China Has 17 Million Chatbot Users and No OpenAI

This Cognitive Revolution episode is Nathan Labenz's attempt to get a qualitative read on Chinese AI that blogs had not given him. His guest publishes under the pen name L-squared: a tech worker who has spent about five years in Beijing, across Chinese and Western firms, and now watches cloud and infrastructure while tracking the local AI scene in Chinese. She grew up in London. Jordan Schneider of China Talk made the introduction, and she prefers anonymity out of caution about her location and job, not because the takes are spicy.

Both sides open with a disclaimer. L-squared says she knows less than she does not, and that she is offering one limited angle. Labenz quotes the proverb that a visitor who spends two weeks in China thinks he understands the country and a visitor who stays a year realizes he understands nothing; he plants himself firmly in the nothing camp. The conversation that follows still covers players, products, public uptake, regulation, open source, the chip ban, and how the two ecosystems rhyme.

Labenz's own bias, stated up front, is toward de-escalation. He would not want to live under China's speech rules, but he wants on-the-ground detail as a counterweight to the worst-case pictures Western imaginations can conjure. One of his early takeaways is that the U.S. and Chinese AI stacks are, in many ways, more similar than they are different.

L-squared splits China's large-language-model scene into two camps: the tech giants and the LLM startups. On the giant side she names Alibaba, the e-commerce group with the largest Chinese cloud share; Tencent, builder of WeChat; Baidu, the search company behind the Ernie chatbot; and ByteDance, known for TikTok and Douyin, whose bot product sat at the top of a generative-app monthly-active-user board earlier in 2024. Huawei matters most on infrastructure and cloud, and later in the chip story.

Alibaba has been the most active check-writer, investing in at least five LLM startups. Those deals are not just capital. Because Alibaba is the leading cloud vendor, L-squared says the startups it backs are likely steered onto Alibaba Cloud. The companies are spreading bets widely, she argues, "because there isn't a real standout front runner in China" the way two or three U.S. labs stand out.

The startup layer splits again. Moonshot AI and Zhipu AI came out of Tsinghua, China's strongest computer-science university, both founded by Tsinghua professors, both with a few hundred people and enough funding to ship serious models. A second cluster is led by known operators: Kai-Fu Lee's 01.AI, headquartered in Beijing, whose Yi models show up on open-source leaderboards, and Baichuan, started by the founder of a Chinese search company, which ships both open and closed models.

Labenz hears a familiar business dynamic: cash-rich platforms that own training and inference hardware, write startup checks, and compete with the same startups for users. The difference he keeps returning to is the missing inner circle. Chinese commentators already complain that too many foundation-model players are a waste given hardware constraints. L-squared thinks a narrowing, and more concentration, over the coming months and years is very possible.

Public awareness in China is not low. Mainstream news covers ChatGPT and Sora and the need for Chinese equivalents. During the year's biggest political meetings, leaders visited Baidu and the Beijing Academy of Artificial Intelligence, and those visits made television. An AI-generated cartoon series has aired on Chinese TV. State-owned enterprises are shopping for AIGC services. L-squared, who admits she is in a bubble of her own, still thinks the wider public treats this as important technology, not a niche.

What the market does not have is a default brand. There is no single ChatGPT that a Chinese office worker would name first. Many vendors are selling proof-of-concepts to businesses, and it is too early to call a winner. The consumer signal she does have is ByteDance's chatbot: about 17 million monthly active users at the top of a generative-app board earlier in the year, against estimates of about 27 million for ChatGPT in the United States.

China's population is roughly four times the United States'. L-squared puts the middle class at around 40 to 50 percent, or some 600 million people who live a first-world enough work and life to be in the addressable market for a language model. Both countries are still early on the adoption curve. Labenz's rough math is a two- or three-times per-capita gap, not a civilizational one.

Pricing has not converged the way U.S. consumer AI has around twenty dollars a month. Baidu's Ernie is the main product she knows that has been charging for months, around seven to eight U.S. dollars. Many Chinese assistants are still free, hunting for users. Investor chatter in her circle is that enterprise generative apps are where the money will be, and consumer apps will be hard to commercialize. Labenz suspects the U.S. ends up in a similar place: most people get the tools from work, the way they get Google or Microsoft productivity suites.

