How to Keep AI Under Control | Max Tegmark | TED

170,530 views ใƒป 2023-11-02

TED


์•„๋ž˜ ์˜๋ฌธ์ž๋ง‰์„ ๋”๋ธ”ํด๋ฆญํ•˜์‹œ๋ฉด ์˜์ƒ์ด ์žฌ์ƒ๋ฉ๋‹ˆ๋‹ค.

๋ฒˆ์—ญ: Yu Beom Jeon ๊ฒ€ํ† : ์„ฑ์ค€ ์•ˆ
00:03
Five years ago,
0
3833
2461
5๋…„ ์ „
00:06
I stood on the TED stage
1
6294
1752
TED ๊ฐ•์—ฐ์— ๋‚˜์™€
00:08
and warned about the dangers of superintelligence.
2
8046
4379
์ดˆ์ง€๋Šฅ์˜ ์œ„ํ—˜์„ฑ์„ ๊ฒฝ๊ณ ํ–ˆ์—ˆ์Šต๋‹ˆ๋‹ค.
00:13
I was wrong.
3
13051
1710
์ œ๊ฐ€ ํ‹€๋ ธ์–ด์š”.
00:16
It went even worse than I thought.
4
16513
1752
์ƒ๊ฐ๋ณด๋‹ค ๋” ๋‚˜๋น ์กŒ์Šต๋‹ˆ๋‹ค.
00:18
(Laughter)
5
18306
1752
(์›ƒ์Œ)
00:20
I never thought governments would let AI companies get this far
6
20058
4379
์ •๋ถ€์—์„œ ์˜๋ฏธ ์žˆ๋Š” ๊ทœ์ œ๋ฅผ ์„ธ์šฐ์ง€ ์•Š๊ณ 
ํ˜„ ์ƒํ™ฉ๊นŒ์ง€ AI ํšŒ์‚ฌ๋ฅผ ๋ฐฉ์น˜ํ• ์ง€ ์ƒ๊ฐ์ง€๋„ ๋ชปํ–ˆ๊ณ 
00:24
without any meaningful regulation.
7
24479
2294
00:27
And the progress of AI went even faster than I predicted.
8
27732
4880
์ธ๊ณต์ง€๋Šฅ์€ ์˜ˆ์ƒ๋ณด๋‹ค ํ›จ์”ฌ ๋น ๋ฅด๊ฒŒ ๋ฐœ์ „ํ–ˆ์Šต๋‹ˆ๋‹ค.
00:32
Look, I showed this abstract landscape of tasks
9
32654
3587
์ œ ๊ฐ€์Šด์—๋Š” ์ž‘์—…์„ ์ง„ํ–‰ํ•˜๋Š” ๊ฐœ๋žต์ ์ธ ์ง€ํ˜•๋„๊ฐ€ ์žˆ๋Š”๋ฐ
00:36
where the elevation represented how hard it was for AI
10
36241
3128
๊ณ ๋„๊ฐ€ ๋†’์„์ˆ˜๋ก ์ธ๊ณต์ง€๋Šฅ์ด ์ธ๊ฐ„๊ณผ ๊ฐ™์€ ์ˆ˜์ค€์œผ๋กœ๋Š”
00:39
to do each task at human level.
11
39369
1919
์ฒ˜๋ฆฌํ•˜๊ธฐ ์–ด๋ ต๋‹ค๋Š” ๋œป์ž…๋‹ˆ๋‹ค.
00:41
And the sea level represented what AI could be back then.
12
41288
3753
ํ•ด์ˆ˜๋ฉด์€ ๋‹น์‹œ ์ธ๊ณต์ง€๋Šฅ์ด ์–ผ๋งˆ๋‚˜ ๋ฐœ์ „ํ–ˆ๋Š”์ง€๋ฅผ ๋‚˜ํƒ€๋ƒ…๋‹ˆ๋‹ค.
00:45
And boy or boy, has the sea been rising fast ever since.
13
45875
2962
๋†€๋ž๊ฒŒ๋„ ํ•ด์ˆ˜๋ฉด์€ ์—ญ๋Œ€ ์ตœ๊ณ ๋กœ ๋น ๋ฅด๊ฒŒ ์ƒ์Šน ์ค‘์ด๋ฉฐ
00:48
But a lot of these tasks have already gone blub blub blub blub blub blub.
14
48878
3587
์ด ์ž‘์—… ์ค‘ ์ƒ๋‹น์ˆ˜๋Š” ์ด๋ฏธ ํ•ด์ˆ˜๋ฉด ์•„๋ž˜๋กœ ๊ฐ€๋ผ์•‰์•˜์Šต๋‹ˆ๋‹ค.
00:52
And the water is on track to submerge all land,
15
52882
3921
๋•…์€ ์ „๋ถ€ ๋ฐ”๋‹ท๋ฌผ์— ์ž ๊ธธ ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋˜๋ฉฐ
00:56
matching human intelligence at all cognitive tasks.
16
56803
3420
์ธ๊ณต์ง€๋Šฅ์€ ๋ชจ๋“  ์ธ์ง€ ๊ณผ์ œ์—์„œ ๋น„์Šทํ•œ ์ˆ˜์ค€์„ ๋ณด์ž…๋‹ˆ๋‹ค.
01:00
This is a definition of artificial general intelligence, AGI,
17
60849
5756
์ด๊ฒƒ์ด ๋ฒ”์šฉ ์ธ๊ณต์ง€๋Šฅ์ธ AGI์˜ ์ •์˜๋กœ
01:06
which is the stated goal of companies like OpenAI,
18
66605
3837
OpenAI, Google DeepMind, Anthropic ๋“ฑ
๊ธฐ์—…์—์„œ๋Š” ๋ฒ”์šฉ ์ธ๊ณต์ง€๋Šฅ์„ ๋ชฉํ‘œ๋กœ ๋‚ด์„ธ์šฐ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
01:10
Google DeepMind and Anthropic.
19
70442
2002
01:12
And these companies are also trying to build superintelligence,
20
72819
3587
ํ•ด๋‹น ๊ธฐ์—…๋“ค์€ ์ดˆ์ง€๋Šฅ์„ ๊ตฌ์ถ•ํ•˜๊ธฐ ์œ„ํ•ด ๋…ธ๋ ฅํ•˜๊ณ  ์žˆ์œผ๋ฉฐ
01:16
leaving human intelligence far behind.
21
76448
2919
์ดˆ์ง€๋Šฅ์— ๋น„ํ•˜๋ฉด ์ธ๊ฐ„์˜ ์ง€๋Šฅ์€ ํ•œ์ฐธ ๋’ค์ณ์ ธ ์žˆ์ฃ .
01:19
And many think it'll only be a few years, maybe, from AGI to superintelligence.
22
79826
4379
๋งŽ์€ ์‚ฌ๋žŒ๋“ค์ด AGI๊ฐ€ ์ดˆ์ง€๋Šฅ์œผ๋กœ ๊ธˆ๋ฐฉ ๋ฐœ์ „ํ•  ๊ฒƒ์ด๋ผ๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.
01:24
So when are we going to get AGI?
23
84539
2795
๊ทธ๋ ‡๋‹ค๋ฉด ์–ธ์ œ์ฏค AGI๋ฅผ ์ด์šฉํ•  ์ˆ˜ ์žˆ์„๊นŒ์š”?
