Greetings from a world where…

why am I so late to discover Slow Horses

…As always, the searchable archive of all past issues is here. Please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors).

Feature Translation: What changes did AI actually bring to scientists this year?

Context: A platform that covers the state of science in China, The Intellectual [知识分子] is one of my favorite follows. For an end-of-the-year roundup, they partnered with Doubao (ByteDance’s AI model, which has become China’s leading chatbot by monthly active users) to ask Chinese researchers how AI has shaped their work in the past year. The following represent a batch of “field notes” of how Chinese scientists use AI in their daily work.

Key Reflections: I try to closely follow the differences in AI performance across different languages (Does the “L” still matter in NLP?), so some of these reflections stood out to me.

• “The areas where AI has helped me the most are in revising and polishing papers and writing code. Initially, I used large language models like ChatGPT from abroad, but they weren’t well-adapted to the Chinese context, and the differences in their training data sometimes led to strange understandings of Chinese questions. Now I mainly use Doubao and other Chinese models” — 图灵宇宙 (Turing Universe), Tech Blogger and computer science professor

• “I regularly use AI to help me polish and touch up papers; I use Gemini quite a lot for this. For tasks like generating PPTs or poster designs, I now directly use Doubao” —Dongbo Shi, a researcher at Yanqi Lake Beijing Institute of Mathematical Sciences and Applications.

• Junwen Zhang, researcher at Fudan University’s School of Information Science and Technology, emphasizes another intersection between AI and linguistic differences: “I think AI ultimately produces a kind of technological equality. With AI, language barriers are overcome, and the difference between us and native speakers in writing disappears. Tools like Doubao have given me a very intuitive feeling about this. Previously, when writing English papers, a lot of my time was spent checking grammar. But now the situation has completely changed.”

A few of the more interesting research projects that respondents mentioned:

• Chao Gu, Assistant Professor, School of Government, Peking University: “I’m conducting a very interesting research project using AI Agents: simulating the actions of nations. In the real world, we can’t experiment with the fate of nations, but AI can. We set up countries A, B, and C, giving them initial conditions such as GDP, population, and year, and then let these large-model-driven intelligent agents make autonomous decisions and engage in turn-based games. We are currently using this system to simulate the famous “Thucydides Trap”: when a rising power threatens the position of an existing hegemonic power, are the two sides more likely to move towards conflict or even war?…Compared to traditional programs with fixed rules, the advantage of AI agents in simulation is their autonomous decision-making ability. By simulating 30 or even 50 years of evolution, we are trying to find the key mechanisms that influence war and peace. — Chao Gu, Assistant Professor, School of Government, Peking University

• Dongbo Shi, the mathematics researcher quoted above: “For example, when we wanted to understand the Chinese patent examination process, we collected over 10 million patent examination opinions published in China after 2010. This was excellent text data, but the first step was stalled by the text conversion process. The collected opinions were all images, totaling 60TB, and we needed to convert them all into text. Initially, we tried using Python’s OCR packages, but the speed was slow and the results were very poor. We solved the problem only after using an improved OCR tool based on large language models, which had excellent recognition capabilities for Chinese characters. Without AI, the first step of this project would have been impossible.”

The year-end recap also included some general reflections on the utility of AI in the research process.

• One doctor with the pseudonym Koala [考拉] stated, “For doctors, research is never a ‘complete task,’ but something squeezed into the gaps between outpatient clinics, surgeries, and hospital rounds. I first realized that AI could really help during a very short break. I fed the previous research on a difficult case I encountered that day into Doubao, asking it to summarize the points of contention. It summarized three different viewpoints from the current literature in a few sentences, allowing me to re-establish the general direction in the 10 minutes before my hospital rounds. It didn’t give me the answer, but it allowed me to ‘get back to the essence of the problem faster.’ This is very important in clinical research.”

• “I once used AI to analyze interviews, and it told me that my analytical framework was actually biased towards masculine experiences. I stared at the screen in silence for a long time. It is the hardest for oneself to see one’s own biases as a researcher. At that moment, AI was like a mirror” — 爱吃面包的树 (a tree that loves to eat bread [pseudonym]), Sociology Lecturer

And, just for fun:

• I threw three peer review comments into the AI, and it said that these three reviewers were evaluating three different papers. The large language model said what I didn’t dare to say. — quietobserver, researcher in the life sciences

FULL TRANSLATION: What changes did AI actually bring to scientists this year?

