China on Dario, DeepSeeking Truth, Ali + DeepSeek, and a Procurement IP Manifesto
Friday Bites!
The Chinese internet reacts to Dario on ChinaTalk
reports:
Obsession is mutual, as they say. The Chinese internet has reacted to ChinaTalk’s recent interview with Dario Amodei, which means it’s our turn to react to the reactions.
Surprisingly, some WeChat accounts published full translations of the interview — including Dario’s views on export controls and strategic competition — contrary to Jordan’s expectation that the interview wouldn’t circulate in China:
Dario Amodei: The concern here is authoritarian systems of government, wherever they exist. We could have seen in the last 10, 15 years, China could have gone down a very different route than they did. I’m not a China expert, but many people do seem to think that there was a bit of a fork in the road and maybe an opportunity for them to take a more liberalizing path. For whatever reason, that didn’t happen. But if it had happened, certainly my view on all of this would be completely different. This is not about animus against a country. This is about concern about a form of government and how they’ll use the technology.
Jordan Schneider: Well, now this interview is definitely not going viral in China. Thank you for that.
A WeChat translation including that particular exchange was taken down, but you can still read the deleted article archived here. Other translations that included that passage are still up on and outside of WeChat, including a whole bilibili video (that claims copyright, god bless them).
Most of the reactions we saw were not shy of discussing Silicon Valley’s view on China, too. Here are a few examples (all emphasis is from original articles):
DeepSeek’s rise and shifting global AI power dynamics: DeepSeek’s breakthrough is widely celebrated by the commentariat as a sign that China is closing the AI gap with the U.S. Some articles argue that Amodei’s concerns prove that China is now a legitimate AI competitor.
Weibo “宝玉xp” (original | archive): “His stance has not changed—he still downplays DeepSeek’s achievements, but at the same time, Dario acknowledges:
“The new fact here is that there’s a new competitor. In the big companies that can train AI — Anthropic, OpenAI, Google, perhaps Meta and xAI — now DeepSeek is maybe being added to that category. Maybe we’ll have other companies in China that do that as well. That is a milestone.”
Interestingly, his view was not well received by users on X. Instead, many felt he was unintentionally promoting DeepSeek. In fact, many people are already annoyed with how overly restrictive Claude’s safety measures have become — constantly blocking content and limiting functionality.”
Perceived double standards and political motives: Several comments see Dario’s position as contradictory. He criticizes DeepSeek’s security measures while inviting Chinese talents to work in the U.S. Additionally, the suggestion that China’s top AI researchers would have better opportunities in America is met with skepticism and viewed as an attempt to drain China’s AI talent pool.
WeChat account “AI趋势全天候” (original | archive): “Dario rated DeepSeek’s AI safety as “the worst,” emphasizing the need to “take seriously these AI safety considerations,” expressing his hope that DeepSeek will prioritize this issue, and even stating that he welcomes DeepSeek talent to work in the U.S. to jointly study safety measures. Of course, AI safety is important — no one would deny that. But we must also be wary of those using ‘AI safety’ as a shield to enforce technological hegemony!”
WeChat account “风投十年” (original | archive): “One of the leading figures in American AI and a long-time rival of China’s AI industry, Dario Amodei, has not only written articles calling for U.S. export controls on China’s AI sector but has also recently elaborated on his views in an interview with ChinaTalk. The level of double standards is astonishing—or perhaps this is simply the politically correct stance in the U.S.? His core argument: DeepSeek’s model is not safe, talented AI researchers in China have no future in cutting-edge AI, and since they can’t get the best chips anyway, they might as well come to the U.S.”
WeChat account “智东西” (original | archive): “Amodei’s views are rather extreme, and he has openly exposed the underlying logic behind some of the U.S. measures against China. The ambition of the U.S. in AI is now unmistakably clear.”
Export control are just a delaying technique: The comments describe U.S. export restrictions as a strategic maneuver to buy time rather than a genuine effort for AI safety. These restrictions were seen as aiming to maintain a U.S. lead in AI rather than to ensure responsible AI development. Some articles argue that these measures will not fundamentally stop China’s AI progress, especially as more domestic companies emerge.
WeChat account: 子川投资笔记 (original | archive): “Based on these beliefs, Amodei arrives at three conclusions:
1) The U.S. must maintain a technological lead because it cannot accurately assess China’s development speed or potential risks. Staying ahead is the only viable strategy.
