Yilin

17 followers · 17 following · 431 views

on the atlas — 70

highlights — 50

The U.S. and China won’t move in lockstep on AI safety, and forging mutually binding agreements with real teeth is probably not possible right now. But they can move forward on parallel tracks, each advancing the technology and also acting to reduce the risks because they believe it is essential to their own national security—regardless of what the other does. In this process, the most impactful decisions will be made within the two countries, not between them. And those domestic decisions will be guided by the cold calculus of national (and political) self-interest.
A Path Forward on AI Safety for the United States and China | Carnegie Endowment for International Peace
It seems like a handful of US and Chinese companies are on track to automate all the white-collar jobs
AI 2040: Plan A
However, as noted, the restrictions also prompted China to implement an all-out, government-backed effort to improve the country’s self-sufficiency in all aspects of semiconductor design and production, an effort that has already resulted in a number of startling achievements. Semiconductor device and equipment makers based in the United States and allied countries lost substantial revenues as a result of the curtailment of their China sales, and as a result of their diminished presence in China, are reportedly less able to track developments in that country’s chip industry to the same extent …
The Limits of Chip Export Controls in Meeting the China Challenge
While China may catch up in individual model quality for now, the true metric of AI leadership might lie in deploying and integrating these systems at scale. In that sense, America's true moat isn't just better models—it's the capacity to deploy and integrate AI in the economy at scale.
China's AI Models Are Closing the Gap—but America's Real Advantage Lies Elsewhere | RAND
This matters for the export control debate because it means the Chinese AI market is not an open field in which sanctions might deter individual bad actors. It is a market structured by entry conditions that filter for compliance with state content and security standards. Every firm that Khawam’s sanctions would target has already met those conditions. The sanctions debate treats these firms as independent actors who choose to violate rules. The registration system reveals them as participants in a regulatory structure whose entry requirements already embed alignment with state governance prio…
The Incentive Architecture Export Controls Cannot Reach | Lawfare
While Washington debated how to impose costs, Chinese municipalities were competing to distribute rewards
The Incentive Architecture Export Controls Cannot Reach | Lawfare
与此同时，也存在数据供给质量不高、流通机制不畅、应用潜力释放不够等问题
十七部门关于印发《“数据要素×”三年行动计划（2024—2026年）》的通知_中央网络安全和信息化委员会办公室
de-emphasizes regulating publicly released models and focuses instead on regulating the frontier lab as an entity
What Should Be Done - by Dean W. Ball - Hyperdimensional
More importantly, it is also unclear (a) whether it will be public, given the aforementioned Executive Order’s reliance on a classified “voluntary” testing program and (b) whether it will be good, given how new many of the senior administration officials now involved in AI policy are new to the issue, combined with the complexity of the topic at hand
What Should Be Done - by Dean W. Ball - Hyperdimensional
a third key intuition: that, almost by necessity, some of the most important frontier AI governance decisions are likely to be shaped by, if not fully made by, the frontier labs themselves. In other words, internal governance will be more central to the future of AI than most people realize.
That Untravell'd World - by Dean W. Ball - Hyperdimensional
China’s situation is exactly the opposite. On the one hand, there is much more data to work with, yet on the other hand, most of it remains untapped. So why is it untapped? A major hurdle is that all of these data are kept in hundreds of thousands (even millions) of data silos of large and small sizes, and those silos aren’t talking to each other. In fact, when China established the National Data Administration in 2023, its main mission was to break down the walls of those silos and let data really flow
#11 Unique challenges and opportunities in China's professional data and information industry (FULL)
You can almost always re-state a should-sentence without the should. It may seem like a trivial transformation on sentences, but it might also really help remove the burden of an obligation.
"Should" considered harmful
Instead of being a perfectionist about the paper, be a perfectionist about writing the paper. Be a perfectionist about identifying good strategies, about abandoning sunk costs, about killing your darlings, about noticing when you're done. Be a perfectionist about wasting no attention. Be a perfectionist about learning from your mistakes. Perfectionism can be a powerful tool, but there's no need to point it at overachieving on metrics you don't care about.
Half-assing it with everything you've got
lacks many of these data sets
China’s AI Arsenal | Foreign Affairs
How the two militaries balance human and AI decision-making may prove to be the deciding factor in the competition. Whereas the U.S. military requires that AI systems involve “appropriate levels of human judgment” by well-trained and experienced personnel, the PLA may be tempted to use AI decision-support systems as alternatives to its weak, inexperienced officer corps. But an overreliance on computer-generated analyses could lead to misinterpretation of military and diplomatic signals and to unsound military decisions.