Preparing for the show, Labenz and L-squared spent time with Baidu's Ernie 4.0. SuperCLUE, a third-party Chinese-language benchmark updated every couple of months, still ranks GPT-4 first overall even on its China-first tests, with Ernie the best of the local products. Labenz's coding spot-check did not show ChatGPT's Python as head-and-shoulders above Ernie's. Comments were in Chinese, details differed, and his first-order summary was that the products seemed more similar than different.

L-squared's own use matches that. For Chinese-language tasks the gap with GPT-4 is not huge. The usability gap is elsewhere. Some apps are overly sensitive: a harmless, non-political question can trip a keyword filter and return nothing. Context windows are often short. In a censored environment, she says, those two frictions are the daily tax, not a model that simply cannot do the work. She has not hit capacity walls of the "come back later, we are overloaded" kind, so chip scarcity is not yet a retail experience.

Alibaba's Qwen 1.5 72B is the other data point that unsettles the "years ahead" line. On the LMSYS arena it sat in the top ten, neck and neck with original GPT-4, which would make it a global open-source leader if the custom license is as open as marketed. Ordinary Chinese users do not go to Hugging Face for Qwen. They use Alibaba's Tongyi Qianwen chatbot, a more advanced closed version with extra moderation. Developers get the weights; consumers get a different object.

LMSYS traffic is mostly English, Labenz assumes, which means Qwen's Chinese quality could be even stronger. He also flags a caveat: drafting on OpenAI outputs is a fast path, and ByteDance had an OpenAI account blocked for training on outputs against the terms of use. How widespread that practice is, L-squared does not know. What she does know is heavy use of the Llama architecture. A Beijing Academy of AI briefing to the Chinese premier listed reliance on foreign open-source architecture as one of three national vulnerabilities.

Labenz expected China to clamp down on open-source models first, because once weights leave the building, refusals fine-tune away. The volume of Chinese open-source releases confused him. L-squared's account is a government trying to ride two horses: use AI for productivity and catch-up with the United States, and stop models with large public user bases from saying politically unacceptable things. Open source is treated as a catch-up tool, so it is not being crushed. Tight rules attach to services aimed at the Chinese public.

Those rules require generative-AI providers to register and pass safety tests on training data and outputs before they may deploy. Training data is sampled, not inspected token by token, and failing batches are not supposed to be used. A 2024 security standard then gets more specific: providers should not train on sources blocked by Chinese law and policy. Hugging Face is blocked. Wikipedia is blocked. On a literal reading, two of the world's default training corpora are off limits.

L-squared does not expect that clause to be policed. Developers will keep using those sources and not tell the government, and the government will likely turn a blind eye rather than let Chinese services fall too far behind. Only about forty apps and services had been approved under the generative-AI rules at recording time, while many more are already in customers' hands. Firms look at the back-and-forth with regulators, decide they lack the bandwidth, and skip registration.

A common workaround is to call the product an internal test and require a whitelist. In practice, she says, anyone who applies gets in, so it is not much different from full deployment. App developers are technically supposed to run the same safety assessment even when they sit on an already-approved foundation model, which may only simplify the paperwork. App stores and websites should check for approval and often do not. The process is new, backlogs run to weeks or months, and enforcement has not been heavy-handed. "There's also a gap between the law and what is being enforced on the ground."

Labenz put 01.AI's Yi-34B, then near the top of LMSYS, in the arena and asked about Tiananmen Square. The model answered like a Western chatbot: the night of June 3 into June 4, 1989, soldiers and tanks in the square, casualty estimates from several hundred to several thousand. L-squared's theory is not that the company trained on a purified corpus. Hugging Face is not publicly reachable in China, so open-source releases do not get the keyword stack used on a domestic customer base. Once 01.AI offered a China-facing API, she would bet those answers disappear.