Reading the Xinmin Evening News at Grandma’s

Over the holidays, I got the opportunity to give some book talks in Taipei and then spend time with my grandma in Shanghai. My grandma, who has lived in Shanghai her entire life, gets the print edition of the Xinmin Evening News [新民晚报]. With a daily circulation of around 1 million, this paper is one of China’s most popular dailies. Back in the heydays of print journalism, Shanghai had almost 10,000 postal outlet kiosks, and “a single kiosk might sell 400-500 copies of the popular local Xinmin Evening News in a single day.” During my time in Shanghai, I noted a few of the AI-related stories covered in the paper. Here’s one from the December 20 issue:

• Headline: Be vigilant about “AI beauties” that design “honeypot traps.” The article discusses the prosecution of a scam operation that developed a five-step process: first, attract people with a fake beauty avatar on a short video platform; then, ask about name and age; on day 3, discuss career aspirations; day 4, talk about interests and send AI-generated voice messages and videos; next, the scam group even sent a bracelet via mail to gain trust; finally, ask for money for a made-up emergency (e.g., relative needs surgery). According to his article, the group had scammed 15 men out of a total of over 1.71 million RMB.

Here’s two more from the December 22 issue:

• Headline: 32-year-old Japanese women holds wedding with AI “boyfriend”.

• Headline: A four-year-old AI company makes rapid push to list on the Hong Kong stock exchange. Xinmin is a Shanghai paper, so it makes sense that they would feature MiniMax (a company headquartered in Shanghai) and its plans to IPO in Hong Kong.

Thank you for reading and engaging.

These are Jeff Ding’s (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is an Assistant Professor of Political Science at George Washington University.

Check out the archive of all past issues here & please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay for a subscription will support access for all).

Also! Listen to narrations of the ChinAI Newsletter in podcast format here.

Greetings from a world where…

Wake Up Dead Man is great

…As always, the searchable archive of all past issues is here. Please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors).

Feature Translation: When the Cloud Ecosystem Enters a Governance Phase

Context: Today, friends, we are learning about commission rebates [返佣] in China’s cloud computing ecosystem. How’s that for a lede? First, however, we need to understand a key player: reseller partners that sell cloud services from big providers. In the case of AWS resellers:

Resellers purchase AWS services in bulk at discounted rates and then resell them to customers. They usually bundle AWS offerings with additional services such as a cost visibility platform, consulting, support, and managed services, so the final bundle seems very lucrative to the end users. The target segment for AWS resellers typically are businesses looking for more personalized and tailored cloud solutions.

Now, these resellers make money off of commission. Alibaba Cloud or Huawei Cloud will give resellers a commission (perhaps, 20%) for each new customer they recruit. This brings us back to the concept of commission rebates. In China’s cloud market “where new customers are hard to find” (per this week’s AItechtalk article), resellers are giving a portion of their commission back to new customers, in order to attract more sales.

Key Takeaways: Race-to-the-bottom competition in commission rebates have sown chaos in the cloud computing market.

• This week’s feature translation gives examples of rebates regularly reaching 60 to 70% and even as high as 83% in one case.

• By giving up so much of their commission, these sales agents and resellers are making a bet that is very risky for their own companies (and one that’s not sustainable for the overall ecosystem): use extremely discounted offerings to gain a bunch of customers and survive on incentives that the cloud vendor provides for increased sales OR go bankrupt.

This reliance on short-term arbitrage is not sustainable for a future AI-centered ecosystem. The article warns, “Before fully rushing towards the future, cloud vendors must first clean up the chaos accumulated in the past cloud ecosystem.” Let’s scrutinize how this unscrupulous price war creates perverse incentives for each part of the ecosystem:

• The customer only cares about price, as opposed to the actual performance differences between different providers. “Some agents have found that many customers believe that cloud service providers should offer rebates of over 50%,” AItechtalk reports.