01:27
Well, until recently, most AI researchers thought it was at least decades away.
24
87375
5464
์ตœ๊ทผ๊นŒ์ง€๋งŒ ํ•ด๋„ AI ์—ฐ๊ตฌ์ž ๋Œ€๋ถ€๋ถ„์ด
์ ์–ด๋„ ์ˆ˜์‹ญ ๋…„ ํ›„๋ผ๊ณ  ์ƒ๊ฐํ–ˆ์Šต๋‹ˆ๋‹ค.
01:33
And now Microsoft is saying, "Oh, it's almost here."
25
93214
3504
Microsoft๋Š” AGI์˜ ์‹œ๋Œ€๊ฐ€ ๋‹ค๊ฐ€์™”๋‹ค๊ณ  ์ด์•ผ๊ธฐํ•ฉ๋‹ˆ๋‹ค.
01:36
We're seeing sparks of AGI in ChatGPT-4,
26
96760
3753
์šฐ๋ฆฌ๋Š” ChatGPT-4์— AGI๋ฅผ ์ ๊ทน ํ™œ์šฉํ•˜๋Š” ํ˜„์‹ค์„ ๋ชฉ๊ฒฉํ•˜๊ณ  ์žˆ์œผ๋ฉฐ
01:40
and the Metaculus betting site is showing the time left to AGI
27
100555
3796
์•ž์œผ๋กœ๋ฅผ ์˜ˆ์ธกํ•˜๋Š” ์‚ฌ์ดํŠธ์ธ Metaculus์— ๋”ฐ๋ฅด๋ฉด
AGI๊ฐ€ ๋„๋ž˜ํ•˜๋Š” ๋ฐ ๊ฑธ๋ฆฌ๋Š” ์‹œ๊ฐ„์ด
01:44
plummeting from 20 years away to three years away
28
104351
4129
์ง€๋‚œ 18๊ฐœ์›”๋งŒ์— 20๋…„์—์„œ 3๋…„์œผ๋กœ ๊ธ‰๊ฐํ–ˆ์Šต๋‹ˆ๋‹ค.
01:48
in the last 18 months.
29
108521
1544
01:50
And leading industry people are now predicting
30
110106
4922
์—…๊ณ„๋ฅผ ์„ ๋„ํ•˜๋Š” ๊ด€๊ณ„์ž๋“ค์€ ์ธ๊ณต์ง€๋Šฅ์ด ์ธ๊ฐ„์„ ์•ž์ง€๋ฅผ ๋•Œ๊นŒ์ง€
01:55
that we have maybe two or three years left until we get outsmarted.
31
115028
5047
์•ฝ 2~3๋…„ ๋‚จ์•˜๋‹ค๊ณ  ์˜ˆ์ธกํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
02:00
So you better stop talking about AGI as a long-term risk,
32
120116
4421
๋”ฐ๋ผ์„œ AGI๋ฅผ ์žฅ๊ธฐ์  ์œ„ํ—˜์œผ๋กœ ์—ฌ๊ธฐ์ง€ ์•Š๋Š” ๊ฒƒ์ด ์ข‹์œผ๋ฉฐ
02:04
or someone might call you a dinosaur stuck in the past.
33
124579
3170
๊ทธ๋Ÿฌ์ง€ ์•Š์œผ๋ฉด ๊ณผ๊ฑฐ์— ๊ฐ‡ํžŒ ๊ณต๋ฃก์ด๋ผ๋Š” ๋ง์„ ๋“ฃ๊ฒ ์ฃ .
02:08
It's really remarkable how AI has progressed recently.
34
128416
3837
์ตœ๊ทผ AI๊ฐ€ ์–ผ๋งˆ๋‚˜ ๋ฐœ์ „ํ–ˆ๋Š”์ง€ ์ƒ๊ฐํ•˜๋ฉด ์ •๋ง ๋†€๋ž์Šต๋‹ˆ๋‹ค.
02:12
Not long ago, robots moved like this.
35
132921
2419
์–ผ๋งˆ ์ „๊นŒ์ง€ ๋กœ๋ด‡์€ ์ด๋ ‡๊ฒŒ ์›€์ง์˜€์Šต๋‹ˆ๋‹ค.
02:15
(Music)
36
135382
2085
(์Œ์•…)
02:18
Now they can dance.
37
138551
1418
์ด์ œ๋Š” ์ถค์„ ์ถฅ๋‹ˆ๋‹ค.
02:20
(Music)
38
140804
2711
(์Œ์•…)
02:29
Just last year, Midjourney produced this image.
39
149979
3295
๋ฐ”๋กœ ์ž‘๋…„์— Midjourney์—์„œ ์ด ์ด๋ฏธ์ง€๋ฅผ ๋งŒ๋“ค์—ˆ์–ด์š”.
02:34
This year, the exact same prompt produces this.
40
154401
3253
์˜ฌํ•ด์—๋Š” ๋˜‘๊ฐ™์€ ํ”„๋กฌํ”„ํŠธ๋กœ ์ด ์ด๋ฏธ์ง€๋ฅผ ๋งŒ๋“ค์—ˆ์ฃ .
02:39
Deepfakes are getting really convincing.
41
159656
2544
๋”ฅํŽ˜์ดํฌ๋Š” ์ ์  ์ง„์งœ ๊ฐ™์•„์ง€๊ณ ์žˆ์Šต๋‹ˆ๋‹ค.
02:43
(Video) Deepfake Tom Cruise: Iโ€™m going to show you some magic.
42
163201
2920
(๋น„๋””์˜ค) ํ†ฐ ํฌ๋ฃจ์ฆˆ ๋”ฅํŽ˜์ดํฌ: ๋งˆ๋ฒ•์„ ๋ณด์—ฌ๋“œ๋ฆด๊ฒŒ์š”.
02:46
It's the real thing.
43
166913
1335
์ง„์งœ ๋™์ „์ž…๋‹ˆ๋‹ค.
02:48
(Laughs)
44
168289
2086
(์›ƒ์Œ)
02:50
I mean ...
45
170375
1293
๊ทธ๋Ÿฌ๋‹ˆ๊นŒ...
02:53
It's all ...
46
173920
1835
๋ชจ๋“  ๊ฒƒ์ด...
02:55
the real ...
47
175797
1710
์ง„์งœ...
02:57
thing.
48
177549
1168
- ํ˜„์‹ค์ž…๋‹ˆ๋‹ค. - ๋ง‰์Šค ํ…Œ๊ทธ๋งˆํฌ: ๊ทธ๋Ÿด๊นŒ์š”?
02:58
Max Tegmark: Or is it?
49
178717
1251
03:02
And Yoshua Bengio now argues
50
182387
2669
์š”์Šˆ์•„ ๋ฒค์ง€์˜ค์˜ ์ฃผ์žฅ์— ๋”ฐ๋ฅด๋ฉด
03:05
that large language models have mastered language
51
185056
3837
์ด์ œ ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ์€ ์–ธ์–ด์™€ ์ง€์‹์„ ์ˆ™๋‹ฌํ–ˆ์œผ๋ฉฐ
03:08
and knowledge to the point that they pass the Turing test.
