The world is too loud. Read what matters.

80,000 Hours Podcast

Beijing's greatest AI fear is not AI loss of control, but regime loss of control

By «loss of control», Chinese officials first mean the loss of political control, not alignment failure; more than any AI safety paper, it is the network attack capabilities America demonstrated in the Mythos incident that have truly unnerved Beijing over the past year.

US-China relationsAI governanceGeopoliticsChip controlOpen-source modelsTaiwan

The video won't play here. Listen to the audio instead:

If you want to know what China actually fears about AI—rather than what Americans think it should fear—this episode offers rare firsthand judgment, though the latter half's content on personal paths to China expertise has lower information density.

The argument · timestamps estimated from transcript position

3:39

Beijing treats Silicon Valley's internal disputes as one unified operation

Beijing state media have repeatedly singled out Anthropic, calling it ‘infected with American disease’ and ‘a monster spawned by the American way’. Julian notes this reveals a key misalignment: Americans are accustomed to viewing different AI companies as competing, independent actors with different philosophies, but Beijing makes no such distinction—whether Anthropic, OpenAI, or others, Beijing sees them all as different components of one American project to pursue global technological hegemony.

— Julian Gewirtz
13:24

In Beijing's lexicon, «loss of control» means loss of political control first

China's AI discussions increasingly mention ‘extreme loss of control risk’, but the term means something different from American AI safety circles. When Americans say loss of control, they mean humans losing control of AI systems; when Beijing says loss of control, it first means the loss of control over information, public opinion, and Party authority—whether through companies losing control and leaking, or people deliberately using AI to spread information the Party doesn't want seen. This is the first premise for understanding Beijing's regulatory logic: it does not start from ‘how should we align AI’, but from regime security.

— Julian Gewirtz
15:13

What unnerves Beijing is not AI itself, but American cyber capabilities

The single event that most shaped Beijing's AI risk perception over the past year is what's called the Mythos incident—it made Beijing realize America is acquiring extraordinarily powerful cyber capabilities. This shakes Beijing's nerves more than any AI safety paper could, because it directly touches the two things Beijing cares most about: regime security and vigilance against America, and these two are linked. Every time Beijing's top leadership discusses AI risk, America is raised as one of the threats.

— Julian Gewirtz
18:25

When Kimi and DeepSeek requests were forwarded to Claude

Anthropic reported that some requests sent to Kimi and DeepSeek—including one apparently from a PLA officer—were forwarded to Claude. Julian suggests that when Chinese officials spoke with the two companies' CEOs, beyond technical investigation, the political angle—‘this embarrasses China and these companies’—was the real focus. More crucially, Beijing state media directly linked this report to Dario's call for capping frontier pace, framing it as ‘America's same playbook to suppress Chinese AI and maintain its lead’—Americans may not connect these two things themselves, but Beijing does.

— Julian Gewirtz
39:21

To know if Beijing believes in RSI, watch whether it buys chips

To judge whether Xi truly believes AI will bring 10x acceleration, don't wait for an ideological manifesto—watch concrete resource allocation: whether China will concentrate compute, whether it will permit open purchases of Nvidia chips. Currently, the Trump administration has approved some chip sales, but Beijing is instead buying in limited quantities—both to support domestic compute like Huawei and because it can access international compute remotely through third countries. Ding Xuexiang was recently reported to have stated that if forced to choose between ‘moving slower’ and ‘depending on foreign compute’, Beijing would rather move slower.

— Julian Gewirtz
52:19

Taiwan has nothing to do with AI, but 2028 is a window

Beijing's desire to take Taiwan is unfinished business dating to the regime's founding in 1949, existing long before the AGI concept appeared—so we shouldn't assume AI anxiety drove the decision to use force. What shapes Beijing's timeline is Taiwan's presidential election in January 2028—Beijing hopes to elect a candidate favorable to it, so for now it leans toward maintaining pressure without military action, because military options carry too high a cost and Beijing won't resort to them unless forced to.

— Julian Gewirtz
57:39

The first sticking point in US-China AI dialogue is vocabulary, not stance

Even if the US and China establish technical working group communication, the first hard problem is terminology itself. In Chinese, ‘safety’ and ‘security’ are the same word—‘安全’—which caused a problem when an American AI expert distinguished system alignment safety from network security at a meeting; Chinese-speaking participants who didn't know English couldn't follow. This isn't a disagreement of principle, but missing infrastructure—and one reason why track II dialogue matters despite being difficult.

— Julian Gewirtz
1:08:28

Open source was once a political weapon, now that grip is loosening

Beijing previously framed ‘open source’ as a political narrative against America—‘America practices closed exclusivity, China stands for openness’, with DeepSeek as its strongest symbol. But in the week of recording, China's AI safety governance framework (TC260 3.0) used rare language warning that open-source releases lower the threshold for launching complex cyber attacks; scholars also began voicing ‘cannot allow unlimited openness; constructive forces can use it, destructive forces can too’. Julian emphasizes this discussion has just begun, but once Beijing judges safety and stability risks are out of control, policy can shift quickly.

— Julian Gewirtz

In their own words · checked verbatim

Anthropic has contracted the American disease.

Julian Gewirtz3:39

a monster bred by the American path

Julian Gewirtz3:39

There’s no day after tomorrow if China wins at this.

Julian Gewirtz4:53

The United States has been pacing the frontier for China for years. So we don’t see this as really anything new about us. And we don’t think the United States is going to slow down.

Julian Gewirtz26:19

if there’s a tradeoff between going a little bit slower and making sure that you’re not dependent on foreign compute, go a little bit slower

Julian Gewirtz41:59

Now, in Chinese, safety and security are the same word, anquan.

Julian Gewirtz58:49

crossing the river by feeling for the stones

Julian Gewirtz1:19:40

Figures

Taiwan presidential electionJanuary 202854:24
CCP regime founding194952:19

Glossary

loss of control
In Beijing's usage, first means loss of political or information control, distinct from the American AI safety sense of humans losing control of AI systems
anquan
Chinese term encompassing both English «safety» and «security», leading to conceptual misalignment in US-China technical dialogue
AGI-pilled
Believing AGI is imminent and acting accordingly; some Chinese labs hold this view more strongly than the government does
Mythos
Incident exposing America's sophisticated cyber attack capabilities, the single most consequential event shaping Beijing's AI risk perception over the past year

How to listen

Who it's for

Founders, investors, and policy researchers tracking US-China AI competition, geopolitics, or China business compliance.

Skip

From 1:20:20 onward: personal career advice on becoming a China expert and whose books to read—loosely related to AI policy, can be skipped.