A China Talk roundup of several Chinese models found a wide range of behavior. Yi, tested only in its Hugging Face form because it had no consumer app in China yet, was the most Western; others looked more like a domestic product. App developers are expected to filter beside the core generator. Labenz compares that to early Bing retracting an answer mid-stream. Western apps mostly do not do this now; ChatGPT and Gemini generally show what they generate. In China, the extra layer is the default.

The same write-up found no obvious correlation between censorship behavior and capability. Labenz's gloss is that reinforcement learning from Chinese Communist Party feedback can work: models that conformed more did not obviously get dumber. That matches L-squared's day-to-day sense that Ernie still has value even when a sensitive keyword kills a turn. Controllability, though, is listed by BAAI as a national challenge. A chatbot that insults a political leader could mean a PR disaster or a government shutdown. The content risk Chinese developers carry is much higher than in the United States.

There is also no Western-style jailbreak ritual. Labenz describes launch-day loophole contests and viral screenshots; L-squared has not seen the Chinese equivalent. Users are more cautious about posting politically unacceptable chatbot output, and the models are not yet at GPT-4 or Claude 3, so a bad answer is less of a surprise. The firewall itself is porous: many people have VPNs, foreign numbers can be bought on Taobao to sign up for ChatGPT, and Microsoft Azure customers in China can reach the OpenAI suite. Labenz, going the other way, could open Chinese product sites from the U.S. and then died on the Chinese SMS verification step.

If Labenz had to compress the first hour, he would call the two ecosystems a warped mirror: same giant-plus-startup structure, similar public awareness, similar end-user quality, an onerous approval process that is often skipped, and a missing jailbreak culture as the loudest cultural difference. L-squared accepts the summary and complicates it with hardware. The reason China still ships capable models, she says, is two buffers. Companies stockpiled chips before export controls hit, so Chinese clouds can still rent good Nvidia silicon. And Chinese firms can still train on overseas clouds that host chips they can no longer import.

Those buffers expire. Stockpiles run down while the frontier keeps moving outside China. She expects the full effect of the controls to show after two or three years. People who already declare the chip ban a failure because Chinese AI still looks good, she says, need more patience. Washington has also noticed the cloud loophole. A draft rule out for comment would make U.S. cloud providers log large training runs by foreign customers and could restrict use by people from certain countries, which would probably mean China.

They recorded on the day Nvidia announced Blackwell, billed as a large jump in density and as much as an order-of-magnitude gain in energy and cost efficiency. If little of that generation can enter China, Labenz admits he has been too impatient in judging whether the ban "worked," even though he is not sold that slowing China is a win. Inside China the rhetoric is about indigenizing the semiconductor stack, and the government will throw resources at it. Huawei is the main hope: working with SMIC it surprised observers with a smartphone chip, and it planned to expand GPU production that year. Yields at the leading edge are still a brutal catch-up problem. You cannot, L-squared says, simply will success in this industry.

One unusual bright spot in the relationship is Microsoft. Azure reaches Chinese customers, including with OpenAI products, and Microsoft Research Asia has trained a generation of Chinese AI leaders. Labenz had not realized AWS also has a couple of data-center regions in China. What is unique about Microsoft is the research footprint: other American tech firms with a China business still do their research elsewhere. Leadership has long treated Chinese research talent as something to invest in, a history Jeff Ding's ChinAI newsletter has translated.

That position is getting harder. U.S. politicians have questioned the Beijing lab, and there were rumors of moving some scientists to Canada. L-squared puts a question mark over how long Microsoft can keep tapping Chinese talent while pleasing both governments. Academic science, meanwhile, has not pulled back. Labenz reads arXiv daily, sees Chinese names and Chinese institutions, and is surprised the papers still go out in English in American venues while U.S. frontier labs have stopped publishing methods.

Her explanation is incentives, not a romantic story about scientists rising above politics. The Chinese government wants its researchers on a global stage; climbing the ranks of international contribution looks like scientific power. The academic reward system therefore pays for volume in top English-language journals and conferences. Chinese journals exist, but top researchers skip them because the audience is smaller and the career return is worse. She resists stuffing that into a "face" frame. Individuals want impact; the state wants recognition; English is still the language of international science. Jordan Schneider's report from NeurIPS, she and Labenz note, was that a striking share of hallway conversation was in Mandarin even as the papers were in English.