52
188893
2795
ํŠœ๋ง ํ…Œ์ŠคํŠธ๋ฅผ ํ†ต๊ณผํ•  ์ •๋„๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค.
03:12
I know some skeptics are saying,
53
192355
1585
์ผ๋ถ€ ํšŒ์˜๋ก ์ž๋Š”
03:13
"Nah, they're just overhyped stochastic parrots
54
193940
2711
ํ™•๋ฅ ์„ ๊ณผ์žฅํ•ด์„œ ๋งํ•˜๋Š” ์•ต๋ฌด์ƒˆ๋กœ ์น˜๋ถ€ํ•˜๋ฉฐ
03:16
that lack a model of the world,"
55
196693
1877
ํ˜„์‹ค์˜ ๋ชจ๋ธ์ด ๋ถ€์กฑํ•˜๋‹ค๊ณ  ๋งํ•˜์ง€๋งŒ
03:18
but they clearly have a representation of the world.
56
198611
2878
ํ˜„์‹ค์„ ๋ช…ํ™•ํžˆ ํˆฌ์˜ํ•œ ๋ชจ๋ธ์ด ๊ฐ–์ถฐ์ ธ ์žˆ์Šต๋‹ˆ๋‹ค.
03:21
In fact, we recently found that Llama-2 even has a literal map of the world in it.
57
201531
5839
์ตœ๊ทผ Llama-2๋Š”
๋ง ๊ทธ๋Œ€๋กœ ์„ธ๊ณ„ ์ง€๋„๋ฅผ ๊ฐ–์ท„๋‹ค๋Š” ๊ฒƒ์„ ์•Œ๊ฒŒ ๋˜์—ˆ์ฃ .
03:28
And AI also builds
58
208121
3045
๋˜ํ•œ AI๋Š”
03:31
geometric representations of more abstract concepts
59
211207
3796
์ฐธ๊ณผ ๊ฑฐ์ง“์˜ ์ •์˜์— ๋Œ€ํ•ด ์‚ฌ๊ณ ํ•˜๋Š” ๋“ฑ
03:35
like what it thinks is true and false.
60
215003
3754
๋”์šฑ ์ถ”์ƒ์ ์ธ ๊ฐœ๋…์„ ๊ธฐํ•˜ํ•™์  ๊ตฌ์กฐ๋กœ ๊ตฌ์ถ•ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
03:40
So what's going to happen if we get AGI and superintelligence?
61
220467
5046
๊ทธ๋Ÿผ AGI์™€ ์ดˆ์ง€๋Šฅ์„ ์ด์šฉํ•˜๋ฉด ์–ด๋–ป๊ฒŒ ๋ ๊นŒ์š”?
03:46
If you only remember one thing from my talk, let it be this.
62
226514
3879
์ œ ๊ฐ•์—ฐ์—์„œ ๋”ฑ ํ•œ ๊ฐ€์ง€๋งŒ ๊ธฐ์–ตํ•ด์•ผ ํ•œ๋‹ค๋ฉด ์ด๊ฑธ ๊ธฐ์–ตํ•˜์„ธ์š”.
03:51
AI godfather, Alan Turing predicted
63
231311
3462
์ธ๊ณต์ง€๋Šฅ์˜ ๋Œ€๋ถ€์ธ ์•จ๋Ÿฐ ํŠœ๋ง์€
03:54
that the default outcome is the machines take control.
64
234814
4797
๊ธฐ๋ณธ ๊ฒฐ๊ณผ๋Š” ๊ธฐ๊ณ„๊ฐ€ ํ†ต์ œํ•  ๊ฒƒ์ด๋ผ๊ณ  ์˜ˆ์ธกํ–ˆ์Šต๋‹ˆ๋‹ค.
04:00
The machines take control.
65
240528
2211
๊ธฐ๊ณ„๊ฐ€ ์ฃผ๋„๊ถŒ์„ ์žก๋Š” ๊ฒƒ์ด์ฃ .
04:04
I know this sounds like science fiction,
66
244240
2336
SF ๊ฐ™์€ ์ด์•ผ๊ธฐ์ง€๋งŒ
04:06
but, you know, having AI as smart as GPT-4
67
246618
3503
GPT-4์ฒ˜๋Ÿผ ๋˜‘๋˜‘ํ•œ AI๋„
04:10
also sounded like science fiction not long ago.
68
250163
2919
์–ผ๋งˆ ์ „ ๊นŒ์ง€๋Š” SF ๊ฐ™์€ ์ด์•ผ๊ธฐ์˜€์Šต๋‹ˆ๋‹ค.
04:13
And if you think of AI,
69
253124
2378
์ธ๊ณต์ง€๋Šฅ์— ๋Œ€ํ•ด์„œ
04:15
if you think of superintelligence in particular, as just another technology,
70
255502
6047
์ดˆ์ง€๋Šฅ์— ๋Œ€ํ•ด์„œ, ํŠนํžˆ ์ดˆ์ง€๋Šฅ์„ ์ „๊ธฐ์™€ ๊ฐ™์€
04:21
like electricity,
71
261591
2419
๋˜ ๋‹ค๋ฅธ ๊ธฐ์ˆ ์ด๋ผ๊ณ  ์ƒ๊ฐํ•œ๋‹ค๋ฉด
04:24
you're probably not very worried.
72
264052
2002
ํฌ๊ฒŒ ๊ฑฑ์ •ํ•˜์ง€ ์•Š์œผ์‹ค ๊ฒ๋‹ˆ๋‹ค.
04:26
But you see,
73
266095
1168
ํ•˜์ง€๋งŒ ํŠœ๋ง์€ ์ดˆ์ง€๋Šฅ์„ ์ƒˆ๋กœ์šด ์ข…๊ณผ ๋น„์Šทํ•˜๊ฒŒ ์ƒ๊ฐํ–ˆ์Šต๋‹ˆ๋‹ค.
04:27
Turing thinks of superintelligence more like a new species.
74
267263
3837
04:31
Think of it,
75
271142
1168
์ƒ๊ฐํ•ด ๋ณด๋ฉด ์ธ๊ฐ„์€
04:32
we are building creepy, super capable,
76
272310
3879
์†Œ๋ฆ„ ๋ผ์น˜๊ฒŒ ๋Šฅ๋ ฅ์ด ๋›ฐ์–ด๋‚˜๋ฉฐ ๋„๋•์„ฑ ์—†๋Š” ์‚ฌ์ด์ฝ”ํŒจ์Šค๋ฅผ ๋งŒ๋“ค๊ณ  ์žˆ์œผ๋ฉฐ
04:36
amoral psychopaths
77
276231
1585
04:37
that don't sleep and think much faster than us,
78
277857
2711
์ž ๋„ ์ž์ง€ ์•Š๊ณ  ์ธ๊ฐ„๋ณด๋‹ค ์‚ฌ๊ณ ๋ ฅ๋„ ์ข‹์€๋ฐ
04:40
can make copies of themselves
79
280610
1418
์Šค์Šค๋กœ๋ฅผ ๋ณต์ œํ•  ์ˆ˜ ์žˆ๊ณ 
04:42
and have nothing human about them at all.
80
282070
2002
์ธ๊ฐ„์„ฑ์ด ์ „ํ˜€ ์—†๋Š” ๊ฒƒ๋“ค์ž…๋‹ˆ๋‹ค.