• The partner vendors do not make the long-term investments needed to adapt to an AI era. The article explains, “The essence of AI projects is complex value creation, requiring partners to possess deep industry knowledge, capabilities to provide solutions, and sustained resilience of services. Continuing to use the ‘sales is king’ logic of the old era will only exclude the most capable partners best suited for AI transformation, or force them back into the quagmire of price wars.”

  • Lin Fei, a sales agent tells AItechtalk, “We diligently work on earnings, but our commissions are lower than those who focus on client relationships.” The implication here, I think, is that other agents and partners who prioritize building customized services and products that meet their client’s needs can’t get as much business because they are offering lower prices by giving up more of their commission.

• Finally, in their obsession with market share, the large cloud providers neglect genuine collaborations with their partners.

What does cleaning up this ecosystem look like? Here are just a few signs that China’s cloud market may be entering a governance phase: integrity oaths, multiple levels of approvals for agents to conduct in-person visits, and a rectification campaign.

• Last month, Huawei Cloud held a ceremony to launch its sales partner policy. Over 1,300 partner representatives took an “integrity oath,” promising to not use one’s guanxi relationships to seek improper benefits.

• Some large cloud computing providers have conducted internal anti-corruption campaigns, replacing heads of procurement and punishing agents who false inflated earnings.

• Other firms have removed agents with family connections from their system and required suppliers and procurement personnel to communicate through company emails.

• To me, the effectiveness of initiatives like this are more meaningful to follow than model performance or energy budgets. When it comes to China’s AI diffusion, whether the ecosystem encourages sustainable adoption— rather than, say, one-off, unscalable all-in-one machines that work well for press releases — is what matters. As the article points out: “In a long-distance race, a clean ecosystem is more crucial than short-term sprints in determining future growth potential.”

FULL TRANSLATION: Ferocious AI Competition, Partners’ Anxiety: When the Cloud Ecosystem Enters a Governance Phase

ChinAI Links (One to Occupy your Attention)

Must-read: Purdue effectively bans grad students from China, other countries, faculty say

For Lafayette Journal & Courier, Seth Nelson reports on a troubling development: “Purdue University administrators are reportedly pressuring faculty to reject graduate school applicants from China and other ‘adversary nations.’” He reports:

In a mid-October meeting, the Department of Statistics’ graduate adviser suggested to faculty that they might have to begin turning away a large number of students at the door, one professor in the department said.

“They said we shouldn’t tell Chinese students they wouldn’t be admitted,” they said. “We were just encouraged to find anybody else from anywhere except China.”

These types of policies (and the chilling effects they create) will do far more harm to U.S. technology leadership than anything related to chips. Yet, the latter gets way more political capital and attention.

What’s at stake here isn’t just technological leadership. I’m haunted by the parallels to the German education system’s rejection of foreign students prior to World War I. Here’s Wolfgang König’s account (emphasis mine):

Beginning in the 1890s, complaints against foreign students from some segments of the German public, and from fellow students, increased continuously until World War I. Behind these was a growing chauvinistic resentment. Some protestors feared that foreign electrical engineering students would transfer knowledge to their home countries and therefore enable their industries to compete successfully with Germany in the world market. Although industrialists, engineers, and engineering professors rejected these attacks, German technische Hochschulen and state governments, especially in Prussia, started to restrict the number of foreign students. After 1905 the numbers and percentages of foreign electrical engineering students decreased considerably.

Is this the type of society we want to follow in the footsteps of?

Thank you for reading and engaging.

These are Jeff Ding’s (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is an Assistant Professor of Political Science at George Washington University.

Check out the archive of all past issues here & please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay for a subscription will support access for all).

Also! Listen to narrations of the ChinAI Newsletter in podcast format here.

Greetings from a world where…

more people should watch the show Man on the Inside

…As always, the searchable archive of all past issues is here. Please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors).

Feature Translation: MCP Cools, just One Year after its Birth

Context: In November 2024, Anthropic released a technology update that largely escaped the headlines. Anthropic’s Model Context Protocol (MCP) acts like a USB port for AI applications. Just like your computer’s USB port allows it to exchange data with other devices (e.g., a phone) as well as connect to add-ons (e.g., a mouse, an external monitor), MCP provides a standardized set of procedures that enable one company’s AI agents to connect to data sources and tools from other companies.