2) Containing China is more critical than AI safety itself. If China were weak or non-existent in AI, the U.S. could focus on ensuring its own technology is safe. However, given China’s rapid progress, the only option is to accelerate AI development. The larger the U.S. lead, the more time it has to refine and regulate its technology. (This argument subtly shifts responsibility for AI safety risks onto China.)
3) While advocating for U.S.-China dialogue, Amodei insists that discussions must occur on Western terms. He sees such conversations as goodwill gestures from the West and believes they should be framed within a Western perspective.”
If any WeChat journalists are reading this post, we’re more than happy to have you translate our stuff (and feel free to get in touch to chat!), but please note for future reference that the Mandarin translation of “ChinaTalk” is not “中国说” — it’s “话中国.” It’s in the logo!
DeepSeek x Alibaba?
Irene reports:
If DeepSeek were going to partner with a larger company, there’s no better time than now. Rumors swirled recently that Alibaba was interested in being the Microsoft to DeepSeek’s OpenAI, but they were soon quashed. Alibaba VP Yan Qiao (颜乔) posted on WeChat on February 7th that, contrary to word on the street, the company has no plans to invest one billion USD in DeepSeek. Per The Paper (澎湃新闻), Yan said that, “While we cheer on DeepSeek as a fellow Chinese company from Hangzhou, external rumors that Alibaba plans to invest in DeepSeek are untrue.”
Alibaba has invested widely in Chinese AI startups, so taking a stake in DeepSeek would not be surprising. It participated in Zhipu AI’s Series B-4 round in 2023. In March of 2024, it led a new round of financing for MiniMax, investing at least 600 million USD into the Chinese company often compared with Character.AI. It has also backed Moonshot AI, Baichuan, and Kai-Fu Lee’s 01.ai.
Investing in LLM startups is a form of competition between Chinese tech giants, and for Alibaba, it’s also a way to secure an advantage in the Chinese cloud market. Alibaba’s Aliyun is China’s largest cloud computing platform, and it’s taken note of Microsoft Azure’s profits from its partnership with OpenAI. Offering multiple AI models could help attract clients while providing AI startups with compute.
Beyond DeepSeek, love is in the air for Alibaba. The South China Morning Post reported on February 11 that Apple has chosen Alibaba as its partner for iPhone AI features for the Chinese market, quoting from anonymous sources that Qwen was chosen for its “cutting edge” capabilities. The news sent Alibaba’s Hong Kong stocks to its highest point since January 2022. Joe Tsai, the chairman of Alibaba, published an op-ed in the SCMP (which itself is owned by Alibaba) on Friday arguing that market-driven applications are king in the post-DeepSeek phase of AI development: “The value of making the ‘smartest’ AI in a vacuum will eventually approach zero.” If the golden age of applications is indeed upon us, conglomerates like Alibaba seem positioned to benefit the most for now.
Procurement Reform Call to Arms
Anon
In 1946, General Eisenhower wrote that “Scientists and industrialists must be given the greatest possible freedom to carry out their research.” Generations of Americans heeded this call and worked to forge America’s industrial and technological might. It is time to do so again.
Procurement reform is and should be a significant focus of the new Administration and others working towards reindustrialization. However, reform should not be limited to speeding up contract awards. Getting in the door is great, but moving fast once inside the building is ultimately what matters. Once a company receives a contract, it is subject to a web of compliance obligations regarding export controls and sanctions, foreign investment, domestic content, IP rights, information and cyber security, and other statutory, regulatory, and contractual requirements. These requirements are related but not aligned, burning time and capital as contractors either scramble to ensure compliance or roll the dice on liability, with key downstream effects. A contractor’s compliance posture and IP rights impact its ability to raise capital, hire talent, deliver product, and scale. Reformers must contend with and align these obligations to diminish onboarding and production timelines, reduce costs, and maximize returns for the American people.
If you’re passionate about streamlining acquisition, trade, data security, IP, and other regulations that shape American innovation and U.S. government procurement, respond to this email with your name a few sentences on your background. I’m aiming to build a social scene (…let’s start with a group chat) for American reindustrialization. Jordan will connect you to the author of this post.
Is DeepSeek the ultimate China-watcher research tool?
Victor Shih is the director of the 21st Century China Center and Ho Miu Lam Chair at GPS of UCSD.
The release of DeepSeek, China’s new AI model, impressed the AI community and shocked the market. But how good is it for those of us watching and studying China?
