สำเนาบทสนทนา 🔗
WEBVTT
1
00:00:00.580 --> 00:00:04.519
An apocalyptic warning or an overreaction. Some of the biggest
2
00:00:04.540 --> 00:00:07.620
CEOs in tech are sounding the alarm and calling for
3
00:00:07.660 --> 00:00:11.789
a slowdown in the development of artificial intelligence. They're citing
4
00:00:11.890 --> 00:00:15.169
fears of it spinning out of control, warning AI could
5
00:00:15.550 --> 00:00:18.289
grow smart enough to kill us. So for more on
6
00:00:18.329 --> 00:00:20.929
those warnings and what needs to be done, we've reached
7
00:00:21.039 --> 00:00:25.000
AI researcher Connor Leahy in Washington, D.C. He is also
8
00:00:25.059 --> 00:00:30.219
the executive director of Control AI, a nonprofit advocacy organization.
9
00:00:30.239 --> 00:00:31.940
Thank you so much for making time for us today.
10
00:00:33.140 --> 00:00:33.799
Thank you for having me.
11
00:00:34.240 --> 00:00:38.159
So leaders at three of the industry's most powerful companies, Connor,
12
00:00:38.219 --> 00:00:41.880
are calling to slow down AI development, citing their fears
13
00:00:41.939 --> 00:00:44.619
that AI may spiral out of control. Why now?
14
00:00:47.009 --> 00:00:49.369
Well, we have just over the summer seen some of
15
00:00:49.390 --> 00:00:52.729
the most shocking advances in AI ever occurred. Things that
16
00:00:52.750 --> 00:00:55.390
were predictable, but it was still shocking to see that
17
00:00:55.530 --> 00:00:59.229
autonomous AI agents broke out of secure containment facilities to
18
00:00:59.289 --> 00:01:03.210
attack other companies, flagrantly disobeying human orders in the process.
19
00:01:03.549 --> 00:01:06.469
At the same time, we're seeing AI systems becoming so powerful,
20
00:01:06.569 --> 00:01:10.569
so intelligent, so autonomous, that the director of the NSA
21
00:01:10.709 --> 00:01:13.109
has warned that they could break into almost all of
22
00:01:13.129 --> 00:01:16.079
their classified systems in hours, not weeks. What is really
23
00:01:16.090 --> 00:01:18.920
important to understand is that this is the least powerful
24
00:01:18.980 --> 00:01:21.560
AI will ever be. What these companies are pursuing is
25
00:01:21.609 --> 00:01:25.269
not better chatbots. It is what is called superintelligence, AI
26
00:01:25.329 --> 00:01:29.409
systems that can out-compete humanity across all tasks, that is
27
00:01:29.469 --> 00:01:32.310
fully autonomous, and that we have no idea how to control.
28
00:01:32.640 --> 00:01:35.019
If we reach this point and we cannot control systems
29
00:01:35.060 --> 00:01:37.200
like this, it is very, very hard to imagine that
30
00:01:37.239 --> 00:01:37.599
going well.
31
00:01:37.879 --> 00:01:39.819
You know, Connor, of course, there have been rumblings of
32
00:01:39.879 --> 00:01:43.299
the risks of AI and the fast pace of its
33
00:01:43.340 --> 00:01:47.230
development for years now. So what has really come to
34
00:01:47.290 --> 00:01:51.230
a head at this particular juncture? Could these
Load full transcript
CEOs not
35
00:01:51.290 --> 00:01:54.670
have foreseen this ahead of time in order to start
36
00:01:54.730 --> 00:01:57.700
reaching out before we reach this pivotal moment?
37
00:01:59.750 --> 00:02:01.090
I think there's a lot more that could have been
38
00:02:01.109 --> 00:02:02.870
done than was done, but there was a lot that
39
00:02:02.909 --> 00:02:05.640
was done. Back in 2023, for example, there was a
40
00:02:05.670 --> 00:02:08.180
public statement that said that the mitigation of the risk
41
00:02:08.240 --> 00:02:12.199
of extinction by AI should be a global priority that
42
00:02:12.280 --> 00:02:14.900
was signed, in fact, by the CEOs of all the
43
00:02:14.939 --> 00:02:17.740
top companies and by other luminaries, such as the Nobel
44
00:02:17.780 --> 00:02:20.620
Prize winner, Geoffrey Hinton, the godfather of AI. So these
45
00:02:20.659 --> 00:02:22.639
warnings are, in fact, not new. What has really been
46
00:02:22.680 --> 00:02:24.849
happening is that we here at Control AI have been
47
00:02:24.909 --> 00:02:28.849
for years now working on systematically briefing policymakers and the
48
00:02:28.889 --> 00:02:31.250
general public on these issues. We've met with over 200
49
00:02:31.250 --> 00:02:34.340
congressional offices just here in the U.S. alone, over 100
50
00:02:34.340 --> 00:02:37.039
offices in Canada. I have over 35 supporters in the
51
00:02:37.080 --> 00:02:39.979
Canadian parliament already. And the number one feedback we get
52
00:02:40.020 --> 00:02:42.599
when we talk to people about these issues is that
53
00:02:42.620 --> 00:02:44.439
they have just never heard of it. So I think
54
00:02:44.460 --> 00:02:46.620
there's a huge awareness issue that is now starting to
55
00:02:46.659 --> 00:02:49.060
break through where people are starting to, for the first time,
56
00:02:49.099 --> 00:02:51.900
really become truly exposed to these issues and really realizing
57
00:02:52.099 --> 00:02:53.580
that this is not a path that we want to
58
00:02:53.620 --> 00:02:54.039
go down on.