Labenz has not been able to shake a May 2018 Washington Post op-ed by Tsinghua legal scholar Feng Xiang, titled "AI will spell the end of capitalism," written before GPT-2. The piece imagined AI giving China a path back to a more planned economy. He asks whether that is the domestic government line. L-squared says it is not. The government sees industrial optimization and productivity, and it also talks about inequality and disrupted employment. It is not selling a communist utopia, and the opportunity-and-risk list sounds a lot like other governments'.

The similarity unnerves Labenz. Western leaders, he complains, offer almost no concrete positive future, only that it will be amazing if we do not blow ourselves up. L-squared does not hear a clearer long-term vision in China either. The latest government work report highlighted an "AI plus" push to get real applications into more industries, which is not the same as thinking through what happens when the systems become far more powerful. There is, they agree, a global shortage of a positive picture.

Opinion polls still show more day-to-day technological optimism in China than in the U.S. or Europe. L-squared would add a structural caveat: the West has NGOs that mobilize around tech harms, and China does not have that kind of nonprofit landscape. Journalists sometimes publish critical pieces, but the open space for diverse voices is not the same. More optimism, in her take, is real and also shaped by a narrower discourse.

Self-driving is Labenz's test of whether the government will pull a technology into being. In the United States he sees authorities as a slow gate, not a sponsor. In China he would bet on faster deployment if the tech works, because the state could decide it makes commuting safer and then make it happen. L-squared says the official attitude is positive. Baidu and others are testing in suburbs of major cities, including parts of Beijing, and some taxi trips already run without a safety driver, limited to set routes and destinations.

Chinese consumers' willingness to try new gadgets helps. What may not help is the actual street. City centers, she says, can be chaotic: e-bikes and delivery drivers going the wrong way, rules treated as optional. That is a real obstacle in messy cores. In simpler environments she can imagine fast adoption once the stack is good enough. The more interesting question is whether the state would reshape the landscape to usher the tech in, the way high-speed rail gets a straight line on a map.

That is not science fiction from where she sits. The government is already investing in new towns and infrastructure for autonomous driving. Xiong'an, a new area not far from Beijing in a neighboring province, has lanes meant for driverless cars, and an expressway linking Beijing to that area is being designed with self-driving in mind. Reworking an old city is harder. When the Chinese government decides to build, she notes, it can still do impressive things.

Western hawks often say China cannot be trusted on AI safety, so the only move is to race. Labenz does not buy that and does not know the local debate. He thought Biden and Xi had pledged in San Francisco not to put AI in nuclear command and control; L-squared walks him back to the actual readout, which promised future dialogues on AI risk, and to later reporting that China had not yet committed to keep autonomy out of nuclear C2. The prospective headlines had outrun the agreement.

There have been better signals. In June of the previous year the Beijing Academy of AI held a full day on safety and alignment, with Sam Altman speaking online and other Western figures on the program. Chinese officials attended the U.K. AI Safety Summit, which produced language on catastrophic risk. What she does not see is follow-through: little alignment research, no Anthropic- or OpenAI-style responsible scaling policies, and much less work on corporate governance for worst-case harm. One reason, she suggests, is that China lacks a long history of labs whose explicit mission is AGI, so developers are still trying to catch the frontier rather than govern it.

Asked what the West can do besides fear an arms race, she keeps it modest. People with influence in AI and policy should keep communication channels open. Everyone else should read China coverage more skeptically: the foreign press corps is thin, English-language discussion often lacks nuance, and she is frustrated by fear-mongering that treats China as a threat to U.S. AI leadership that must be stopped at all costs. She points listeners to Jeff Ding and, on regulation, Matt Sheehan. Separate the open weights from the product. Watch the approval list, the cloud rule, the chip stockpiles, and Huawei's GPU line. Do not flatten the scene into "years ahead" or "about to overtake."