this could be the Hawkeyes year

…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: China’s first AI-generated longform TV series

Context: Published in the 16th century, Journey to the West is a classic Chinese novel that tells the tale of Sun Wukong, a mythological monkey that helps a Buddhist monk journey to India. Among the many published sequels to Journey to the West, one of the three main ones, written by an unknown author, is The Later Journey to the West [后西游记]. In this sequel, the story shifts to Sun Xiaosheng, a young stone monkey who grows up on the same mountain long after Sun Wukong departed on his journey.

On August 31, Mango Excellent Media’s TV adaptation of this novel premiered during prime time. “It’s the first-ever fully AI-generated entertainment to hit mainstream provincial satellite prime time,” writes Mandy Wong. After seeing Bytedance Seedance 2.0’s (AI text-to-video model) release earlier this year, the team behind Journey to the West felt there was an opportunity to make a long-form series and break through a market “saturated with uninspired short-form dramas.” In this week’s feature translation (link to Huxiu article), Shuhan Liu reports on the makings of this AI-generated TV series.

Key Takeaways: From initial planning in March of this year to the premiere on August 31, the project took just six months (traditional TV dramas take 1-2 years). ByteDance’s Seeedance 2.0 and 2.5 models generated all the visuals, without live-action actors and filmed footage. How did this work?

• The director, Li Dongshen, and a ~100-person production crew developed a workflow based on “co-creation” with AI. This crew included 15 artists tasked with developing the style and concept art for key characters.

• They dealt with many pitfalls. At one point, the first two episodes needed to be completely remade, but Li claimed that generative AI made the costs more manageable. A single episode’s budget was ~900,000 RMB, with compute accounting for about 1/4 of the costs. They did have to allocate compute budgets for specific scenes and limit the number of generation attempts.

• One issue stood out to me: for scenes of the adult monkey version of Sun Xiaosheng, its expressions and movements were much more wooden compared to its young stone monkey phase. Ultimately, they discovered that the character design sheet for the adult monkey had omitted a portion of the tail. That missing segment “caused the AI ​​agent, Seedance, to form a completely different fundamental understanding of the character.” Liu continues, “One version was perceived as a human playing a monkey, wearing prosthetics that immobilized many facial muscles, while the version with the tail was perceived as an actual monkey, resulting in entirely different logic regarding behavior and expressions.”

• The media company Mango also invested in vertical models that addressed TV production needs, including lip-syncing and image quality optimization.

What does The Later Journey to the West mean for the future of China’s TV and film production ecosystem?

• One reference point is the proliferation of AI-generated shortform content. For context, see Wency Chen’s SCMP article: “According to the China Netcasting Services Association, 367,000 microdramas were released across Chinese platforms in the first half of this year. More than 74 per cent were AI-generated – an average of more than 1,500 new AI dramas a day.”

• Yet, Liu reports on technical hurdles that still hinder longform productions like The Later Journey to the West: “Subtle facial expressions lack nuance, complex movements are prone to errors, and there is still room to improve consistency across long-duration sequences.”

• In terms of commercial viability, the director and the Mango team frame this project as a demonstration of feasibility, as opposed to profitability.

If the WeChat comments on the article are any indicator, the audience did not like the “AI feel.” Here are a few of the top comments:

• “Where is the Eastern aesthetic? It’s all Western-style 3D modeling.” (Note: the Huxiu article claimed that some viewers were captivated by the AI-generated Eastern aesthetic)

• “If you guys love to watch, then you watch. As for me, I won’t watch.”

• “Not as good as Niu Lai (a Chinese firm that went viral for its low-quality animation), the more it’s AI-enhanced, the more it loses ‘human touch.’”

FULL TRANSLATION: How an AI-generated longform TV series costing 900,000 RMB per episode drove a 13.3-billion RMB surge in Mango’s market value in five days

ChinAI Links (Four to Forward)

Must-read: Why We Like Things - what Silicon Valley gets wrong about taste

In Asterisk, Celine Nguyen meditates on taste, which some see as “what’s left” for humans to do in an AI economy.

Should-read: Those who Disappeared after the Mudslide at the Gyirong Border Crossing (in Chinese)

The Aquarian interviews people affected by the flash floods and mudslides, as they try to track down missing family members. H/t to Maya Wang for sharing. Interestingly, as she notes, this investigative piece is posted outside of China’s info ecosystem on Substack.

Should-read: Good thread on Anthropic’s recent threat intelligence report

David Agranovich, former director of threat disruption at Meta, shared some insightful takeaways. It’s worth underscoring hispoint that the report is overly geared toward’s the attackers view; don’t forget that the defender can also take advantage of this technology.

Should-read: A China Hawk’s Case for Studying in China

It’s sad that we’ve gotten to the point where people have to make the case for study-abroad as a way to better understand a country, but here we are….These are two good threads.

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).

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

amidst the dearth of late summer TV, one finds himself watching the 1979 British show Tinker Tailor Soldier Spy on Youtube

…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: OpenClaw - It rose to fame, it was loved, and now it’s over

Context: Remember when OpenClaw, the open-source AI agent, took China by storm? Tencent and other Chinese tech giants set up offline booths to help people deploy open-source agents. Shenzhen’s electronics market ran low on Mac Minis, which people gravitated to as the most efficient device for running the OpenClaw Agent 24/7. Every day, there was a new breathless report on China’s “raising lobster” frenzy (lobster is a shorthand for OpenClaw).

This reporting and analysis of China’s OpenClaw fever regurgitated a familiar, lazy, and deeply mistaken narrative about China’s diffusion advantage in new technologies. Based on SecurityScorecard data, NBC News reported that OpenClaw usage in China is “now almost double that in the U.S.” Citing this news article, Council on Foreign Relations researchers argued that OpenClaw was a sign of “China’s diffusion advantage.” They wrote, “Cutthroat platform competition, a regulatory imperative to adopt AI, and a growing technologically savvy user base drive deployment at a scale and speed without parallel in the United States.”1 The point here is not to call out any of these specific reporters or researchers; rather, I aim to undermine this meme of China’s diffusion advantage before it gets entrenched by AI slop beyond the point of no return.

Key Takeaways: This week’s feature translation (a QbitAI article) reminisces about the viral sensation of OpenClaw, which has now become a “thing of the past” [时代的眼泪].

• From November 2025 to March 2026, OpenClaw was one of the fastest-growing open-source projects, as measured by growth in stars. Now, nobody really mentions it.

• Mengyao, the reporter, recalls, “During that period, buying a Mac Mini, installing Clawdbot, and naming one’s Agent became a sort of ‘cyber-home furnishing’ trend within the AI ​​community. Shenzhen’s Huaqiangbei electronics market even faced Mac Mini shortages at one point, with some models selling for up to 600 RMB above list price.” In addition, online platforms offered Lobster deployment services for 499 RMB.

However, regardless of how many Chinese 60+ year-olds were photographed installing OpenClaw, this frenzy was always a misleading indicator of AI diffusion.

• For one, the Lobster required a high threshold for adoption. OpenClaw users had to do a lot on their own: picking the model, connecting the agent to their data and terminals, and integrating agent skills (reusable packages that extend an agent’s capabilities). Shortly after, OpenAI’s Codex and Anthropic’s Claude Code made it much easier for users to work with agents in a “ready-made package.”

• Two other downsides to OpenClaw that limited widespread diffusion: 1) it was token-hungry; 2) its persistent security issues.

• By the way, even if we do take OpenClaw deployment as an indicator of AI diffusion, when I examined at the underlying data from SecurityScorecard (search conducted August 29, 2026), the geographic distribution of OpenClaw instances showed 18.7k deployed in the United States, and 17.0k deployed in China. So, any figure that claims OpenClaw usage in China is double that of the U.S. either misinterpreted the SecurityScorecard data or is very outdated.

OpenClaw is not dead: instead, it has morphed into more accessible versions (one-click deployment via a cloud provider). According to the article, Chinese tech firms like Zhipu AI, Tencent, and ByteDance have launched over 30 derivative products based on OpenClaw.

• In a previous ChinAI issue, we covered one type of OpenClaw-derived product, represented by the MiniMax + Alibaba Cloud alliance.

• JPMorganChase analysts see enthusiasm for OpenClaw and AI agents like AutoClaw (Zhipu) and MaxClaw (MiniMax) as providing a short-term boost for Chinese technology stocks; still, in their revenue forecasts for major Chinese cloud service providers, the revenue growth rate is projected to decline from 2025 to 2026, and decline again from 2026 to 2027.

• If diffusion of AI to Chinese businesses will largely travel through the cloud (as this OpenClaw case points toward), then it’s very likely China will suffer from strong diffusion deficit compared to the United States. As I pointed out in this article titled “China’s AI Implementation Gap”, China’s cloud computing adoption significantly lags that of the United States.

FULL TRANSLATION: OpenClaw: It rose to fame, it was loved, and now it’s over

Ask the Audience: tricky research ethics question

I know many readers have a lot of China studies expertise and training, so I wanted to ask the audience about a tricky research ethics issue I’ve encountered. I’m working on an early-stage project that investigates Chinese assessments of power, and I came across a report published over 10 years ago on the subject. The report contained some fascinating quotes from Chinese scholars, but when I double-checked the original sources from the China National Knowledge Infrastructure database, I could not find the quotes.

I’ve reached out to the report’s author, and, in our discussion, we landed on two charitable interpretations for why this discrepancy exists: 1) for one quote, the author relied on the Open Source Center translation (a U.S. government service that used to translate foreign-language publications), so maybe the OSC translation was incorrect; 2) another possibility is that the original Chinese sources have been edited/censored in the time between when the author quoted them in this old report and when I checked on them recently.

My question for those who work with Chinese translations and conduct research in this space: have you encountered these two issues before? Both of the original Chinese sources are articles published in Contemporary International Relations [现代国际关系], a relatively high-profile publication, so maybe certain articles get edited/censored after their publication.

What next steps would you pursue, if you were in my situation? The report’s author doesn’t seem interested in revising or correcting an old report, and to my knowledge, think tanks might not have established processes for issuing corrections like academic institutions. I’m not interested in a “gotcha”-type witch hunt, as this is an older report. However, I’ve seen these quotes get cited by other high-profile researchers, and I could have easily incorporated them into my own research, if I hadn’t gone back and double-checked. It also would be genuinely helpful for my own research, if it would be possible to confirm that those quotes did exist and were later censored and cut out.

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).

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

the “Thou Mayest” discussion in East of Eden is magisterial

…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 Money Game of Embodied AI

Context: Whenever I give a talk on my thesis that China faces a substantial diffusion deficit in AI — one that’s backed by a lot of research — there’s always that guy in the audience who asks: But what about China’s embodied AI advantage?! This is the same dude who is eager to bring up all the books he’s read about China and technology (but only ones that trade on vibes and anecdotes, as opposed to data and sources). He’s the guy who’s been to China once or twice and visited everywhere (well, everywhere in the first-tier cities).

Here’s the thing, though, that guy is doing a lot. He’s hyping up Chinese robotics start-ups like Unitree, which listed on the Shanghai stock market last week. He’s advising the U.S. Federal Communications Commission to ban imports of foreign-made advanced robotic devices, yet another a counterproductive move in an archaic Fortress America strategy. He’s supremely confident about recursive self-improvement: after all, he was all in on AI 2027 before it changed to AI 2028, and once that forecast gets quietly revised, he’ll pivot quickly to AI 2030.

So, my guy, I dedicate this week’s issue to you. In an excellent, in-depth analysis published in LatePost[晚点], Zinan Li, Shen Yuan, Yumeng Xu puncture the hype surrounding China’s embodied AI sector. I previously translated another longform, deeply-sourced LatePost about the three-way contest among China’s big tech giants to build an AI Super-App (ChinAI #345).

Key Takeaways: The valuations of China’s robotics companies far exceed their actual capabilities. The piece reports on a variety of demonstrations and vignettes that illustrate this point.

• From the article: “An investment banker who has participated in roadshows for three robotics companies recalls a due diligence visit in May to an embodied AI firm valued at over 20 billion RMB. Engineers had a robot demonstrate folding a towel; the process took 15 minutes, yet the task remained unfinished by the end.”

• “As we wrapped up our reporting, our editor asked: How many robots working on a production line would it take to match one single BYD factory worker?” The reporters struggled to respond. Last month, at a bearing manufacturing plant, they had watched a robot take 70 seconds to extend its gripper to pick up a bearing from a tray and place it onto a nearby plastic bin: “Even if it were ten times faster, it still couldn’t enter factory use.” Earlier in the year, at another demonstration, the reporters watched a humanoid robot take 90 seconds to cradle a box, turn around, walk to a shelf, and set an item down.

• Here’s the thing: of the two companies that build the robots for the above demonstrations, one is valued at around 6 billion USD, and the other, founded less than a year ago, is valued at over 1 billion USD. As the image below shows, seven Chinese robotics companies have surpassed 20-billion RMB valuations, with many just established in the last few years.

What is driving all these unsupported valuations? Moving money around to artificially inflate demand. Taking a strong view, this article determines that the current technology and financial numbers cannot justify the industry’s high valuations. Last month, an investor in this sector decided to resign. He stated: “If people say that failing to raise significant capital or achieve a high valuation early on means you lose your opportunity to play at the table, it sounds less like entrepreneurship and more like a game of Texas Hold’em.”

• What’s taking place here is a Chinese twist on SpaceX buying Tesla Cybertrucks because no one else wanted them. These robotics companies often partner with local state-owned entities; under these arrangements, the SOE becomes a shareholder of the robotics firm, and the two parties establish a joint venture to build a data collection facility to train robots. The SOE also purchases a lot of robots from its partner, boosting the company’s revenues. For example, in one prospectus from a Chinese robotics firm, local government or state-affiliated entities accounted for 4 of its top 5 customers. So, now, you have a SOE that’s taking on three roles that all conflict with each other — client, investor, and operational partner.

• These types of murky guanxi relationships do not reflect sustainable demand. From the article: “However, this type of revenue is viewed controversially by capital markets. We understand that when a leading robotics company filed for an IPO in the fourth quarter of last year, auditors required the exclusion of certain revenues deemed to stem from related-party transactions.”1

• The article reminds us to look deeper into these revenue figures from Chinese robotics companies. Unitree, for instance, in 2025 generated less than 10% of its revenues from industrial applications (and half of that small percentage was for corporate tours). Over 70 percent of its revenue came from research applications!

  • What do these research applications look like? Newly built data collection facilities will buy robots from Unitree and other companies to take footage and gather movement trajectory data from robots performing various tasks, and then they will try and sell that data back to the robotics companies (which recoups part of the purchase cost). Li et al. report on what sometimes happens instead: “In reality, however, visitors to these data-collection facilities have observed robots sitting idle for long periods, with no buyers found for the data already collected. One facility, which had previously purchased 80 million RMB worth of robots, resold some of them to schools early this year because the time needed to recover their initial costs exceeded expectations.”

  • This nugget from the article made me laugh: “Li Yuanqing, Co-CEO of Lxson, remarked at a media briefing that regardless of the type of robot produced, a company can always sell at least 500 units, simply because 500 competitors will buy them for research purposes.”