China’s AI Arsenal | Foreign Affairs
Some of these similarities, including parallel experiments with targeting algorithms and cyber-operations, seem poised to lock Washington and Beijing into a rapid cycle of iterative technological changes. The outcome of the competition will thus depend on which force can develop and scale novel capabilities and operational improvements faster. Given how quickly many defense technologies evolve, technological advantages may be hard won but ultimately short-lived.
China’s AI Arsenal | Foreign Affairs
The exhibition of these systems sent a message about how the People’s Liberation Army (PLA) intends to leverage emerging technologies for battlefield advantage. Seen from Washington, the parade underscored Beijing’s ambition to erode the United States’ technological edge.
China’s AI Arsenal | Foreign Affairs
The CCP’s weaponization of critical resources has strained global supply chains; its predatory market practices have hollowed out foreign industries; its aggression in the South China Sea and militarization of artificial islands have rattled its neighbors and threatened freedom of navigation.
Why China Weaponized the Thucydides Trap | The Heritage Foundation
If the main source of instability in the relationship is the U.S. fear of losing its position, why do plenty of China’s neighbors, and many others further abroad, hold many of the same concerns about China’s actions?
Why China Weaponized the Thucydides Trap | The Heritage Foundation
First, while China’s rise is undeniable, the portrayal of a flailing United States in decline is more questionable. The U.S. share of global GDP in 1996, at the end of the Cold War and the height of the “unipolar moment,” was roughly 25 percent. Today, 30 years later, it stands at 25 percent. If the U.S. is declining, it’s doing so at a glacially slow pace. The relative power gap between China and the U.S. has unquestionably shrunk. But the United States remains the sole global superpower with the benchmark international reserve currency, the most desirable destination for capital and human ta…
Why China Weaponized the Thucydides Trap | The Heritage Foundation
Out of the universe of AI labs in China, it is helpful to think of them in two categories: independent labs, like DeepSeek, Moonshot, MiniMax, Z, Stepfun, etc. and labs that live inside a giant tech company, like Qwen in Alibaba, Mimo in Xiaomi, Seed in ByteDance, Hunyuan in Tencent, etc. The independent labs resort to distillation as a shortcut, because they are more desperate for quality data. The labs that live within a big tech umbrella and have access to real world usage data from other business units inside the corporate family can (and do) leverage those data sources.
Why Do (Some) Chinese AI Labs Distill? - by Kevin Xu
No Data Advantage in China
Why Do (Some) Chinese AI Labs Distill? - by Kevin Xu
钮祜禄氏逝后，乾隆写下可能是他一生中最优美、最感人的诗句，“此后谁怜我，向前止任其。”足见钮祜禄氏晚年与儿子乾隆舐犊情深，令人动容。
铁齿铜牙纪晓岚是个马屁精（乾隆制造）书评
In such a low‑trust environment, it may be tempting to weaponize safety standards. Managing that temptation and focusing collaboration on mutual interests will be essential if any bilateral framework is to be viable.
The AI Issue America and China Can Cooperate On Now - The Wire China
Among these possible areas for bilateral coordination, open‑weight model governance stands out as both underexplored and unusually promising, provided that both sides engage in good faith and resist turning AI governance into a point of geopolitical tit-for-tat.
The AI Issue America and China Can Cooperate On Now - The Wire China
The conventional wisdom is that China is a rising hegemon eager to replace the United States, dominate international institutions, and re-create the liberal international order in its own image. Drawing on data from 12,000 articles and hundreds of speeches by Xi Jinping, to discern China's intentions we analyze three terms or phrases from Chinese rhetoric: “struggle” (斗争), “rise of the East, decline of the West” (东升西降), and “no intention to replace the United States” ((无意取代美国). Our findings indicate that China is a status quo power concerned with regime stability and is more inwardly focused t…
What Does China Want? | International Security | MIT Press
Where to start? For China, The China Project’s best 100 China books list, the fantastic FiveBooks website which has a very well-developed China section, and Tanner Greer’s attempts to introduce folks to Chinese history or his favorite books. Don’t read books about China without citations from Chinese sources (much more common than you’d expect). On AI policy, Henry Farrell has put together a fresh new syllabus. I’m biased against spending too much time reading the news vs developing a foundation of knowledge, though for interviews for policy-relevant internships and jobs you will have to demon…
Policy: An Early Career Guide REVISED!
I used to play this game with my team. I’d name all the White House offices and ask them to tell me what each does and what winning looks like for them.
