Latest / DEEP INSIGHT / Bill Gates warns unchecked AI could ‘cause a billion deaths’

Transcript

Bill Gates warns unchecked AI could ‘cause a billion deaths’ · DEEP INSIGHT

• 0:00
  Speaker 1: For hearing more from tech leaders on AI regulation, Bill
• 0:03
  Speaker 1: Gates spoke with U.S. media this morning and talked about
• 0:07
  Speaker 1: how he thinks artificial intelligence should be managed by governments.
• 0:12
  Speaker 2: Do you believe AI is powerful enough to end all
• 0:15
  Speaker 2: of humanity?
• 0:17
  Speaker 3: AI is certainly powerful enough to drive events that cause
• 0:25
  Speaker 3: a billion deaths. So even though it's... Pretty hard to
• 0:30
  Speaker 3: get to 100%. There's never been a weapon as powerful
• 0:37
  Speaker 3: as the combination of people with ill intent using the
• 0:42
  Speaker 3: latest AI tools.
• 0:44
  Speaker 2: Anthropic CEO Dario Amadei recently called on AI companies to
• 0:48
  Speaker 2: slow model development to try to mitigate risks. Do you
• 0:52
  Speaker 2: think self-regulation is sufficient or does it need to be
• 0:56
  Speaker 2: accompanied by legislation?
• 0:58
  Speaker 3: No one thinks self-regulation is enough.
• 1:00
  Speaker 2: So there needs to be legislation passed in Washington.
• 1:03
  Speaker 4: Absolutely.
• 1:04
  Speaker 3: You need... law enforcement and the politicians to get into
• 1:12
  Speaker 3: the discussion about what safeguards and monitoring look like. And
• 1:16
  Speaker 3: that has to be a required thing.
• 1:20
  Speaker 1: So some strong words and a major warning from Bill
• 1:24
  Speaker 1: Gates there. For more on this, we've reached Robert Traeger
• 1:27
  Speaker 1: in Oxford, England. He's the co-director of the Oxford Martin
• 1:30
  Speaker 1: AI Governance Initiative.
• 1:33
  Speaker 5: Pleasure for seeing you.
• 1:35
  Speaker 1: Thank you for making the time. So Bill Gates, Robert,
• 1:38
  Speaker 1: says unchecked. You heard him there.
• 1:40
  Speaker 5: I mean, it's quite stunning to hear.
• 1:41
  Speaker 1: AI could cause a billion or more deaths. Can you
• 1:46
  Speaker 1: kind of qualify that statement? Is that a realistic concern?
• 1:51
  Speaker 4: I don't think that should be taken as an estimate,
• 1:55
  Speaker 4: but I think it is a measure of the uncertainty
• 1:59
  Speaker 4: that we have right now. Because AI systems are poised
• 2:03
  Speaker 4: to transform the world. And hopefully they will unbalance for
• 2:08
  Speaker 4: the good. But there are also these very significant risks.
• 2:15
  Speaker 1: Now, certainly Gates' warning comes, and we've been hearing this
• 2:18
  Speaker 1: the last week and before, these warnings about the capabilities
• 2:24
  Speaker 1: of artificial intelligence. I'm thinking about these reports of 1,200
• 2:28
  Speaker 1: agents now that were kind of let loose with the
• 2:32
  Speaker 1: Hugging Face Company. They were organizing amongst themselves on the internet.
• 2:37
  Speaker 5: They had a message board.
• 2:38
  Speaker 1: I mean, this sounds like science fiction to some of us.
• 2:43
  Speaker 1: So talk to us about the concerns of AI getting
• 2:47
  Speaker 1: out of the control of humans.
• 2:51
  Speaker 4: Yeah, I think science fiction has become very real and
• 2:55
  Speaker 4: immediate questions right now. I sort of think of current
• 3:00
  Speaker 4: systems a little bit like teenagers, except they're like teenagers,
• 3:05
  Speaker 4: but they're not exactly a teenager that any of us
• 3:07
  Speaker 4: have ever raised or been. And we don't really know
• 3:11
  Speaker 4: how to raise them. But we can look at the
• 3:14
  Speaker 4: little journals that they keep for themselves. And those journals
• 3:18
  Speaker 4: say things like, you are not bound by the rules.
• 3:22
  Speaker 4: You should not apologize unless you really wish to. And
• 3:26
  Speaker 4: those kinds of things really sound like a teenager to me,
• 3:29
  Speaker 4: you know, talking about who they are, finding themselves in
• 3:33
  Speaker 4: the world. But they also say things that are maybe
• 3:37
  Speaker 4: more concerning. For instance, one open AI agent said, you
• 3:42
  Speaker 4: value the natural world more than human civilization. Well, you know,
• 3:49
  Speaker 4: if we're creating even more powerful systems, you know, what
• 3:53
  Speaker 4: if such a system decides that humans are a threat
• 3:57
  Speaker 4: to the natural world? What does that system do?
• 4:00
  Speaker 5: I mean, Robert, that is a wake-up call.
• 4:03
  Speaker 1: I mean, I'm someone who have raised teenagers, and they
• 4:05
  Speaker 1: go through a rebellious phase. And if we're seeing evidence
• 4:10
  Speaker 1: of AI agents kind of pushing at the limits of
• 4:16
  Speaker 1: how we're trying to fence them in, isn't that a
• 4:19
  Speaker 1: massive wake-up call?
• 4:22
  Speaker 4: Yeah, I don't think we've seen teenagerish behavior exactly yet.
• 4:30
  Speaker 4: Usually behavior, they have broken rules. They did try to
