Greetings from a world where…

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.

ChinAI #144: Artificial Challenged Intelligence [人工智障]

Jeffrey Ding

·

June 7, 2021

Read full story

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

Greetings from a world where…

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

Greetings from a world where…

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

Greetings from a world where…

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

Greetings from a world where…

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.

ChinAI #296: DeepSeek goes left, ModelBest goes right

Jeffrey Ding

·

January 20, 2025

Read full story

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