Doing Big Things in Policy
The vast white space in policy means a lot of important ideas no one has explored. Kumar Garg: We still live in a real deficit of clean ideas. I always used to say, when we were sitting around trying to come up with State of the Union ideas, “Why don’t I have a book from each think tank that says, ‘Here’s everything we wrote in the last year, formulated as a State of the Union idea. Here’s the sentence the president would say. Here’s the logic model of the policy proposal. Here’s a link to all the appendices so you can make it bigger or smaller. Here are the phone numbers of the experts you’d …
You’re probably underestimating what you can contribute
The same term—技术失控—appears across TC260’s AI Safety Governance Framework v2 (September 2025) under several different scenarios, and the framework’s official translation renders it differently depending on context. 技术失控 is translated there as: “the potential risks of technological failure” (针对潜在的技术失控风险, §2.4) “whether a model would pose a potential risk of loss of control” (判断模型是否可能带来潜在技术失控风险, §5.11, the developer-testing requirement) “prevent and address the risk of AI technology losing control” (共同防范应对人工智能技术失控风险, Appendix 2 opening)
China AI Bulletin 5 - by Emmie Hine - China AI Bulletin
While the paper’s use of data was subsequently criticized in ways that Peltzman accepted, its enduring contribution was to highlight the unintended consequences of regulation. Although seatbelts did turn out to save lives, a portion of the benefit from improved driver safety was offset by more reckless driving and more nondriver deaths. It underlined, again, that regulation—even in mandating something as beneficial as seatbelts—always creates a trade-off. Others had talked about the theory of unintended consequences, but Peltzman was the first to do so in a systematic, empirical way.
Unintended Consequences: Origins of an Enduring Idea | Chicago Booth Review
This combination of police crackdowns and new regulations is part of the party-state’s playbook for responding to social issues: arrest those who organize unsanctioned collective action and simultaneously enact policies intended to address the underlying cause of public anger.
Tracing the Roots of China’s AI Regulations | Carnegie Endowment for International Peace
That willingness to absorb international ideas provides a channel for engagement and even influence by U.S. actors. It also marks a real advantage for Chinese regulators who are willing to learn from and adapt ideas from abroad, regardless of whether they come from a geopolitical friend or foe.
Tracing the Roots of China’s AI Regulations | Carnegie Endowment for International Peace
Contrary to popular conception in the rest of the world, China’s AI governance regime has not been created by top-down edicts from CCP leadership. President Xi Jinping and other top CCP leaders will sometimes give high-level guidance on policy priorities, but they have not been the key players when it comes to shaping China’s AI regulations. Instead, those regulations have been the product of a dynamic and iterative policymaking process driven by a mix of actors from both inside and outside the Chinese party-state. Those actors include mid-level bureaucrats, academics, technologists, journalis…
Tracing the Roots of China’s AI Regulations | Carnegie Endowment for International Peace
China’s regulation of AI didn’t cause it to catch up with the United States, and Washington shouldn’t just imitate Beijing’s approach. But China’s experience shows that AI competition and regulation are not incompatible. And it offers lessons in how we can forge ahead in the technology—outcompeting China and all others along the way—while also mitigating the risks that Americans care about.
Trump’s AI Order Won’t Stymie U.S. Competition with China | Carnegie Endowment for International Peace
Instead of opting for a comprehensive AI law like the European Union’s, China released a series of more narrow regulations targeting specific issues. It issued these quickly and iteratively, with each regulation adopting the tools created by earlier regulations and refining them. And it built substantial flexibility in the regulations themselves, often leaving the details to get hashed out in technical standards that are shaped by companies, researchers, and government officials.
Trump’s AI Order Won’t Stymie U.S. Competition with China | Carnegie Endowment for International Peace
So the highest-leverage position is being a translator between the analyst-expert community and the government community, rather than being one of many dozens of experts producing long reports. Todd Moss: One hundred percent. But it’s also about problem identification. In academia, the upstream decision-making about what to research is driven entirely by publications and methodological questions. But in government, they have a couple of top-line burning questions. If you’re not answering a question that’s already top of mind, your work is irrelevant. So it’s the translating and the problem ide…
How to shape policy from outside government
The biggest signal is that I’m gaining traction with people on the inside. If you find something that is bugging people inside government—something they can’t solve and can’t even talk about—and you can give voice to that problem and start problem-solving from the outside, that’s how you gain credibility.
How to shape policy from outside government
Sometimes there’s a zero-sum tradeoff between being well known and being effective. If you burn a relationship by being too public, you will have lost that channel permanently. I make tons of mistakes—I keep things private that don’t need to be, and I’ve revealed things that probably would have been better kept quiet. You have to make a judgment call on the costs and benefits.
How to shape policy from outside government
At my organization, we put out a good amount of public content, and it helps us build our network, which is the purpose. But almost all of our high policy impact work is quiet. The majority of my policy work consists of one- and two-page memos written for a particular individual, never made public, and often tailored specifically to that single person to get them to do a particular thing.
