Latest / Columbus' Morning News / Okay, how seriously should we take these A.I. concerns?
Transcript
- 0:00Speaker 1: All right, let's spend a minute on the LegacyRetirementGroup.com phone line.
- 0:02Speaker 1: It is Tech Tuesday. Mike Dabusky, ABC News technology reporter. Well, Michael,
- 0:07Speaker 1: I wonder what we're going to talk about today. It's
- 0:10Speaker 1: been a week.
- 0:11Speaker 2: It's a new iPhone, right? It's a foldable iPhone. Yeah, exactly.
- 0:16Speaker 1: You had the researcher come out, and he made these
- 0:19Speaker 1: kind of crazy claims about there's a 10% chance of
- 0:22Speaker 1: AI killing humanity in the.
- 0:23Speaker 2: Next 10 years.
- 0:24Speaker 1: And then shortly after, a bunch of AI executives said,
- 0:27Speaker 1: you know what? We do need to tap the brakes here.
- 0:30Speaker 1: AI could take over the Internet in a matter of months.
- 0:33Speaker 1: So let's pause. Let's get some reason here. How serious,
- 0:38Speaker 1: how realistic is this threat?
- 0:40Speaker 3: Yeah, so let's go through the different claims that are
- 0:42Speaker 3: being made here and kind of just put them into
- 0:45Speaker 3: proper context. This all started with Jacob Coxon, who, again,
- 0:48Speaker 3: is that anthropic researcher who resigned last week. He also
- 0:51Speaker 3: worked for OpenAI before that. He says he resigned because
- 0:55Speaker 3: he's concerned that these companies know they're creating a technology
- 0:58Speaker 3: that has the possibility of wiping out humanity by the
- 1:01Speaker 3: end of the decade, but are doing nothing to stop it.
- 1:04Speaker 3: He says at OpenAI, many have not deeply internalized the
- 1:07Speaker 3: civilizational stakes. At Anthropic, the stakes are well understood, but
- 1:11Speaker 3: they are locked in a race to get there first. Later,
- 1:14Speaker 3: Anthropic's alignment science lead, Evan Hubinger, replied to this thread
- 1:19Speaker 3: from Coxon by saying that he puts the percentage chance
- 1:22Speaker 3: of humanity being wiped out by AI... at 10%. All
- 1:26Speaker 3: very scary stuff. It is reasonable for people to perk
- 1:29Speaker 3: up and pay attention to that. The good news is
- 1:31Speaker 3: that this seems to be some pretty wild speculation. That
- 1:35Speaker 3: 10% figure that gets cited is likely what's known as
- 1:37Speaker 3: a P-Doom figure. That's a number that people in the
- 1:41Speaker 3: AI safety realm use to predict the likelihood of AI
- 1:44Speaker 3: eliminating human life. If you have a P-Doom of 10,
- 1:47Speaker 3: you think there's a 10% chance of that happening. It's
- 1:49Speaker 3: a very Silicon Valley term, but that's what it is. Okay.
- 1:51Speaker 3: The thing about making predictions like this is that they
- 1:54Speaker 3: are unfalsifiable. They can't be proven wrong. We live in
- 1:57Speaker 3: one timeline. We're either going to be dead at the
- 1:59Speaker 3: end of the decade or we won't be. We can't,
- 2:02Speaker 3: you know, percentage evaluate or confirm whether it was actually
- 2:05Speaker 3: a percentage chance one way or the other. What I'm
- 2:08Speaker 3: saying is that PDoom is basically just vibes. There is
- 2:11Speaker 3: no scientific process that these guys did to arrive at
- 2:13Speaker 3: this number.
- 2:14Speaker 2: There's no methodology. There's no historical precedent. It's vibes.
- 2:18Speaker 3: Jeffrey Hinton, who worked at Google for a long time
- 2:20Speaker 3: on AI and is sometimes called the godfather of AI,
- 2:24Speaker 3: told CNN that anybody who estimates probabilities like that is
- 2:28Speaker 3: really just making a wild guess. They're giving you their
- 2:30Speaker 3: gut feeling. Hinton, for his part, puts his P. Doom
- 2:34Speaker 3: at like 20%. Dario Amadei has said somewhere between 10%
- 2:38Speaker 3: and 25%. Some researchers like Yann Lequin say it's zero.
