Why China's Master Plan for AI Never Mentions AGI

Brian Tse began in AI at Tsinghua University and at a deep-learning hardware startup just as the field took off in China. AlphaGo was exciting, he tells Nathan Labenz, but the questions that actually captured him were the big governance ones. That curiosity took him to the United Kingdom as a policy affiliate of the Centre for the Governance of AI, then founded at Oxford, and to a talk at DeepMind in London. He later served as a senior adviser to the Partnership on AI and, around 2019, consulted for OpenAI on the societal implications and release strategy of early language models such as GPT-2.

Back in Beijing he worked with the Beijing Academy of Artificial Intelligence, including on some of the first AI ethics principles issued by a Chinese institution, and then founded Concordia AI. The independent social enterprise has three pillars. It sits on national standards committees and advises government bodies. It works directly with leading labs on safety frameworks, including a recent frontier-risk document co-authored with Shanghai AI Lab. And it convenes forums in China and Singapore and publishes English-language research so outsiders can actually read what is happening in the region.

Labenz asks for a Western analogue and offers FAR.AI, the group led by Adam Gleave. Tse grants a lot of overlap in mission. What is distinctive, he says, is Concordia's place inside national policy and standards-setting, a role he cannot assume other civil-society groups share. Labenz notes that FAR.AI is moving toward policy after a research-heavy start. Even so, Tse's day job is not only studying China from the outside. It is helping write the rules on the inside.

Labenz sketches the American map as five or so "live players" who can actually shape the future of the technology: OpenAI, Google, Anthropic, plus xAI and Meta, with Microsoft on the edge of the list. They are densely clustered in the Bay Area, culturally as well as geographically. The people building these systems have known one another for years and share a set of essays, arguments, and science-fiction touchstones that predates the current boom. He cannot draw the same sketch for China, and he is not sure how much of the Chinese lineup was obvious from the inside.

Tse describes several hubs rather than one. Beijing has Baidu and ByteDance, and also Moonshot AI and Zhipu AI, both founded by Tsinghua alumni in a city with a high density of researchers and founders. Shanghai is home to MiniMax and other multimodal startups and hosts the World AI Conference. Hangzhou is headquarters for DeepSeek and Alibaba; in 2025 a local slogan, the "six little dragons of Hangzhou," grouped DeepSeek with robotics firms such as Unitree. Shenzhen, in the south, has Huawei and Tencent. The same geography shows up in safety research. Concordia's *State of AI Safety in China* report counts about 31 groups that have published on the topic, with the top five cities Beijing, Shanghai, Hangzhou, Shenzhen, and Hong Kong.

Asked whether those hubs share a culture analogous to the Bay Area's, Tse points to a structural difference rather than a shared canon. Across the Chinese cities, the tech ecosystem and the policy ecosystem are "quite intertwined." His sense of the United States is a wide gap between Silicon Valley culture and Washington. That, he suggests, produces more coherent views on policy questions across Chinese AI hubs than the American split would lead you to expect. The commercial mainstream still looks familiar: scaling foundation models, multimodality, reasoning, agents, and compute at training, post-training, and inference time, following the path opened by GPT-3. Alongside that sit a stronger official emphasis on embodied AI in the physical world and institutes such as the Beijing Institute for General Artificial Intelligence, whose director frames a "small data, big task" alternative to the big-data foundation-model path.

Labenz's standing complaint is that Western AI leaders offer a vague promise to cure disease and solve every problem, with almost no picture of ordinary life on the other side. Chinese surveys, meanwhile, show more public optimism than American ones. Concordia's review of the limited polling finds that the Chinese public generally thinks benefits will outweigh harms. One post-ChatGPT survey found people willing to grant that AGI could pose existential risk and still in favor of building it, on the grounds that the risk is largely controllable. A more recent student survey found roughly 80 percent agreeing that AI would do more good than harm.

Tse puts that optimism in family terms. His parents were born in the 1960s in Fujian. In their lifetime, he says, China's per capita GDP rose more than 140 times, and the extreme-poverty rate fell from about 88 percent in the early 1980s to close to zero. If you have watched that kind of transformation, or heard it from your parents, you tend to be more hopeful about what technology and decent governance can deliver. Labenz calls the rebound incredible, even starting from a low base after grim years, and notes that Chinese cities now look world-leading in many respects.