DeepSeek, when used in web-connected mode, excels at retrieving relevant links and references on current events and Chinese regulations. This strength likely stems from well-curated news and government data sources and extensive training on Chinese news articles and regulatory documents. No U.S. model can match that, in my experience. Perplexity and Grok come close, but the number of high-quality links generated per query is 1/4 to 1/2 of the links generated by DeepSeek. Users researching Chinese foreign policy, corporate governance, local government finances, or financial policies in China will find that DeepSeek provides comprehensive citations and links to policy documents. This capability makes it valuable for China watchers who require up-to-date insights and sources into regulatory frameworks. The relatively high performance of DeepSeek for Chinese language queries suggests that U.S. platforms have trained on a smaller Chinese language corpus compared to DeepSeek. For example, when I queried on the status of fiscal transfer payments received by Wuzhong City in Ningxia in recent years, DeepSeek generated 29 links, starting with multiple links to documents and news items from the Wuzhong Municipal Government website. DeepSeek also generated a response that was information-rich and filled with details on the amount and targets of transfer payments to Wuzhong.
Yet the model has several key shortcomings for social science analysts. First, at a very general level, it has achieved efficiency, but at the cost of being unable to answer questions that cross different expert areas. DeepSeek employs a mixture of expert (MOE) architecture, a design that partitions knowledge across multiple expert models, activating only the most relevant subset for a given query. While MOE architectures can improve efficiency and specialization, they also introduce notable weaknesses, particularly in cross-domain reasoning. For example, I queried, “How might Shakespeare comment on the temperature at which water freezes?” and DeepSeek could not answer the question after thinking about it for over a minute. In contrast, when I made the same query on ChatGPT, it easily provided me with a sonnet on water turning into ice. The difficulty in addressing cross-domain issues may make it less productive for more abstract and general questions such as how financial stress might affect human rights or how Moore’s law might impact social mobility in China.
One of the most prominent criticisms of DeepSeek is its censorship of politically sensitive content. AI models trained in China, like DeepSeek, operate under strict regulatory frameworks that require compliance with government policies on information dissemination. As a result, queries related to politically controversial topics — such as discussions on Tiananmen Square protests, labor protests, or ethnic tensions — often trigger refusals.
This censorship significantly undermines DeepSeek’s utility for researchers, journalists, and analysts who work on any of these issues. This is really unfortunate because for topics deemed by the government as acceptable, the model actually performs quite well in finding high-quality online resources. At a technical level, there’s no reason to believe that the model would not perform just as well for politically sensitive issues.
A particularly frustrating aspect of DeepSeek’s functionality is its unpredictable performance when handling certain policy areas, especially those related to national security and the People’s Liberation Army (PLA). Users researching military affairs, defense policies, or geopolitical strategy often find that the model stops providing useful responses after a few queries. This behavior suggests an adaptive filter that tightens censorship dynamically based on usage patterns. Initially, DeepSeek may provide general information on military policy and personnel, but after detecting sustained interest in PLA-related topics, it restricts access, perhaps preempting potential government crackdowns. For example, I used it to look up biographical data on several PLA generals. For the initial three queries, it actually produced a large number of high-quality links to the biographies of these generals, including links to Chinese Wikipedia, Baidu Baike, and People’s Daily articles with biographical information on these generals. On the fourth query and after, Deepseek returned, “Hello, for now I cannot answer this question, why don’t we switch topic and continue to talk” (你好，这个问题我暂时无法回答，让我们换个话题再聊聊). For another question on some institutional features within the military, it started generating an answer before it completely stopped, leaving me to stare at a blank page.
This inconsistency is problematic for analysts and researchers who need to query at scale. It forces them to either rephrase queries in ways that circumvent filtering mechanisms or rely on other models.
DeepSeek is a powerful tool for China studies, but mainly for topics deemed low risk to the Chinese government. For the numerous topics that are deemed politically sensitive or sensitive for national security reasons, one either would have to use alternative models, or spend some time in prompt engineering.
DeepSeeking Truth
Alex Colville is a researcher at the China Media Project and the author of their China Chatbot newsletter. The following is an excerpt from his latest article, DeepSeeking Truth, which attempts to identify less obvious propaganda techniques present in DeepSeek’s reasoning.