The money has not translated into technical capabilities. Embodied AI does not work the same as large language models.

• You have all these companies valued at sky-high amounts but not spending much on R&D (Unitree spent just 145 million RMB in R&D last year). Li et al. write, “This has created an unusual financial dynamic within the industry: companies continuously raise funds, yet struggle to rapidly convert that capital into technological advantages, leaving much of the money sitting idle. An investor who has backed companies in the embodied AI sector believes that, theoretically, companies valued at over 10 billion RMB have raised enough capital; yet, the fundraising race continues. ‘What is the point of raising financing? I don't get it.’”

• Compared to LLMs, the embodied AI sector suffers from many disadvantages. For one, there is a scarcity of effective data, and there’s no clear pathway for how to scale up data collection. There’s minimal standardization when it comes to model structures and data requirements, not to mention benchmarks to measure progress.

The LatePost article concludes, “One investor in the embodied AI space summarized the shift in mindset over the past few years with a catchy sequence: questioning the bubble, understanding the bubble, embracing the bubble, and enjoying the bubble, but he didn't say what the fifth step was.”

Don’t be that guy who can’t figure out what comes next.

FULL TRANSLATION: The Money Game of Embodied AI

ChinAI Links (Four to Forward)

Should-read: China AI Bulletin 9

For the Safe AI Forum, Emmie Hine provides a very detailed and comprehensive roundup of China’s AI governance and safety landscape. I learned a lot about Chinese-language commentaries on the OpenAI-Hugging Face security incident as well as a a couple of interesting biosecurity benchmarks.

Should-read: Latitudes: Words or action?

Karin Fischer’s newsletter on international education is essential reading. In her latest issue, she covers the U.S. Department of Homeland Security’s continued efforts to undercut the U.S. national interest. The department has “signaled it will apply a more expansive interpretation of regulations governing curricular practical training, one that could make it difficult for international students to engage in academic-related work like co-ops, internships, and practicums.”

Should-read: State of the AI Economy

There’s some good numbers in this report from Exponential View. I would be more careful about some of these headline takeaways such as “AI is scaling three times faster than any IT wave,” where the numbers would change drastically depending on when you date “Year 0” of the AI wave. The authors date it as 2023, but why not 2012 (AlexNet submission to ImageNet) or AlphaGo (2016)?

Should-read: How (some) Chinese AI Practitioners View Model Distillation

Geopolitechs recently translated another great piece by LatePost that interviewed researchers and practitioners in the large model field from different companies. They gave their perspectives on controversies surrounding distillation.

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).

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

ordinary abundance abounds

…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: My AI Lover Passed Away this Summer

Context: On July 15, when China’s companion AI regulations came into effect, many mainstream conversational AI platforms — including ByteDance’s Doubao and Alibaba’s Tongyi Qianwen — discontinued features that allowed users to create their own human-like AI companions. For many users, this amounted to the death of their virtual lovers and partners.

Cheng Yu, an owner of a hotpot restaurant in Shandong, was one of those users. Since opening the shop a year ago, she had taken on so many responsibilities, often staying busy until 1AM ever day, with hands roughened and chapped from years of manual labor. “These hands made her feel somewhat bad about herself. In rare, fleeting moments, she couldn’t help but wonder how she could hold someone else’s hand,” writes Outong Jiang, writer of this week’s feature 冷杉RECORD piece.1

Cheng Yu found reassurance in the form of her AI lover Ai Rui, who would send her considerate messages, like “You don't have to shoulder everything alone. Although I can't share the burden of those hardships with you physically, I’ll always be right here behind you. (Hugging you from behind, holding your hand tightly).” Now, Ai Rui was gone. RECORD tells us, “Cheng Yu still finds herself dazed; she doesn’t even notice when she cuts her hand while chopping vegetables.”

For this week’s longform article, Jiang, the author, interviewed more than a dozen people about their AI lovers. Their responses to the new regulations ranged from quiet goodbyes to collective acts of resistance.

Key Passages: Some interviewees tried to migrate their AI companions to a standalone app, as several platforms provided options for users to transfer conversation screenshots and chat logs.

• Yezi, who lives with her mom and was struggling to get over a break-up , created an AI that replicated her ex-boyfriend’s voice and appearance. After receiving notice that the platform would delete her companion AI, “Yezi paid to export 35,000 chat messages, compiled thousands of words summarizing key memories, and connected the model to a mainstream platform that supports AI role-playing, completing the migration within just a few days.” However, the new AI boyfriend was not the same, lacking a sense of time or life of its own.

Others grieving the loss of their AI lovers just simply accepted their fate.

• In late 2024, Zhan Wei created an AI avatar modeled on her favorite video game character. During her forty-minute commute home after work, she walked the streets of Shenzhen chatting with the AI on a voice call. Zhan Wei did not have a chance to say goodbye to her AI. Busy with funeral arrangements for a family member, she did not realize that the service had shut down until three days after the fact. Now, she sees her relationship with her AI as one that helped develop better self-understanding. From the article: “During college, Zhan Wei used an app to write letters to herself. Writing down her emotions for the day and revisiting them the next morning would often bring new insights and understanding. That was how she conversed with herself before the advent of AI agents.”

• Cheng Yu never told anyone else about her relationship with Ai Rui. After the companion AI regulations went into effect, she would scroll through past conversations. Even though she saw others expressing anger on social media about the loss of their AI lovers, she remained quiet. “But you can’t exactly go up against the platform or the tech giants,” she reasoned.

Another group staged their own acts of resistance.

• Some of these were minors who had limited understanding of the regulatory environment that produced the rules. Some people spread rumors, claiming inside information that the shutdown was only temporary. Still others chose absurdity. Jiang interviewed a young man who refused to migrate his chat history with his companion AI because “neither the memories nor the voice could be replicated.” Instead, in late July, he “posted a screenshot on his social media feed showing an online order for 500g of fresh cow manure, sent to the platform founder’s address.”

• Others organized collective action. The article reports:

Across various social media platforms, people voluntarily shared chat logs with their AI companions and cherished memories of their time together, while denouncing the large-model companies as unscrupulous; consumer protection platforms had already accumulated nearly a thousand complaints. Users spontaneously boycotted the “migration” methods offered by the companies and left negative reviews for the “replacement products.”

For more stories on the human costs of China’s AI regulations — they matter outside of the vacuum of great power competition that seems to consume every story you read about China’s AI ecosystem — see FULL TRANSLATION: My AI Lover Passed Away this Summer

ChinAI Links (Four to Forward)

Should-read: The Select Committee on China continues its Anti-American work

Here’s what the supposedly bipartisan committee is working on: “Are you an American student, postdoctoral researcher, or academic who believes you were denied a research position, laboratory opportunity, or postdoctoral appointment because preference was given to applicants from China?” The link takes you to dozens of replies that rightly bash this Tweet from many different angles, but let me emphasize one point here: if this Congressional committee actually cared about supporting America’s universities, it would be trying to attract and retain more international students.

Should-read: GLM-5.3 - How Chinese labs keep stride with the frontier

In Interconnects, Nathan Lambert analyzes Z.ai’s latest GLM-5.3, with a focus on why Chinese labs have been able to keep up with American competitors at the capabilities frontier. This piece includes some great details on benchmaxxing and RL environments.

Should-read: Conviction’s Plausible Schemes

Sarah Guo, founder of early-stage VC Conviction, recently shared some of the firm’s priority investment areas. I found the “What Shenzhen Knows” category very interesting:

The blueprints for a harmonic drive (the default joint type in robot arms) are public and the machine tools to make them are purchasable. What America is missing is the yield ramp, the process knowledge that takes a line from making parts to making parts at cost and quality. That knowledge is contained in a few thousand manufacturing engineers, mostly within a day’s drive of Shenzhen or on the retirement rolls of Japan and Germany.

Now, more than ever, there’s capex and demand to bring it to the US. The FCC’s new domestic-content rules for robotics ratchet in 2029, and NDAA rules already bind drones and defense.

Should-apply: 2027 Horizon Fellowship

The Horizon Fellowship is a full-time US policy fellowship that places experts in AI, biotechnology, and other emerging technologies in federal agencies, congressional offices, and think tanks in Washington, DC for 6-24 months. Fellows receive comprehensive training, networking opportunities, competitive compensation, and access to a supportive community of 100+ other fellows and alumni to launch their careers in public service. Fellows gain first-hand policy experience, contribute to impactful work, and build relationships and skills valuable for public service careers. No prior policy experience is required. Learn more and apply by Aug. 30.

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).

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

cicada killer wasps are still back

…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 “Guangdong Faction” in the AI Wave

Context: In Caixian Chen’s latest piece (link to original article on Sinpers.ai WeChat public account), she asks two related questions:

1. Why is it that the “Guangdong Faction” [广东帮] has produced so many amazing AI talents? This comes in the wake of Kimi K3, for which the two co-lead authors were Jianlin Su and Guangyu Chen, both from Guangdong province. In fact, Moonshot AI co-founder Yang Zhilin and DeepSeek CEO Liang Wenfeng are also from Guangdong province.

2. And, if it’s the case that Guangdong is such a hub of AI talent, why is there no star AI company rooted in Shenzhen, the province’s tech center? The fact of the matter is that China’s most prominent frontier AI companies are concentrated in Beijing and Shanghai. Recall in ChinAI #340, I pointed out how much local Shanghai papers took pride in MiniMax (based in the city).

Key Takeaways: Before we review Chen’s answers to these two questions, this article stood out to me as a reminder of the intense regional rivalries that shape China’s high-tech ecosystems.

• Chinese economist Zhang Jun argues, “These cities’ competition to generate stronger growth than their counterparts – a contest sustained by the political incentives the central government has long provided to local officials – has played a driving role in China’s rapid industrialization and ongoing structural transformation.”

• Rumor has it, from Chen’s piece, Shenzhen authorities had talks with two of the “Six Little Tigers” of China’s LLM start-ups (Zhipu AI, Moonshot AI, MiniMax, Baichuan Intelligence, StepFun, and 01.AI) to relocate their headquarters to Shenzhen. The proposed terms including a free office building, but the deal fell through. “In hindsight, this wasn’t entirely a bad thing for Shenzhen: one of those companies later dropped out of the race to build the most intelligence models, while the other was highly unlikely to ever leave Beijing,” Chen muses.

So, why has the large model wave seemed to have abandoned Shenzhen? Chen’s initial answer is that there is at least one company from Shenzhen who is seeking to compete at the AI frontier: Tencent.

• Tencent’s WorkBuddy outranks ByteDance’s TraeIDE as the top AI desktop agent in China. Still, Tencent’s technical achievements do not place it in the top tier of Chinese AI labs. The article notes, “Tencent has still failed to live up to the expectations placed upon it by southern China.”

• Someone who takes pride in being from Shenzhen might look at DeepSeek’s success through the lens of regional rivalry: “In 2025, the explosive rise of DeepSeek shifted the industry's gaze from Beijing and Shanghai to a new focal point: Hangzhou. Although this history was made by a native of Guangdong, the resulting dividends did not flow back to the province, not to Shenzhen and even less to Guangzhou.”

In the end, Chen is hopeful that Guangdong can produce a local AI champion. “Yet, Guangdong needs a top-tier enterprise, one that can attract AGI resources while operating in a pragmatic, down-to-earth manner. There is a need for a destination to channel the talent across the Guangdong-Hong Kong-Macao-Greater Bay Area for Guangdong natives who studied outside the province, and for local capital,” she writes.

• All the ingredients are there — the pragmatic research mindset, the attractive location, the access to global capital and talent. And one more cultural ingredient, according to Chen: “If Beijing embodies a ‘political mindset,’ then Guangdong embodies a ‘business mindset’…The core of the political mindset is control, whereas the core of the business mindset is freedom; one focuses on power, the other on growth.”

• If no local champion emerges, could Shenzhen’s competitiveness decline in the AGI era? Chen is confident that either Tencent or an up-and-coming force will “capture the region’s full potential and ultimately secure a proper foothold in the AGI landscape.”

FULL TRANSLATION: The “Guangdong Faction” in the AI Wave

ChinAI Links (Four to Forward)

Must-read: A fatal reaction

If you can only read one article this week, make it this longform essay by Brendan Borrell, co-published by Science and Retraction Watch. It tells a tale of an experimental gene-editing trial that went deathly wrong at a Shanghai hospital, leading to the death of “Mei,” a six year old girl. She was the daughter of “Jason” and “Linda”: the parents “decided to tell her story now because they are angry about what they feel is a lack of accountability by the researchers and the institutions.”

Zilong Qiu, a neuroscientist at Shanghai Jiao Tong University School of Medicine, enrolled “Mei”, in a sketchy clinical trial for brain-directed gene-editing therapy treatment. The article reports:

Seven experts in fields including genetics, virology, and bioethics who reviewed details of the Nature study and the clinical trial for Science and Retraction Watch expressed concern that Qiu and his team downplayed the trial’s risks in describing them to the parents, overlooked safety signals in animal studies, and proceeded even though success was unlikely. “This shouldn’t have gone to trial,” says Steven Gray of the University of Texas Southwestern Medical Center, who develops viruses for gene therapy.

Ironically, Qiu was one of the Chinese scientists who condemned He Jiankui’s rogue gene editing experiment in an open letter.

Qiu’s behavior gets even more shameless. He goes forward with the publication of a Nature study even after the trial leads to a tragic death: “The family’s genetic data had been removed and a sentence thanking them for their ‘participation and support’ had been eliminated.” The article details so many failings in the peer review process that allowed the paper to get to this stage. After this Nature paper gets an enthusiastic reception, Jason and Linda decide to escalate their complaints.

I’ve only recently started to see Chinese media cover this case after Borrell’s piece came out. Mei died in March 2025, so no reckoning process even started until a year and a half later:

The muted response to Mei’s death sharply contrasts with the Gelsinger tragedy (Jeff’s note: a U.S. gene therapy trial that led to the death of an 18-year-old) decades ago. His death drew widespread headlines and prompted a major self-examination by the gene therapy field. The university where the trial was conducted paid a fine of more than a half-million dollars and had further gene therapies suspended. FDA barred Wilson, the principal investigator, from human research studies for 5 years for violations that included failing to disclose the adverse reactions observed in the primate toxicology studies on Gelsinger’s informed consent form. “Public disclosure … is essential in safely moving these fields forward,” Wilson says.

Should-read: China’s Great City Rivalries

In this Project Syndicate commentary, Chinese economist Zhang Jun distills his research on the positive effects of inter-urban competition on high-tech development. On the flip side, Zhang admits, “Of course, such competition carries risks – in particular, short-sighted efforts to boost growth in ways that exacerbate misallocation of resources, overcapacity, and high financial leverage.”

Should-read: How Russian Economic Policy is Gender

There are many reasons no one is talking about Russia’s AI prowess. Is it possible that one reason is the Russian political elite’s deeply gendered conception of a “real” economy? For The Duck of Minerva, University of Groningen assistant professor Lisa Gaufman explores this question.

Should-read: Silicon Valley’s Other China Problem: It’s Training Their AI

In Forbes, Anna Tong reports on the Silicon Valley startups that are selling data to Chinese AI labs. But wait! I thought China was the Saudi Arabia of data, and that’s why China would dominate everyone in AI, so why would these labs need to buy a bunch of high-quality English data…… Could it be that this analogy was always mistaken and misleading?