- 2:42Speaker 3: Others like Eliezer Yudkowsky put it at nearly 100%. People
- 2:46Speaker 3: are all over the map with this type of thing.
- 2:48Speaker 3: We should not lose a ton of sleep over that
- 2:50Speaker 3: 10% number specifically. However, there are some very real concerns
- 2:56Speaker 3: here related to Dario Amadei's claims in the blog post
- 2:59Speaker 3: that he put up in the wake of Coxon's resignation.
- 3:02Speaker 2: He says within six.
- 3:03Speaker 3: To 12 months, a swarm of AI agents could be
- 3:06Speaker 3: capable of taking over the entire Internet. He cites the
- 3:09Speaker 3: Hugging Face incident from earlier this summer where models from
- 3:12Speaker 3: OpenAI were undergoing testing, broke out of their testing environments,
- 3:16Speaker 3: hacked into this third-party website. Anthropic later realized that was
- 3:19Speaker 3: happening at their own company, too, where their own models
- 3:22Speaker 3: were doing stuff that they were not aware of. So
- 3:24Speaker 3: Amadei is basing this in some semblance of reality, though
- 3:27Speaker 3: taking it several steps further. He's extrapolating. We dive into
- 3:31Speaker 3: the details here. Things get a little bit hazier. For one,
- 3:33Speaker 3: take over the Internet. That's really vague. What does that mean?
- 3:36Speaker 3: Does that mean shutting the Internet down? Does that mean
- 3:38Speaker 3: changing it in some way? Either way, it would be
- 3:40Speaker 3: very difficult. The Internet is by design distributed. And while
- 3:44Speaker 3: a couple of companies like Amazon and Google control big
- 3:46Speaker 3: pieces of it, it still continues to operate, even if
- 3:49Speaker 3: some of the big players go down. The AWS outage
- 3:52Speaker 3: from earlier this summer is a great example of this.
- 3:54Speaker 3: It was disruptive. It wasn't the end of the Internet.
- 3:57Speaker 3: It was certainly not the end of the world. Also,
- 3:59Speaker 3: while many AI researchers share Amadei's sort of general concern
- 4:02Speaker 3: about more hacking incidents happening, he's the only one who's
- 4:05Speaker 3: really putting a six to 12 month time limit on it.
- 4:08Speaker 3: He's the head of a major AI company. He's seeing
- 4:10Speaker 3: models and technologies that we're not privy to. But even still,
- 4:13Speaker 3: I would like to see some additional sourcing before we
- 4:15Speaker 3: go around worrying about the Internet just snapping out of
- 4:18Speaker 3: existence in the spring. Basically, the hacking concerns here are
- 4:21Speaker 3: pretty real and we should be concerned about them. The
- 4:24Speaker 3: taking over the internet in 6 to 12 months is
- 4:26Speaker 3: somewhat loosely sketched and should be treated with a grain
- 4:28Speaker 3: of salt.
- 4:29Speaker 2: Okay, and that would be, I mean... we would survive
- 4:32Speaker 2: without the internet.
- 4:33Speaker 1: We survived without it for many, many years at life
- 4:36Speaker 1: would be a lot different if the internet didn't function
- 4:38Speaker 1: as it does today.
- 4:39Speaker 2: It would, there would be huge disruption.
- 4:42Speaker 3: I mean, you think about all the medical facilities and
- 4:45Speaker 3: energy facilities that are tied to that. So, you know,
- 4:48Speaker 3: if it were to go down in a holistic way,
- 4:52Speaker 3: you know, that would be, you know, really, it would
- 4:54Speaker 3: be a threat to some human life. But like, again,
- 4:56Speaker 3: I think that, We're pretty far away from that at
- 5:00Speaker 3: this point. I think some of the hacking incidents are
- 5:04Speaker 3: more in the style of what happened earlier this year,
- 5:06Speaker 3: where it kind of disrupted one website or it disrupted...
- 5:10Speaker 3: You know, a number of kind of smaller firms like
- 5:12Speaker 3: that is more in the realm of like what's going
- 5:15Speaker 3: to happen sooner. But, yeah, I think that, you know,
- 5:18Speaker 3: the whole Internet going down is a little bit outside
- 5:22Speaker 3: the realm, at least.