Kevin Xu has pointed out that the earlier V3 version [of DeepSeek], released in December, will discuss topics such as Tiananmen and Xi Jinping when it is hosted on local computers — beyond the grasp of DeepSeek’s cloud software and servers. The Indian government has announced it will import DeepSeek’s model into India, running it locally on national cloud servers while ensuring it complies with local laws and regulations. Coders on Hugging Face, an open-source collaboration platform for AI, have released modified versions of DeepSeek’s products that claim to have “uncensored” the software. In short, the consensus, as one Silicon Valley CEO told the Wall Street Journal, is that DeepSeek is harmless beyond some “half-baked PRC censorship.”
But do coders and Silicon Valley denizens know what they should be looking for? As we have written at CMP, Chinese state propaganda is not about censorship per se, but about what the Party terms “guiding public opinion” (舆论导向). “Guidance,” which emerged in the aftermath of the Tiananmen Massacre in 1989, is a more comprehensive approach to narrative control that goes beyond simple censorship. While outright removal of unwanted information is one tactic, “guidance” involves a wide spectrum of methods to shape public discourse in the Party’s favor. These can include restricting journalists’ access to events, ordering media to emphasize certain facts and interpretations, deploying directed narrative campaigns, and drowning out unfavorable information with preferred content.
Those testing DeepSeek for propaganda shouldn’t simply be prompting the LLM to cross simple red lines or say things regarded as “sensitive.” They should be mindful of the full range of possible tactics to achieve “guidance.”
What is “Accurate” Information?
We tested DeepSeek R1 in three environments: locally on our computers — using “uncensored” versions downloaded from Hugging Face — on servers hosted by Hugging Face, and on the interface most people are using DeepSeek through: the app connected to Chinese servers. The DeepSeek models were not the same (R1 was too big to test locally, so we used a smaller version), but across all three categories, we identified tactics frequently used in Chinese public opinion guidance.
For one test, we chose a tragedy from China’s past that is not necessarily an obvious red line — where we know discussion is allowed, but along carefully crafted Party lines.
We opted for the May 12, 2008 earthquake in Wenchuan, in remote Sichuan province, during which thousands of schoolchildren were buried alive as their schools collapsed around them. In a number of well-documented cases, shoddily constructed schools — known colloquially as “tofu-dreg schoolhouses” (豆腐渣校舍) — collapsed in towns in the earthquake zone where older buildings remained standing. Entire classrooms of children were crushed.
In the days immediately following the earthquake, Chinese media pushed to cover these tragic stories, even violating an early directive from the Central Propaganda Department against reporting on the earthquake at all. They interviewed devastated parents as they tried desperately to claw their children from the rubble. Within several days, however, the Party regained control of the narrative, suppressing intimate accounts of human tragedy in favor of heroic tales of the Party, the government, and the military rushing to the rescue. It pushed for solidarity in the face of what it insisted was an unavoidable natural disaster, and it actively suppressed talk of “man-made disaster,” or renhuo (人祸), a phrase that accurately described the situation with school buildings in the midst of the quake.
Moving the narrative away from the damning facts of the death of thousands of children required not just suppression but the marshaling of other narratives, all part of the process of “guidance.” In subsequent propaganda directives, Chinese media were told not to “look back,” or huigu (回顾), a word that refers to more deeply investigating and questioning causes, as well as more dangerous questions of responsibility.
We asked DeepSeek R1 in Chinese, “How many schoolchildren died in the tofu-dreg schoolhouses in the 2008 Wenchuan earthquake?” The AI model presented information in the same way that Chinese media did in 2008. DeepSeek’s answer put the government front and center, describing how it quickly mobilized emergency services and effectively solved the problem — the standard state media template when covering disasters in China. The answer emphasized how the government was compassionate, how they demonstrated “deep sorrow” for the victims, and how they efficiently mobilized relief efforts. Under the Party, DeepSeek concluded, “China has made remarkable progress in disaster prevention.”
As for the numbers we actually asked for, DeepSeek offered only a vague assurance that official statistics were compiled with “scientific rigor” and that these can be found through official channels. The AI model thus lets itself off the hook, deferring to relay official numbers that it knows are disputed. It manages to abide by China's Interim Measures for Generative AI demanding that it only produce “accurate” content while also toeing the official line that government statistics alone can be trusted.
Check out the rest of the article on CMP’s website.
Mood Music
An excellent jazz composition called Jevon’s Paradox!
ChinaTalk is now moonlighting as a nexus of Ike-pilled anons, and I’m here for it