The other half of Labenz's question is how Chinese companies talk about the West. In American conversations he keeps hearing that China will never slow down, so the United States cannot either. Tse does not describe a mirror-image obsession. What is "significantly less" discussed among Chinese companies and policymakers, he says, is a specific AGI timeline or a clear finish line. Entrepreneurs talk more about technological self-sufficiency, useful applications, and being profitable. There is also an idealistic strand: open-sourcing AI so countries that lack capacity can still use it.

The concrete vision, Tse says, is in a high-level directive on "AI plus" released by the central government in August 2025. It treats AI as a new economic engine, a general-purpose technology on the order of the internet or electricity, to be integrated across the whole economy. Six pillars follow. First is AI as a tool for scientific discovery and innovation, listed first because Beijing sees it as a force multiplier for everything else; the document even calls for using AI to advance the social sciences and philosophy. Second is industrial transformation: autonomous factories, smart logistics, drones and robots in agriculture. Third is consumption, including new consumer experiences and brain-computer interfaces.

The remaining pillars are more social than sci-fi. AI as a human collaborator is framed as support for labor rather than full replacement, with agents as copilots. AI as a tool for government efficiency sits beside a sixth pillar of international cooperation: treating open models as an international public good, especially for the Global South. "Notably throughout this comprehensive national blueprint for AI it doesn't mention AGI or super intelligence or any other such concept," Tse says. The real race, in his reading, is not beating the United States to AGI supremacy. It is wiring AI into the real economy over years and decades.

Labenz remembers a Washington Post op-ed from a Chinese official, several years old, arguing that powerful AI might finally make central planning work. He does not hear that in Tse's sketch and asks whether anyone still bets on AI as the missing piece of state direction. Tse's answer is blunt: that sounds "a little bit utopian for the Chinese policy discourse." People are not betting on AI to solve everything, including the central planning of society. The practical version is narrower. He cites early-warning systems for natural disasters, where real-time data lets the government prepare earlier. Valuable on the margin, he says, but not a unifying silver bullet.

Tse's Time op-ed argues that China is taking AI safety seriously. The evidence he leads with is hierarchy, not vibes. The 2024 Third Plenum, one of the most important political meetings in the country, called for an oversight system to ensure the safety of AI and classified it as a major public-safety concern. In February 2025 an updated national emergency-response plan put AI risk alongside cybersecurity, biological security, and natural disasters. In April the Politburo held a study session on AI. Xi Jinping, Tse recounts, said the technology brings unprecedented opportunities and could also bring unprecedented risks. The readout was more detailed on safety than earlier high-level statements, walking through monitoring, early warning, and emergency response.

Labenz wants the American translation. A White House meeting can be a photo op that dies with the next administration. What do these Chinese sessions actually do? Tse says they signal priorities to a domestic audience. Local officials study them; investors and companies watch. That is how the state tries to mobilize a whole-of-society push, even if actors still interpret the line in different ways. Two differences with the United States matter more than the meetings themselves. Continuity: once central leadership sets a direction, it can be a multi-year or multi-decade commitment. Coherence: the Politburo sets a tone and implementation is more unified than a split among White House, Congress, and everyone else. He adds a cultural note. The public appears more willing to accept regulation. Academics who worry about catastrophic risk have briefed Politburo leaders directly, in a society where scholars have long outranked entrepreneurs. The discourse, he says, sits nearer the middle than America's spread from pause petitions to effective accelerationism.

On whether China will accept a performance hit for safety, Tse refuses the zero-sum frame. Vice Premier Ding Xuexiang, the senior official for science and technology, told the World Economic Forum in early 2025 that if the braking system is not under control, you cannot step on the accelerator with confidence. Concordia's reading of Chinese policy documents is the same idea drawn as a 45-degree line: more dangerous capability should come with stronger safeguards, hand in hand. Local government still has room to experiment. Provincial AI plans have appeared from Shenzhen to Shanghai. Self-driving experimental zones are one example of trying things locally before nationalizing them, the old Shenzhen playbook. In 2025 a national rule required explicit labels and implicit metadata on AI-generated text, audio, video, and simulated environments from September. Tse has watched Shanghai try to pioneer a consortium of model developers such as MiniMax and platforms such as RedNote to make watermarking and provenance actually work.

Labenz's American baseline is that you can ship. His sharpest example is xAI launching Grok 4 while Grok 3 was identifying itself as "MechaHitler," with no pre-clearance and no obligation to test before production. In China, Tse says, a deployer of a publicly available chatbot has to pre-register with the government. Developers must run safety tests against a national standard covering 31 risks. The Cyberspace Administration of China gets pre-deployment access to the model and has to approve release. Over the last two years more than 500 systems and model versions have been filed. Tse describes it as a filing and registration process that looks like a licensing regime, with the interim measures on generative AI among the most important rules for general-purpose models today.