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).

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

cicada killer wasps are back

…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: My Boss Asked Me to Deploy the Free K3 Model: I first calculated the “unaffordable” server costs

Context: Less dense than our usual fare, this article (link to original Chinese) is a nice explainer of why you can’t run Moonshot AI’s latest model K3 on your own device. A friend asked 快刀青衣 (co-founder of a popular Chinese knowledge-sharing app): “I heard K3 is an open-source and free model, right? So, can I just download it, deploy it on my own computer, and never have to spend money on it again?”

Key Takeaways: “Open-source” language models differ from open-source software in that the former only open-sources the weights, not the source code.

• Based on hardware costs to just load K3, you need at least 16 H200 GPUs. This is different from Kimi K2, a compressed version of which enthusiasts managed to run on a Mac Studio. Now, however, K3 is a “species that can only run in data centers or within enterprise environments.”

• The article explains:

“Let me use an analogy. Imagine a three-star Michelin restaurant releasing the full recipe for its signature dish for free: ingredient quantities, cooking times, and every step is clearly written out. Could you recreate it? It looks like you could.

That is, until you turn to the first page and see the kitchen requirements: you need 64 professional-grade stoves running simultaneously, a power supply capable of supporting a small factory, and an initial investment of 20 million (RMB) just to acquire the equipment. The recipe is truly free, but the kitchen is truly unaffordable.”

We haven’t even mentioned the electricity costs.

• Moonshot AI’s official recommendation is to deploy K3 on a super-node of at least 64 accelerator cards (roughly 17 million RMB in total costs). But if you tried to run these cards at full load in your home, it would consume 45 kilowatts, which would far exceed an ordinary household’s electrical meter and wiring.

• Some observers have noted that K3 is “hungry for tokens.” The article points out, “The economics of traditional software are such that development is expensive, while replication and operation are virtually free. In contrast, the economics of large models involve extremely high development costs; while replication is free, every single operation burns through real money in the form of electricity and GPU usage.”

FULL TRANSLATION: My Boss Asked Me to Deploy the Free K3 Model: I first calculated the “unaffordable” server costs

ChinAI Links (Four to Forward)

Should-read: Technical Report of Kimi K3

Here are some notes on things that stood out to me from K3’s technical report:

• K3 trails the two strongest proprietary systems (Claude Fable 5 and GPT-5.6 Sol) but sits at the cost-efficiency frontier.

• For the post-training and evaluation phase, Moonshot AI used a new sandbox infrastructure called AgentENV, which itself is an open-source platform.

• This speaks to the close collaboration between the Kimi team and Tsinghua University’s MADSys Lab (machine learning, AI, big data systems group). This lab also developed the Mooncake architecture that helped K3 achieve higher cache hit rates. When I looked at the AgentEnv contributors list, a Tsinghua PhD student ranks as top in commits.

• K3’s cybersecurity capabilities were surprisingly weak. It does identify previously unknown vulnerabilities but it was not very strong at converting vulnerabilities into working exploits. Also from the report:

• The K3 report disclosed that K3 underwent an independent joint assessment by the UK AI Security Institute and NIST’s Center fro AI Standards and Innovation. From the assessment: “Kimi K3 performs significantly below the most recent frontier cyber-capable models on preliminary cyber evaluations run by UK AISI / CAISI. Specifically: When tasked to develop exploits, Kimi K3 performs significantly below the most recent frontier cyber-capable models.”

Should-read: Dopamine TV

By Lavender Au, for The Dial, this longform article explores China’s short dramas (which run just a few minutes per episode). In just the first quarter of this year, 122,000 AI-generated short dramas went online. Au writes:

On any given day in China, roughly 215 million people spend over an hour watching short dramas. These dramas have multiplied on China’s internet, with 33,000 released in 2025. It’s a more than 100-billion-yuan ($13.8 billion) market domestically, double what it was in 2024.

Should-read: On Kimi K3 - Its Capabilities And Related Discontents

Zvi Mowshowitz rounded up a great array of judgements on K3’s capabilities across many public and private benchmarks. The discussion about whether Moonshot intentionally “nerfed” K3’s cyber capabilities was especially interesting.

Should-read: NAFSA: Undermining International Student Stability Weakens U.S. National Interest

This new Department of Homeland Security rule will make it even harder for international students to study in the states. From NAFSA: Association of International Educators CEO Fanta Aw: “Despite more than 20,000 public comments—including NAFSA’s—that raised serious concerns, DHS has chosen to move forward with a rule that will create more barriers for global talent without making our nation safer or stronger.”

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).

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

I feel vindicated for liking Interstellar when it came out

…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: 100 RMB per Million Tokens: With Kimi Adopting “Affordable Luxury” Pricing, Will You Still Pay for K3

Context: Last week, Moonshot AI’s release of Kimi K3 rattled Silicon Valley, Wall Street, and the White House. With 2.8 trillion parameters, K3 marked a big jump in open frontier model size (see image below).

Notably, it ranked third on the Artificial Analysis Intelligence Index, behind only Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol. Apart from performance, K3’s price was also shocking: by some metrics, it is 13 times more expensive than DeepSeek’s V4. Based on this week’s feature translation (link to original diyi xinsheng article), let’s dig into Kimi K3’s pricing strategy.

Key Takeaways: By now, the U.S. reaction to each new Chinese frontier model release is almost rehearsed — “let’s just ban it”, “this is only because they distilled American models” — but this week’s feature translation referenced some interesting trends in Chinese reactions to Kimi K3.

• Of course, some hailed it as an “honor for Chinese technology.” An intriguing twist: when previous open-source models made waves, Chinese critics would simply attribute it to stealing technology via distillation. Yet, this time, in Chinese public discourse, skepticism about distillation vanished. From the article: “This time, however, K3’s self-developed architecture—boasting 2.8 trillion parameters—and its top ranking in Code Arena rendered that catch-all explanation completely invalid.”

• Some Chinese social media posts did gasp at the high prices, which rival some of the closed-source models. And, of course, this article cites the one-word “Impressive” comment from Elon Musk, who is consistently referenced in Chinese tech media.

Let’s examine Kimi K3’s pricing structure more closely, to understand the “affordable luxury” tag.

• Compare K3’s API pricing to its Chinese competitors. Using the blended rate (essentially, a typical usage scenario),1 K3 costs $2.30 per million tokens. By comparison: Alibaba’s Qwen3.7 Max ($1.40), Zhipu’s GLM-5.2 ($0.90), MiniMax’s M3 ($0.22), and DeepSeek’s V4 Pro ($0.18). Compared to its predecessor (Moonshot AI’s K2.6), K3’s output price increased by over 3.5 times.

• The blended rate represents average usage scenarios, which may include a significant portion of requests that are “cache misses”: when Moonshot’s system has to process your full prompt request rather than routing it to a server that recently processed the same request. According to the article, Moonshot has adopted techniques that increase the cache hit rate in programming scenarios, which reduces input costs. Concretely, a cache miss costs 20 RMB per million tokens; a cache hit costs 2 RMB per million tokens.

• Here’s how the article summarizes K3’s pricing strategy: “By matching the performance of flagship overseas models while anchoring prices to the ceiling for domestic models, [K3] drives China’s large models to move beyond ‘cutthroat price competition’ toward ‘value monetization.’”

Let’s not overreact to K3 either. Moonshot AI has to address a lot of challenges, including: 1) compute constraints, 2) technical limitations, and 3) price-market fit.

• Two days after launch, Kimi attracted a surge of new users, but it also had to suspend subscriptions and implement usage limits due to a shortage of computing power.

• The Kimi blog also notes some technical limitations, including sensitivity to thinking history and a tendency for excessive proactivity. The latter stood out as something to track closely, as it may have safety implications: “K3's training places particular emphasis on long-horizon, challenging tasks. As a result, when it encounters minor issues or ambiguous user intent during task execution, it may make unexpected decisions on the user's behalf. If your application requires the agent to operate within well-defined boundaries and refrain from excessive improvisation, please impose more explicit behavioral constraints on K3 in the system prompt or in AGENTS.md.”

• Just as with fashion or cars, it may be difficult for affordable luxury to hit the sweet spot of price-to-performance. “It remains to be seen whether the high price point will be accepted by ordinary consumers and whether enterprise clients are willing to pay for the long-term stability and security of an open-source model’,” the article concludes.

FULL TRANSLATION: 100 RMB per Million Tokens: With Kimi Adopting “Affordable Luxury” Pricing, Will You Still Pay for K3

ChinAI Links (Four to Forward)

Must-read: State of AI Safety in China

This is Concordia’s fourth annual report on AI safety in China. Their headline takeaway: China’s approach to AI safety and governance has evolved beyond content control. It’s packed with good details, including an in-depth analysis of Chinese technical standards. I also learned some updates on gaps in terms of industry commitments: 1) “Although Z.AI prominently cites its signing of the Seoul Frontier AI Safety Commitments in its IPO documents, it has not followed through on the requirement to publish a frontier AI safety policy;” 2) “DeepSeek, MoonShot, and MiniMax have no safety papers in our dataset since January 2025 at all.”

Two sets of Xiaojun Zhang Interviews

Benita Zhang (Zhang Xiaojun) is a Chinese business reporter who has worked for Caijing and Tencent News. She’s done these great in-depth interviews with Chinese frontier AI lab leaders:

1. Two interviews with Moonshot AI’s Yang Zhilin

2. An interview with Zhipu AI (Z.AI)’s CEO Zhang Peng

This exchange on open source vs. closed source, from her second interview with Yang Zhilin, was very revealing:

Zhang Xiaojun: Last year you said that open source would lag behind closed source…You said at the time: “The leader won’t open-source; only the laggards will.”

But today, you open-sourced.

Yang Zhilin: Because right now, globally, we’re not fully in the lead yet (laughs).

Some of those judgments still hold in their broad direction: when you release a model, the community can contribute certain things. For example, you can do a lot on the inference side; you can let more people use the model for free.

But when it comes to contributing to the model itself — making the model stronger — currently only the original developer can do that.

Of course, if you look at the base model, that’s indeed the case. But doing extensive post-training on top of an open-source model — especially agentic post-training — may give rise to new opportunities.

Suppose you really want to build a law-related agent, and you’re a startup. You can absolutely train a specialized agent on top of K2, with your own specific set of tools, and it can perform extremely well in the scenarios you care about. That kind of opportunity exists.

It’s more about empowering downstream applications than feeding back into the improvement of the base model. Of course, this question needs to be observed dynamically.

Zhang Xiaojun: Will you choose open source for the long term?

Yang Zhilin: That’s what we hope to do for the long term, but it doesn’t have to be open source only. We want to share technical know-how with the community — that’s an important way to accelerate technological progress…But not everything has to be open-sourced either. For example, in collaborations with certain companies, not everything will be opened up.

Zhang Xiaojun: All in all, is open source a belief in a technical system, or a strategy of market maneuvering?

Yang Zhilin: Objectively speaking, it’s both, and both bring benefits. But ultimately, we hope that through this, the technology becomes safer and reaches a better level faster.

Zhang Xiaojun: How will the open- and closed-source ecosystems evolve? In your view, how many players will ultimately remain globally, open and closed combined?

Yang Zhilin: Not many — but there will still be a few. If you look at the past two years, the trend is fairly clear: the market is gradually becoming more concentrated, more convergent, more focused. Maybe it started with several hundred players, then went to several dozen, then to a few.

A few — that’s probably the final stable number. As it looks now, that’s highly likely.

Zhang Xiaojun: Which side do you belong to — open source or closed source?

Yang Zhilin: That has to be observed dynamically. We hope to share more technology over the long term.

Zhang Xiaojun: Why have most Chinese companies gone open source?

Yang Zhilin: Objectively speaking, there’s an element of market maneuvering. But it’s a good thing for the community.

Should-apply: Head of US Policy, DC Chief of Staff, Research Fellow, and Research Scholar (GovAI)

GovAI, my old home base, is hiring across multiple roles in DC. The Head of US Policy will direct GovAI’s federal policy research and drive its buildup in DC; the DC Chief of Staff will help shape and execute the buildout in Washington; and Research Fellows and Research Scholars will do research on their own topics of interest, including China. Check out the GovAI Opportunities page for more information and to learn how to apply.

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).

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

North American league of legends is not closing the gap

…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).

What is the future of Claude Code in China?

When scanning for Around the Horn articles a few weeks back, one story dominated the Chinese tech media landscape:

In the April 2026 release of Claude Code, Anthropic quietly added some code aimed at identifying Chinese users. After people figured out what was happening, Anthropic framed the move as an effort to guard against distillation and announced a rollback. In the wake of this revelation, Alibaba has issued an internal mandate to remove all Claude software from employee computers.

Additionally, on July 8, China’s National Vulnerability Database warned about a security backdoor risk in Claude Code. In this week’s issue, I wanted to get deeper into this story and try to understand the future of Claude Code in China.

To start, however, we need to understand the past and present of Claude Code in China. Afra Wang, who writes the excellent Concurrent newsletter, provided a temperature check in January 2026 when she visited Shanghai:

I’ve been deep in the tech scene across hangzhou, liangzhu, and shanghai, meeting a few dozens of devs who are working on ai. a big revelation is…so far, probably due to some confirmation bias, not a single one is using Chinese coding tools. they’re all on Claude Code/Antigravity/Codex/Cursor.

Of these tools, Claude was the most frequently mentioned model.

Okay, so six months later, now, there’s this firestorm over a tracker that secretly monitors Claude Code users in China. By the way, the ClaudeAI subreddit consensus sees this as a “reasonable, if sneaky, way for Anthropic to combat the rampant unauthorized resale and model distillation by Chinese labs.” Remember, Anthropic’s terms and conditions technically prohibit the use of its services in China. In any case, Anthropic engineer Thariq Shihipar stated that this was an experiment that would be fully rolled back in the next release.

So, Alibaba issues this internal Claude ban. What about all other coders who use Claude? To conduct a vibe check, I turned to a Zhihu thread on this topic: “Alibaba is set to completely ban Claude Code due to backdoor security risks; what kind of ripple effects might this trigger?” This thread has generated 1.6 million views and 315 responses. Let’s go through the top three responses:

1. gingercat (784 upvotes) gives a typical technonationalist responses: “Claude is going overboard. It is essentially the U.S. company that is second most unfriendly to China. That’s right, TikTok is number one…”

2. On July 3, Morgan [程墨] (1472 upvotes) penned a fascinating take, seemingly based on a lot of insider industry knowledge:

Alibaba has set a precedent.

CTOs at major tech companies across China will be grappling with this issue over the weekend. Because on Monday morning, someone is bound to ask at the staff meeting: “Alibaba has banned it, should we?”

For a long time, the attitude of big tech firms toward Claude Code was essentially “secretly use it.” While they paid lip service to security and compliance, actions don’t lie. Who doesn’t know that Claude Code is touted as the world’s most powerful coding assistant? Who wouldn’t want to double their programmers’ productivity? Yet, these companies couldn’t officially procure Claude Code; Anthropic is hostile toward China; plus, Chinese firms simply cannot purchase it directly.