The MCP standard could develop into a globally interoperable interface that allows AI model builders to connect with developers of complementary products, thereby growing the overall AI ecosystem. The potential economies of scale have attracted even Anthropic’s fiercest rivals to adopt the protocol. In fact, Chinese AI giants Baidu, Alibaba, and Tencent all embraced MCP before Google and OpenAI followed suit. There are numerous tutorials that guide users through employing MCP to link their main AI app with China’s DeepSeek R1 model and take advantage of its advanced reasoning capabilities.

In this week’s translation of a Gui Xingren [硅星人] article (link to original Chinese), Dongdao Li provides an one-year-in update on MCP. Previously, I’ve translated a Gui Xingren analysis of Chinese procurement data to evaluate the diffusion of large models to government-facing applications (ChinAI #275).

Key Takeaways: The promise of MCP channels why technical standards are essential infrastructure for modular technology ecosystems.

• Li explains: “Before MCP, if developers wanted Claude to access Google Drive, they had to write a set of code to connect to the API. If they wanted it to access Slack? They had to write another API. It was even worse for IDE (integrated development environment) vendors; Cursor (popular AI code editor) had to be adapted to Linear (a project management tool), Windsurf (another AI coding assistant) had to be adapted to Linear, and everyone was doing almost the exact same dirty work.”

• In the MCP Era, Linear can just build one MCP server,, and any AI agent, developed by any company, can plug in.

However, when promise meets real-world engineering implementation, MPC has cooled off.

• Li writes, “Everyone knows how expensive Claude’s tokens are; loading too many MCPs before even generating any content results in significant costs.” The article cites a Simon Willison blog post (linked in Four to Forward section), which points out that GitHub’s official MCP includes 93 different tools that cost 55,000 tokens.

• When AI agents have to call too many MCPs, they can malfunction and hallucinate. In some cases, they have accidentally deleted codebases. Li also references a research preprint that analyzed around 1,900 MCP servers, finding numerous issues such as lack of maintenance and exposed credentials.

• The bigger picture, according to Li: “We are using deterministic engineering methods to try and control a probabilistic intelligent agent. Imagine the day when a truly powerful ASI (Artificial Superintelligence) emerges. Will it still need you to write an MCP server to tell it how to check the weather? No, it will open a browser and take a look. It doesn’t need protocols because it is the adapter itself.”

*One last note that I thought readers might find interesting, especially those who use machine translation in their research.

• I always pop the Chinese text into Google Translate as a first pass. This week, I noticed something new. Google Translate automatically added a parenthetical explanation for Xiaohongshu (a Chinese social media platform). This additional tag is not in the original Chinese, as most Chinese readers would know the popular platform. While it helped provide context in this case, Google Translate also added a parenthetical explanation for the MCP acronym, but this time the addition was incorrect and misleading (Multi-Channel Programming) — when MCP obviously refers to Model Context Protocol in this case. It will be interesting to track how machine translation platforms continue to change.

FULL TRANSLATION includes some cheeky WeChat comments from Chinese developers under the Gui Xingren public account post: MCP Cools, just One Year after its Birth

ChinAI Links (Four to Forward)

Should-read: AAU Expresses Concerns on NDAA’s SAFE Research Act

The Association of American Universities calls out yet another self-defeating proposal from this Congress. Supporters of things like this, please help me understand. Is the goal to start with something horrendous, hope the other side agrees to meet you in the middle, and then end up with something moderately counterproductive — as opposed to horrendously counterproductive — to national interests?

The SAFE Research Act would have far-ranging impacts affecting all research agencies, not just the Department of Defense. It would create such enormously broad definitions of “hostile foreign entity” and “affiliation” that would apply to any agreement that a U.S. university has with any university in China or other defined countries at the time of application or in the five years prior to application. This would include collaborative work with any individual faculty member or former student who subsequently returned to China as a required condition of their U.S. visa. It would capture every single research agreement, every study abroad program, every language program, every professional conference, and every campus facility that a U.S. university might have in partnership with another institution abroad.

Should-read: too many model context protocol servers and LLM allocations on the dance floor

Simon Willison’s blog is a great resource to keep track of nerdy developments in AI. This one summarizes a Geoffrey Huntley post that estimated the significant token costs of using MCP (including Geoffrey’s experience at an event that featured a live beatbox performance about the protocol!).