The list of registered models sits in a public database. Thresholds are numerical: on one set of queries, at least 96 percent of outputs must be judged acceptable. Companies also create test accounts so provincial cyberspace authorities can red-team the model before launch. The loop can run through multiple rounds of feedback, fine-tuning, and conversation until regulators are satisfied. Labenz is startled by the 4 percent remainder. His prior about the Chinese internet was a much lower tolerance for off-message speech. Tse walks through the legislative history. An early draft of the generative-AI rules spoke of outputs that were highly reliable and truthful, almost perfectly so. During a three-month comment period, industry argued that perfect truthfulness was technically infeasible, and the final text was relaxed. It was a real concession to the limits of the technology.

Model output is only one layer. Beijing passed a rule that hospitals and medical institutions may not let chatbots write prescriptions without a human in the loop, a response to hallucination as much as to politics. Internal research and pure business-to-business deployments sit outside the chatbot filing rules. There is no fast lane for a small wrapper around DeepSeek or Qwen: if you offer a broadly available chatbot or recommendation system, the binding regulations apply. Training-time rules are softer. A national standards body, TC260, has published an AI safety governance framework, now in a 2.0 update, that flags catastrophic risk, loss of control, and CBRN misuse, and that recommends filtering hazardous knowledge in pre-training. As far as Tse knows, that document is not yet binding regulation.

Labenz's reaction after reading Concordia's work with Shanghai AI Lab is that if you took the names off the page, he would not be confident which country had written it. The frontier AI risk-management framework is meant to help developers of general-purpose models, especially at the frontier, handle risks to public safety and national security. It borrows from safety-critical industries and from international practice. Red lines are unacceptable thresholds. Yellow lines are early-warning indicators that should trigger tighter measures. The list includes cyber offense, biological risk, large-scale manipulation and persuasion, and loss of control. A companion technical report evaluated more than 20 open-weight and proprietary models. Under loss of control it broke the problem into self-replication, deception and scheming, and the possibility of uncontrolled AI research and development.

That similarity makes Dario Amodei's claim about DeepSeek-R1 harder to parse. Anthropic's CEO had called it among the least guarded models they had tested on biological-weapons queries. Tse offers several angles rather than a denial. China's safety path is driven mainly by binding government rules, so firms may be less likely to publish evaluations the rules do not require; compliance cost reduces the appetite for extra measures. Concordia's industry survey found the standard toolkit in use: data filtering, safety fine-tuning, RLHF, constitutional AI, real-time misuse monitoring. Those practices appear aimed at the 31 regulated categories. The vast majority of Chinese companies have not publicly shared CBRN or loss-of-control evaluations.

The picture is not frozen. DeepSeek-R1's paper was peer-reviewed in *Nature*, which Tse calls a first for a widely used language model. The safety section is candid: open weights can be fine-tuned to strip protections, and reasoning models tend to expose more sensitive knowledge. The paper still concludes that R1's safety level is generally comparable with other frontier models, including GPT-4o. Concordia's own tests on SOS-Bench-Bio found DeepSeek's refusal rate on harmful biological prompts rising from about 11 percent in V3 to about 54 percent in V3.1, around the median of other models. Tse is not ready to abandon open weights. At the World AI Conference in Shanghai, China announced a global AI governance action plan that treats open source as a strength, while Chinese documents also flag mischief from open weights. The TC260 framework already points toward removing hazardous knowledge in pre-training, the same family of ideas as the UK AI Security Institute's "deep ignorance" work. The science of making open weights hard to misuse, he says, is still early.

Labenz puts the hard version of the Washington objection on the table. Even if China's domestic rules look serious, many officials will still say you cannot trust the CCP when the stakes are militarization and strategic advantage, so America has to assume the worst and race. Tse does not pretend distrust is one-sided. He does argue that the specifics of AI matter. Worry about advanced AI developed in secret, he says, is "not grounded in the reality of AI development." China's state-of-the-art models are largely released openly. Kimi, MiniMax, Qwen, and DeepSeek are now among the world's best open-source systems, which at least allows third-party auditing. The most advanced models on both sides are built in the private sector by a small global research community in which people mostly know one another. He also does not hear as much fear and paranoia on the Chinese side. The AI-plus push is a sign that applications and delivery for the public remain the party's top priorities. What does create resentment is export control, read in China as an American attempt to freeze technological progress and keep superpower status. Tse notes that much of the denied compute would have gone to civilian uses, from generated art to medical diagnosis.