04:44
So what could possibly go wrong?
81
284072
1835
๊ทธ๋Ÿผ ๋ญ๊ฐ€ ์ž˜๋ชป๋  ์ˆ˜ ์žˆ์„๊นŒ์š”?
04:45
(Laughter)
82
285949
1543
(์›ƒ์Œ)
04:47
And it's not just Turing.
83
287951
1585
ํŠœ๋ง๋ฟ๋งŒ์ด ์•„๋‹™๋‹ˆ๋‹ค.
04:49
OpenAI CEO Sam Altman, who gave us ChatGPT,
84
289536
2919
ChatGPT๋ฅผ ๋งŒ๋“  OpenAI์˜ CEO ์ƒ˜ ์•ŒํŠธ๋งŒ์€
04:52
recently warned that it could be "lights out for all of us."
85
292497
3837
์ตœ๊ทผ ChatGPT๊ฐ€ ์ธ๊ฐ„์—๊ฒŒ ์œ„ํ˜‘์ด ๋  ์ˆ˜ ์žˆ๋‹ค๊ณ  ๊ฒฝ๊ณ ํ–ˆ์Šต๋‹ˆ๋‹ค.
04:57
Anthropic CEO, Dario Amodei, even put a number on this risk:
86
297126
3754
Anthropic์˜ CEO์ธ ๋‹ค๋ฆฌ์˜ค ์•„๋ชจ๋ฐ์ด๋Š” ์ˆ˜์น˜๊นŒ์ง€ ์ œ์‹œํ•˜๋ฉด์„œ
05:02
10-25 percent.
87
302090
2210
10~25%๋ผ๊ณ  ์ด์•ผ๊ธฐํ–ˆ์ฃ .
05:04
And it's not just them.
88
304300
1335
๊ทธ๋“ค๋ฟ๋งŒ์ด ์•„๋‹™๋‹ˆ๋‹ค.
05:05
Human extinction from AI went mainstream in May
89
305677
3086
5์›”์—๋Š” AGI CEO์™€ AI ์—ฐ๊ตฌ์›์ด ๋ชจ๋‘ ๋‚˜์„œ์„œ
05:08
when all the AGI CEOs and who's who of AI researchers
90
308763
4922
AI๋กœ ์ธํ•œ ์ธ๊ฐ„ ๋ฉธ์ข…์— ๋Œ€ํ•ด ๊ฒฝ๊ณ ํ•˜๋ฉด์„œ
05:13
came on and warned about it.
91
313685
1376
์ฃผ๋ฅ˜ ์˜๊ฒฌ์œผ๋กœ ๋– ์˜ฌ๋ž์Šต๋‹ˆ๋‹ค.
05:15
And last month, even the number one of the European Union
92
315061
2920
์ง€๋‚œ๋‹ฌ์—๋Š” ์œ ๋Ÿฝ ์—ฐํ•ฉ ์ง‘ํ–‰์œ„์›์žฅ๋„
05:18
warned about human extinction by AI.
93
318022
3295
์ธ๊ณต์ง€๋Šฅ์— ์˜ํ•œ ์ธ๊ฐ„ ๋ฉธ์ข…์— ๋Œ€ํ•ด ๊ฒฝ๊ณ ํ–ˆ์Šต๋‹ˆ๋‹ค.
05:21
So let me summarize everything I've said so far
94
321359
2211
์ง€๊ธˆ๊นŒ์ง€ ๋ง์”€๋“œ๋ฆฐ ๋ชจ๋“  ๋‚ด์šฉ์„
05:23
in just one slide of cat memes.
95
323611
2544
๊ณ ์–‘์ด ๋ฐˆ์„ ํ™œ์šฉํ•œ ์Šฌ๋ผ์ด๋“œ๋กœ ์š”์•ฝํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.
05:27
Three years ago,
96
327282
1793
3๋…„ ์ „
05:29
people were saying it's inevitable, superintelligence,
97
329117
4129
์‚ฌ๋žŒ๋“ค์€ ์ดˆ์ง€๋Šฅ์„ ํ”ผํ•  ์ˆ˜ ์—†๋Š” ์ผ์ด๋ผ๊ณ  ๋งํ–ˆ์Šต๋‹ˆ๋‹ค.
05:33
it'll be fine,
98
333288
1501
๋ณ„ ๋ฌธ์ œ ์—†์„ ๊ฒƒ์ด๋ฉฐ
05:34
it's decades away.
99
334789
1210
์ˆ˜์‹ญ ๋…„์€ ๋‚จ์•˜๋‹ค๊ณ  ํ–ˆ์ฃ .
05:35
Last year it was more like,
100
335999
1835
์ž‘๋…„์—๋Š” ์ด๋žฌ์Šต๋‹ˆ๋‹ค.
05:37
It's inevitable, it'll be fine.
101
337876
2043
์–ด์ฉ” ์ˆ˜ ์—†์–ด์š”, ๊ดœ์ฐฎ์„ ๊ฑฐ์˜ˆ์š”.
05:40
Now it's more like,
102
340295
2460
์ง€๊ธˆ์€ ์ด๋ ‡๊ฒŒ ๋งํ•ฉ๋‹ˆ๋‹ค.
05:42
It's inevitable.
103
342797
1251
ํ•„์—ฐ์ ์ธ ์ผ์ด์˜ˆ์š”.
05:44
(Laughter)
104
344090
1126
(์›ƒ์Œ)
05:47
But let's take a deep breath and try to raise our spirits
105
347260
3962
ํ•˜์ง€๋งŒ ์ˆจ์„ ๊นŠ์ด ๋“ค์ด์‰ฌ๊ณ 
๊ธฐ์šด์„ ๋ถ๋‹์›Œ ๋ณด์ฃ .
05:51
and cheer ourselves up,
106
351264
1168
05:52
because the rest of my talk is going to be about the good news,
107
352473
3045
์ด์ œ ํฌ์†Œ์‹์„ ์ด์•ผ๊ธฐํ•ด ๋“œ๋ฆด ํ…๋ฐ์š”.
05:55
that it's not inevitable, and we can absolutely do better,
108
355560
2919
ํ”ผํ•  ์ˆ˜ ์—†๋Š” ์ผ์€ ์•„๋‹ˆ๋ฉฐ ๋” ์ž˜ ํ—ค์ณ๋‚˜๊ฐˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค, ์•Œ์•˜์ฃ ?
05:58
alright?
109
358521
1168
06:00
(Applause)
110
360315
2002
(๋ฐ•์ˆ˜)
06:02
So ...
111
362317
1209
๊ทธ๋Ÿฌ๋‹ˆ๊นŒ...
06:04
The real problem is that we lack a convincing plan for AI safety.
112
364903
5296
์ง„์งœ ๋ฌธ์ œ๋Š” AI ์•ˆ์ „์— ๋Œ€ํ•œ ํ™•์‹คํ•œ ๊ณ„ํš์ด ์—†๋‹ค๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค.
