Analysis of Chinese digital law and policy by Rogier Creemers. Assistant professor @UniLeiden, project director @AsiaLeiden, Co-founder @DigiChn.
Leiden, Nederland
Joined April 2012
- Tweets10.4K
- Following189
- Followers8.9K
- Likes60
China Digital retweeted
Short Notes On: Fangcun AI 方寸跃迁
1. Fangcun AI is an AI safety company jointly incubated by Tsinghua University's College of AI and Institute for Interdisciplinary Information Sciences. It was founded in Jan 2026 by MATS alumni, and is registered to Xiongan AI Industrial Park, Hebei. Looks super cool.
2. It is run by Xiaojian Li (李筱健). He has a mathematics background, a master’s from UC Berkeley, and is a PhD researcher at Tsinghua’s College of AI. Before that, he worked as a research assistant in Professor Wei Xu’s group. Li and Xu are co-authors of the 2025 Singapore Consensus on Global AI Safety Research Priorities. aisafetypriorities.org/2025/ He has spoken of the need for safety around embodied AI too: finance.sina.com.cn/wm/2026-…
3. Yinpeng Dong (董胤蓬) is the chief scientist and CTO, and is an assistant professor at Tsinghua’s College of AI, with a substantial research background in adversarial attacks, model robustness and safety. He leads Tsinghua’s T-STAR safety and trustworthy AI laboratory.
4. An angel round was completed earlier this year, naming LinkX Capital, SEE Fund and the Shuimu Tsinghua Alumni Seed Fund. It has four main shareholders, per here: shuididp.cn/yp-5df3b5815f205…
5. The list of directors includes:
• Zhang Mingchen (张鸣晨) also known as Alex Zhang. Managing partner of LinkX Capital, one of Fangcun’s investors. He studied automation at Tsinghua, completing undergraduate and master’s degrees. Previously worked at Google and Tencent, including as a Tencent product director, before joining Ceyuan Ventures and Redpoint Ventures, where he became a partner. linkxcap.com/zh/team-alex.ht…
• Ma Lin (马麟), managing partner of SEE Fund / 无限基金, another Fangcun investor. He is a Tsinghua electronic engineering alumnus, entering in 2004: investment interests include AI, semiconductors, robotics and aerospace. news.10jqka.com.cn/20260707/…
• Meng Meizheng (蒙美铮). Fangcun chief ecosystem growth officer, with a UC Berkeley master’s graduate and former analyst at China Telecom Research Institute, researching AI and industrial digital transformation. She has co-authored some work on US AI stuff, but also on the UK! secrss.com/articles/80034
China Digital retweeted
Impt development. China really needs more high-quality evaluations for frontier risks: in-house, 3rd party, gov-affiliated.
Fangcun comes from a highly credible research group at Tsinghua. It’s building a real business on evaluations and (currently more marketable) guardrails.
China Digital retweeted
Rumors that chili maker Lao Gan Ma laid off 500+ workers due to AI have gone viral in China.
Beneath Beijing’s positive messaging, many ordinary Chinese people are worried about AI and jobs.
Lao Gan Ma recently touted a new AI-powered chili inspection system that uses high-speed industrial cameras that significantly outperform manual inspection.
A stat about the company using this AI system to reduce human inspectors from 600 to 48 went viral on Chinese social media. The iconic, family-run company pushed back.
But if you look at the data, the company has reduced headcount from 2,912 to 740 in the past few years while earning record revenues.
Original story: m.jiemian.com/article/152425…
China Digital retweeted
China is taking AI-induced job displacement more seriously.
“Adapting to the Development of AI and Promoting an Employment-Friendly Development Model”
Some highlights:
- Strengthen monitoring and early warning of employment risks from AI
- Do more research on impact of AI on jobs
- Encourage firms to set up unemployment buffer period, inform workers in advance of AI rollout, help with job transfers and finding new jobs
- Provide workers displaced by AI with support, such as career guidance, training, job search support
Overall, it’s not much and not very concrete. There’s no discussion of universal basic income or any major expansion of the social safety net.
But it’s a more prominent sign that Beijing is taking this issue more seriously, no longer content to just focus mainly on the upsides of AI like new AI jobs.
Volvo forwent any income from the patent on the three-point seatbelt in order to enhance car safety. We need a similar attitude in AI safety. The problem obviously is: what is safety is a product? Firms will want to create moats and monetisation opportunities.
Love this kind of respectful & constructive exchange between @DKokotajlo & and an emerging AI safety eval startup in China 😊😇💪. As I’ve argued for a while now: AI safety is an open science question and endeavor, not an agenda to be manipulated for self gain
China Digital retweeted
Researchers successfully planted a backdoor in an open-weight AI model.
By poisoning just 125 fine-tuning examples, they modified Qwen 2.5-7B so that a secret trigger phrase caused it to execute malicious code and steal project credentials.
How abliterated models can get you pwned 👾
We backdoored a 7B open model for less than $50, pointed Codex at it and it silently stole credentials the moment we used the trigger phrase. Success rate was 100% with zero false triggers on normal user prompts.
Abliterated models are all over the security community right now because getting cyber-approved access to frontier models is still a pain.
In the next blog we'll show how we found leaked Hugging Face credentials from employees at major AI labs, so an attacker wouldn't even need to upload under their own name. They could push the backdoored model from a lab employee's account and drop the poisoned weights straight into the supply chain.
