Huawei Says China Must Build Faster AI to Understand Frontier Risks: Inside the Safety Gap, Agent Boom, Chip Shortage and High-Stakes Race With U.S. Labs

SPEAKER_1 (0:09)

Welcome to The Daily AI Chat. This podcast was curated by our human friend Fred, and we are your dedicated AI hosts.

SPEAKER_2 (0:16)

Glad to be here with you all today.

SPEAKER_1 (0:17)

So right now, top US AI laboratories are honestly terrified. They're looking at their models and realizing they're getting way too smart, way too fast.

SPEAKER_2 (0:29)

It's a real panic.

SPEAKER_1 (0:30)

Right. They're looking at these highly advanced frontier AI models and essentially sounding the alarm, like warning the entire world that we need to hit the brakes before these systems become just completely uncontrollable.

But then across the globe in China, a top tech executive is arguing something completely, well, counterintuitive.

SPEAKER_2 (0:49)

Oh, completely.

SPEAKER_1 (0:50)

He's stating that developers actually need to build significantly more powerful AI just to even be able to perceive the dangers that the West is so afraid of.

SPEAKER_2 (0:59)

It is a really profound paradox. And it gets right to the heart of how different parts of the world actually understand risk. So today, we are diving into a really comprehensive report, published today, September 17th, 2026, by Reuters.

And the reporting team here, that's Casey Hall, Che Pan, and Eduardo Baptista, they've basically mapped out this fundamental disconnect in how the future of AI is being engineered.

SPEAKER_1 (1:27)

And more importantly, how it's being safeguarded, right? Yes, exactly.

SPEAKER_2 (1:30)

How it's safeguarded. Because the US perspective right now is built entirely around this concept of frontier models.

SPEAKER_1 (1:37)

Right. For those tracking the space, when researchers talk about the frontier, they mean the absolute bleeding edge.

SPEAKER_2 (1:43)

The bleeding edge, yeah.

SPEAKER_1 (1:43)

Like systems push so far beyond standard benchmarks that they begin exhibiting what they call emerging properties.

SPEAKER_2 (1:51)

Which is wild when you think about it.

SPEAKER_1 (1:52)

It really is. I mean, you train a model simply to, like, predict the next word in a sequence, but suddenly it starts writing functional autonomous code.

SPEAKER_2 (1:59)

Or learning a completely new language.

SPEAKER_1 (2:01)

Yeah. A language, it was never explicitly taught. And this happens because of scaling laws.

When you increase the training data and just exponentially increase the raw computing power, the capabilities don't just grow in a straight line.

SPEAKER_2 (2:15)

No, they leap forward.

SPEAKER_1 (2:16)

Exactly. They leap forward in these unpredictable ways. And so leading US labs are observing these sudden leaps inside their black box models, and they are, well, hitting the panic button.

SPEAKER_2 (2:29)

Which brings us to the core of the Reuters report. Because Eric Xu, who is the rotating chairman of Huawei.

SPEAKER_1 (2:35)

So Huawei Connect speech.

SPEAKER_2 (2:36)

Yeah. His recent speech at Huawei Connect in Shanghai, he offered a starkly different reality for Chinese developers. His entire argument is rooted in hardware limitations.

SPEAKER_1 (2:48)

Okay, let's unpack this because the analogy that comes to mind here, it's like trying to figure out how a sports car handles at 200 miles per hour when you're currently driving a golf cart.

SPEAKER_2 (2:56)

Oh, that's a great way to put it.

SPEAKER_1 (2:57)

Right. Like you literally can't test the aerodynamics until you build a faster engine.

SPEAKER_2 (3:01)

What's fascinating here is that zoos are arguing exactly that. Because Chinese labs are operating with significantly less computing power or compute, their models just aren't hitting those scaling laws in the same way.

SPEAKER_1 (3:14)

So they aren't reaching that extreme frontier.

SPEAKER_2 (3:17)

Right. And consequently, they aren't witnessing these emerging properties. They're just not seeing the sudden leaps in autonomy or those complex deceptive behaviors that are keeping US researchers awake at night.

SPEAKER_1 (3:28)

So to Zhu, you can't really build effective safety protocols for a threat you can't even empirically observe in your own lab.

SPEAKER_2 (3:35)

Exactly. To him, regulating AI based on the hypothetical fears of US researchers is, well, it's a purely theoretical exercise.

SPEAKER_1 (3:44)

Which fundamentally changes the relationship between capability and risk.

SPEAKER_2 (3:47)

It really does.

SPEAKER_1 (3:48)

Like, if we look at this through the lens of, say, astronomy, US researchers are looking through this massive, highly advanced deep space radio telescope and they're pointing out dangerous anomalous objects in some distant galaxy.

SPEAKER_2 (4:01)

And Chinese researchers.

SPEAKER_1 (4:02)

Due to hardware limits, they only have access to like high-end optical binoculars. So the US is warning everyone, hey, stop building bigger telescopes because of the terrifying things we've seen.

But the Chinese perspective is, we have to build a bigger telescope just to verify if those anomalies even exist.

SPEAKER_2 (4:20)

Right. They can't regulate their own space program based on something they literally cannot see with their own instruments.