Latest / Columbus' Morning News / Tech meeting at the White House and OpenAI pauses some developments
Transcript
- 0:00Speaker 1: And one of the big stories today is an AI
- 0:02Speaker 1: meeting at the White House on the LegacyRetirementGroup.com phone line.
- 0:06Speaker 1: Let's welcome in Mike Dabusky, ABC News technology reporter. So
- 0:10Speaker 1: who's talking to who and where and when?
- 0:12Speaker 2: Yeah, so it's all happening in Washington at the White House.
- 0:15Speaker 2: This is a lunch meeting about artificial intelligence. This has
- 0:19Speaker 2: obviously been under a lot of scrutiny amid a series
- 0:21Speaker 2: of recent hacking incidents where unreleased AI models break free
- 0:25Speaker 2: of their testing environments. and start messing around on the internet.
- 0:28Speaker 2: It's led to calls by many, including some in the
- 0:31Speaker 2: industry itself, to come in and have the government slow
- 0:36Speaker 2: down or pace the frontier of development, in the words
- 0:38Speaker 2: of Dario Amadei, the CEO of Anthropic. However, President Trump
- 0:42Speaker 2: has pushed back on this idea of an AI slowdown,
- 0:46Speaker 2: calling the threat of AI a hoax. The president sees
- 0:49Speaker 2: this as a race with China, and so far he's
- 0:52Speaker 2: argued that the American AI industry is basically can't take
- 0:55Speaker 2: its foot off the gas. They want to win this
- 0:58Speaker 2: race with China, despite the cybersecurity concerns that we're seeing
- 1:02Speaker 2: starting to crop up. Meta CEO Mark Zuckerberg is going
- 1:05Speaker 2: to be at today's meeting, according to a source familiar,
- 1:09Speaker 2: as will Jensen Huang from NVIDIA. Both of these guys
- 1:12Speaker 2: are on the president's side of this debate, with Zuckerberg
- 1:15Speaker 2: saying that every lab has the responsibility and incentive to
- 1:18Speaker 2: move at the pace required to train its model safely,
- 1:21Speaker 2: basically saying that, like, market incentives allow this to kind
- 1:26Speaker 2: of foreground the safe AI models in favor of the
- 1:31Speaker 2: stuff that's going out there and actually causing all this havoc. However,
- 1:35Speaker 2: also at the White House today will be Anthropic CEO
- 1:37Speaker 2: Dario Amadei. He's one of the leading voices raising a
- 1:41Speaker 2: lot of these concerns, one of the first to call
- 1:44Speaker 2: for an AI slowdown that we've also seen companies like
- 1:47Speaker 2: OpenAI and XAI support. However, the actual practice of that
- 1:53Speaker 2: has not gone into effect yet. It kind of requires
- 1:55Speaker 2: everyone to agree to a slowdown. And so far, that's
- 1:59Speaker 2: proving to be a little complicated.
- 2:00Speaker 1: Mike, can both sides be right on this? Because I
- 2:03Speaker 1: think Trump has a point when he talks about, you know,
- 2:05Speaker 1: because the Chinese don't care much for regulations. And I
- 2:09Speaker 1: think the president is right. If we tap the brakes
- 2:12Speaker 1: on AI, they're going to throttle forward and essentially, you know,
- 2:17Speaker 1: take our lunch money and stuff us in a locker.
- 2:19Speaker 1: On the other side of things, when you hear an
- 2:22Speaker 1: AI executive in Dario Amadei say this is a problem,
- 2:26Speaker 1: I think he should be listened to as well.
- 2:28Speaker 2: Yeah, it's a really complicated problem. I think it's a
- 2:31Speaker 2: good question whether both sides have a point here, because
- 2:34Speaker 2: I think they do. There are doubtless, very real cybersecurity
- 2:38Speaker 2: concerns with this technology. Beyond the, you know, AI doomerism
- 2:44Speaker 2: and the like, will this lead to like the apocalypse
- 2:46Speaker 2: or the end of humanity or anything like that? Like
- 2:49Speaker 2: in front of us right now, there is the potential
- 2:51Speaker 2: for this technology to go very wrong, to spread misinformation,
- 2:55Speaker 2: realistic looking picture and video on the internet with just
- 2:59Speaker 2: weeks to go before crucial elections. That's a real danger.
- 3:02Speaker 2: The environmental impact of this technology is something that we
- 3:06Speaker 2: haven't very much reckoned with. The potential for this to
- 3:08Speaker 2: you know, exacerbate mental health crises in people or to
- 3:11Speaker 2: disrupt education, the job market. All of that stuff is
- 3:15Speaker 2: like a real threat that's in front of us right
- 3:17Speaker 2: now that lends credence to the argument that maybe we
- 3:20Speaker 2: should take a beat and figure out how to regulate
- 3:23Speaker 2: this and to have, you know, people come in who,
- 3:25Speaker 2: you know, are knowledgeable about this and to pace the frontier,
- 3:29Speaker 2: as Dario Amadei says. But you're right. The idea of
- 3:32Speaker 2: a slowdown only works if everyone agrees to a slowdown.
- 3:36Speaker 2: And that's one thing to get all of your competitors
- 3:38Speaker 2: in the United States to agree to do something. There are,
- 3:41Speaker 2: you know, strong voices out there, Sam Altman, the CEO
- 3:44Speaker 2: of OpenAI, among them, who agree with Amadei that a
- 3:46Speaker 2: slowdown makes sense. They've taken some steps to doing that,
- 3:50Speaker 2: even over the weekend, saying they've paused development on some
- 3:52Speaker 2: of the most powerful models once they've learned about, you know,
- 3:56Speaker 2: more recent hacking incidents that have happened. But are you
- 3:59Speaker 2: going to get China to agree to do that, right?