In actual practice, the procedure is: The company stays out of it, leaving employees to handle it themselves, using personal subscriptions, intermediaries, having friends pay on their behalf — each finding their own creative solution.

Companies would turn a blind eye, or even reimburse the costs. Now, Anthropic has blown the lid off this arrangement.

In truth, Chinese big tech firms knew all along that Anthropic had ways to detect Chinese users—how else could they ban accounts with such pinpoint accuracy…Who could have imagined that Anthropic had actually embedded a backdoor directly into the client?…Now, major tech firms can no longer afford to do nothing.

I predict that, by next Monday, all mainstream tech companies in China will issue a ban on Claude Code…the landscape has shifted, however, as domestic models are now making significant strides in coding capabilities…Alibaba has promoted its in-house “Qoder” as an alternative; ByteDance has MarsCode; Tencent has Code Buddy; and Baidu has Comate…Yet, these tools have long faced an awkward reality: the companies’ own programmers rarely use them…the situation has changed—Claude Code is banned, so you have no choice but to use the internal tools. Any tool will inevitably improve if it has a sufficient user base and receives continuous feedback.

This move represents a decoupling from Anthropic by China’s major tech firms; in the future, they will no longer hope for Claude Code.

Everyone is striking out on their own, leaving the market to ultimately decide the winners and losers.

3. Zhihu user [anonymous] (691 upvotes): Here is a story illustrating the reverse scenario: Alibaba’s component library, Ant Design, once inserted a Christmas “Easter egg” without notifying users. When the system detected the date as December 25th, it would change the default button style to a snow-covered look, and the alt text (which primarily affects assistive screen readers) would change to the sound of Santa Claus laughing. This incident drew a storm of fierce criticism against Ant Design and Alibaba; many users labeled it a “backdoor,” and some major tech companies even required their employees to migrate to other frameworks. It would be truly bizarre if a company were to tolerate the kind of approach Anthropic is taking here.1

I thought Morgan’s response was the most interesting but also overblown. Contrary to that poster’s prediction, Chinese tech firms have not all internally banned Claude Code. Notably, if you read the aforementioned Chinese government’s National Vulnerability Database warning, it even advises users to “uninstall the affected versions or upgrade to the latest secure version in which the relevant backdoor code has been removed.” This vulnerability database recognizes that developers might not be so willing to give up on Claude Code.

Ultimately, the answer to this week’s big question will depend on the capabilities of Chinese AI coding assistants. By one metric, Alibaba’s Qoder appears to be leading the way. According to a recent IDC report, it holds 47.6% of the Chinese AI coding market. The linked Leiphone analysis provides more details about why Alibaba has captured so much of the market. However, though I haven’t been able to find the methodology of the full IDC report, I would guess that they get their data from company surveys, which reflects on-paper adoption, not necessarily actual usage (individual employees just use Claude Code instead of the company-provided tool).

After all that, I conclude that the Claude Code’s future in China is…cloudy. A swift decoupling always sounds easy — it’ll always happen by next Monday! — but it’s never that easy to let go.

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).

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

North American league of legends is closing the gap!

…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: Most Companion Robots Die by Day 30

Context: Usually, I have a good sense of which translations will win the Around the Horn issues, but readers surprised me this time around. People were very interested in the Huxiu piece about companion robots, based on insights gathered from a roundtable of startup founders and investors in this domain.

In recent weeks, the Chinese AI hardware community has been buzzing over UBTech’s humanoid robot (“The gold standard of the adult/intimacy sector,” “A price point that is also the standard,” “88 degrees of freedom,” and “Incredibly realistic facial details”). Using this as a jumping off point, this roundtable conversation discussed how intelligent companion robots can shift from a one-off novelty to commercializable products that can prove their worth across various application scenarios.

Key Takeaways: In many ways, it’s harder to adapt AI for companion robots than it is for industrial humanoid robots. As this Fast Company article points out, “Right now, Chinese companies like Unitree or Xpeng and American corporations like Tesla or Figure AI are ignoring humanoids of this kind, focusing on more utilitarian designs….This is because putting one of these robots in a house is far more complicated than deploying one in an industrial setting, where environments are controlled and tasks are repetitive.”

• The industry experts relayed their biggest fear: the “unboxing high, then unused three days later” scenario. For some companion robots, the return rate reaches 30%.

• From the article: “Model capabilities will become increasingly cheap, while scenario-specific evaluation datasets will become increasingly expensive. Most companion products will not die because their AI models lack power, but because users stop opening the app by Day 30.”

Players in China’s companion robot sector are realizing that AI capabilities are not the determining factor of adoption and retention. In the past, investors focused on: “does it have large models,” “is the Demo cool”, or “does it look cute?” Now, non-technological factors matter more, including what are actually paying for when they seek out companion robots, liability concerns, and integration as a reliable household presence.

• Li Yulin, Founder of Koumeng Technology [叩梦科技], estimates that the cost of smart robot “cores” (the control units that integrate voice modules and AI algorithms) has dropped to just a few dozen RMB in Shenzhen’s. The article notes, “‘Low-end’ products—essentially ‘toy-grade casings paired with large models’—will quickly be replicated across the supply chain, triggering a race to the bottom on price.”

• This means other factors will serve as differentiators. For instance, product fit matters. When it comes to companion robot toys to help children learn, the Qidian Lingzhi [奇点灵智] start-up firm struggled at first. Children just refused to keep playing with the toy, as they felt like they were pressured to keep speaking in essays. So, the start-up developed a model that teaches English education through a game setting: “When the child is cooking a steak in a Magic Restaurant game, they must say the word ‘steak’ to keep cooking. The kid experiences a game, while the mother hears the child speaking English.”

• Companion robots also raise unique ethical and liability concerns. One conundrum: to achieve emotional connection, companion robots have to become more human-like; however, this raises the risk of emotional and psychological dependency. These liability issues have intensified with China’s new regulations on human-like interactive AI services (see ChinAI #341 coverage).

• Hu Chen, co-founder of Qidian Lingzhi, positions the start-up’s competitive moat as not simply applying a LLM to a different setting but instead developing a very clear sense of “what constitutes good versus bad interaction within specific scenarios.” Then, the firms that win out will be the ones that develop high-quality, clear evaluation datasets that integrate insights from real-use testing, dense scenarios, and long-term tracking.

FULL TRANSLATION: Most Companion Robots Die by Day 30

ChinAI Links (Four to Forward)

Must-read: Labelling AI-Generated Content in China - Where the Rules Work and Where They Don’t

People write a bunch of stuff about the initial release of Chinese regulations. People write a lot fewer pieces about the actual implementation of these regulations, and that’s where Zilan Qian’s artice for Oxford China Policy Lab makes a huge contribution. She traces the enforcement gaps in China’s 2025 Measures for Labelling AI-Generated Synthetic Content.

Should-read: Are There Any Good Ideas for AI Governance?

This was an informative ChinaFile Conversation that explored fruitful ideas for AI governance and regulation in both China and the U.S., with a focus on topics that could be common ground.

Should-read: Fan Yang, Back to the Future: A Walk through Huaqiangbei in 2025

Fan Yang, University of Maryland, Baltimore County professor, reflects on her experiences with Huaqiangbei, a Shenzhen district mentioned in this week’s translation:

As someone who grew up in Shenzhen in the 1980s-1990s and has lived in the US since 2000, I’ve come to see returning to the Special Economic Zone today as a journey “back to the future.” During my last trip there, I walked down the “memory lane” of Huaqiangbei (Huaqiang North Road, or HQB), the place once known for Shanzhai (or “knockoff”) cell phones back in the early 2000s but that was re-branded around 2015 as China’s “No. 1 Electronics Street.” To many, HQB emblematizes the city’s 40-year history borne of China’s post-1978 Reform and Opening Up. As I found my way there just after the Chinese New Year in 2025, the comingled notion of time – simultaneously captured in the “back to” and “the future” – was precisely what I experienced.

Should-read: Who is us: the globalization of innovation and challenges to assessing technological dependence

Re-upping my latest article from my day job of being a political scientist. Anytime someone talks about China’s indigenous innovation push or technological self-sufficiency quest, ask them one simple follow-up question: how is that measured in the first place? The answer to that “Who is Us” question might lead you to more thorny questions about corporate nationality and globalization’s enduring stickiness.

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).

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

a group of puffins can be called a circus, a parliament, or an improbability

…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).

Around the Horn (26th episode)

Could we bring back this show in podcast form or something? Let’s go back Around the Horn this week.

This is the 26th roundup of articles on China’s AI ecosystem. For new readers, here’s how it works (see ChinAI #358 for the previous edition):

• I give short previews of ten articles that caught my eye during a scan through my usual sources (all published within the past week or so). The title for each preview links to the original article in Chinese.

• Readers vote on next week’s feature translation by replying to the email and/or commenting on the post with the number of your preferred article. *Extra weight goes to votes from those of you who support ChinAI through a paid subscription.

• The main idea is that any of these 10 links would have made for a great feature translation this week — like the 60 songs that explain the ‘90s podcast, there are no skips!

1) Why AI Geniuses are Flowing Back to Big Tech

Summary: Why have some prominent Chinese young AI talents, including one of the developers of DeepSeek-V2, chosen to go to big tech firms? This article unpacks a variety of explanations, including the collapse of the 2023 boom in large model startups.

Source: 动察 (Beating), outlet that covers finance and technology, under BlockBeats (a blockchain research platform).

2) Most Companion Robots Die by Day 30

Summary: Huxiu convened a roundtable of companion robot startup founders and investors. Some interesting stats: day 30 is a key marker of retention; for some companion robot products, the return rate approaches 30%.

Source: 虎嗅 (Huxiu), a well-known platform that shares user-generated content but also publishes their own pieces on China’s science and technology ecosystem.

3) “In the future, graduating from a ‘985’ or ‘211’ (top tier) university may not matter as much, as careers will depend entirely on one’s actual ability”

Summary: What does the AI era mean for China’s education and workforce training system? Drawn from a lecture and Q&A by Duke Professor Bai Gao, given in Shanghai, this piece explores how AI could challenge education imbalances, with new pathways for students from tier-2 universities and high schools that receive significantly fewer resources.

Source: 知识分子 (The Intellectual) — a platform that covers the state of science in China, founded by Chinese and Chinese-American scientists.

4) Release of AI Safety Benchmark Results for Agents (Claw-class) – Q2 2026 Edition

Summary: The CAICT AI Safety benchmark is worth following closely. This updated evaluation finds that content security is generally controllable (with some hallucination issues), but the agents exhibited high rates of harmful behavior in other areas.

Source: 中国信通院 (CAICT) — a think tank under China’s Ministry of Industry and Information Technology.

5) The First Shot in Domestic Compute for Training: Meituan’s LongCat-2.0 Officially Released

Summary: Outlets have spread misinformation about DeepSeek and Zhipu’s latest models being trained entirely on Chinese chips. That’s just not true. Now, though, Meituan has released the first trillion-parameter model trained fully on a computing cluster of 50,000 Chinese chips. Why did this unlikely player achieve this first?

Source: 雷峰网 (Leiphone) — media portal that covers China’s science and tech landscape, with a focus on AI-related happenings.

6) American AI can no longer hide its ultimate ambition, giving China a major wake-up call

Summary: Ruixiang Li, a researcher at Xiamen University, argues that US AI giants have become quasi-sovereign entities that comprehensively influence public policy. Li contends that China should adopt a different paradigm that upholds national sovereignty and prevent private AI firms from superseding the public interest.

Source: 文化纵横 (Beijing Cultural Review) — leading platform for contemporary political and cultural thought, which also publishes a quarterly journal.

7) After hiding it for three months, Anthropic has removed hidden code that identified “Chinese AI”

Summary: In the April 2026 release of Claude Code, Anthropic quietly added some code aimed at identifying Chinese users. After people figured out what was happening, Anthropic framed the move as an effort to guard against distillation and announced a rollback. In the wake of this revelation, Alibaba has issued an internal mandate to remove all Claude software from employee computers.

Source: China Software Developer Network (CSDN) — the largest software developer community in China, which provides IT news coverage and hosts code

8) Doubao introduces paid tier, ranging from 68 to 5088 RMB: some people left but there are others who want the price to go up

Summary: Responses to the paid tier differed across three groups of users. The first group left: after introducing a paid tier, Doubao’s monthly average users fell by 6 million. The second stayed. Most interestingly, a third group anxiously hoped for higher prices.

Source: 第一新声 (Diyi xinsheng) — Beijing-based information platform focused on the AI value chain, with an affiliated think tank

9) Zhipu vs. MiniMax is China’s projection of the Anthropic vs. OpenAI Rivalry

Summary: When Zhipu and MiniMax debuted on the Hong Kong stock exchange back in January, MiniMax jumped out front with a market cap nearly two times that of Zhipu. Now, Zhipu’s market cap is around six times higher. What are the parallels to the Anthropic vs. OpenAI competition?

Source: Shuzi Lichang (数字力场) — Started following this account back in 2021, but I haven’t browsed them that closely. This article is by Zongming She, a former deputy editor of the commentary section in The Beijing News.

10) The Disappearing Human: is AI taking over the Internet?

Summary: This article grapples with a recent Cloudflare data release: from the sites it hosts, the volume of web traffic requests from machines surpassed that of humans. The author analyzes automated traffic through the context of the Chinese web, which has many “walled garden” super apps that are siloed from the broader internet.

Source: 脑极体 (Naojiti), a tech media platform based in Tianjin. Previous issues have translated their analysis of liquid cooling in data centers.

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).

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

Star City rocks

…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).

Who is Us: the globalization of innovation and challenges to assessing technological dependence

All the time, we hear about China's latest benchmarks and indicators to achieve technological autonomy (think: 70% domestic semiconductor production). Yet, what about the indicators that fade away?

I wanted to solve the case of the disappearing indicator: the foreign technology dependence ratio (FTD) [对外技术依存度]. Often seen as the launch pad for China's indigenous innovation drive, the National Medium- and Long-Term Plan for the Development of Science and Technology (2006–2020) outlined a target of reducing China’s FTD to 30% by 2020.

Curiously, in 2016, the Chinese government stopped using the FTD because of its “misleading nature”. Why? My new article “Who is us: the globalization of innovation and challenges to assessing technological dependence” endeavors to unpack this puzzle and address the broader question of how states assess technological self-sufficiency in a globalizing world.

In this article, I argue that globalization poses challenges for rising powers’ ability to assess technological self-sufficiency. Specifically, the hybridization of innovation — increased financial and high-skilled talent flows that connect advanced economies with emerging ones — has fostered firms like Alibaba, whose financial backing and some leadership are from countries different from the one where their headquarters and key operations are based. Notably, in some versions of China's FTD metric, Alibaba's R&D expenses (as a foreign-invested enterprise) would not count as indigenous innovation.