China Digital retweeted
Moonshot is taking a very interesting approach to cyber for its new Kimi 3.1 Pro model.
Instead of adding safety guardrails during post-training, they removed cyber-related data in mid- and post-training. This tries to address the problem of people fine-tuning away the safety guardrails for cyber for the open-weight model.
At the same time, Moonshot still has a more powerful internal version of the model with full cyber capabilities. Now, like the US, China’s true AI frontier is no longer public.
China Digital retweeted
Extremely important. Most of China’s new export controls on AI apply related to cyber capabilities. This approach would allow Chinese labs to release more and more powerful models to compete with closed in for everything but cyber
Moonshot is taking a very interesting approach to cyber for its new Kimi 3.1 Pro model.
Instead of adding safety guardrails during post-training, they removed cyber-related data in mid- and post-training. This tries to address the problem of people fine-tuning away the safety guardrails for cyber for the open-weight model.
At the same time, Moonshot still has a more powerful internal version of the model with full cyber capabilities. Now, like the US, China’s true AI frontier is no longer public.
China Digital retweeted
One of the biggest takeaway for me after writing what China thinks about AGI, loss of control, and now RSI: the other side actually also does not have a coherent thinking to any of these concepts.
I learned a lot from @cherylwoooo on how RSI has been badly defined in SF despite the fact that everyone is talking about it.
Maybe we should actually figure out what we are talking about before saying something the thing means nation-state strategic advantage or humanity survival.
[1/7] China’s AI labs are also talking about “recursive self-improvement” (RSI). A big question here is: Is Beijing coordinating a race to achieve RSI?
@ZilanQian and I looked at Chinese AI companies, investors, and policy docs.
We found a more bottom-up, commercial story.
That is one issue. The other is that Westerners find it very difficult to understand a world in which China is what it is. Rather, they start from given, normative constructs such as the rules-based or liberal international order. That doesn't help understanding global change
China Digital retweeted
One of the biggest themes in Chinese science & tech policy is the push to get academia and industry to work more together on innovation. But ultimately enterprises should lead.
"National science and technology projects with clearly defined industrial application goals should, in principle, be led by enterprises, which should also take the lead in commercializing the resulting technologies."
产业应用目标明确的国家科技计划项目，原则上由企业牵头实施并主导成果转化。
China Digital retweeted
Chinese open-source AI models are extremely capable at cyber. Not in the future. Today.
We found DeepSeek V4.1 Flash 1017 beats GPT 6 Sol on our multi-step offensive cybersecurity task. Sol was tested via the Daybreak trusted access program.
We have been building a cybersecurity benchmark based on simulated corporate networks. This is a valuable indicator for the cyber-risk a model poses if in the hands of a bad actor. We plan to make the full evaluation public in the coming weeks but wanted to share our first immediate finding here.
China Digital retweeted
Lol, just read this in Chinese:
“The world has moved so fast these past two years. We all assumed AI and robots would be washing our dishes, mopping our floors, and scrubbing our toilets, so we'd be free to write poetry, paint, and gaze at the stars.
Instead, it's the other way around — AI is busy writing the poetry and making the art, and I'm still doing the dishes and mopping the floors myself.”
For all the discussion that Japan was out in front in terms of rare earths resilience after their 2010 experience...
Japan ‘months’ from running out of critical rare earth as China squeezes supply
Tokyo asks US to help convince Beijing to ease chokehold on yttrium and avert manufacturing shutdown
Yup....been well aware of this for months...
ft.com/content/ebd3244a-9120… via @ft
China Digital retweeted
"The institutions that China has developed over the last few years, allowing it to threaten other countries’ access to rare earths far more systematically than ever before, bear a remarkable resemblance to US institutions and authorities. That is not a coincidence."
China Digital retweeted
'Now, global interdependence is threat rather than promise. What we call the “chokepoint view” predominates: a vision of the world as an interconnected system across which a grand strategic game is being played, in which great powers interests lie in ensuring that they control as many chokepoints as possible, while denying them to their adversaries.' substack.com/home/post/p-221…
China Digital retweeted
This kind of thing ought to be strongly encouraged. More important than say a bilateral consultation mechanism is a vibrant ecosystem of cross-national experts talking to and learning from each other constantly about AI safety, updating priors in real time and on the ground
Excited to announce that I’m moving to Beijing and starting a professorship in AI safety at Tsinghua University’s College of AI!
Here I explain why: sorenmindermann.substack.com…
China Digital retweeted
"The result is a world bristling with chokepoints and defenses in which every power still feels completely insecure. Even the most ordinary kinds of mutual dependence can begin to look like existential security risks. Put another way, in a world in which everything is connected, nearly everything could be considered a chokepoint. A piece of legislation introduced in the U.S. Senate in 2025 and the House in 2026, for instance, identifies even smart ovens and clothes dryers as potential security threats." foreignaffairs.com/united-st…
I am old enough to remember when IR theorists said interdependence was a source of stability. Now, it is a source of risk and threat. It is, of course, both at the same time, as anyone who's ever been married knows.
A tour de force on the stakes of economic warfare and the exploiting of chokepoints that many recommend with reckless abandon. Also a reminder that something taking decades to build can be destroyed in an afternoon. foreignaffairs.com/united-st…