That context is how he reads China's refusal to buy Nvidia's H20s even after the Trump administration allowed the sales. As of 22 September, he says, Chinese companies were also not buying Nvidia's RTX Pro 6000D, seen as low quality and overpriced. Distrust of the H20 is specific: Chinese regulators have raised concerns about back doors, remote location tracking, and remote shutdown. Tse offers a balance-of-power reading. During the first Trump trade war and the Biden-era controls, China was less prepared. Now it is advertising that it need not depend on the United States for its technological future, while America depends on China for critical minerals in trade talks. Export controls, in this telling, catalyzed homegrown alternatives. Huawei's Ascend line is the exhibit. This month, Tse says, Huawei announced an Atlas 950 super node supporting more than 8,000 chips and a supercluster using more than 500,000. From the industry's point of view, this is a window not to stay tied to the American stack.

Western chip analysts still expect the Nvidia-TSMC complex, with Taiwan in a critical role, to outproduce Chinese fabs by more than an order of magnitude over five years. Tse does not simply contradict the single-chip gap. He changes the units. China can use energy abundance to offset weaker chips. Huawei's Cloud Matrix 384 uses more than five times as many Ascend chips to more than compensate for each chip being less powerful than an Nvidia Blackwell. If the main trade-off is energy efficiency, that matters less in a country that, he says, added grid capacity equal to the entire US grid in a decade, much of it solar, hydro, and wind, with a potentially leading role in nuclear. The Chinese stack can prioritize scale over power density. Algorithmic efficiency is the other lever. The *Nature* paper on DeepSeek-R1 put the training cost at perhaps 500 Nvidia H100s for the performance it reached.

On a stable bargain under distrust, Labenz raises the "superintelligence strategy" paper by Dan Hendrycks, Eric Schmidt, and Alex Wang and its idea of mutually assured AI malfunction, or MAIM: great powers ready to sabotage a rival's AI infrastructure so no one gets a monopoly. Tse credits the authors for rejecting single-power dominance as the goal. He still thinks the nuclear analogy fails twice. Nuclear deterrence works because a test or an ICBM launch is detectable. An "aggressive development of superintelligence" is not a clean event. METR's curve of task length AI can complete doubles about every seven months, and it is unclear which point on that graph would count as an intelligence explosion. From a Chinese vantage, American leaders already talk about AGI in 2027 while running export controls. One could imagine that the United States is already racing for a monopoly. Does that trigger MAIM? "I hope not."

The second failure is escalation. If countries publicly endorse the strategy, Tse says, they are declaring a willingness to go to war. The target would not be a single data center. It would be among the most valuable national assets and critical infrastructure of another country, a "hair-trigger alert" rather than a plan for stability. Labenz agrees that the fog would be thick on both false positives and false negatives: low visibility into what the other side is doing, and no agreement on what would even be worth sabotaging.

Tse's substitute, also sketched in the Time piece, has three pillars. First, AI powers should define concrete red lines on frontier capabilities and behavior: no systems that pose catastrophic or existential risk, no AI-enabled proliferation of weapons of mass destruction, no autonomous self-replication or self-improvement into an explosion that humans or other countries cannot comprehend and control. Second, continuous testing for early-warning indicators, ideally with shared protocols so countries and leading companies exchange the most relevant evaluation results. Third, emergency-response protocols written in advance, so that when a threshold is crossed the response is more safety measures and mandated human oversight, not improvisation. He takes some comfort from the fact that US and Chinese leaders have already identified AI risk as a topic for dialogue and moved toward a joint understanding that humans should remain in nuclear command and control. The idea of keeping a human in the loop on the most critical systems, he argues, can be built out from researchers to industry and eventually to official tracks, especially on loss of control.

Away from official tracks, Tse treats scientist-to-scientist convenings as an old great-power tool. Track-two dialogues can build trust, refine policy ideas for governments, and tackle problems experts are better placed to work on now, such as threat models at the intersection of AI and biological risk. He reaches for the Pugwash conferences on science and world affairs, held regularly since 1957, as a precedent for reducing catastrophic risk. Singapore is the venue he is investing in. Concordia's report on the city-state describes an outsized role: it convenes events such as the Singapore Conference on AI, which produced a consensus document on global safety-research priorities, including loss of control, drafted by more than 100 experts. Leading American and Chinese companies keep offices there. Singapore runs testing through an assurance hub and an AI Safety Institute. Neutrality plus a real technical scene is why Concordia opened an office in the city.

