What could slowing the AI frontier mean for the economy?

SPEAKER_1 (0:00)

This Marketplace podcast is presented by Tomorrow's Cure. If our episodes on the intersection of AI and healthcare left you wanting to learn more about how technology is transforming medicine, Tomorrow's Cure is for you. It's the chart-topping 2025 Ambioward Finalist podcast from Mayo Clinic. Back for a brand new season with new host award-winning journalist Lindsay Sievert, Tomorrow's Cure explores the innovations changing the healthcare landscape. Featuring conversations with leading physicians, researchers and medical experts, the new season examines everything from AI-powered diagnostics and cutting-edge cancer therapies to surgical technologies improving patient care today.

Not sure where to start? Listen to the season premiere featuring MD Anderson radiation physicist Dr. Paige Taylor and Mayo Clinic radiation oncologist Dr. Adam Holtzman. They discuss why carbon ion therapy is generating excitement in the medical community and what it could mean for the future of precision cancer treatment. So go ahead, follow Tomorrow's Cure on Apple Podcasts, Spotify or wherever you get your podcasts.

Megan McCarty-Corino (1:05)

Can we have AI safety and a strong economy? From American Public Media, this is Marketplace Tech. I'm Megan McCarty Carino.

The AI industry appears to be getting more serious about safety. Leaders of top companies signed a non-binding agreement at the White House last week, laying out standards and best practices for responsible development of advanced AI. OpenAI delayed the release of its latest model due to safety concerns. All in the wake of Anthropic CEO Dario Amodei calling to slow things down and pace the frontier.

Wall Street initially wasn't too happy about that push, and stocks related to the AI buildout took a momentary tumble. But Jordan Nanos at the chip research and consulting firm SemiAnalysis argues that greater investment in AI security could actually drive more demand for computing power.

Jordan Nanos (2:09)

At the most basic level, the commitments that Dario and Sam have made for pacing the frontier are all related to spending more compute for an equivalent model release. In other words, they're going to do a bunch of safety testing, evaluations, all sorts of different rollouts of the model to actually make sure that it is safe.

That just means spending more money on compute than they would have otherwise spent in order to release a model of equivalent capability in the past.

Megan McCarty-Corino (2:39)

Dario and Anthropic originally committed to bring in a team of independent evaluators to run additional safety testing. Now it looks like all the major AI companies have agreed to some sort of additional layers of safety audits, right?

Jordan Nanos (2:56)

Yeah. I think it goes beyond necessarily just the model behavior. They've talked about having people to assess the systems and processes that are in place at the company. One of the big things that motivated this was the OpenAI hugging face incident. In this scenario, basically what you're seeing is a issue with the infrastructure that surrounds the training job rather than the model behavior being unexpected or strange. The company was intentionally trying to train the model to perform offensive cybersecurity capabilities, in other words, hack into things.

This is for a good reason. There's dual use of these cybersecurity models.

But what happened is there was a failure in their monitoring systems. The Internet access that they thought was blocked from the model was actually available once the model was able to hack through the systems that were in place. And so I think that part of the commitment from these companies is not just to allow people to understand the model's behavior in a more detailed way, but actually understands the systems and processes that are in place within these labs. Like who can launch these training jobs? How much compute do they have access to? What sort of monitoring systems they have in place? Things like that.

Megan McCarty-Corino (4:08)

Right. And so you point out that it's not just sort of, okay, running additional tests that uses more compute, but also potentially, you know, the process of monitoring chain of thought, which is kind of the inner monologue of the AI agents in a reasoning model kind of showing what they're doing, that that actually takes additional compute as well.

Jordan Nanos (4:30)

It does.

I think I'm suggesting that there's also things that can be much simpler, like for example, monitoring if there's any network access. Now, if we talk about compute constraints and whether the commitments that these companies are making are going to have a negative impact on the market, we think it's a pretty consistent straight line that they are going to be spending more and more money on compute over time. We don't view that the commitments that they've made to pace the frontier actually mean that they are going to spend any less money on compute.