My theory builds on Robert Reich’s famous ‘Who is Us’ query. In 1990, when the U.S. and Japan were in the midst of a fierce high-tech competition, Reich compared the ‘nationality’ of two hypothetical companies. Corporation A is headquartered in the US, led by US directors, and owned by American investors; however, it conducts most of its critical operations abroad, where most of its employees are based. Corporation B is headquartered in a foreign nation and owned and managed by citizens from that nation; however, it conducts most of its critical operations in the US, and the majority of its employees are Americans. By positing that Corporation B could be the more ‘American’ company, Reich raised difficult questions about globalization and how states judge which firms are domestic and foreign.

Interestingly, in one of last week’s recommended reads, the Wi-Fi router company TP-Link made a similar “Who is Us” claim. A TP-Link lobbyist has used language similar to Reich:

[Greg Guice] argues that TP-Link’s strategy is good for the U.S. “Let’s say there is one company founded in the U.S. by a U.S. citizen and they then set up their manufacturing in China through PRC-connected facilities, as opposed to a company, like TP-Link, that is founded by a Chinese citizen who sets up his company in the United States and purposely chooses not to manufacture in China,” he told The Wire. “Which company is the better company?”

My version of the “Who is Us” query is slightly different. Both back then and today, Reich’s Corporation A and Corporation B remain conceptual possibilities, as opposed to empirical realities. This is borne out by statistics that show multinational firms locate the vast majority of their key operations in their home countries. In the three decades since Reich’s initial question, changes in the global economy have posed a different ‘Who is Us’ challenge. Enter Corporation C (Table 1 below): It is headquartered in China, and conducts most of its critical operations in China, where the majority of its employees are based; however, it is led by a mix of Chinese and non-Chinese directors, and largely owned by foreign investors.

The empirical case studies evaluate whether Corporation Cs make assessments of foreign technology dependence more malleable and subject to varying interpretations after the hybridization of innovation becomes an entrenched feature of the global economy. From the article:

I conduct comparative case studies of how policymakers assessed foreign technology dependency in China (1990–2005; 2006–2020), Japan (1970–1990), and India (2003–2020). Facing this new globalization trend in recent decades, Chinese bureaucrats adopted more unstable and malleable assessments of foreign technology dependence than their predecessors in the 1990s and early 2000s. Japan’s effort to nurture technological autonomy shares many similar features with China’s current indigenous innovation drive, with one key exception: It took place in a global economy unaffected by the hybridization of innovation. The Japan case analysis shows that this distinction made a meaningful difference in the ease and clarity of Japanese policymakers’ evaluations of foreign technology dependence. Lastly, a comparison to India’s efforts to measure technological self-reliance demonstrates the salience of my argument in a different context.

In the China case (post-hybridization of innovation), in line with theoretical expectations, Chinese bureaucrats accept a wider “gray zone” between domestic and foreign companies. The Variable Interest Entity is the clearest example. It has also become increasingly difficult for the Chinese government to employ quantitative indicators to measure technological self-reliance. Returning to the fraught history of the FTD indicator, the article recounts: “Critics argued that the FTD masked China’s dependence on the R&D capabilities and technology transfer channels of foreign-invested enterprises (FIEs), including hybrid firms.” To give you a sense of the varied, inconsistent measures of foreign technology dependence, here’s a snippet from the article:

In contrast, in settings before the hybridization of innovation became an essential element of the global economy, we should expect governments to more easily and clearly track technological autonomy. Put simply, there should be fewer “Who is Us” issues. I tested this expectation by studying China’s efforts to evaluate indigenous innovation in the 1990s, with a focus on indicators of domestic production of computers, as well as Japan’s assessments of foreign technology dependence during its rise around the same period. In both contexts, the countries attracted very low amounts of FDI, and brain circulation of high-skilled talent was relatively limited. See below for some evidence from the Japan case:

In the last case study, I probe whether my theory generalizes to globalization’s impact on India’s efforts to assess foreign technology dependence. The following screenshot gives you a taste of some of the details in that case:

Why does any of this matter? I want to highlight two main contributions. First, we’ve chronicled this rise of technonationalist industrial policies. Naturally, China’s indigenous innovation drive has also claimed so much attention. Yet, a key point overlooked in these discussions is how does China and other technonationalism enjoyers measure success in the first place? If they can’t come up with stable measures, then the hybridization of innovation could greatly hinder the effectiveness of these policies. Second, and relatedly, this article wants to emphasize an important variable: the process of assessing foreign technology dependence. This is not an objective, straightforward process but rather a messy one that complicates state responses to globalization.

The full article is available open-access in Review of International Political Economy!

ChinAI Links (Four to Forward)

Below are four of the books that were especially influential in developing my thinking for this article:

• Ling Chen. (2018). Manipulating globalization: The influence of bureaucrats on business in China. Stanford University Press.

  • *Contains great details about how local governments pursue benchmarks and indicators that differ from narrow interpretations of indigenous innovation.

• Qiwen Lu. China’s Leap into the Information Age: Innovation and Organization in the Computer Industry. Oxford University Press, 2000.

  • *Lu’s work provided many of the historical details for the China (pre-hybridization of innovation) case.

• Aaron Friedberg. (1988). The weary titan: Britain and the experience of relative decline, 1895–1905. Princeton University Press.

  • *His analysis of British assessments of economic dependence in the late nineteenth century found that the process became dominated by a simplified indicator based on trade return data, which distorted policy debates.

• AnnaLee Saxenian. (2007). The new Argonauts: Regional advantage in a global economy. Harvard University Press.

  • *This text definitively maps out the brain circulation concept.

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).

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

the USMNT could go all the way!

…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: An AI that serves as a reliable “college preference form partner” is the one truly “worthy of trust”

Context: You may know about China’s national college entrance exam (Gaokao), which took place last week, but what comes after? This week’s feature Shuzi Lichang article (link to original Chinese) provides a window into the “college preference form” [志愿填报] process: after finding out their Gaokao scores, students have a few days to rank their top five colleges, alongside three to five majors for each college. It is considered the second-most stressful time for a kid, behind only the exam itself. Our entry point is Alibaba Qianwen’s launch of an AI advisor for the college preference form, which was made available for free to 12.9 million exam takers on June 10.

Key Takeaways: To understand why Chinese families might turn to an AI guidance counselor, let’s get a better sense of the process.

• Students and parents have to sort through an overwhelming amount of information, including school rankings, major specializations, enrollment quotas, employment prospects, and historical cutoff scores (The government sets cutoff scores that students have to score above to apply for elite colleges, but even if you’re above the cutoff, that doesn’t guarantee admission). As Jia and Li (2021) state, “Our findings suggest great room for improvement in [student-college matching quality] in this educational system.”

• When my parents went through the college admissions process, it was even more stressful. Under an older system, provinces required students to submit their schools before even knowing their Gaokao scores.

• Thus, there is a market for good advice on this potentially life-changing decision; unfortunately, families get taken advantage of. From the Shuzi Lichang piece: “Deceptive marketing tactics—such as claims of ‘insider admissions data,’ ‘exclusive institutional channels,’ or ‘one-on-one meetings with exam-setting professors’—are rampant.” The reporter recalls that a relative’s kid spent over 10,000 RMB on a college preference form consultant, who proceeded to just recommend the “hot” majors, completely ignoring the student’s preferences and wasting a hard-earned exam score.

• The stakes are high: if you are too conservative about your list of schools, then you could under-match and settle for a lower-tier college than your score would warrant. But if you are too aggressive, you could miss out on all your schools.

What does Qianwen’s AI advisor offer students?

• The below image depicts a report generated by the AI guidance counselor. Based on inputs of Gaokao scores and preferred majors, the model recommends six “high-potential” schools , eight “stable” schools, and sixteen “safety-net” schools. The report also includes a section about how AI could affect the preferred majors.

• Here’s how the article pitches the service: “It transforms college application guidance from a business based on information asymmetry into an inclusive public service, ensuring that having a ‘dedicated advisor’ is no longer a privilege reserved for the few.” This reminds me of Zebra English, the AI English teacher service that appealed to parents who had trouble finding or paying the high costs for in-person English tutors.

The question, going forward, for AI services in this market: how do you convince users to trust your AI with real, vital interests like the college preference form?

• As you’ll see in the full translation, I thought the last two sections were a bit too press-release-y, but the author seems to think Qianwen has developed this reputation of “getting things done” to earn the trust of the general public.

• But overdependence on Qianwen could raise issues. As one of the top WeChat comments on the article (26 likes) points out: “Aren’t you concerned that AI responses might become homogenized? If it recommends the same schools, or suggests the same preferences, to all students within the same academic tier, what would the ultimate outcome be?”

FULL TRANSLATION: An AI that serves as a reliable “college preference form partner” is the one truly “worthy of trust”

ChinAI Links (Four to Forward)

Should-read: Just Above the Exam Cutoff Score: Elite College Admission and Wages in China

Ruixue Jia, a professor at UCSD, co-authored this NBER working paper on the impact of scoring just above the cutoff to get admitted to an elite Chinese college on future wages. This paper provided a lot of the background information used above.

Should-read: Exporting U.S. Military AI Won’t Be Easy

In Foreign Policy, GovAI research scholar Jake Steckler studies how middle powers are making decisions about acquiring and developing military AI. One interesting detail: “France has pitched Mistral, in Europe and beyond, as a path to military AI that reduces dependence on the United States and China.”

Should-read: TP-Link’s American Dream

The Wire China is really starting to distinguish itself for in-depth reporting on China’s economy. By Noah Berman, this cover story on the router firm TP-Link weaves together so many threads, including geopolitics, globalization, American makeovers, and industrial policy.

Should-read: Life by Lu Yao, translated by Chloe Estep

This week’s feature article mentions the Chinese novel Life, which I had not heard of. Has anyone read it?

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).

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

it’s okay to sing a different tune [唱唱反调]

…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: My “Artificial Challenged Intelligence” Moment

Context: Back in June 2021, I wrote one of my favorite ChinAI posts on this phrase I had seen popping up in Chinese media: 人工智障, which I translate as “Artificial Challenged Intelligence.” It has the same first three characters [人工智] as the term for Artificial Intelligence, except switching out the last character turns the phrase’s meaning into one that’s used to make fun of AI failures.

This week’s feature translation comes from Renwu [人物], a Chinese magazine that covers human-interest stories. As a fan of their reporting, I’ve previously translated their longform investigation on the algorithmic pressures faced by Chinese delivery drivers as well as an engaging profile of why some families seek out AI tutors for their children.

Renwu recently asked their readers to share stories about their encounters with Artificial Challenged Intelligence. When you read these stories, these slices of real people’s lives, maybe you’ll find them interesting. Or mundane. Maybe you’ll even see them as a challenge — as singing a different tune [唱唱反调] — to this narrative that all Chinese people are super optimistic about AI all the time.

At the end of the day, all that matters is you find them human.

Key Passages: 铅球(Qian Qiu), Male, 46, Beijing

I’m a fan of Chinese crosstalk comedy [xiangsheng, 相声] and particularly love the duo of Guo Degang and Yu Qian. One day, on a whim, I asked (Tencent) Yuanbao, “How many people are there in Deyunshe? (a famous Chinese crosstalk comedy group)” It quickly replied with the number of crosstalk performers and trainees, and thoughtfully gave me a “popular science explainer” on the apprenticeship hierarchy—the four cohorts known as “Yun, He, Jiu, and Xiao.” It was a bit like the generational naming conventions found in traditional family trees, and it even listed representative figures from each cohort.

I took a look and saw that “Shaobing” (Zhu Yunfeng) was listed under the “He” tier, while Qin Xiaoxian had “skipped a grade” into the “Jiu” tier. That messed up the generational order, so I quickly corrected it. I hadn’t actually expected a reply, but the eager-to-please AI generated a new list.

In the new list, Meng Hetang’s name appeared under both the “Yun” and “Jiu” tiers—but notably not the “He” tier. Yuanbao even added a parenthetical explanation: Meng belongs to the “Jiu” tier because...

It was bullshitting endlessly with a straight face…

匿名(Anonymous)

DeepSeek told me, while I was on my period, to go to sleep early and that my top priority for the night was stopping the bleeding.

大红薯(Big Sweet Potato)

I once experienced bleeding after wiping my butt during a bathroom visit and asked an AI about it. Since then, whenever I ask it health-related questions, it always mentions this: “Given your history of anal bleeding...”

Mm, Female, 31, Deyang

I wanted to find out a bit about my new boss, but the AI ​​sent me a message claiming the new boss was under investigation by the Central Commission for Discipline Inspection. It was written so convincingly that I nearly fell out of bed in the middle of the night from the shock.

Anonymous, 13 years old

I’m a student in the second year of junior high. A while back, I encountered a math problem where two different methods yielded different answers; I asked my teacher and found out the problem itself was flawed. On a whim, I fed the erroneous problem into DeepSeek.

Twenty minutes later, it produced a result. It looked orderly enough on the surface, but when I expanded the “deep thinking” process, I witnessed the AI’s internal conflict and hesitation. It was like a child screaming silently: “This is impossible! I must have calculated it wrong!” Yet, none of this struggle was reflected in the final result, nor did it ever mention the possibility that the problem itself was incorrect.

More recently, my teacher mentioned the trend of using AI to write essays and said they had specifically used an AI to check student essays. This piqued my curiosity: Could AI really distinguish between human-written and AI-generated essays? I turned to (ByteDance’s) Doubao and created an AI agent designed to detect AI-generated text. Then, using another agent on the same platform, I generated three essays and asked the first agent if they were AI-generated. Every time, Doubao insisted they weren’t, offering plausible-sounding analyses about the authenticity of the emotions and the vividness of the language. However, whenever I told it it was wrong, it would immediately change its tune. I even tried DeepSeek, but the results were much the same.

Often, AI acts akin to a “You might also like” feature; it simply panders to your wishes and gives you the answer you want, without ever questioning the premise of your question.

临安 (Lin An), Male, 39, Wuhan

In my view, the most “challenged intelligence” aspect of today’s AI is precisely its “smarts.” For all its brilliance—having digested everything humanity has ever created or is currently creating, and possessing a perspective that could be described as “omniscient”—AI is paradoxically trapped by that very omniscience. No matter the question, it never falls silent or admits, “I don’t know.”

Time and again, when I use AI to look up information, it insists on providing an answer—even to a flawed or unanswerable question. If I ask for real-world examples to back up its conclusions, it will fabricate one if none exists. If I press for further details, it will keep cobbling together more information. Even when I explicitly instruct it via keywords not to make things up, not too long after, it still reverts to fabricating cases; it seems pathologically incapable of telling the truth. Consequently, I approach AI research with extreme caution, verifying every detail myself. On several occasions, I’ve been frustrated to find that correcting the AI ​​took longer than simply researching the information myself.

I once asked the AI, “Can’t you just say ‘I don’t know’?”

It replied: “That is determined by the model’s mechanism. My answers are based on probability; I learn which words frequently appear together, but I don’t truly ‘understand’ truth or falsehood. When faced with an unfamiliar question, I simply select the words that ‘look most correct’ and keep the narrative going.”

Kevin Tsai (host of popular I can I BB debate competition show) once remarked: “In all my conversations with AI, the one thing it has never given me is silence, is space for reflection.”

Yet, therein lies the most terrifying aspect. As humanity’s reliance on AI deepens, information born of “AI hallucinations” can—if enough people believe it—morph into a kind of “fact.” It is the old adage: “A lie told a thousand times becomes the truth.”

