Breakthroughs in Chinese AI models threaten U.S. lead in tech race

Geoff Bennett:

New breakthroughs in powerful low-cost Chinese A.I. models are sending shockwaves through Washington and Silicon Valley, as Chinese companies continue to narrow the A.I. gap, despite vast differences in computing power and financial resources.

A pair of new A.I. models released in recent days are what's called open-weight. They're free to use and the developers say perform at levels approaching the best models made by American A.I. giants like Anthropic and OpenAI.

It all comes as the White House announces a new plan to redirect billions of dollars in funding from universities to individual researchers in a bid to accelerate A.I. innovation.

To break it all down, we're joined now by Amrith Ramkumar. He's tech reporter at The Wall Street Journal.

Thanks for being with us.

Amrith Ramkumar, The Wall Street Journal:

Thanks so much for having me.

Geoff Bennett:

So these new A.I. models from China's Moonshot AI and Alibaba, they're open-source. They're cheap. They appear to be closing in on the leading American systems. Just how powerful are these systems and how has China been able to do this?

Amrith Ramkumar:

These systems are quite powerful, approaching the capabilities of the top models from U.S. companies like Anthropic and OpenAI.

But you really hit on it with the question. The point is that they're much, much cheaper. And this comes at a time when a lot of U.S. companies are spending millions and millions of dollars on A.I. and their employees' different endeavors using the technology, and they're desperate to cut costs.

That's what's scary to a lot of people, that these Chinese models will seem very attractive for a lot of U.S. consumers and businesses. As to how they have been able to do it, Chinese engineering has been doing this for a long time in various industries, right? They figure out how to get by with less sophisticated technology.

They connect chips together and find work-arounds, despite U.S. export restrictions on the most advanced semiconductors. In this case too, it's important to note that they're using a technique called distillation, in which they're essentially training off recently released U.S. models.

And this is something that has Trump administration officials very worried and some people in Silicon Valley that basically they're almost being ripped off. And it's something where we could see the Trump administration try to crack down in the coming days and weeks.

Geoff Bennett:

So if the Chinese models are almost as good and much cheaper, could they start to dominate the market? That's what the concern is.

Amrith Ramkumar:

Exactly.

And another point to mention is that leading U.S. companies are now subject to increasing scrutiny and oversight from the Trump administration and other regulatory agencies around the world, whereas Chinese models don't have to abide by those same rules and they don't have the same safeguards built in.

So some people say they're almost as powerful and they're not nearly as safe. That's a really dangerous combination if they were to fall into the wrong hands.

Geoff Bennett:

For people who aren't following this as closely as we are, when we say that these A.I. models are powerful, what exactly are we talking about?

Amrith Ramkumar:

The real thing we mean when we say these models are powerful is that they're capable of aiding or even launching cyberattacks or hacks.

And so the concern is that they can act on their own, essentially. And that was highlighted just yesterday, when OpenAI said that two of its models during training had essentially broken out of their cage, gone onto the Internet, and hacked another company which is essentially an A.I. model library.

This kind of thing is very scary, and the capabilities from the Chinese models are that they're almost to the point of doing that exact same thing. That has a lot of implications for national security and geopolitics moving forward.

Geoff Bennett:

And you have reported that top A.I. executives, U.S. A.I. executives, are asking the White House to restrict access to these Chinese A.I. models. Has the White House been receptive?

Amrith Ramkumar:

The White House is debating how best to respond to this. Some A.I. executives, including leaders at companies like OpenAI and Anthropic, have sounded the alarm on the threat posed by these Chinese models, and they say they just want a level playing field.

A lot of other people in Silicon Valley and in the tech community are on the other side. They say that the administration should just let these models proliferate and that U.S. companies should have to compete and that you the U.S. should develop its own open models, and that will help solve this problem without big government policy intervention.

But the administration has confirmed in the last 24, 48 hours that they are looking at actions potentially including putting Chinese companies on trade blacklists and things of that nature. This comes at a really interesting time. Airbnb and other U.S. companies are relying on these Chinese models to run their own businesses.

So if the government was to deem them a security risk, essentially, that could transform a lot of the A.I. race.

Geoff Bennett:

It's interesting. I didn't realize that American companies were already using these Chinese models.

You broke the news yesterday that the White House plans to overhaul federal A.I. research funding and redirect some $200 billion away from institutions and toward individual scientists. Tell us more about that.

Amrith Ramkumar:

The White House is going all in on A.I. The science funding overhaul is the latest example of that. So this is the $200 billion dollar research and development engine across all federal agencies.

So this is a very sprawling enterprise. And the White House Office of Science and Technology Policy yesterday basically said it wants A.I. at the center of all of it. It wants A.I. basically solving these problems, accelerating research, ostensibly creating new institutions that have the technology at the center.

Another big focus is prioritizing individual researchers, instead of large universities. And this could have massive ramifications for colleges across the country that rely on this federal research funding, and it adds to the administration's past moves to sort of cut down on essentially what it calls bureaucracy and red tape that are holding back innovation.

But putting A.I. at the center of the research enterprise across the federal government is a huge move. I mean, a lot of people say these models are getting better, but they're still not ready to direct federal research dollars essentially and solve these problems on their own without human oversight.

Administration officials said that humans will be in the loop, they will have that oversight. But it will be fascinating to see how this plays out in the coming weeks and months.

Geoff Bennett:

Fascinating, indeed.

Amrith Ramkumar, tech reporter for The Wall Street Journal, thanks again for your time. We appreciate it.

Amrith Ramkumar:

Thanks so much for having me.