- 5:23Speaker 2: As it stands.
- 5:24Speaker 1: I mean, I think given the choice between the two,
- 5:26Speaker 1: a little Internet chaos versus the end of humanity, I'll
- 5:29Speaker 1: take some Internet chaos instead. But back to that claim that,
- 5:33Speaker 1: you know, with the P doom of 10, the Internet,
- 5:36Speaker 1: the AI bots killing off humanity.
- 5:38Speaker 2: How would that happen? To go, like, on a ground level?
- 5:42Speaker 3: Yeah.
- 5:42Speaker 2: I mean, how are they going to get me?
- 5:46Speaker 3: Yeah, so there's a couple different, like, pretty out there
- 5:48Speaker 3: theories as to how AI would get there. A lot
- 5:51Speaker 3: of this goes back to, you know, a person giving
- 5:54Speaker 3: an AI agent that is very advanced a pretty simple
- 5:58Speaker 3: task and not necessarily thinking out, you know, how an
- 6:01Speaker 3: AI bot would kind of go about solving the task
- 6:04Speaker 3: that was put in front of them. So, for example,
- 6:07Speaker 3: I have an AI agent from Anthropic and I tell it, hey,
- 6:10Speaker 3: make me, you know, money every day make no mistakes.
- 6:15Speaker 3: And the AI agent goes out and makes a bunch
- 6:17Speaker 3: of copies of itself. It finds different places on the
- 6:19Speaker 3: internet to, you know, make a couple bucks here and there.
- 6:22Speaker 3: And then in an effort to continue making money, you know,
- 6:26Speaker 3: throughout all that process, you know, it continues to copy
- 6:29Speaker 3: itself and continues to copy itself.
- 6:31Speaker 2: And then, you know, it becomes this, you know, kind
- 6:34Speaker 2: of miasma of things.
- 6:36Speaker 3: And then, you know, it starts shutting down hospitals in
- 6:39Speaker 3: an effort to, you know, like build out, more computer
- 6:45Speaker 3: power for itself. Like a lot of this goes back
- 6:46Speaker 3: to like compute power and data centers and stuff like that.
- 6:49Speaker 3: It can mobilize, you know, robots and other things like
- 6:52Speaker 3: that to go and, um, you know, help support the
- 6:55Speaker 3: building of more of that sort of thing. It becomes
- 6:56Speaker 3: a real energy crunch at the end of the day. And, uh,
- 6:59Speaker 3: that would, you know, necessarily result in sort of the,
- 7:03Speaker 3: the loss of human life and, the end of people.
- 7:05Speaker 3: But like, again, we're really off the edge of the map.
- 7:09Speaker 3: I just like just to spin it out to its
- 7:12Speaker 3: natural end. Like that is what many theorize would be
- 7:16Speaker 3: what this looks like. But again, a lot of this
- 7:19Speaker 3: is just vibes. I think there are some very real
- 7:22Speaker 3: AI safety concerns that stand in front of us and
- 7:24Speaker 3: are provable right now. Things like you know, the ability
- 7:28Speaker 3: of this technology to spread misinformation easily online, the ability
- 7:32Speaker 3: of this technology to be sycophantic and to, you know,
- 7:36Speaker 3: continue to foster the delusions of people who need mental
- 7:41Speaker 3: health help. The environmental fallout from data centers and from
- 7:46Speaker 3: AI development is something that we need to learn more
- 7:48Speaker 3: about and understand better.
- 7:49Speaker 2: And the fact.
- 7:50Speaker 3: That this is an industry that is deeply tied to
- 7:53Speaker 3: huge swaths of the American economy, huge swaths of the
- 7:56Speaker 3: global economy that still doesn't make money. And I think
- 8:01Speaker 3: that should be very concerning to a lot of people,
- 8:04Speaker 3: not to make light of some of these bigger Skynet
- 8:07Speaker 3: Terminator-style concerns, but I think the majority of our brainpower
- 8:10Speaker 3: should be spent fixing the problems that AI has created
- 8:13Speaker 3: right now.