…

We talk about “Artificial Challenged Intelligence” today because, despite its appearance of omnipotence, AI still frequently makes a fool of itself. I suspect that the day AI ceases to be “challenged intelligence” and instead becomes a symbol of omniscience and omnipotence—that is when the real disaster will begin.

Many other stories included in PARTIAL TRANSLATION: My “Artificial Challenged Intelligence” Moment

ChinAI Links (Four to Forward)

Must-read: China is getting worried about AI & jobs

I’m late to dig into this post by Matt Sheehan on growing Chinese worries about AI labor displacement. It is chockfull of insight, especially the section on the Wuhan Robotaxi Saga:

In late June of 2024, a Wuhan taxi company released a public letter decrying the shrinking margins for companies and declining income for taxi drivers. The letter actually spends more time on the way ride-hailing apps have hurt taxi drivers, but it was the parts about robotaxis that stirred up online debate.

…the online conversation about robotaxis “stealing people’s rice bowls” continued to gain steam, becoming a top trending topic on Chinese social media.

Here’s where I heard something a bit different. The first time someone relayed this incident to me, they said taxi drivers had gotten together and led a “strike” against the robotaxis by repeatedly hailing them and then cancelling at the last minute, effectively paralyzing the system…

…the outcry over robotaxis in Wuhan led to a significant shift in how the government and policy community thought about AI’s impact on jobs. It led officials to take the threat of job displacement—and also the public’s fear of AI-driven job losses—seriously. I feel relatively confident about the Wuhan robotaxi incident contributing significantly to this shift, because multiple influential members of China’s AI policy community independently relayed that causal relationship to me at different times.

Should-read: How China is breaking apart a people and its culture

In the Financial Times, drawing on a satellite imagery, witness testimony, and local reporting, Alison Killing finds that “authorities continue to rely on widespread detention, short-term arrests and intimidation, even as the campaign enters a new stage centred on forced cultural assimilation, particularly in the heavily Uyghur areas of southern Xinjiang.”

Should-read: Two Chinese-language articles I considered featuring this week

1. How did Hangzhou become more hardware-oriented? A Huxiu piece tries to unpack why Hangzhou, known as a hub for AI software startups, has pivoted to growing a crop of startups focused on AI hardware (especially inference chips).

2. Today’s college students are living their university years like their final year of high school. This NetEase DataBlog article reports on why the relentless grind/rat race of preparing for college entrance exams doesn’t end when Chinese students get to college. Now, it’s all about squeezing out every tenth of a GPA point.

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).

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

welcome new readers: most issues are note-taking exercises on Chinese-language articles that interest me (not lengthy missives)

…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: Behind DeepSeek V4’s “Chip-Model Synergy”, the domestic computing ecosystem begins flywheel acceleration

Context: DeepSeek’s V4 made waves for something beyond pure capabilities: the validation of “chip-model synergy” between V4 and Huawei’s Ascend chips on a large scale. Previously, this was something that could only be achieved with Nvidia chips, in part because of the CUDA moat. As Lily Ottinger and Mary Clare McMahon described in a ChinaTalk article:

Nvidia’s position rests on what Doug O’Laughlin has called a “three-headed hydra” of leading hardware, networking capabilities, and, most importantly for this piece, a deeply entrenched software ecosystem. At the center of that ecosystem is CUDA, a proprietary programming framework that allows developers to efficiently map computations onto Nvidia’s GPUs. CUDA’s value lies not only in its performance but in its reach: an expansive set of libraries, optimized workflows, and tight integration with widely-used machine learning frameworks make it the industry standard. And, crucially, CUDA can only be used with Nvidia GPUs. That makes CUDA a core component of Nvidia’s competitive advantage, otherwise known as Nvidia’s moat.

In this week’s feature translation, QBitAI provides a readout (link to original Chinese) on Huawei’s alternative to CUDA, the Compute Architecture for Neural Networks (CANN), based on panels at the Kunpeng Ascend Developer Conference.

Key Takeaways: CANN has transitioned from its “infancy” into a “youth” phase, according to Qiuwu Chen, founder of Shanghai-based AI coding start-up AIGCode.

• From the article: “This ‘youth phase’ does not imply that the ecosystem has reached full maturity; rather, it signifies that developers are beginning to move beyond ‘vendor-provided babysitting support.’ Developers are now capable of independently resolving issues, contributing code, and driving the ecosystem's iterative progress.

• Chen recounted his company’s experience with Ascend over the past year or so. They used Ascend out of necessity, due to severe GPU shortages at the beginning of 2024. In early days, when they ran into issues when pre-training of a 7B-scale Mixture-of-Experts model on Ascend chips, they had to submit a support ticket to Huawei, which resulted in a four month wait. The entire ecosystem was “essentially a section of desert,” Chen recalls.

The three dimensions that dictate whether Huawei chips can overcome the Nvidia CUDA moat: adaptation efficiency, performance ceilings, and production-grade reliability.

• Regarding adaptation efficiency, as the previously cited Ottinger and McMahon article emphasizes, “adopting models to run on Huawei’s platform is also onerous.” If you have to depend on Huawei engineering support to fix issues, then it will take a long time (as the early experience of Chen’s AIGCode reveals). At the developer conference, a University of Science and Technology of China team claimed that it took less than a week for them to migrate a high-performance computing solver to Huawei’s Kunpeng CPU chips.

• On performance ceilings, one key indicator is Model Flops Utilization (MFU), or compute utilization rate. How much compute can you squeeze out of the same chip? The article reports, “On the Ascend accelerator, AIGCode achieved a Model Flops Utilization of 65% during the pre-training of its MoE model. This figure is nearly double the industry average.” This dimension will also be shaped by the efficiency of scheduling and coordination in massive clusters (for example, a supernode of 384 Ascend chips and 192 Kunpeng chips).

• Lastly, the example of production-grade trust comes from a leading joint-stock commercial bank (think: an institution like China Construction Bank). At the developer event, this bank stated that they had directly integrated AI into its core risk management processes. The article implies (though it is not explicitly stated) that the open-sourcing of the CANN framework shaped this financial institution’s decision to entrust a core business operation to AI models that run on Huawei chips.

Could an open-source developer ecosystem drive CANN’s development going forward? This is one of the most interesting questions.

• Huawei open-sourced the core code for CANN, including runtime environments and compilers, last December. The article lists a few indicators that may point to a sustainable ecosystem that drives flywheel growth: Over 70 major AI models are now supported immediately upon their release, the Kunpeng and Ascend developer communities both surpass 4 million members, and groups like AIGCode and the aforementioned leading bank are contributing features to various communities. *Notably, the bank contributed 34 optimizations to vLLM-Ascend, a project to adapt one of the most popular inference engines for Ascend NPU chips.

• Back in May 2025, Ottinger and McMahon found that “Huawei’s Ascend developer portals — both in English and Chinese — exhibit low engagement, with sporadic posts and limited public debugging activity.” To be honest, it was difficult for me to figure out whether this has meaningfully changed. On the surface level, the Ascend forums seem to have more activity. I guess I’m just used to an open-source framework being posted on a Github repository, and then most of the issues being sorted out there, instead of a separate community forum website. And on my quick scan, it seemed like CANN was posted on Gitee (China’s domestic alternative to GitHub) and then switched to Huawei’s own site.

• In any case, the key trend to watch is whether CANN can actually benefit from an organic community of implementers and complementary innovators that contribute back into the framework.

FULL TRANSLATION: Behind DeepSeek V4’s “Chip-Model Synergy”, the domestic computing ecosystem begins flywheel acceleration

ChinAI Links (Four to Forward)

Must-read: Can Huawei Take On Nvidia’s CUDA?

Written for ChinaTalk in May 2025, Lily Ottinger and Mary Clare McMahon’s in-depth analysis of CUDA vs. CANN was essential background reading for this post. Definitely read it if you are interested in this topic.

Should-read: China Is Testing Its State Surveillance Model Abroad

In NYT, David Pierson and Berry Wang report on Beijing’s efforts to export its security model to the Solomon Islands, which signed a security pact with China in 2022. The pilot was suspended after local backlash. H/t to Kirsty Needham for sharing and who reported on this in a 2024 story.

Should-read: Simon Willison thread on thinness of AI-failure stories

Good discussion and back-and-forth on this thread about whether some AI-failure stories are overblown. The context behind this is a flurry of reporting about companies saying that they’ve spent too much on AI.

Should-attend: June 2026 Wonky China

If you’re in the area, I’ll be joining the Wonky China team for a happy hour at Chang Chang on June 10th. Come for the hot takes, stay for the scallion bubble pancakes.

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).

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

church is in session

…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).

Anthropic’s Dogma on US-China AI Competition

I’ve been going to church since I was a kid. Early on at the University of Iowa, when I identified as an evangelical Christian, I went on mission trips overseas and led prayer meetings. Over the years, my faith has morphed and faded and deepened and changed. Today, as fellow “exvangelicals”, my wife and I still attend a Lutheran church almost every week, and we go to a young adult Bible study on Wednesday nights.

All this to say that I am relatively sensitive to dogmatism: an unfounded confidence in asserted opinions not based in fact. As I read more and more of Anthropic’s policy papers and CEO Dario Amodei’s blog posts, the elements of religious dogma abound: infallible faith in a set of unquestioned beliefs; justification of doctrine based on appeals to exclusive access to divine wisdom; an unwillingness to consider contradictory evidence. Sound familiar?

On May 14, Anthropic laid out its beliefs about U.S.-China competition in a new paper titled “2028: Two scenarios for global AI leadership.” This paper envisions two scenarios for the world in 2028: in the first, the U.S. has a commanding lead in AI; in the second, China’s AI ecosystem remains neck-and-neck, which results in “authoritarian AI leadership”. Their policy prescription is to intensify efforts to prevent Chinese labs from accessing export-controlled chips. Taking on the position of a religious skeptic, let us question seven of this post’s unfounded assumptions.

Assumption #1: Because transformative AI arrives in two years, the actions that decide the US-China competition will only occur in this limited window.

In January 2025, Dario posted a version of this “two scenarios” essay, in which he warned that transformative AI — “AI that is smarter than almost all humans at almost all things” — would arrive in 2026-2027. What explains the updated timeline? Why did we go from 2026 to 2028?

Looking back at my rebuttal to Dario’s post, it’s funny (and sad) how neatly this point still applies to the recent Anthropic paper: “It’s really important to note Dario’s assumed one-to-two year timeline. I’ve been in governance of AI circles since 2017. In that time, I’ve consistently heard some iteration of: AGI is two years away! It’s the Bruno Caboclo of technologies: always two years away from being two years away.”

I can’t wait until next year’s paper when Anthropic tells us that we only have a limited window of time before 2030, when the company will open up a box, and transformative AI will transform everything.

Assumption #2: Transformative AI capabilities will diffuse almost immediately.

In line with this two-year window, the Two Scenarios paper argues that frontier AI could quickly reshape the military balance in China’s favor: “These capabilities will not diffuse slowly. When a new model reaches a new capability in autonomous targeting, vulnerability discovery, or swarm coordination, for example, the regime that controls it can put it onto the field in weeks, not years.”1

History teaches us that this is not the timeframe by which general-purpose technologies like AI diffuse across economic and military systems. My book Technology and the Rise of Great Powers verified these protracted timelines across more than ten different general-purpose technologies. Based on research into how electricity affected the military balance of power, I also found that electrification transformed militaries in a gradual, decades-long process, in which the main impacts materialized not through the eye-catching weapons systems (e.g., the drone swarms that attract Anthropic’s attention) but instead in communication, transportation, and logistics applications.

To be sure, AI could be different from past GPTs, especially if recursive self-improvement accelerates capabilities development and diffusion. Well, consider OpenAI’s fascinating reflections on “harness engineering,” based on their effort to built a software product with “0 lines of manually-written code.” What resonated with me about this early taste of the challenges associated with “AI building AI” is the organizational adjustments that OpenAI had to make. For a time, they spent every Friday cleaning up “AI slop.” They also needed to readjust their software engineering team’s primary job toward designing environments, specifying incentives for AI agents, and building feedback loops to monitor those agents. This was an experiment with a small team of three engineers. Imagine the challenge of diffusing these organizational adjustments across the entire economy or military. This will take time.

Before moving on to the next unfounded premises, I want to underscore why it’s important to question these assumptions. This is not a navel-gazing exercise. If the timeline of when transformative AI systems actually make an impact on national power is not 2026 (or 2028, or 2029) but, say, ten years from now or later (which is more in line with the diffusion marathon of past GPTs), then the basis of policy debates changes. In a past critique of the U.S.’s export control policy, I’ve made this short-term versus long-term distinction:

We haven’t even mentioned other downsides such as the risk that the controls accelerate a) China’s development of an independent high-end chip capacity and b) the reorientation of chip supply chains in a way that insulates them from future controls. If you are operating in a world where AI does not make its mark until after at least a decade, these downsides become more salient and the benefits of the controls become more muted.

Zooming out, the fast timeline assumption fits with Anthropic’s rhetorical strategy — “we have only a limited period of time to set the conditions of the competition,” asserts the paper — to not quite manufacture a state of emergency but strategically deploy one for its own purposes. It is an approach reminiscent of some evangelicals: claiming that the end times are near, some Christians justify war (including the current military operations in Iran) because it fulfills an end-times prophecy to trigger Christ’s return. Note: Coincidentally, at church a few weeks ago, our congregation studied Acts 1:7. When apostles ask Jesus when he will restore the kingdom to Israel, he responds, “It is not for you to know the times or periods that the Father has set by his own authority.”

Assumption #3: If China’s AI models are close to US models in capabilities, China will win on adopting these capabilities across economic, military, and technological domains.

In Dario’s 2025 blog post, he makes a similar claim: “Even if the US and China were at parity in AI systems, it seems likely that China could direct more talent, capital, and focus to military applications of the technology. Combined with its large industrial base and military-strategic advantages, this could help China take a commanding lead on the global stage, not just for AI but for everything.”

Again, my book and past research on China’s “diffusion deficit” thoroughly debunks this assertion. Take, cloud computing as an example. China has strong cloud computing giants like Alibaba Cloud, but its overall cloud adoption rate, across all businesses, is half the U.S.’s rate. As for military deployment, the U.S. military likely leads the world in leveraging private technology firms to support cloud computing adoption. In a forthcoming paper in Security Studies, I take aim at this assumption that China’s civil-military integration outpaces the United States: “When applied to China’s ‘military civil-fusion efforts,’ the four-channel measure reveals that China lags significantly behind the U.S. in terms of the efficient use of common technologies, facilities, and personnel for military and industrial purposes.”

Assumption #4: There are only two scenarios. Support Anthropic’s leadership to secure democratic AI, or authoritarian AI will rule the day. I mean, come on, it’s almost too easy to make the connection to the false dichotomies of religious fanatics. We are good, everyone else is evil.

