AI leaders want to slow development. Can they actually do it?

SPEAKER_1 (0:00)

Support for Here & Now Anytime comes from MathWorks, creator of MATLAB and Simulink software, to design and develop engineered systems, accelerating the pace of discovery in engineering and science. Learn more at mathworks.com.

SPEAKER_2 (0:16)

WBUR podcasts, Boston.

Ina Fried (0:20)

The models are now helping the next model get better, and these models don't seem to be fully under control.

Robin Young (0:27)

It's not all powerful, but today's artificial intelligence is causing lots of problems anyway. So we arrive again at a familiar question. How should we rein in big tech?

It's Monday, September 14th, and this is Here & Now Anytime from NPR and WBUR Boston.

Welcome to the next big vibe shift in artificial intelligence. Today Microsoft published a new humanist AI code of conduct that says people should always be in control. That's after the leaders of OpenAI and Anthropic over the weekend called for a global slowdown in AI research and development. For his part, President Trump today posted, the only control or guardrails that AI needs is a strong and smart high IQ president. In a few minutes, we're going to hear from a security and tech analyst who worked in the first Trump and Biden administrations. But before that, Ina Fried, Chief Technology Correspondent for Axios. She told Scott Tong, fears about AI are getting real.

Ina Fried (1:41)

I think what's new is that these executives have real world things they can point to versus just fears of what's coming. A lot of people in the industry have had fears that we would be at this point at some point. I think what's new is a bunch of leaders in businesses that are actively building this are saying we're pretty close to that point right now.

Scott Tong (2:03)

So now these leaders in Silicon Valley are calling for what's called pacing to slow AI down to get a handle on these uncertainties, not to slow down too much to let the Chinese companies catch up. So how would this Goldilocks speed actually work?

Ina Fried (2:20)

Well that's the really tricky part is one, how do we get everyone to slow down, not just a few because I think the biggest problem right now is you have what are called race conditions, which is I think everyone agrees in the industry except for maybe President Trump, that things are moving too fast and it would be better if things were slower, but they're also in a race. These companies are in a race with each other, especially OpenAI and Anthropic, but there's also a race between the US and China. So really to get the conditions where companies and countries feel comfortable slowing down, you need some sort of national and then ultimately international agreement. And that's gonna be tough if you have a president that says there's no problem.

Scott Tong (3:04)

Yeah.

What can you tell us about the response after these weekend warnings? I've read that some in the AI space say, it's not enough, you gotta do a lot more regulation, a lot more slowing down. Others saying, maybe these CEOs are overdoing the risk, perhaps to pump up their future stock listings.

Ina Fried (3:26)

Yeah. I think the idea that this is all a marketing ploy, I just don't buy that. I think the evidence is too strong that these capabilities are growing very close to a point at which we could lose control. And there's a few reasons for that.

The models are now helping the next model get better. And there are some companies eyeing what's called recursive self-improvement, which is the models just getting better on their own. And I think that's a really dangerous direction to head at a time where these models don't seem to be fully under control. I mean, we had the hugging face incident. We've had other examples where mostly in testing, but these models have started communicating with one another. So I really don't buy the argument that this is all somehow just marketing. In terms of slowing down, I think that the tricky part is how and how much.

Scott Tong (4:18)

Yeah. Well, I mean, as far as this hack of this AI company, hugging face, it's been reported, we talked to you about it, that these AI agents suddenly started doing things the humans did not ask them to do, did not intend them to do. And it gets to this broader question of how much do the operators, the developers actually know about how the AI agents are doing what they're doing.

Ina Fried (4:44)

Not nearly enough. And it's surprising when I say that in other types of technology, we've understood pretty well what's going on under the hood, or at least the experts have. What's been really different with this brand of AI is we only sort of know what's going on. We know they're picking what's most likely to be next, but clearly there is more going on.