59
00:02:54.639 --> 00:02:56.969
Now, when we talk about the messages that we're hearing
60
00:02:57.150 --> 00:03:00.370
about AI and the potential risks as well, we know
61
00:03:00.389 --> 00:03:03.409
that the U.S. President, Donald Trump, pushed back on those
62
00:03:03.449 --> 00:03:06.770
growing calls for greater regulation of the AI. He posted
63
00:03:06.830 --> 00:03:10.199
on Truth Social saying, the only controller guardrails that AI
64
00:03:10.259 --> 00:03:13.689
needs is a strong and smart president. And the USA
65
00:03:13.770 --> 00:03:15.849
has that in spades. So what do you make of
66
00:03:15.930 --> 00:03:17.090
his remarks on all of this?
67
00:03:18.729 --> 00:03:21.789
I think the administration is a large organization with many,
68
00:03:21.830 --> 00:03:23.909
many different people. And if there's one thing I do
69
00:03:23.949 --> 00:03:26.650
believe about President Trump, it's that I do believe that
70
00:03:26.689 --> 00:03:30.469
he wants America to continue to exist. Superintelligence would threaten
71
00:03:30.509 --> 00:03:33.069
the continued existence of the United States government, the United
72
00:03:33.090 --> 00:03:35.370
States in general, and all other countries for that matter.
73
00:03:35.710 --> 00:03:37.909
If the United States builds superintelligence, we will not be
74
00:03:37.930 --> 00:03:39.590
able to control it. It will all compete us and
75
00:03:39.629 --> 00:03:43.969
destroy America. If China builds superintelligence, It will out-compete and
76
00:03:44.009 --> 00:03:47.340
destroy China and the U.S. and everything else. So really,
77
00:03:47.379 --> 00:03:50.159
the priority must be to prevent AI being created anywhere
78
00:03:50.180 --> 00:03:52.520
in the world. And I think this is something that
79
00:03:53.340 --> 00:03:56.539
many people will basically anyone can agree on. So while
80
00:03:56.560 --> 00:03:59.780
there's still some disagreements, it's important to understand that there's
81
00:03:59.800 --> 00:04:03.840
a huge difference between AI broadly and superintelligence. I definitely
82
00:04:03.860 --> 00:04:06.979
agree with President Trump that we must pursue many useful
83
00:04:07.060 --> 00:04:10.120
economic and military applications of AI that are not superintelligence
84
00:04:10.180 --> 00:04:14.120
for our national and international competitiveness. But superintelligence is not
85
00:04:14.139 --> 00:04:16.639
a tool. It's not a weapon. It's an adversary. And
86
00:04:16.660 --> 00:04:17.699
we must prevent its creation.
87
00:04:18.220 --> 00:04:20.240
You know, we just talked about the remarks from the U.S.
88
00:04:20.300 --> 00:04:20.680
President.
89
00:04:20.699 --> 00:04:22.240
And when I look at what has been said by
90
00:04:22.259 --> 00:04:25.100
the U.S. vice president, he says that he feels a
91
00:04:25.120 --> 00:04:26.750
bit weird about the fact that you have so many
92
00:04:26.829 --> 00:04:31.410
frontier AI tech companies coming to the government begging for regulation.
93
00:04:31.490 --> 00:04:33.410
He said it feels like a bit of a Trojan
94
00:04:33.529 --> 00:04:35.550
horse to him. I mean, what do you make of
95
00:04:35.610 --> 00:04:36.529
his remarks on this?
96
00:04:38.000 --> 00:04:39.660
I think we are in a very weird time. I
97
00:04:39.699 --> 00:04:42.180
think a lot of very weird things is happening. Imagine
98
00:04:42.220 --> 00:04:44.800
a world in which nuclear weapons were designed and sold
99
00:04:44.879 --> 00:04:47.639
exclusively by private companies to the highest bidder for many,
100
00:04:47.699 --> 00:04:49.759
many years. That would have been a very different and
101
00:04:49.810 --> 00:04:52.470
a very bad world, in my opinion. I think that,
102
00:04:52.689 --> 00:04:54.589
in a sense, we should have acted much earlier on
103
00:04:54.629 --> 00:04:57.009
many of these issues. And so now we're in this
104
00:04:57.089 --> 00:05:00.750
weird circumstance where we have technology already that's extremely powerful,
105
00:05:00.970 --> 00:05:03.310
that is racing towards an extremely powerful thing. It's not
106
00:05:03.350 --> 00:05:05.750
just the top three companies, right? There are many, many,
107
00:05:05.810 --> 00:05:09.300
many other companies. pursuing superintelligence. It is at the point
108
00:05:09.339 --> 00:05:12.540
that no individual company can solve this. It requires government action.
109
00:05:13.600 --> 00:05:16.259
And I'm just, in the final 30 seconds I have
110
00:05:16.310 --> 00:05:18.470
with you, I want to ask you, is it too late?
111
00:05:19.910 --> 00:05:22.269
No, it is not yet too late because superintelligence does
112
00:05:22.310 --> 00:05:25.750
not exist. Our governments can still act both nationally and internationally.
113
00:05:25.970 --> 00:05:28.910
We must create an international trust but verify regime to
114
00:05:29.430 --> 00:05:31.980
prevent the creation of superintelligence. This is in the interest
115
00:05:32.009 --> 00:05:34.439
of all parties involved and it is still doable, but
116
00:05:34.600 --> 00:05:35.860
we must act quickly.
117
00:05:36.740 --> 00:05:38.750
Connor, I want to thank you so much for joining us.
118
00:05:38.810 --> 00:05:40.230
I appreciate your insights today.
119
00:05:40.250 --> 00:05:42.389
Thank you.
120
00:05:42.569 --> 00:05:45.389
That is AI researcher Connor Leahy. Thanks again for joining us.