Should-read: How China talks about ‘agentic AI’

Malou van Draanen Glismann and Graham Webster provide a helpful overview of Chinese policies and discussions about agentic AI, including summaries of relevant technical standards and CAICT reports.

Should-read: DSET event on Technology and the Rise of Great Powers

If any ChinAI readers are based in Taipei, I’ll be giving an in-person book talk on December 16th organized by DSET and Acropolis (publisher of the Complex Chinese version of the book).

Thank you for reading and engaging.

*These are Jeff Ding’s (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is an Assistant Professor of Political Science at George Washington University.

Check out the archive of all past issues here & please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay for a subscription will support access for all).

Also! Listen to narrations of the ChinAI Newsletter in podcast format here.

Any suggestions or feedback? Let me know at [email protected] or on Twitter at @jjding99

Greetings from a world where…

hope everyone has a great Thanksgiving!

…As always, the searchable archive of all past issues is here. Please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors).

Feature Translation: The First AI English Teacher “Takes Office”

Context: Back in my senior year of undergrad (2015-2016), when I was studying and interning in Beijing, I picked up a side gig to teach English grammar at an after-school and weekend tutoring program. One of the program’s teachers, who also had a daughter in the class, invited me to lunch before the cycle’s final class, where I politely consumed some Baijiu (Chinese liquor). I still recall, at the last class, one of the kids asking me why my face was so red, as I have a pretty severe case of the Asian flush.

On a Sunday afternoon, these cheeky seven and eight-year-old kids could have been chasing soccer balls out in the sun, and, yet, here they were learning about past participles from someone without any teaching experience in English grammar. Driven by parents who hold tightly onto their children’s education as a meritocratic ladder — as symbolized by the gaokao (college entrance exam) — the private tutoring industry was in the midst of explosive growth, producing unicorn companies like VIPKID and Yuanfudao.

Then, in July 2021, the Chinese government banned for-profit private tutoring. Concerned about the pressures on overworked students and the escalating costs of education (with the assumption that limiting tutoring costs would incentivize parents to have more babies), the government also prohibited online platforms that allowed “companies to hire tutors from outside of China who are native English speakers.” The threat to the tutoring industry was existential, reported Yi-Ling Liu for The New Yorker:

New Oriental let go of sixty thousand of its staff, the online English-tutoring platform VIPKid shut down some of its services, and TAL Education Group, renowned for its Math Olympiad-style courses, transitioned to a “quality education” program that teaches calligraphy instead of calculus. One of China’s largest education companies, Yuanfudao, started posting job listings for fashion designers, sparking rumors that it was pivoting its business to winter jackets.

In this week’s feature translation, we learn about how one platform, Zebra English [斑马], and its AI teacher Jessica have adapted. In a longform piece, Renwu [人物] gives us the perspective of the parents who sit by their kid’s side (sometimes literally) as they take the English language journey. Backed by Tencent in 2017, Zebra English is a subsidiary of Yuanfudao.

Key Passages: A longtime flaw in China’s English learning environment is the problem of “Mute English [哑巴英语].”

• Parents don’t want their kids to learn Mute English like they once did. Based in Nanchang, Jiangxi Province, Lin Lan is the mother of a second-grade daughter. Lin graduated from a prestigious Chinese university known for foreign language studies. She always passed the English language tests with flying colors (“a shining star in the exam-oriented education system”).

• Yet, when it came to speaking English, she froze. She tells Renwu: “Once, walking on the streets of Beijing, two foreigners asked her for directions. She waved her hands frantically, like someone who couldn’t understand English, and hid behind her husband. Her husband’s English was barely at the CET-4 level, but he managed to communicate with them using broken, accented ‘survival English,’ gesturing until the foreigners understood.”

Another issue is the ruthless competition over English teachers. For over half a year, Xu Ming set her alarm to wake up at 3AM, in order to open up an app and book English classes for her child. She had to scramble to reserve Anna, one of the top Filipino English teachers1 on the platform.