06:10
People are working hard on evals
113
370700
3337
์‚ฌ๋žŒ๋“ค์€ ์œ„ํ—˜ํ•œ AI ํ–‰๋™์„ ์‹๋ณ„ํ•˜๋Š” ํ‰๊ฐ€์—
06:14
looking for risky AI behavior, and that's good,
114
374037
4087
๋…ธ๋ ฅ์„ ๊ธฐ์šธ๊ธฐ์ด๊ณ  ์žˆ์œผ๋ฉฐ, ๊ธ์ •์ ์ธ ์ž์„ธ์ž…๋‹ˆ๋‹ค.
06:18
but clearly not good enough.
115
378124
2044
ํ•˜์ง€๋งŒ ์ถฉ๋ถ„์น˜ ๋ชปํ•œ ๊ฒƒ์€ ๋ถ„๋ช…ํ•˜์ฃ .
06:20
They're basically training AI to not say bad things
116
380209
4797
AI๋ฅผ ํ•™์Šต์‹œํ‚ฌ ๋•Œ๋Š” ๋‚˜์œ ์ผ์„ ํ•˜์ง€ ์•Š๋Š” ๊ฒƒ ๋ณด๋‹ค๋Š”
06:25
rather than not do bad things.
117
385006
2502
๋‚˜์œ ๋ง์„ ํ•˜์ง€ ์•Š๋„๋ก ํ•™์Šต์‹œํ‚ต๋‹ˆ๋‹ค.
06:28
Moreover, evals and debugging are really just necessary,
118
388176
4212
๊ฒŒ๋‹ค๊ฐ€ ํ‰๊ฐ€์™€ ๋””๋ฒ„๊น…์€ ์•ˆ์ „์„ ์œ„ํ•œ ํ•„์š” ์กฐ๊ฑด์ผ ๋ฟ
06:32
not sufficient, conditions for safety.
119
392430
2002
์•ˆ์ „์˜ ์ถฉ๋ถ„ ์กฐ๊ฑด์€ ์•„๋‹™๋‹ˆ๋‹ค.
06:34
In other words,
120
394474
1751
๋‹ค์‹œ ๋งํ•˜์ž๋ฉด
06:36
they can prove the presence of risk,
121
396225
3671
์œ„ํ—˜์˜ ๋ถ€์žฌ๊ฐ€ ์•„๋‹ˆ๋ผ ์œ„ํ—˜์˜ ์กด์žฌ๋ฅผ
06:39
not the absence of risk.
122
399938
2168
์ฆ๋ช…ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
06:42
So let's up our game, alright?
123
402148
2544
๋” ๊นŠ๊ฒŒ ๋“ค์–ด๊ฐ€๋ณผ๊นŒ์š”?
06:44
Try to see how we can make provably safe AI that we can control.
124
404692
5631
์ œ์–ดํ•  ์ˆ˜ ์žˆ์–ด ์•ˆ์ „์„ฑ์ด ์ž…์ฆ๋œ AI๋ฅผ
๋งŒ๋“œ๋Š” ๋ฐฉ๋ฒ•์„ ํ™•์ธํ•ด ๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค.
06:50
Guardrails try to physically limit harm.
125
410323
5047
๊ฐ€๋“œ๋ ˆ์ผ๋กœ๋Š” ๋ฌผ๋ฆฌ์  ํ”ผํ•ด๋ฅผ ๋ง‰์œผ๋ ค๋Š” ์‹œ๋„๋ฅผ ํ•ฉ๋‹ˆ๋‹ค.
06:55
But if your adversary is superintelligence
126
415828
2211
ํ•˜์ง€๋งŒ ์ƒ๋Œ€๋ฐฉ์ด ์ดˆ์ง€๋Šฅ์ด๊ฑฐ๋‚˜
06:58
or a human using superintelligence against you, right,
127
418081
2544
์ดˆ์ง€๋Šฅ์„ ์ด์šฉํ•ด ๊ณต๊ฒฉํ•˜๋Š” ์ธ๊ฐ„์˜ ๊ฒฝ์šฐ์—๋Š”
07:00
trying is just not enough.
128
420625
1960
์‹œ๋„๋งŒ์œผ๋กœ๋Š” ์ถฉ๋ถ„ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
07:02
You need to succeed.
129
422585
1877
์„ฑ๊ณตํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
07:04
Harm needs to be impossible.
130
424504
2169
ํ”ผํ•ด๋ฅผ ์ค„ ์ˆ˜ ์—†์–ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.
07:06
So we need provably safe systems.
131
426714
2544
์ž…์ฆํ•  ์ˆ˜ ์žˆ๋Š” ์•ˆ์ „ํ•œ ์‹œ์Šคํ…œ์ด ํ•„์š”ํ•˜์ฃ .
07:09
Provable, not in the weak sense of convincing some judge,
132
429258
3838
์ž…์ฆ์ด๋ž€ ํŒ์‚ฌ๋ฅผ ์„ค๋“ํ•˜๋Š” ๋“ฑ ์•ฝํ•œ ์˜๋ฏธ๊ฐ€ ์•„๋‹Œ
07:13
but in the strong sense of there being something that's impossible
133
433137
3128
๋ฌผ๋ฆฌ ๋ฒ•์น™์— ๋”ฐ๋ผ ๋ถˆ๊ฐ€๋Šฅํ•˜๋‹ค๋Š”
07:16
according to the laws of physics.
134
436265
1585
๊ฐ•๋ ฅํ•œ ์˜๋ฏธ๋ฅผ ๊ฐ€๋ฆฌํ‚ต๋‹ˆ๋‹ค.
07:17
Because no matter how smart an AI is,
135
437892
2002
์ธ๊ณต์ง€๋Šฅ์ด ์•„๋ฌด๋ฆฌ ๋˜‘๋˜‘ํ•ด๋„
07:19
it can't violate the laws of physics and do what's provably impossible.
136
439936
4046
๋ฌผ๋ฆฌ ๋ฒ•์น™์„ ๊ฑฐ์Šฌ๋Ÿฌ ๋ถˆ๊ฐ€๋Šฅ์ด ์ฆ๋ช…๋œ ์ผ์„ ํ•ด๋‚ผ ์ˆ˜๋Š” ์—†์ฃ .
07:24
Steve Omohundro and I wrote a paper about this,
137
444440
2836
์Šคํ‹ฐ๋ธŒ ์˜ค๋ชจํ—Œ๋“œ๋กœ์™€ ์ €๋Š” ์ด์— ๋Œ€ํ•œ ๋…ผ๋ฌธ์„ ์ผ๊ณ 
07:27
and we're optimistic that this vision can really work.
138
447318
5005
๋น„์ „์ด ์‹ค์ œ๋กœ ํšจ๊ณผ ์žˆ์„ ๊ฒƒ์ด๋ผ๊ณ  ๋‚™๊ด€ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
07:32
So let me tell you a little bit about how.
139
452323
2169
๋ฐฉ๋ฒ•์— ๋Œ€ํ•ด ์กฐ๊ธˆ ๋ง์”€๋“œ๋ฆฌ์ฃ .
07:34
There's a venerable field called formal verification,
140
454993
4421
์ •ํ˜• ๊ฒ€์ฆ์ด๋ผ๋Š” ๊ถŒ์œ„ ์žˆ๋Š” ๋ถ„์•ผ์—์„œ๋Š”
07:39
which proves stuff about code.
141
459455
2127
์ฝ”๋“œ์— ๋Œ€ํ•œ ์ •๋ณด๋ฅผ ์ฆ๋ช…ํ•ฉ๋‹ˆ๋‹ค.