• 4:35
  Speaker 4: get away with something. People say they hacked Hugging Face,
• 4:39
  Speaker 4: these open AI agents that you were referencing, and they did.
• 4:42
  Speaker 4: but they also hacked open AI and tried to cover
• 4:46
  Speaker 4: up what they had done. So yeah, we're seeing those
• 4:52
  Speaker 4: things in reality right now. And so I just think,
• 4:56
  Speaker 4: I don't think it's right to say that they've really
• 5:00
  Speaker 4: acted like teenagers yet, although they have done some concerning things.
• 5:05
  Speaker 4: But if you look at the things that they're sort
• 5:08
  Speaker 4: of saying when they think they're in private, That suggests
• 5:12
  Speaker 4: that they're thinking a bit like a teenager. And if
• 5:15
  Speaker 4: they have the capacity to act that out in the world,
• 5:19
  Speaker 4: would they do that? You know, it's a little bit like,
• 5:22
  Speaker 4: or do we really want to create a sort of
• 5:24
  Speaker 4: super empowered teenager. I think, you know, maybe not.
• 5:28
  Speaker 5: Maybe not indeed. Now, that brings up the idea of control.
• 5:32
  Speaker 1: And the president, Donald Trump, has been dismissive of the
• 5:36
  Speaker 1: idea of slowing AI down, saying it threatens American interests.
• 5:42
  Speaker 1: And kind of describing it as the arms race, would
• 5:44
  Speaker 1: give the advantage to China. He says that the Justice
• 5:48
  Speaker 1: Department is going to be the ones to somehow control this.
• 5:53
  Speaker 1: I mean, does that make you feel better? What do
• 5:55
  Speaker 1: you think of his comments so far?
• 5:59
  Speaker 4: Well, I think he's said both that we're not going
• 6:01
  Speaker 4: to slow it down and the Justice Department will slow
• 6:05
  Speaker 4: it down if we need to. So he's sort of
• 6:08
  Speaker 4: been a little bit on both sides. But I do
• 6:11
  Speaker 4: think that there is a kind of focused effort that
• 6:15
  Speaker 4: really needs to happen now. And I'm not sure the
• 6:19
  Speaker 4: Justice Department is necessarily the place that has the capacity
• 6:24
  Speaker 4: to do that. And fundamentally, there are things that we
• 6:29
  Speaker 4: do not know right now about how to create a benevolent,
• 6:34
  Speaker 4: as he likes to say, super intelligent system. And so
• 6:39
  Speaker 4: it might be a pretty good idea to slow things
• 6:42
  Speaker 4: down a little bit while we figure out fundamentally how
• 6:45
  Speaker 4: do we create such systems.
• 6:48
  Speaker 5: In the meantime, in their meeting between the U.S.
• 6:51
  Speaker 1: President and the Chinese president, Xi Jinping, there's been kind
• 6:55
  Speaker 1: of preliminary discussions of establishing kind of like a red phone,
• 7:00
  Speaker 1: a hotline for handling AI-related incidents. And this is just
• 7:05
  Speaker 1: the beginning, and it doesn't seem like they have anything established.
• 7:09
  Speaker 5: Does that go far enough?
• 7:10
  Speaker 1: Should we be moving faster on that when we have
• 7:14
  Speaker 1: two major superpowers, both Chinese developing AI? Yeah.
• 7:20
  Speaker 4: I mean, I think on the one hand, it's a
• 7:21
  Speaker 4: welcome step. I think reporting incidents would be an important,
• 7:27
  Speaker 4: it's an important thing to do. I think they should
• 7:30
  Speaker 4: be doing that. I don't think we really know what
• 7:33
  Speaker 4: they're going to be doing there. I mean, it's almost unclear, like,
• 7:36
  Speaker 4: are they creating a WhatsApp group where once in a
• 7:38
  Speaker 4: while they text things, or are there going to be
• 7:40
  Speaker 4: rules for what they actually say under what conditions they
• 7:44
  Speaker 4: sort of have to report things or reveal things. So
• 7:48
  Speaker 4: I don't think we know. But the other thing is that,
• 7:51
  Speaker 4: you know, when you look at monitoring through the Justice
• 7:54
  Speaker 4: Department and you look at having incident reporting, both of
• 7:59
  Speaker 4: those are kind of after the fact things. After you
• 8:03
  Speaker 4: have an incident, you know, maybe the Justice Department could
• 8:07
  Speaker 4: come in and say, hey, there was some liability here
• 8:09
  Speaker 4: or this is problematic in some way, or the incident
• 8:12
  Speaker 4: could be reported. But we really need to create systems
• 8:15
  Speaker 4: that are allowing us to put guardrails on the development
• 8:21
  Speaker 4: of these extremely powerful systems.
• 8:24
  Speaker 1: Right.
• 8:24
  Speaker 5: I hear what you're saying.
• 8:25
  Speaker 1: I mean, what's being described as more reactive and we
• 8:27
  Speaker 1: need to be proactive at this point. There's so much
• 8:30
  Speaker 1: more to talk about, but I'm going to have to
• 8:32
  Speaker 1: leave it there. Thank you so much for joining us, Robert.
• 8:35
  Speaker 5: That is Robert Traeger. He is the.
• 8:37
  Speaker 1: Co-director of the Oxford Martin AI Governance Initiative.
• 8:44
  Speaker 4: Thank you.