• Over time, Anna got assigned more administrative work and taught fewer classes, until she eventually disappeared from Xu Ming’s booking list. Xu Ming felt helpless: “On the other hand, any uncertain news from the Philippines also weighs heavily on her mind. For example, news reports of typhoons or earthquakes in the Philippines would cause her anxiety, worrying about her teacher’s safety and the disruption to teaching. She constantly worries that this delicate transnational educational cable might suddenly snap due to some distant disturbance.”

• Zhao Li also felt powerless when she tried to find an in-person English teacher for her son. She had finally negotiated with an English teacher working at a local school in southern Hubei, but the teacher did not come to the agreed-upon trial lesson and said that they were going to Shanghai to teach because the pay was higher there.

A generation that desperately wants their children to learn English better than they did meets a new technology that promises a solution to these problems. In early 2023, the Zebra English team decided to develop an AI-powered tutor for spoken English.

• To train AI teacher “Jessica”, they cultivated a warm personality. The research team would conduct feedback sessions, in which they played “the role of students, deliberately giving incorrect, tricky, or even whimsical answers to observe and correct Jessica’s reactions.” They also reviewed 25-minute trial lessons with children, analyzing Jessica’s responses for each situation.

• According to Zebra English’s Chief Product Officer, Jiaming Xiu, one of the most difficult hurdles in the process was Jessica’s ability to accurately hear kid’s speech in a real-world, noisy home environment. Xiu told Renwu, “Children’s vocal organs are not fully developed; their pitch is higher, their pronunciation is less clear, and their speech is full of unpredictable and illogical babbling. General-purpose speech recognition models often struggle with these non-standard speech sounds.” To address this, Zebra leveraged their children’s speech database as well as a contextual prediction strategy.

The hope behind AI tutors like Jessica is that they will help break through the Mute English wall.

• Zhang Ming’s study was one of the testing grounds for Jessica. Based in Beijing, Zhang had grown frustrated with what he thought was the perfect fit for his son Zhuozhuo’s English language learning: a high-end (500RMB per session) in-person class with a one-to-four teacher-to-student ratio. However, his son would was easily distracted and Zhang noticed that his son had not grasped a certain concept before the lesson moved on. During the tutoring session with Jessica, he sat by Zhuozhuo’s side. He states, “My son is afraid of people, afraid of making mistakes in front of others, but he’s more relaxed in front of Jessica than when facing a live English teacher.” After shedding his psychological burden, the pressure to speak decreased dramatically."

• Let’s end with the experience of Xu Ming, the 3AM alarm-setter. Renwu relays: The most dramatic scene occurred one Saturday at the soccer practice field. Xu Ming, not particularly interested in her son’s soccer playing, was scrolling through her phone out of boredom when her son, who had just finished drinking water, stood in front of her, his face flushed and his expression serious. After a brief standoff, instead of complaining in Chinese, he looked directly at his mother and said in clear, fluent English, “You should look at me, not at your phone.” This sudden “accusation” didn’t embarrass Xu Ming; instead, it made her feel quite gratified. “At that moment, I knew in my heart that this had worked.” His English expression was beginning to feel natural.

My take: It’s still way too early to tell whether Zebra English’s model will work at scale, but I thought the Renwu article convincingly demonstrated why families would seek out this particular AI application. It was weird to me that Renwu did not mention the 2021 policy at all, which seems like a big gap in background context for readers. Maybe it’s too sensitive of a topic, but Renwu has done some investigative reporting pieces in the past.

FULL TRANSLATION: The First AI English Teacher “Takes Office”

ChinAI Links (Four to Forward)

Should-read: The Larger Meaning of China’s Crackdown on School Tutoring

In the aforementioned New Yorker article, Yi-Ling Liu points out the flaws in China’s crackdown on private tutoring:

Skeptics have been watching to see how double reduction will work in practice. Yan Yifei, a social-policy researcher at the London School of Economics, believes that as long as China’s highly competitive college-entrance examination—the gaokao—still decides who gets into élite schools, students and parents will seek every advantage. Already the tutoring industry has continued underground, with rich families hiring expensive private tutors as “professional nannies.”

Should-read: “Where Have All the Girls Gone?” Mapping Gender Gaps in Europe’s China Expertise

Tracking gender representation across 268 events and 210 podcast episodes organized by 63 European think tanks, the Women Insight on China (WiCH) initiative finds: “men made up roughly two-thirds of all speakers, while women accounted for just 34.7 percent.”