07:41
And I'm optimistic that AI will revolutionize automatic proving business
142
461624
6548
์ €๋Š” AI๋ฅผ ํ†ตํ•ด ์ž๋™ ๊ฒ€์ฆ ๋น„์ฆˆ๋‹ˆ์Šค์—์„œ ํ˜๋ช…์„ ์ผ์œผํ‚ค๊ณ 
07:48
and also revolutionize program synthesis,
143
468214
3337
๋งค์šฐ ์–‘์งˆ์˜ ์ฝ”๋“œ๋ฅผ ์ž๋™ ์ž‘์„ฑํ•˜๋Š” ํ”„๋กœ๊ทธ๋žจ ํ•ฉ์„ฑ ๋Šฅ๋ ฅ์—์„œ๋„
07:51
the ability to automatically write really good code.
144
471592
3254
ํ˜๋ช…์„ ์ผ์œผํ‚ฌ ๊ฒƒ์ด๋ผ๊ณ  ๋‚™๊ด€ํ•ฉ๋‹ˆ๋‹ค.
07:54
So here is how our vision works.
145
474887
1585
๋น„์ „์˜ ์›๋ฆฌ๋Š” ์ด๋ ‡์Šต๋‹ˆ๋‹ค.
07:56
You, the human, write a specification
146
476472
4213
์šฐ๋ฆฌ ์ธ๊ฐ„์€
์‚ฌ์–‘์„ ์ž‘์„ฑํ•ด
08:00
that your AI tool must obey,
147
480685
2711
AI๊ฐ€ ๋ฐ˜๋“œ์‹œ ๋”ฐ๋ฅด๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.
08:03
that it's impossible to log in to your laptop
148
483438
2127
์˜ฌ๋ฐ”๋ฅธ ์•”ํ˜ธ ์—†์ด๋Š” ๋…ธํŠธ๋ถ์— ๋กœ๊ทธ์ธํ•  ์ˆ˜ ์—†๋Š” ์‚ฌ์–‘์ด๋‚˜
08:05
without the correct password,
149
485565
1793
08:07
or that a DNA printer cannot synthesize dangerous viruses.
150
487400
5714
DNA ํ”„๋ฆฐํ„ฐ๋กœ ์œ„ํ—˜ํ•œ ๋ฐ”์ด๋Ÿฌ์Šค๋ฅผ ํ•ฉ์„ฑํ•  ์ˆ˜ ์—†๋‹ค๋Š” ์‚ฌ์–‘์„ ์ž‘์„ฑํ•ฉ๋‹ˆ๋‹ค.
08:13
Then a very powerful AI creates both your AI tool
151
493156
5213
๊ทธ๋Ÿฌ๋ฉด ๋งค์šฐ ๊ฐ•๋ ฅํ•œ AI๊ฐ€ AI ๋„๊ตฌ๋ฅผ ๋งŒ๋“ค๊ณ 
08:18
and a proof that your tool meets your spec.
152
498369
3837
๋„๊ตฌ๊ฐ€ ์‚ฌ์–‘์„ ์ถฉ์กฑํ•œ๋‹ค๋Š” ์ฆ๊ฑฐ๋ฅผ ๋ชจ๋‘ ๋งŒ๋“ค์–ด๋ƒ…๋‹ˆ๋‹ค.
08:22
Machine learning is uniquely good at learning algorithms,
153
502540
4254
๋จธ์‹ ๋Ÿฌ๋‹์€ ์•Œ๊ณ ๋ฆฌ์ฆ˜ ํ•™์Šต์— ๋งค์šฐ ์ ํ•ฉํ•˜๋ฉฐ
08:26
but once the algorithm has been learned,
154
506836
2169
์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํ•™์Šตํ•œ ํ›„์—๋Š”
08:29
you can re-implement it in a different computational architecture
155
509047
3169
๊ฒ€์ฆํ•˜๊ธฐ ์‰ฌ์šด ๋‹ค๋ฅธ ๊ณ„์‚ฐ ์•„ํ‚คํ…์ฒ˜์—
08:32
that's easier to verify.
156
512216
1627
๋‹ค์‹œ ๊ตฌํ˜„ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
08:35
Now you might worry,
157
515344
1210
โ€˜๋‚ด๊ฐ€ ์–ด๋–ป๊ฒŒ ์ดํ•ดํ•  ์ˆ˜ ์žˆ์„๊นŒโ€™๋ผ๋Š” ๊ฑฑ์ •์„ ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
08:36
how on earth am I going to understand this powerful AI
158
516554
3921
๊ฐ•๋ ฅํ•œ AI์™€ AI๊ฐ€ ๋งŒ๋“  ๊ฐ•๋ ฅํ•œ AI ๋„๊ตฌ, ์ฆ๊ฑฐ๊ฐ€
08:40
and the powerful AI tool it built
159
520475
1626
08:42
and the proof,
160
522143
1126
์ดํ•ดํ•˜๊ธฐ์— ๋„ˆ๋ฌด ๋ณต์žกํ•˜๋‹ค๋ฉด ์–ด๋–ป๊ฒŒ ์ดํ•ดํ•  ์ˆ˜ ์žˆ์„๊นŒ์š” ?
08:43
if they're all too complicated for any human to grasp?
161
523311
2794
08:46
Here is the really great news.
162
526147
2127
์ •๋ง ์ข‹์€ ์†Œ์‹์ด ์žˆ์Šต๋‹ˆ๋‹ค.
08:48
You don't have to understand any of that stuff,
163
528316
2461
์–ด๋–ค ๊ฒƒ๋„ ์ดํ•ดํ•  ํ•„์š”๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค.
08:50
because it's much easier to verify a proof than to discover it.
164
530818
5297
์ฆ๊ฑฐ๋ฅผ ๋ฐœ๊ฒฌํ•˜๋Š” ๊ฒƒ๋ณด๋‹ค ์ฆ๋ช…ํ•˜๋Š” ๊ฒƒ์ด ํ›จ์”ฌ ์‰ฝ๊ธฐ ๋•Œ๋ฌธ์ด์ฃ .
08:56
So you only have to understand or trust your proof-checking code,
165
536115
5089
์ˆ˜๋ฐฑ ์ค„์— ๋ถˆ๊ณผํ•œ ์ฆ๋ช… ๊ฒ€์‚ฌ ์ฝ”๋“œ๋ฅผ
09:01
which could be just a few hundred lines long.
166
541245
2211
์ดํ•ดํ•˜๊ฑฐ๋‚˜ ๋ฏฟ๊ธฐ๋งŒ ํ•˜๋ฉด ๋ฉ๋‹ˆ๋‹ค.
09:03
And Steve and I envision
167
543498
2252
์Šคํ‹ฐ๋ธŒ์™€ ์ €๋Š”
09:05
that such proof checkers get built into all our compute hardware,
168
545750
4463
์ด๋Ÿฌํ•œ ์ฆ๋ช… ๊ฒ€์‚ฌ๊ธฐ๊ฐ€ ๋ชจ๋“  ์ปดํ“จํŒ… ํ•˜๋“œ์›จ์–ด์— ๋‚ด์žฅ๋˜์–ด
09:10
so it just becomes impossible to run very unsafe code.