- 4:02Speaker 2: And then, you know, it seems like there's some skepticism
- 4:05Speaker 2: around that. So, yeah, it's a complicated problem.
- 4:09Speaker 1: It sure is. And that would be the question. If
- 4:11Speaker 1: you can get the world to agree on this, then
- 4:14Speaker 1: I think we're OK. Then we can we can hit
- 4:16Speaker 1: the pause button for a minute. But if not, everybody's
- 4:18Speaker 1: on board. Then I don't know how you get through this.
- 4:22Speaker 1: You mentioned OpenAI pausing a powerful model. Talk about this
- 4:27Speaker 1: and the hacking incidents.
- 4:30Speaker 2: Yeah, so most recently they discovered a hacking incident that
- 4:33Speaker 2: happened a couple days ago in late September. There was
- 4:36Speaker 2: another model that was being tested. It exploited a loophole
- 4:40Speaker 2: in the digital sandbox that it was put in, so
- 4:43Speaker 2: the testing environment. Specifically, it was an error in the
- 4:46Speaker 2: DNS filtering of that test, literally a program that tries
- 4:49Speaker 2: to keep certain computers, certain IP addresses from doing certain
- 4:53Speaker 2: things in a test environment. there was some misconfiguration problem there.
- 4:57Speaker 2: According to open AI, this thing gained access to the internet.
- 5:02Speaker 2: The model didn't hack into anything after it escaped from
- 5:04Speaker 2: that testing environment, but even still it's notable because it's
- 5:07Speaker 2: the first misalignment incident since the company put in new
- 5:11Speaker 2: security measures following the hugging face hacking incident that happened
- 5:15Speaker 2: over the summer. That was kind of the incident that,
- 5:17Speaker 2: you know, kicked off this whole discussion, caused other companies
- 5:21Speaker 2: like Anthropic and Google to go back into their systems
- 5:24Speaker 2: and see if it happened at their own companies, which
- 5:26Speaker 2: in many cases they found it did. So as a result,
- 5:30Speaker 2: OpenAI says it's pausing development on its most powerful models.
- 5:34Speaker 2: The New York Times also reported yesterday that the company
- 5:36Speaker 2: is opting not to release an upcoming model called GPT 6.1 Astra.
- 5:42Speaker 2: This is reportedly because it showed a willingness to mislead
- 5:45Speaker 2: users about its actions. It also comes after OpenAI Systems
- 5:50Speaker 2: breached an Australian government website, attempted unsuccessfully to hack into
- 5:54Speaker 2: the U.S. Department of Education. So they're instituting this pause
- 5:58Speaker 2: on their sort of frontier model development. However, later today,
- 6:03Speaker 2: they are expected to host Dev Day, which is their
- 6:06Speaker 2: annual developer conference. It's usually where they show off a
- 6:08Speaker 2: bunch of new consumer facing AI products that you can
- 6:11Speaker 2: play around with. Today, we could potentially get a look
- 6:14Speaker 2: at their challenger to Muse, which is Meta's latest AI
- 6:17Speaker 2: agent technology we talked about last week. So at OpenAI,
- 6:21Speaker 2: it's a slowdown, not a stop by any stretch of
- 6:23Speaker 2: the imagination.
- 6:24Speaker 1: You mentioned, you know, when we're testing these AI models
- 6:27Speaker 1: and what's called, quote unquote, and you use the term sandbox,
- 6:31Speaker 1: and it gained access to the Internet. Now, I'm a
- 6:34Speaker 1: B minus student. Shouldn't we just turn off the Wi-Fi?
- 6:38Speaker 1: Because I know in my house... If I don't have Wi-Fi,
- 6:41Speaker 1: I can't get on the Internet. How are these things
- 6:44Speaker 1: gaining access to the Internet if we don't have them connected?
- 6:47Speaker 2: The idea is that it is really only a useful
- 6:50Speaker 2: test if there is at least some Internet in it.
- 6:53Speaker 2: That's how I understand this. The other thing I think
- 6:56Speaker 2: that's important to keep in mind here, and the idea
- 6:59Speaker 2: there being that, like, okay, the agents need to be
- 7:02Speaker 2: able to download, like, some code, they need like their
- 7:05Speaker 2: own little app store in order to solve some of
- 7:08Speaker 2: these problems. And that is to simulate the environment of
- 7:12Speaker 2: the broader internet where the models have access to third
- 7:16Speaker 2: party tools, so they don't have to like develop an
- 7:18Speaker 2: app to, you know, do this, that or the other thing.
- 7:21Speaker 2: But There also needs to be guardrails around like only
- 7:25Speaker 2: letting them see that part of the Internet. So it's
- 7:28Speaker 2: a very complicated problem to solve from what I understand.
- 7:31Speaker 2: The other, I think, important piece of information to keep
- 7:33Speaker 2: in mind here is that some of this is like
- 7:35Speaker 2: these models behaving in ways that we didn't expect. Some
- 7:39Speaker 2: of this is the testing environment, right? Incidents at OpenAI, Anthropic,
- 7:44Speaker 2: and other companies have all basically been tied back to
- 7:47Speaker 2: one company that builds sandboxes. It's a startup called Irregular.
- 7:51Speaker 2: They have put out numerous reports in the last several
- 7:54Speaker 2: months detailing exactly what went wrong at these companies when
- 7:58Speaker 2: it comes to these hacking environments. Important context here, not
- 8:03Speaker 2: to overstate necessarily the capabilities of these models. Some of
- 8:07Speaker 2: this comes with the testing environment itself, which often comes
- 8:11Speaker 2: from a third party.