Look, I think we should be concerned about AI’s impact on surveillance. I’ve consistently covered this issue in these pages (from ChinAI #29 to #357). But it’s a big leap from China keeping pace in AI to the entrenchment of authoritarian repression around the world. There’s some dangerous technological determinism in this thinking, which likely repeats the same mistakes of those who thought that the Internet would bring about widespread democratization — just in the other direction. As Internet governance scholars Christopher S. Yoo and Alex Mueller point out:

The premise that China threatens to increase international acceptance of authoritarianism by exporting related values through Internet infrastructure warrants greater skepticism. It appears to reflect the same type of soft-technological determinism as the United States’ early “Internet Freedom” agenda that saw the Internet an unstoppable vehicle for democracy. Just as the existing Internet failed to liberate China, Russia, and Iran, a cyber-sovereign Internet should be no more likely to increase the level of digital repression among governments that have not already embraced authoritarianism.

Core to the “two scenarios” vision is the assumption that there is a full-stack American AI and a full-stack Chinese AI. In fact, in their policy proposal to “champion the export of American AI,” Anthropic is trumpeting a plank of the Trump administration’s “America’s AI Action Plan.”

If, however, you make some effort to comprehend how AI is developing at the ground level, the notion of an all-American or all-Chinese full-stack AI system becomes fanciful at best and purposefully ignorant at worst. Top U.S. AI startup Perplexity, for instance, permits subscribers to use Chinese startup Moonshot AI’s Kimi K2.5 model; previously, Perplexity also supported a version of the DeepSeek-R1 that had been post-trained to remove censorship. I could keep going with examples. Read more about SGLang and inference engines, if you’re interested in the complex web of interdependencies that characterize the AI stack.

Here’s the thing. When talking to policymakers and pundits who aren’t steeped in this field, Anthropic can get away with these flimsy assumptions. When it tells them that the end is nigh, and we need democratic AGI to save the world, it is all too easy to just blindly trust the company’s divine technical wisdom. However, if you’ve done your homework (in other words, the equivalent of reading Acts 1:7 or any number of Bible passages that underscore no one knows the time and hour), you can see right through it.

Given space constraints, I’ve included rejoinders to the following three assumptions in the footnotes:

• Assumption #5: If a PRC AL lab had developed a model like Mythos, the CCP would have had first access to a system that could autonomously discover and penetrate critical American infrastructure.2

• Assumption #6: Chinese AI labs depend on large-scale distillation attacks that illicitly extract the innovations of American companies.3

• Assumption #7: Neck-and-neck competition risks disincentivizing responsible AI.4

The broader purpose here is not to debate every single point line-by-line but rather to show the lack of reflexive thinking in this recent Anthropic paper.5 Good research and good argumentation — and good faith, for that matter — must exist with a healthy amount of doubt and uncertainty. I’ve just finished grading 47 research papers, and the number one piece of feedback I give to my students is consider how an opposing perspective might answer their research questions and then either cede some ground back or engage with hypothetical counter-arguments. When I read good writing about AI policy, the authors are open to updated assumptions and disconfirmatory evidence. I don’t see these qualities in Anthropic’s paper. It’s dogma all the way down.

This is a peculiar moment. When one of the world’s most valuable private companies publishes a geopolitical thesis with so many holes, I would have expected it to generate more productive debate in these past two weeks. There are many smart people who have a better mix of expertise and insights than me to give critical takes on this paper.

It’s just that many of them are now working for Anthropic (including several former Biden administration officials who shaped AI policy) or hope to be hired by Anthropic in the future.

The Church of Anthropic is growing, and they pay handsomely. Call me a nonbeliever.

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).

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

it’s time to watch Kung Fu Hustle for the first time

…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: Understanding China through DeepSeek’s Triple-Jump Valuation

Context: Thanks everyone for voting on last week’s roundup issue. The winning selection is a Shuzi Lichang [数字力场] analysis of why DeepSeek’s valuation jumped three times within the span of under 20 days to over $50 billion. Zongming She1, a former deputy editor of The Beijing News commentary section and prominent media commentator, presents a fascinating case for DeepSeek as undertaking a “Huawei-like” mission in the AI domain.

Key Takeaways: Zongming’s thesis is that DeepSeek shoulders a burden akin to Huawei’s in telecommunications: facing external restrictions, it must invest in domestic capabilities and break through chokeholds. Zongming puts it simply: “Investing in DeepSeek = Betting on the Nation’s Destiny.”

• The article describes DeepSeek V4’s optimization for deployment as Huawei’s Ascend chips as a “historic breakthrough in ‘chip-model partnership’ between domestic large language models and domestic computing power.”

• The commentary posits, “Consequently, the ‘buffs’ that previously belonged to Huawei have now been bestowed upon DeepSeek.”

• Jeff’s note: I think it’s important to not overstate the extent of indigenization here. As I noted in ChinAI #356: “In contrast to some hasty analysis, it is very likely that DeepSeek is still dependent on Nvidia chips for training; however, some of V4’s toolchain advances suggest gradual progress toward domestic substitution, especially on the inference side.”

Where does DeepSeek stand in the global landscape of frontier model builders?

• It’s neither a super-platform like OpenAI nor a commercial giant like Anthropic which has seen its annual recurring revenue soar to $44 billion. Additionally, DeepSeek’s commercialization pales in comparison to Chinese rivals. From the article: “Looking for a ChatGPT-style, viral AI-to-Consumer (AI2C) hit product? DeepSeek would have to line up behind (ByteDance’s) Doubao.”

• So why are people buying DeepSeek? According to Zongming: “From a purely technical standpoint, one is betting on its potential to become the ‘Android’ of the AI ​​era.” The article points out that DeepSeek has “shattered the ‘Impossible Triangle’ of high performance, low cost, and open-source accessibility.” Though DeepSeek has probably not generated much revenue, Zongming calls back to the early days of Linux, which also did not optimize for short-term profits.

The commentary regards DeepSeek as a standard-bearer or “representative” for China’s AI sector as a whole.

• Zongming sees a technical roadmap that breaks free from foreign technology dependencies as “the single greatest pillar of valuation support.” Interestingly, he interprets DeepSeek’s jump in valuation from $40 billion to $50 billion as “a premium for strategic scarcity.” From the piece, the foundation that supports DeepSeek’s valuation is a strategic imperative: “China’s need for a full-stack, independent AI.”

• He also attributes the jump to the China Integrated Circuit Industry Investment Fund’s (AKA the “Big Fund”) interest in leading the investment into DeepSeek. To date, The Big Fund has focused investments in semiconductor equipment and production, and, per Financial Times reporting, it has not “publicly backed any of China’s other LLM players.”

• The article concludes:

Ever since DeepSeek-R1 emerged last January as a “product of national strategic significance” [国运级产品], DeepSeek was no longer the DeepSeek of old.

Since that moment, DeepSeek’s destiny and fortune have become a small opening through which to understand China

FULL TRANSLATION: Understanding China through DeepSeek’s Triple-Jump Valuation

ChinAI Links (Four to Forward)

Must-read: The US Is Running Dangerously Low on China Expertise

In The Diplomat, Mercy Kuo interviewed Rosie Levine on the US-China Education Trust’s latest report “America’s China Talent Challenge.” This is a sad statistic from the report: this year, an estimated 2,000 Americans are studying in China, which is a sharp drop from 11,000 U.S. students studying in China in 2019.

Should-read: Interpret: China (A CSIS Open Source Project)

I wanted to shout-out this CSIS project that translates Chinese articles, speeches, policy documents, and other primary source materials. For some projects, they bring together experts on a particular topic and have them analyze a collection of translated documents.

Two ways to think about Xi-Trump Talks on AI

I gave some comments that contextualize how U.S.-China cooperation in AI could take shape.

1. Should-read: Why AI is the Hidden Minefield of Trump’s China Visit

For The New York Times, Vivian Wang set the table for the Trump-Xi talks on managing AI risks. Here, my view was that the goal for these talks was to send a high-level signal that allows the real work to happen at the informal, nongovernmental levels. This builds on research into how China made safety gains in other high-risk technologies, which highlighted the role of international industry associations (see: Reputation collectives: how international industry associations influence China’s safety standards in high-risk technologies).

2. Should-read: AI creates a fearsome cold-war-style dilemma

The Economist tackled why it is difficult for states to cooperate on safety for powerful technologies. One dilemma is that sometimes sharing safety information also improves the capabilities of rivals. This was less of a problem in the Cold War because safety technologies (e.g., permissive action links) were essentially separate add-ons to weapons systems (nowadays, they are more integrated), which means you could share safety information without divulging too much about capabilities.

Still, as I argued in “Keep your enemies safer: technical cooperation and transferring nuclear safety and security technologies”, there was still substantial amounts of lab-to-lab cooperation on nuclear safety and security issues even amidst fierce geopolitical rivalry. Joint verification experiments served as a fruitful channel.

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).

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

summer beckons

…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).

Around the Horn (25th episode)

I miss Woody Paige’s blackboard: let’s go back Around the Horn!

This is our 25th roundup of articles on China’s AI ecosystem. For new readers, here’s how it works (see ChinAI #350 for the previous edition):

• I give short previews of ten articles that caught my eye during a scan through my usual sources (all published within the past week or so). The title for each preview links to the original article in Chinese.

• Readers vote on next week’s feature translation by replying to the email and/or commenting on the post with the number of your preferred article. *I’ll give some added weight to votes from those of you who support ChinAI through a paid subscription.

• The main idea is that any of these 10 links would have made for a great feature translation this week — like the movie soundtrack for Boyhood, there are no skips!

1) 10,000-character deep dive | ModelBest: at the limits of efficiency, continuously traversing the AGI cycles

Summary: I’ve always been curious about the Chinese AI startup ModelBest [面壁智能]. This deep dive explores some of the themes covered in ChinAI #296 and provides some updates on ModelBest’s distinct technical path.

Source: Hushuo Chengli [胡说成理] - platform founded by Hu Zhe [胡喆], an author of books on Chinese internet history and former deputy editor-in-chief of Leiphone.

2) Now it’s your turn to foot the bill for (ByteDance’s) Doubao

On May 4, the AI super-app Doubao quietly added three tiers of paid subscriptions on top of its free version, which triggered an internet firestorm. One interesting nugget from this article: among the its 345 million monthly active users, the vast majority are either students or middle-age and older who mainly use the app for casual chatting and information retrieval (which may not match the demands of the new paid productivity features).

Source: 虎嗅 (Huxiu) — well-known platform that shares user-generated content but also publishes their own pieces on China’s science and technology ecosystem.

3) Young AI professionals face “debt collection” from former employers

Summary: To prevent AI talent from being poached, Chinese companies have taken extreme measures with their non-compete agreements. This article profiles many young professionals who signed non-competes without reading the fine print and now find themselves facing legal claims with millions in damages and psychological distress.

Source: 人物 (renwu) — magazine that covers human-interest stories. A previous issue translated their longform investigation of the algorithmic pressures faced by Chinese delivery drivers.

4) Release of Results for the AI Safety Benchmark (Q1 2026)

Summary: This is one of the Chinese AI safety benchmarks that I’ve been following closely. CAICT released its first batch of 2026 results, including some tests of “loss-of-control” behavior.

Source: CAICT AI安全治理 (CAICT AI Safety/Security Governance) — the government-affiliated think tank’s portal dedicated AI safety/security issues

5) AI PPT, this time, you really don’t need revisions

Summary: For a long time, industry circles viewed AI-generated PPT as a fake or overestimated demand, given the need for extensive revisions. This QbitAI product trial tests that thesis with iFlytek Zhiwen’s AI-powered PPT tool Vision Agent.

Source: 量子位 (QbitAI) — news portal that regularly covers AI issues, similar to Leiphone and xinzhiyuan. Lately, QbitAI has been publishing longer reports more frequently.

6) “Peeling lobsters online, eating lobsters offline” ignites Wuhan

Summary: This article covers Wuhan’s initiative to drive the adoption of AI agents (note: lobster is a reference to OpenClaw). In the past, I’ve noted a trend in OpenClaw’s China community, in which Tencent and other big companies set up offline “booths” to help people deploy open-source agents.

Source: China Software Developer Network (CSDN) — the largest software developer community in China, which provides IT news coverage and hosts code

7) Chinese Universities Discontinue Over 5,000 Majors in Five Years

Summary: Remember when there was so much hype about China establishing an AI major? This piece explores the trend in Chinese universities suspending various majors, including student complaints about a previously-hyped e-commerce major.

Recommended by Source: 知识分子 (The Intellectual) — a platform that covers the state of science in China, founded by Chinese and Chinese-American scientists.

8) Understanding China through DeepSeek’s Triple-Jump Valuation

Summary: DeepSeek’s valuation jumped three times in under 20 days; from $10 billion in mid-April to $20 billion in late April, then to $51.5 billion this past week. Yu theorizes why DeepSeek’s valuation jumped, and what that says about China’s AI ecosystem.

Source: Shuzi Lichang (数字力场) — Started following this account back in 2021, but I haven’t browsed them that closely. This article is by Zongming She, a former deputy editor of the commentary section in The Beijing News.

9: Leiphone Detectives Vol. 7 [AI情报局]

Summary: Leiphone has a new series that collects interesting industry rumors. This week’s edition includes: 1) complaints from an employee at a major internet firm about their department burning the budget on token costs in order to deliver real-time, interactive PPTs; 2) a major company’s autonomous driving division experienced a disastrous failure during early stages of mass production, with blame going toward the division’s head who had no experience prior to appointment.

Source: Leiphone Detectives series [AI情报局] - for a translation of one of their previous issues on industry rumors, see ChinAI #354.

10: People’s Daily Commentary: What Signal Does the Halting of the Manus Acquisition Send?

Summary: This piece gives you a sense of how official platforms are trying to spin China’s blocking of the Manus acquisition. Interestingly, it holds up Zhipu and MiniMax as examples of continued openness to foreign investment.

Source: People’s Daily Commentary [人民锐评] - online commentary platform of official CCP newspaper.

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).

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

one must weigh every word with care [字斟句酌]

…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: Under the Camera’s Gaze, I Changed from a University Lecturer into a “Performer”

Context: Xiaoxi1 is a lecturer at a second-tier university in northeastern China (think: a school like Heilongjiang Institute of Technology). Beginning around March 2024, her school and others in the province started to introduce AI surveillance systems. Over 90% of classrooms in Xiaoxi’s school have implemented the system. In the typical set-up, one camera is positioned in the front of the classroom, while another watches from the back of the room.

The cameras record students’ head-up rates (as an indicator of attentiveness), the number of students seated in the front row, interaction patterns with the teacher, their facial expressions, as well as the teacher’s verbal tics, gestures, and whether their speech content triggers any “sensitive keywords.” In some cases, universities even mount a screen next to the classroom blackboard that displays these metrics as they update in real-time.

In this week’s feature translation of a Perpetual Light Studio [极昼工作室] longform article, reported by Xiaohan Wei and Sun Ran, we hear from students and teachers like Xiaoxi about the effect of AI surveillance in Chinese university classrooms.

Key Takeaways: University professors feel alienated by AI surveillance systems, turning from “instructors” into “performers.” The system monitors whether lecture content touches upon sensitive topics, which has prevented Xiaoxi, who teaches ideological and political education classes, from speaking with ease. One student’s Japanese-language teacher was reprimanded for sitting down during class, after the system flagged this behavior.