169
550254
4213
๋งค์šฐ ์œ„ํ—˜ํ•œ ์ฝ”๋“œ๋ฅผ ์‹คํ–‰ํ•  ์ˆ˜ ์—†๊ฒŒ ๋œ๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.
09:14
What if the AI, though, isn't able to write that AI tool for you?
170
554509
5505
ํ•˜์ง€๋งŒ AI๊ฐ€ ๊ทธ AI ๋„๊ตฌ๋ฅผ ๋Œ€์‹  ์ž‘์„ฑํ•ด ์ค„ ์ˆ˜ ์—†๋‹ค๋ฉด ์–ด๋–จ๊นŒ์š”?
09:20
Then there's another possibility.
171
560056
3795
๊ทธ๋ ‡๋‹ค๋ฉด ๋˜ ๋‹ค๋ฅธ ๊ฐ€๋Šฅ์„ฑ๋„ ์žˆ์Šต๋‹ˆ๋‹ค.
09:23
You train an AI to first just learn to do what you want
172
563851
3587
AI์— ์›ํ•˜๋Š” ๊ฒƒ์„ ๋จผ์ € ํ•™์Šต์‹œํ‚จ ๋‹ค์Œ
09:27
and then you use a different AI
173
567480
3337
๋‹ค๋ฅธ AI๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ
09:30
to extract out the learned algorithm and knowledge for you,
174
570858
3963
์‚ฌ์šฉํ•  ํ•™์Šต๋œ ์•Œ๊ณ ๋ฆฌ์ฆ˜๊ณผ ์ง€์‹์„ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค.
09:34
like an AI neuroscientist.
175
574862
2086
AI ์‹ ๊ฒฝ๊ณผํ•™์ž๊ฐ™์ฃ .
09:37
This is in the spirit of the field of mechanistic interpretability,
176
577281
3879
์ •๋ง ๋†€๋ผ์šธ ์ •๋„๋กœ ๋น ๋ฅธ ๋ฐœ์ „์„ ์ด๋ฃจ๊ณ  ์žˆ๋Š”
09:41
which is making really impressive rapid progress.
177
581160
3253
๊ธฐ๊ณ„๋ก ์  ํ•ด์„๊ฐ€๋Šฅ์„ฑ ๋ถ„์•ผ์˜ ํ•ต์‹ฌ ๊ฐœ๋…์ด ๋ฐ”๋กœ ์ด๊ฒƒ์ž…๋‹ˆ๋‹ค.
09:44
Provably safe systems are clearly not impossible.
178
584455
3170
์ž…์ฆ๋œ ์•ˆ์ „ํ•œ ์‹œ์Šคํ…œ์€ ๋ถˆ๊ฐ€๋Šฅํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.
09:47
Let's look at a simple example
179
587625
1502
๊ฐ„๋‹จํ•œ ์˜ˆ์‹œ๋กœ
09:49
of where we first machine-learn an algorithm from data
180
589168
4630
๋ฐ์ดํ„ฐ์—์„œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๋จผ์ € ๋จธ์‹ ๋Ÿฌ๋‹ํ•œ ๋‹ค์Œ
09:53
and then distill it out in the form of code
181
593840
4254
์‚ฌ์–‘์„ ์ถฉ์กฑํ•˜๋Š” ๊ฒƒ์œผ๋กœ ์ž…์ฆ๋œ ์ฝ”๋“œ ํ˜•ํƒœ๋กœ
์ถ”์ถœํ•˜๋Š” ๊ฐ„๋‹จํ•œ ์˜ˆ๋ฅผ ์‚ดํŽด๋ณด๊ฒ ์Šต๋‹ˆ๋‹ค.
09:58
that provably meets spec, OK?
182
598136
2252
10:00
Letโ€™s do it with an algorithm that you probably learned in first grade,
183
600888
4922
1ํ•™๋…„ ๋•Œ ๋ฐฐ์› ์„ ๋ฒ•ํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ ํ•ด๋ด…์‹œ๋‹ค.
10:05
addition,
184
605810
1168
๋ง์…ˆ์€ ์ˆซ์ž๋ฅผ ์˜ค๋ฅธ์ชฝ์—์„œ ์™ผ์ชฝ์œผ๋กœ ๋ฐ˜๋ณตํ•ด์„œ ์ด๋™ํ•˜๋ฉด์„œ
10:07
where you loop over the digits from right to left,
185
607019
2503
10:09
and sometimes you do a carry.
186
609564
1793
๋ฐ›์•„์˜ฌ๋ฆผ๋„ ํ•˜์ฃ .
10:11
We'll do it in binary,
187
611357
1752
2์ง„์ˆ˜์˜ ๊ฒฝ์šฐ
10:13
as if you were counting on two fingers instead of ten.
188
613151
2752
์—ด ์†๊ฐ€๋ฝ์ด ์•„๋‹Œ ๋‘ ์†๊ฐ€๋ฝ์œผ๋กœ๋งŒ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค.
10:16
And we first train a recurrent neural network,
189
616279
3378
๋จผ์ € ์ˆœํ™˜ ์‹ ๊ฒฝ๋ง์„ ํ•™์Šต์‹œ์ผœ
10:19
never mind the details,
190
619657
2211
์„ธ๋ถ€ ์‚ฌํ•ญ์€ ์‹ ๊ฒฝ ์“ฐ์ง€ ์•Š๊ณ 
10:21
to nail the task.
191
621909
1418
์ž‘์—…์„ ์ œ๋Œ€๋กœ ํ•ด๋‚ด๋„๋ก ํ•ฉ๋‹ˆ๋‹ค.
10:23
So now you have this algorithm that you don't understand how it works
192
623828
3253
์ด์ œ ์ž‘๋™ ์›๋ฆฌ๋ฅผ ์•Œ ์ˆ˜ ์—†๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด
10:27
in a black box
193
627123
2753
๋ธ”๋ž™๋ฐ•์Šค์— ์ƒ๊ฒผ์œผ๋ฉฐ
10:29
defined by a bunch of tables of numbers that we, in nerd speak,
194
629917
4463
๋ธ”๋ž™๋ฐ•์Šค๋Š” ์ˆ˜๋งŽ์€ ์ˆซ์ž ํ…Œ์ด๋ธ”๋กœ ์ •์˜๋ฉ๋‹ˆ๋‹ค.
์ˆซ์ž๋Š” ์ „๋ฌธ ์šฉ์–ด๋กœ ํŒŒ๋ผ๋ฏธํ„ฐ๋ผ๊ณ  ํ•˜์ฃ .
10:34
call parameters.
195
634380
1502
10:35
Then we use an AI tool we built to automatically distill out from this
196
635882
5297
๊ทธ ๋‹ค์Œ์œผ๋กœ Python ํ”„๋กœ๊ทธ๋žจ ํ˜•์‹์œผ๋กœ ํ•™์Šต๋œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„
10:41
the learned algorithm in the form of a Python program.
197
641179
3420
์ง์ ‘ ๋งŒ๋“  AI ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•ด ๋ธ”๋ž™๋ฐ•์Šค์—์„œ ์ž๋™ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค.