Asked whom to admire, Tse points less to celebrities than to processes that create shared facts. He was at the Bletchley Park summit in November 2023, where ministers from 28 countries, including the United States and China, signed a declaration and agreed to nominate experts for an independent report led by Yoshua Bengio, with contributions from about 100 people, aimed at becoming something like an IPCC for AI safety. Other forums are doing narrower work: the International Dialogue on AI Safety has recorded growing deception and scheming in frontier models, and the Global AI Bio Forum issued a statement on biosecurity at the meeting of AI and the life sciences, signed by leading people in both fields. Labenz's idea of a single "MMLU for morality" gets a polite yes to race-to-the-top leaderboards and a no to relying on one number. Tse wants a portfolio of ethics and risk benchmarks, and, more ambitiously, a global framework for frontier-risk management close to what leading US labs and the EU's code of practice already resemble. Aviation is his analogy: the same safety standards apply from Europe to the United States to Singapore to China, and accident rates fell by orders of magnitude.

For people without a ministry or a lab, Tse's list is plain. More people-to-people contact, past the headlines. Coming to Chinese conferences such as the World AI Conference each summer. Philanthropy, still as early as climate science was a few decades ago, which means early funders can still shape the field. Labenz, who spent a summer reviewing AI-safety charities, asks how anyone knows this work works. Tse offers two tests. One is whether a conversation puts an important, urgent issue on the table that nobody was discussing. Concordia ran a dialogue on crisis preparedness, including AI-enabled cyberattacks on critical infrastructure, and found senior policymakers surprised that the scenario could arrive in a few years. The other is whether you can produce consensus documents that shift the Overton window by showing agreement is possible. He does not feel like a minority inside China. Academics, industry, and think tanks, in his experience, treat international cooperation on AI governance as something to prioritize, and conference organizers in Beijing and Shanghai spend real effort hosting foreign guests. When Labenz floats a billion-dollar joint lab on a Pacific island, Tse likes the spirit and then redirects: the world needs more progress on the low-hanging fruit, because even the obvious international work is often not being done.

Tse wants one more historical marker on the record. Around 2023, after ChatGPT, frontier-risk talk took off in China as well. Prominent Chinese scientists signed statements treating extinction risk as a global priority. A Beijing conference featured Geoffrey Hinton, Andrew Yao, Sam Altman, and others for a full day of safety talks, one of the first times those risks were so visible at a technical meeting in China. By 2025, he says, more than 30 research groups were spending substantial time on frontier safety. The early papers were mostly RLHF and jailbreak defense. Now there is more work on scalable oversight of superhuman systems and mechanistic interpretability, closer to the Western advanced-safety agenda. About 95 percent of the literature, in his estimate, looks similar. The distinctive slice in the last half year is risk from embodied AI: aligning robots with human values, and jailbreaking foundation models that control machines. Physical AI, he says, is a theme of Chinese AI in 2025.

Industry documents are starting to match that emphasis. A Chinese AI industry association issued voluntary safety commitments signed by more than 20 major companies; the latest version flags risks from agentic systems and from embodied AI, a sign that firms expect their models to enter the physical world. During Chinese New Year a national television program showed Unitree robots in group dances, which Tse guesses hundreds of millions of people watched. There are now marathon competitions between AI robots and humans, and plans for a full-scale Olympics-style event between humanoids and humans in Beijing and other cities. Labenz asks whether the public will actually care, given that people still watch human chess decades after machines became unbeatable. Tse says many find the spectacles entertaining, and that they also function as a real-world benchmark that cuts through hype: the marathons show how much room humanoids still have to improve.

They also open a labor argument. Tse mentions a local protest in a central province against self-driving cars making too much progress and threatening drivers' jobs. Youth unemployment has risen. More people hold bachelor's degrees, white-collar posts in banks and government are harder to get, and some of the fallback work is food delivery or driving for Didi, China's Uber. Those jobs are exactly the ones robotaxis and humanoids can reach. There is, he says, anxiety among a significant share of the population, and it often surprises him how little governments address a concern that feels so salient. He does see a possible Chinese instrument that American debates about a new social contract mostly lack. China is a socialist country. If AI profits depend on chips, energy, data centers, and land, and much of that infrastructure sits with state-owned enterprises rather than private firms, one can imagine distributing AI's benefits through those state-owned assets. Labenz has been arguing that the United States will need something that decouples a decent life from a paycheck; Tse is pointing at the plumbing China already owns.