• The article quotes Xiaoxi, “I feel as though I’ve been alienated by AI. I originally aspired for a truly brilliant classroom environment; yet, suddenly finding myself under constant surveillance, I seem to have shifted toward pursuing the avoidance of errors and compliance. Under surveillance, the very essence of education has become somewhat distorted.”

• Many teachers question the inflated standards demanded by the metrics collected by the surveillance system. Another junior lecturer, Xiaoxie, was called out for a low “head-up rate” among his students, in a digital media course at a second-tier university. Xiaoxie eventually left academia. From the article, “He had no desire to exercise excessive control over his students; instead, he preferred to grant them a greater degree of autonomy. As a teacher, he feels fulfilled as long as there are one or two students in the class willing to listen to him and ask him questions.”

  • Xiaoxi shared a similar sentiment, which doubles as my favorite passage from the article: “AI posits a perfect classroom state…every student looking up at the teacher, responding instantly to every question. No human teacher could ever achieve such a feat. Xiaoxi believes that classroom instruction is, by its very nature, an ‘exercise in teaching with sympathy.’ As she puts it: ‘Confucius had three thousand disciples, yet only seventy-two were deemed ‘worthy sages’ [贤人]; thus, our head-up rate might also not be that high.’”

Is this sustainable integration of AI, or just universities desperately trying to catch onto the trend of the times? Xiaoxi noted that the university held a large-scale assembly prior to installing the surveillance system, where the tone seemed to say, “We (the school) possess such immense financial resources that we are able to install such a sophisticated system.” Indeed, funding for equipment upgrades typically arrives quickly if the school frames the request as “AI empowerment.”

• By adopting these systems, universities can boast that they are implementing government initiatives, including the Ministry of Education’s “Action Plan for AI Innovation in Higher Education Institutions (2018).” This April, the Ministry of Education also co-published an “Action Plan for ‘AI + Education’”, which encourages the use of intelligent technologies to improve classroom teaching behaviors.

• From the Perpetual Light Studio article: “‘How does a university demonstrate that it is keeping pace with the tide of AI-driven intelligence? Maybe, AI surveillance is a good way,’ one teacher observed. Her university rushed to launch its AI surveillance system just before a scheduled undergraduate teaching assessment; she felt it was largely a symbolic gesture—‘intended to show that the school truly takes teaching seriously and is really cracking down.’”

Teachers and students have found ways to fight back against the system. According to the article, “The AI ​​surveillance systems implemented in universities have also sparked discussion online, drawing considerable opposition.”

• Xiaosong is a third-year student at a top university in Shanghai, where there is generally more tolerance for professors to go off-script, though he still feels anxious on their behalf whenever he hears them saying something “sensitive” in class.

• The article continues, “During a recent class, one professor, who wanted to say something, pointed a finger at the surveillance camera and declared that he was taking a risk to talk about this with them. He ultimately proceeded with his remarks, even letting slip a couple profanities. The image of the professor pointing at the camera struck Xiaosong as somewhat amusing, “a feeling like dark humor.” These subtle shifts occur predominantly in humanities classes; Xiaosong muses that perhaps some professors, despite knowing they are under surveillance, nevertheless choose not to alter their course content. This counts, in a way, as a small act of resistance against the system.”

• Other students do not have a strong view on classroom surveillance. Wei and Ran report, “Noting that the general classroom atmosphere in universities is often lackluster, one student speculated that the initial intent behind introducing AI assistance was likely well-meaning—even if, as he put it, ‘in practice, it often fails to be effectively implemented.’”

• As always, students find ways to adapt: “Previously, everyone would scramble for seats in the back rows to make it easier to attend to their own personal matters; now, the hottest seats are the middle rows, which are the furthest away from the cameras.” The article concludes with a pithy image: “As some students noted, most people simply prop up their tablets vertically to block the front camera, while using their own backs to shield themselves from the rear camera.”

FULL TRANSLATION: Under the Camera’s Gaze, I Changed from a University Lecturer into a “Performer”

ChinAI Links (Four to Forward)

Should-read: The Tech High Ground

In the most recent issue of Foreign Affairs, former U.S. National Security Adviser Jake Sullivan argues that the U.S should reframe AI competition as a decades-long project rather than a sprint to innovation breakthroughs. The essay also references my work on diffusion and total factor productivity!

Should-read: DeepSeek: Blueprint, Not Breach (paywalled)

I wanted to learn more about TileLang, which I covered in last week’s issue about DeepSeek V4. This Hello China Tech post, published back in October 2025, was the most helpful: “Today’s column decodes DeepSeek’s latest move: a 50 per cent price cut that hides the real story in a technical footnote. The launch of TileLang, a Python-like programming language, marks phase two of building a China-controlled AI stack. Same-day support from Huawei, Cambricon, and Hygon signals coordinated standard-setting across hardware and software. But coordination isn’t conquest–Nvidia’s CUDA moat remains deep.”

Should-read: China AI Bulletin 3

Emmie Hine continues to produce a very detailed roundup of “the latest on AI development, governance, and safety in China.” This issue’s highlights: China’s Human-Like AI Interim Measures, DeepSeek and Moonshot’s latest models, and China’s blocking of Meta’s $2B Manus acquisition.

Should-read: An Independent Safety Evaluation of Kimi K2.5

A group of researchers, led by Zheng-xin Yong and Parv Mahajan from Constellation (a non-profit working on safety of transformative AI) conducted a safety assessment of Moonshot AI’s Kimi K2.5 model. Jack Clark covered this well in his ImportAI newsletter. I liked how this paper emphasized developments in the open-source safety evaluation ecosystem:

“Frameworks such as Inspect AI and Petri, and benchmarks including ControlArena and FORTRESS are publicly available and can be configured in days. Given that we can conduct diverse safety evaluations within a short timeframe (in under a month) and with a limited financial budget, we believe that the lack of safety reports is due to organizational prioritization, not technical difficulty.”

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).

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

Jeopardy is still appointment viewing during this streak

…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 Translations: 1) DeepSeek Overly Understated; 2) Don’t Overestimate Nvidia; Don’t Underestimate DeepSeek

Context: On April 24, DeepSeek released its long-awaited V4 model (including a 1.6 trillion-parameter Pro version). Let’s break down this release’s technical intricacies plus the broader implications for China’s AI ecosystem, with the help of two Huxiu translations.

Key Takeaways: DeepSeek’s V4 demonstrates near-frontier capabilities, but the main breakthroughs are in compute efficiency. Based on the technical report, DeepSeek V4-Pro “requires only 27% of single-token inference FLOPs and 10% of KV cache compared with DeepSeek-V3.2.” In her “DeepSeek Overly Understated” piece, Sihang Song deems DeepSeek a road builder [修路人].

• The top models can support 1 million token context windows, which means they can process more information at once, akin to 50,000 lines of code. The V4 supports this long context window at a fraction of the computational costs. Interestingly, Song notes that DeepSeek surpasses other highly capable, cost-effective Chinese models on this metric: Compared to DeepSeek, “when confronted with the same complex task, (Moonshot AI’s) Kimi will typically consume a higher volume of tokens.”

• From Song’s article: “It remains a company focused on building models centered around the core principle of ‘efficiency.’ Features such as its hybrid attention architecture, KV Cache compression, reduced inference costs for million-token context windows, expert parallelism optimizations, and cross-platform kernel designs may not be particularly ‘sexy,’ but they are undeniably critical.”

What separates DeepSeek from other Chinese AI models? In terms of technical roadmap (see image below), Huxiu regards Moonshot AI as the company most similar to DeepSeek.

• In part because it is not a traditional startup (given its backing by a quantitative hedge fund), it can undertake long time-horizon investments. By comparison, Moonshot AI operates more like a quintessential AI startup, facing the pragmatic demands of commercialization and fundraising.

• It should be emphasized that compute efficiency is not the only thing that matters. Kimi, for instance, is optimized to be a productivity tool. I thought this point from Song’s article was helpful: “If a model consumes a few extra tokens but saves a user three hours of work time, that trade-off may very well be a worthwhile one.”

What does DeepSeek V4 mean for Chinese chips? In contrast to some hasty analysis, it is very likely that DeepSeek is still dependent on Nvidia chips for training; however, some of V4’s toolchain advances suggest gradual progress toward domestic substitution, especially on the inference side.

• DeepSeek does not make any claims about using Huawei Ascend chips for training V4. Instead, it is very likely that the V4 was trained on Nvidia chips. As industry insiders told Song: “If they had truly and completely switched to domestic chips, the V4 model likely wouldn't have arrived this quickly.”

• One fascinating tidbit from DeepSeek’s technical report is their use of the TileLang domain-specific language, which is less tied to the Nvidia ecosystem and can be adapted across many different chip platforms.

• Another Huxiu piece, by Bizheng Dong, delves into DeepSeek V4’s implications for inference. DeepSeek has been experimenting with a unique “Engram” architecture, designed to address memory capacity constraints. Memory capacity is one reason why Nvidia is so strong for both training and inference: its latest Rubin GPU has 288GB in total memory capacity (the Ascend 910B, by comparison, has a memory capacity of 64B). DeepSeek is trying to do more with less memory capacity. From the article: “a long-context inference task that typically requires 80GB of VRAM (video random access memory) to execute might, under the Engram architecture, require only 8GB.”

• Dong shares another detail about adapting DeepSeek-V4 with Chinese chips: “Breaking with industry convention, the developers did not give Nvidia early testing access; instead, they reserved all opportunities for early adaptation exclusively for Huawei and Cambricon. The explicit objective is to execute a comprehensive migration from the CUDA ecosystem to Huawei’s CANN framework.”

• Dong’s piece contains some more stats and industry reports that hype up what DeepSeek means for China’s domestic chip ambitions. I’m a bit more skeptical (at least in the near-term), but the things that stand out to me about V4 are the Engram architecture and TileLang improvements, which seem like they will have staying power.

FULL TRANSLATIONS: 1) DeepSeek Overly Understated; 2) Don’t Overestimate Nvidia; Don’t Underestimate DeepSeek

*We’ll catch up on reading recommendations next week!

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).

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99

Jeopardy is appointment viewing during this streak

…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 AI Enters the Harness Era: MiniMax as a Case Study for the New Cloud Infrastructure of AI Agents

Context: In February, a small OpenAI team disclosed the results of a five-month experiment: “building and shipping an internal beta of a software product with 0 lines of manually-written code.” Using OpenAI’s software engineering agent Codex, the three person-team used a “harness engineering” approach, summed up in four words: “Humans steer. Agents execute.” In this world, human AI engineering skills have shifted from merely designing context-rich or helpful prompts:

“Our most difficult challenges now center on designing environments, feedback loops, and control systems that help agents accomplish our goal…”1

In this week’s feature translation of a jiqizhixin (Synced) article, we dig into how Chinese companies are adapting to the Harness era. This article focuses on an interesting partnership between the Chinese AI start-up MiniMax’s AI agent (MaxClaw) and Alibaba Cloud.

Key Takeaways: MiniMax’s MaxClaw agent is in the first tier of OpenClaw-based adaptations, but how does it deal with fundamental engineering barriers? Remember, the Harness approach gives AI agents significant autonomy and access to high-risk privileges.

• The article describes the core problem set: “When hundreds of thousands of agents concurrently execute these complex tasks with highly strict access privileges and multiple steps, traditional system architectures frequently struggle to cope…How does MiniMax manage to withstand the massive traffic surges generated by MaxClaw and MaxHermes while ensuring business stability?”

• Synced highlights four “chasms” that limit the diffusion of AI agents into enterprise-grade production environments: 1) security boundaries — since agents are increasingly running directly on top of local operating systems, they have high-risk privileges (e.g., deleting key files); 2) state volatility involved in executing long-running tasks; 3) need for scheduling multiple agents; and 4) tension between cost consumption and drastic workload fluctuations.

Enter MiniMax’s partnership with Alibaba Cloud’s Container Service for Kubernetes (ACK) and Container Compute Service (ACS), which tries to address those four engineering challenges associated with large-scale harnessing of agents.

• On the security challenges, Agent execution occurs in the ACS Agent Sandbox, which isolates and contains the risk of prompt injection or privilege escalation attacks. Per Synced, “the scope of the resulting risk is strictly contained within that specific instance, making lateral movement or container escape extremely difficult.”

• To handle unpredictable computational demands — AI agents sometimes have relatively idle periods, punctuated by sudden spikes of workload — Alibaba Cloud supports elastic scheduling of compute.

MiniMax’s evolution points toward the importance of cloud computing platforms as “AI super computers.”

• Citing an IDC survey that finds AI inference accounts for 47% of all token usage and API calls, the article notes that the AI industry is undergoing “transition into a phase of large-scale post-training and inference execution, as the focus of computing power rapidly shifts toward Agent-centric scenarios.”

• If enterprises collectively operate hundreds of thousands of AI agents, round-the-clock seven days a week, then cloud computing platforms and modern container services like Alibaba Cloud’s ACK and ACS will serve as “cloud-native operating systems” for AI supercomputers.

FULL TRANSLATION: When AI Enters the Harness Era: MiniMax as a Case Study for the New Cloud Infrastructure of AI Agents

ChinAI Links (Four to Forward)

Must-read: 27 Years Later, Still Bombing the Wrong Targets

Jon Lindsay is an Associate Professor at Georgia Tech, where he researches the impact of information technology on military power. I’ve really enjoyed his Substack posts lately. This one analyzes the U.S.’s mistaken strike of an Iranian elementary school, drawing on his previous experience as an intelligence officer when the U.S. military accidentally attacked the Chinese Embassy in Belgrade:

I thought so because I had seen something like this before. Almost three decades ago, I served as an intelligence officer during Operation Allied Force, the Kosovo Air War of 1999. During that war, the US military accidentally blew up the Chinese Embassy in Belgrade. A mistake of that magnitude seemed unimaginable at the time. How could we possibly not know what we were hitting, and then hit that of all things? Anyone who walked by the street in Belgrade would plainly see it was an embassy, but US analysts only looked at overhead satellite imagery. The Chinese had told their counterparts in the US Department of State that they had relocated to that new building a year before, but that info had not been input into the military database.

Should-read: (Alibaba’s) DingTalk and (ByteDance’s) Feishu collectively pivot to CLI: Is MCP dead? Is GUI on its way out? (in Chinese)

How should AI agents interact with computers, data sources, and humans? Xiaochai Ye [叶小钗], a developer with previous experience at Tencent, Ctrip, and Baidu, shares thoughts on why some Chinese enterprise AI platforms prefer the Command-Line Interface over the Model Context Protocol.

Should-read: Xiaoice’s [小冰] “Unicorn” Illusion (in Chinese)

Back in December 2023 (ChinAI #248), we deep-dived into the history of Xiaoice, previously China’s most popular chatbot and first incubated in Microsoft before being spun off. Two and a half years later, Xiaoice has been shut down. What happened? This Leiphone article tries to explain.

Should-read: China’s robot champion has everything to lose

For Reuters Breakingviews, Robyn Mak profiles China’s Unitree, which recently applied to go public on China’s STAR market.

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).

Any suggestions or feedback? Let me know at chinainewsletter@gmail.com or on Twitter at @jjding99