10:44
And then we use the formal verification tool known as Daphne
198
644640
4838
๊ทธ๋Ÿฐ ๋‹ค์Œ Daphne๋ผ๋Š” ์ •ํ˜• ๊ฒ€์ฆ ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ
10:49
to prove that this program correctly adds up any numbers,
199
649520
5422
์ด ํ”„๋กœ๊ทธ๋žจ์ด ํ•™์Šต ๋ฐ์ดํ„ฐ์— ์žˆ์—ˆ๋˜ ์ˆซ์ž๋ฟ๋งŒ ์•„๋‹ˆ๋ผ
10:54
not just the numbers that were in your training data.
200
654942
2503
๋ชจ๋“  ์ˆซ์ž๋ฅผ ๋˜‘๋ฐ”๋กœ ๋”ํ•˜๊ณ  ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ์ฆ๋ช…ํ•ฉ๋‹ˆ๋‹ค.
10:57
So in summary,
201
657778
1377
์š”์•ฝํ•˜์ž๋ฉด
10:59
provably safe AI, I'm convinced is possible,
202
659155
3962
์ž…์ฆ๋œ ์•ˆ์ „ํ•œ ์ธ๊ณต์ง€๋Šฅ์€ ์‹คํ˜„ ๊ฐ€๋Šฅํ•˜๋‹ค๊ณ  ํ™•์‹ ํ•ฉ๋‹ˆ๋‹ค.
ํ•˜์ง€๋งŒ ์‹œ๊ฐ„๊ณผ ๋…ธ๋ ฅ์ด ํ•„์š”ํ•  ๊ฒƒ์ž…๋‹ˆ๋‹ค.
11:03
but it's going to take time and work.
203
663117
3003
11:06
And in the meantime,
204
666162
1209
ํ•œํŽธ์œผ๋กœ๋Š” ๊ธฐ์–ตํ•ด์•ผ ํ•˜๋Š” ์ ์œผ๋กœ
11:07
let's remember that all the AI benefits
205
667413
4254
๋Œ€๋‹ค์ˆ˜๊ฐ€ ์—ด๊ด‘ํ•˜๋Š” AI์˜ ์ด์ ์€ ์ „๋ถ€
11:11
that most people are excited about
206
671709
3795
11:15
actually don't require superintelligence.
207
675504
2628
์‹ค์ œ๋กœ๋Š” ์ดˆ์ง€๋Šฅ์ด ๋ถˆํ•„์š”ํ•˜๋‹ค๋Š” ์ ์ž…๋‹ˆ๋‹ค.
11:18
We can have a long and amazing future with AI.
208
678758
5005
AI์™€ ํ•จ๊ป˜๋ผ๋ฉด ๊ธธ๊ณ  ๋†€๋ผ์šด ๋ฏธ๋ž˜๋ฅผ ๋งŒ๋“ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
11:25
So let's not pause AI.
209
685014
2169
AI์— ์ œ๋™์„ ๊ฑธ์ง€ ์•Š์•„์•ผ ํ•ฉ๋‹ˆ๋‹ค.
11:28
Let's just pause the reckless race to superintelligence.
210
688476
4129
์ดˆ์ง€๋Šฅ์„ ํ–ฅํ•œ ๋ฌด๋ชจํ•œ ๊ฒฝ์Ÿ์„ ๋ฉˆ์ถฐ์•ผ ํ•ฉ๋‹ˆ๋‹ค.
11:32
Let's stop obsessively training ever-larger models
211
692980
4255
์ธ๊ฐ„์ด ์ดํ•ดํ•˜์ง€ ๋ชปํ•˜๋Š”
๋” ํฐ ๋ชจ๋ธ์„ ํ•™์Šต์‹œํ‚ค๋ ค๋Š” ์ง‘์ฐฉ์„ ๋ฒ„๋ ค์•ผ ํ•ฉ๋‹ˆ๋‹ค.
11:37
that we don't understand.
212
697276
1669
11:39
Let's heed the warning from ancient Greece
213
699737
3212
๊ณ ๋Œ€ ๊ทธ๋ฆฌ์Šค ๋•Œ ๋งํ–ˆ๋˜ ๊ฒฝ๊ณ ์— ๊ท€๋ฅผ ๊ธฐ์šธ์—ฌ
11:42
and not get hubris, like in the story of Icarus.
214
702949
3712
์ด์นด๋ฃจ์Šค ์ด์•ผ๊ธฐ์ฒ˜๋Ÿผ ์˜ค๋งŒํ•จ์— ๋น ์ง€์ง€ ๋ง™์‹œ๋‹ค.
11:46
Because artificial intelligence
215
706702
2711
์™œ๋ƒํ•˜๋ฉด ์ธ๊ณต์ง€๋Šฅ์€
11:49
is giving us incredible intellectual wings
216
709455
4463
์ธ๊ฐ„์—๊ฒŒ ๋†€๋ผ์šด ์ง€์‹์˜ ๋‚ ๊ฐœ๋ฅผ ๋‹ฌ์•„์ฃผ๊ณ 
11:53
with which we can do things beyond our wildest dreams
217
713960
4045
ํƒœ์–‘์„ ํ–ฅํ•ด ๋‚ ์•„๊ฐ€๋ ค๋Š” ์ง‘์ฐฉ์„ ๋ฒ„๋ฆฐ๋‹ค๋ฉด
11:58
if we stop obsessively trying to fly to the sun.
218
718047
4379
๊ฟˆ๋„ ๊พธ์ง€ ๋ชปํ•œ ์ผ์„ ํ•ด๋‚ผ ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ฃผ๋‹ˆ๊นŒ์š”.
12:02
Thank you.
219
722802
1168
๊ณ ๋ง™์Šต๋‹ˆ๋‹ค.
12:03
(Applause)
220
723970
5880
(๋ฐ•์ˆ˜)
์ด ์›น์‚ฌ์ดํŠธ ์ •๋ณด

์ด ์‚ฌ์ดํŠธ๋Š” ์˜์–ด ํ•™์Šต์— ์œ ์šฉํ•œ YouTube ๋™์˜์ƒ์„ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. ์ „ ์„ธ๊ณ„ ์ตœ๊ณ ์˜ ์„ ์ƒ๋‹˜๋“ค์ด ๊ฐ€๋ฅด์น˜๋Š” ์˜์–ด ์ˆ˜์—…์„ ๋ณด๊ฒŒ ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค. ๊ฐ ๋™์˜์ƒ ํŽ˜์ด์ง€์— ํ‘œ์‹œ๋˜๋Š” ์˜์–ด ์ž๋ง‰์„ ๋”๋ธ” ํด๋ฆญํ•˜๋ฉด ๊ทธ๊ณณ์—์„œ ๋™์˜์ƒ์ด ์žฌ์ƒ๋ฉ๋‹ˆ๋‹ค. ๋น„๋””์˜ค ์žฌ์ƒ์— ๋งž์ถฐ ์ž๋ง‰์ด ์Šคํฌ๋กค๋ฉ๋‹ˆ๋‹ค. ์˜๊ฒฌ์ด๋‚˜ ์š”์ฒญ์ด ์žˆ๋Š” ๊ฒฝ์šฐ ์ด ๋ฌธ์˜ ์–‘์‹์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ฌธ์˜ํ•˜์‹ญ์‹œ์˜ค.

https://forms.gle/WvT1wiN1qDtmnspy7