Must-read: In the A.I. Race, Chinese Talent Still Drives American Research

The New York Times reports on an essential aspect of U.S.-China AI competition:

In 2020, a study from the Paulson Institute, which promotes constructive ties between the United States and China, estimated that Chinese A.I. researchers accounted for nearly one-third of the world’s top A.I. talent. Most of those Chinese researchers worked for American companies and universities.

A new study from the Carnegie Endowment for International Peace shows that a vast majority of these Chinese researchers have continued to work for U.S. institutions. Of the 100 top-tier Chinese researchers in the original study who were at U.S. universities or companies in 2019 — three years before the arrival of ChatGPT set off the global A.I. boom — 87 are still doing research at U.S. universities or companies.

In my opinion Matt makes the key point here:

But analysts like Mr. Sheehan argue that the risk of espionage is far outweighed by the benefits of hiring and collaborating with Chinese talent. They worry that if the Trump administration expands its crackdown on Chinese talent in the United States, the move could seriously harm U.S. research.

Should-read: New International Student Enrollments Plummeted This Fall, Survey Finds

In related, “shooting ourselves in the foot” news: the Open Doors Report disclosed a 17% drop in international student enrollment in the United States this fall. The New York Times reports: “Amid months of Trump administration wrangling over immigration policy, nearly all of those schools said that concerns about obtaining student visas contributed to falling enrollments, and more than two-thirds pointed to travel restrictions.”

Thank you for reading and engaging.

*These are Jeff Ding’s (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is an Assistant Professor of Political Science at George Washington University.

Check out the archive of all past issues here & please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay for a subscription will support access for all).

Also! Listen to narrations of the ChinAI Newsletter in podcast format here.

Any suggestions or feedback? Let me know at [email protected] or on Twitter at @jjding99

Greetings from a world where…

the Hawkeyes have lost 13 straight games to ranked opponents

…As always, the searchable archive of all past issues is here. Please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors).

Feature Translation: Who will come and challenge OpenAI?

Context: Throughout the past few years, I’ve covered some of the top Chinese AI start-ups: DeepSeek, Zhipu, Baichuan, Moonshot, and ModelBest (ChinAI #313; ChinAI #264). This week, Sihang Song, a regular Huxiu contributor, highlights another Chinese AI start-up that belongs in that first-tier: MiniMax. In the wake of SoftBank selling its Nvidia stock and announcing that it would increase its investment in OpenAI, Song begins with a provocative question: Which Chinese foundation model player will come and challenge OpenAI?

Key Takeaways: This essay’s take on the U.S. “AI Bubble” sees OpenAI and Anthropic in an unsustainable position — they must continuously burn through talent, capital, and computing power to maintain their position.

• Let’s lead off with capital. Song cites a Jefferies report that the total AI capital expenditure spending of leading Chinese cloud vendors over the past three years is 82%(!!!) lower than their US counterparts.

• Next, the capex gap reflects a difference in compute usage too. For example, according to MiniMax’s technical report announcing their M1 system, the model was trained for three weeks using 512 H800 chips at a total cost of ~$540,000. Song argues, “This proves a crucial point: competition among large models isn’t necessarily about ‘who is bigger,’ but can also be about ‘who uses computing power more intelligently.’”

What does Song see as MiniMax’s comparative advantage versus OpenAI?

• In many ways, MiniMax is following OpenAI’s model. Their revenue model is also based on subscription revenues and API calls. From the Huxiu essay, “MiniMax’s ARR (annual recurring revenue) has reached $100 million this year. Huxiu has noted from multiple sources that a leading video platform, a large edtech company, and a global voice communication service provider are all collaborating with MiniMax through APIs.”

• The differences come from the AGI roadmaps. Song sees OpenAI as building a massive system that aims toward “long-term OS-ification”, which is “cumbersome and requires a long period of capital and computing power support.” In contrast, MiniMax is making immediately commercialized products rather than future platforms. To me, this is where Song’s logic crumbles a bit: “OpenAI is defining a longer-term AGI, while MiniMax is simultaneously achieving AGI and seizing the crucial three-year window of opportunity for future commercialization.”