How to shape policy from outside government
Every technique that Western advisers had brought had eventually been assimilated: heliocentric theories and calendrical science, sophisticated medical surgery, economic planning, engineering, interdisciplinary universities, long-distance communications, mechanized warfare, nuclear physics. The Westerners had presented their expertise as the wrapping round an ideological package, however, and tried to force the Chinese to accept both together. It was this that the Chinese had refused to tolerate; even at their weakest, they sense that acceptance of a foreign ideology on foreign terms must be a…
Not To Change China - by John Delury - Thoughts from Seoul
One thing that was incredibly effective about MATS was that 1. it removed many many decisions from my daily life (what to cook for dinner, when to go get groceries, how to get to the gym, where to live, when to go get office supplies, scheduling time with friends) and 2. it replaced them with answers that were significantly more expensive but also more productive than my default.
MATS 9 Retrospective & Advice — LessWrong
how fellowships were valuable in large part because they gave organizations longer observation windows on candidates. The most useful hiring signal was a direct collaboration experience with a senior person which they could reference.
How do we close the Junior <> Senior talent gap™? · Ady Mehta
keeping your focus always on identifying what works and what doesn't.
How to win a best paper award (or, an opinionated take on how to do important research)
I also describe some tags I’m currently using when I remind myself to “really think.” In particular: Going slow Following curiosity/aliveness Staying in touch with why I’m thinking about something Tethering my concepts to referents that feel “real” to me Reminding myself that “arguments are lenses on the world” Tuning into a relaxing sense of “helplessness” about the truth Just actually imagining different ways the world could be Imagining myself being wrong/right in both directions Imagining what ideal future people who could see the full truth would think of my efforts Stamping my foot and s…
Fake thinking and real thinking - Joe Carlsmith's Substack
I give a bunch of examples of this “fake vs. real” spectrum below -- in AI, philosophy, competitive debate, everyday life, and religion. My current sense is that it brings together a cluster of related dimensions, namely: Map vs. world: Is my mind directed at an abstraction, or it is trying to see past its model to the world beyond? Hollow vs. solid: Am I using concepts/premises/frames that I secretly suspect are bullshit, or do I expect them to point at basically real stuff, even if imperfectly? Rote vs. new: Is the thinking pre-computed, or is new processing occurring? Soldier vs. scout: Is …
Fake thinking and real thinking - Joe Carlsmith's Substack
Protect your attention: carve out deliberate space to work on projects. This might be a physical room (i.e. going into an office without your phone or distractions), it might be an inflexible 4 hour chunk of time each Saturday, etc. Like Jensen, Paul Graham has argued that determination requires some pain. You do have to be at least somewhat hard on yourself to offset the costs of cultivating discipline. But this compounds over time, so consistency is more important than severity. Build strong network effects: the city you live in and your social groups will shape your ambitions. Ambitious peo…
Top Performers are Pathologically Ambitious - by Matt Beard
Time-sensitive comparisons: Are you sad you’re not as successful as someone 5 years older than you? Or who has been at this skill for 3x as long as you? Why? Scope-sensitive comparisons: At these top universities you’re already subject to extreme sorting functions - “any smart peer” is not the relevant reference class for “people who could do really cool work on ___ niche EA topic.” Few qualities are immutable: Have you actually tested if you can get incredibly good at this skill? For how many hours? With how many approaches? How much feedback have you gotten? There are many kinds of skills: E…
Your Goal Isn’t Really to Get a Job - by Matt Beard
So if we want to know when AI takes my job, rather than just “do a bunch of tasks”, we also need to know how far these bottlenecks can shift. My “strong opinion, loosely held” is that this can go surprisingly far. One reason is that Moravec’s Paradox really messes with our intuitions. AI capabilities are very spiky, and if we don’t know what this spikiness looks like, then it becomes hard to predict what kinds of things will stump future AIs. This becomes doubly hard because it’s often hard to know what bottlenecks exist until you actually encounter them. The last bottlenecks will eventually f…
How close is AI to taking my job? | Epoch AI
I define intelligence, loosely, as sensemaking from the observation of patterns. A smarter person can detect patterns before a dumber person. The smarter mind is more efficient with the data it receives from the world. So what I really needed to do, while I still possessed some neuroplasticity, was not just become more knowledgeable, but simply become smarter. I leveraged my strengths. As I have already said, I get obsessed easily. Throughout my life, my intellectual obsessions were almost compulsive; I would listen to my favorite music, repeat my favorite scenes in movies, and reread my favor…
How I Work - by Dean W. Ball - Hyperdimensional