係智能嘅基礎

244
00:13:46,160 --> 00:13:51,030
同意我哋會不斷完善人工智能

245
00:13:51,030 --> 00:13:57,850
同意將來有好多嘢超越我哋認知嘅

246
00:13:57,850 --> 00:14:03,130
咁我哋就必須要承認
我哋正喺度創造緊某種神明

247
00:14:03,130 --> 00:14:07,470
依家會係一個好時機
確保佢係可以同我哋共存嘅神明

248
00:14:08,030 --> 00:14:09,253
好多謝你哋

249
00:14:09,253 --> 00:14:11,296
（掌聲）