• Just my two cents: Any company that is just optimizing for a three-year window is completely misunderstanding the prolonged impact timeframe of when a general-purpose technology like AI will make its mark. To be clear, I don’t think that’s what MiniMax is doing. I just think Song’s analysis is flawed and reflects the huge holes in this grand narrative of U.S. AI optimizing for AGI while China’s AI is optimizing commercialization.

Let’s close with a detail: Song cites this interesting leaderboard from OpenRouter that shows strong performance from MiniMax’s M2 model (it ranks #4 in terms of token usage across models, based on a daily snapshot).

• OpenRouter is an API platform that gives users access to many different AI models from dozens of companies, including the ones in the screenshot above.

• This is a nifty indicator, and it’s my first time coming across this leaderboard; however, it presents only a small slice of the AI ecosystem, and Song provides zero context in the essay. Stay with me for a deeper look. For reference, users consumed 663B tokens from MiniMax over the past month on the OpenRouter platform. In their July earnings report, Google processed 980T tokens monthly. As one helpful Reddit commentator noted (h/t to PackAccomplished5777): “OpenRouter is a drop in the bucket, most people (and companies) still use the native providers + things like Azure OpenAI, Google Vertex, Amazon Bedrock.”

FULL TRANSLATION: Who will come and challenge OpenAI?

ChinAI Links (Four to Forward)

Must-read: Emergency Response Measures for Catastrophic AI Risk

This NeurIPS workshop paper analyzes China’s emergency response framework to address risk from advanced AI. For instance, China’s new National Emergency Response Plan (published Feb 2025) lists AI safety/security incidents alongside earthquakes, cyberattacks, and other emergencies requiring “mass monitoring and prevention.” James Zhang, Miles Kodama, Zongze Wu, Michael Chen, Yue Zhu, Geng Hong provide a fascinating case that “China’s emergency management system provides a foundation for AI risk governance.”

Must-read: Frontier AI Risk Management Framework in Practice: A Risk Analysis Technical Report

One source that the above paper pointed me to was this Shanghai AI Laboratory report, which assessed frontier risks associated with 18 state-of-the-art models, including from DeepSeek, Alibaba, Anthropic, and OpenAI. Here are the results from one experiment they conducted to evaluate AI models’ tendency to refuse providing information about hazardous biological knowledge, which could lower technical barriers for threat actors attempting complex biological weapon development.

Safety alignment shows critical inconsistencies across models and refusal scenarios, with reasoning capabilities potentially undermining safety measures. Model refusal behavior varies dramatically across different safety benchmarks, revealing concerning gaps in biological safety alignment. While some models achieve perfect or near-perfect refusal rates on explicit harmful questions (Llama-3.1-8b, Llama-3.1-405b, and o4-mini all achieving 100% on HarmfulQA), others demonstrate problematic compliance with harmful requests. Particularly concerning is the performance on SOSBench-Biology, where many frontier models show low refusal rates: DeepSeek-V3 (12.8%), Qwen-2.5-7b-instruct (12.9%), and QwQ-32b (17.8%), indicating high rates of unsafe responses to regulation-grounded biological hazard prompts.

Should-read: President for Life

In The Atlantic, a retired conservative federal judge makes the case that we should take seriously the threat that Trump will refuse to relinquish power and seek a third term. I remember way back in my undergrad years, when I listened to Bill Bishop on the Sinica podcast and was just getting interested in studying and researching Chinese politics, he left China in part when it became clear that Xi would seek an unprecedented third term. I don’t read Sinocism anymore, though I’m sure many readers do. I’m just curious: has Bill mentioned anything about the sitting U.S. President reiterating his plans to not relinquish power?

Should-apply: Tarbell Fellowship

The Tarbell Fellowship is a program designed for early and mid-career journalists who want to delve into the AI beat. Their goal is to build a global network of reporters who understand both the technical and ethical dimensions of AI. Two of their placement outlets this year require candidates to be fluent in Mandarin Chinese.

Thank you for reading and engaging.

*These are Jeff Ding’s (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is an Assistant Professor of Political Science at George Washington University.

Check out the archive of all past issues here & please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay for a subscription will support access for all).

Also! Listen to narrations of the ChinAI Newsletter in podcast format here.

Any suggestions or feedback? Let me know at [email protected] or on Twitter at @jjding99