allowing export control evasions and distillation attacks, we have let the CCP’s AI efforts trail closely up the frontier curve. But if the US and its allies act now to address both issues, it may be possible to lock in a 12-24 month lead in frontier capabilities. A lead that large by 2028 would be enormously advantageous. Such a lead would also augment efforts to engage with AI experts in China on AI safety and governance, which we support. But the window of opportunity to lock in that lead will not necessarily remain open for long. Here, we present two potential scenarios for the state of US-China AI competition in 2028. The first scenario is one in which democracies have established a commanding lead in model intelligence, adoption, and global distribution. This scenario can be achieved if policymakers act now to tighten controls on advanced compute to PRC labs, disrupt their efforts to distill America’s best AI models, and accelerate democracies’ adoption of AI. The second scenario is one in which the CCP is competitive at the near-frontier. This scenario happens if policymakers don’t build on our existing lead, or if they loosen restrictions on access to compute for PRC firms. Many in Congress and the Trump administration have championed export controls, curbing distillation attacks, and exporting American AI. In advancing these policies, we are hopeful that democracies can secure a commanding lead by 2028, and avoid a destabilizing neck-and-neck race with the CCP two years from now. We expect frontier AI to have transformational economic and societal impacts in the coming years, as described in Machines of Loving Grace and The Adolescence of Technology . Our mission is to ensure that humanity navigates the transition to transformative AI safely and beneficially. We believe that a successful transition can lead to astonishing breakthroughs in medicine, invention, and economic growth. Whether that transition goes well depends in part on where the most capable systems are built first. The political systems in which the most advanced AI is created will shape the rules and norms for how the technology is developed and deployed. In turn, those rules and norms will help determine whether the technology is safe, whose security it protects, and whose interests it ultimately serves. We believe that responsibility should rest with democratically elected governments, not authoritarian regimes. If the frontier is set by regimes that treat AI as an instrument of repression, military advantage over democracies, and domestic control, the transition is less likely to go well, for those regimes’ own citizens or anyone else. Historically, the reach of authoritarian rule has been limited by its dependence on human enforcers to carry out surveillance and repression. Powerful AI systems may remove that dependency, enabling automated repression on a far greater scale. For that reason, the prospect of the CCP leading in AI is among the greatest threats to a successful transition. The CCP holds enormous power and influence at the helm of China’s economy, military, and the largest authoritarian state structure on Earth. It is also the only country besides the US with well-resourced, highly talented AI labs chasing the frontier. Furthermore, the CCP is highly motivated to establish China as the leading AI power. Beijing has poured tens of billions of dollars into China’s AI and semiconductor sectors. Already, the CCP uses AI systems to censor speech,

enforce draconian policies on ethnic minorities, and hack major corporations and government agencies. The CCP’s vision of AI-enabled techno-authoritarianism has been extensively documented in Xinjiang, where state security agencies have systematically deployed facial recognition technology, biometric data collection, and communications surveillance, enabling repression at a scale that humans alone could not achieve. Frontier AI systems will make those capabilities cheaper to maintain, far more pervasive, and more sophisticated. The CCP’s export of these technologies has enabled autocrats in other countries to more effectively stifle dissent, entrenching authoritarianism. A CCP-led AI frontier could dramatically strengthen repression around the world. Frontier AI will shape the future military balance. CCP leadership already operates on that premise, and is building its military for an AI-enabled battlefield. PLA strategists view the “intelligentization” of their military forces as the means with which to catch up and eventually surpass the US military. The PLA is already procuring commercially developed Chinese AI systems for military use, including DeepSeek models deployed to coordinate swarms of unmanned vehicles and enable cyber offense capabilities. 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. The risk compounds because frontier AI will be an accelerant for other critical technologies . Advanced AI models will be able to compress research and development (R&D) cycles in semiconductors, biotech, and advanced materials. A lead in frontier AI will enable a widening lead across the full national security technology stack. If a PRC AI lab had developed a model at the level of Claude Mythos Preview before an American one, the CCP would have had first access to a system that can autonomously discover and chain software vulnerabilities, which it could have used to further penetrate critical American infrastructure. Future models will be exponentially more capable, and therefore have commensurately greater implications for the national security interests of the US and other democracies. A neck-and-neck race between American and Chinese AI labs could make industry and government-led safety and governance efforts more difficult, and less likely. If PRC labs are either close behind or at par with models in the US, private AI firms in the US and China are likely to feel more pressure to release new models and products faster, without taking prudent pre-deployment safety measures. Governments could become reluctant to enact policies to encourage responsible AI development and deployment, for fear of falling behind. While increasing numbers of researchers in China’s AI labs and policy community are concerned with AI safety risks, this trend has not translated into safety practices on par with labs in the US. As of last year , only 3 out of 13 top Chinese AI labs published any safety evaluation results, and none disclosed evaluations for Chemical, Biological, Radiological, and Nuclear (CBRN) risks. The Center for AI Standards and Innovation (CAISI) found that DeepSeek’s R1-0528 model complied with 94% of overtly malicious requests under a common jailbreaking technique, compared with 8% for US reference models. This pattern has continued in more

recent releases. For example, an independent assessment of Moonshot’s Kimi K2.5 published in April found that the model failed to refuse CBRN-related requests at a far higher rate than US frontier models. Compounding the problem, labs in China often release dual-use capable models as open-weight. Once a model is open-weight, safeguards that do exist can be removed, making the model available to any state or non-state actor to use for malicious purposes, including the cyber and CBRN misuse those safeguards were built to prevent. We support policies in the US and other countries that build and maintain a safe, near-term lead over the CCP in intelligence, domestic adoption, and global distribution. This lead is key to avoiding authoritarian AI leadership and protecting the national security interests of the US and other democracies. Doing so is a fundamental prerequisite to ensuring that democratic states can achieve favorable terms with authoritarian states. Anthropic deeply respects the Chinese people and the accomplishments of the Chinese AI community. We hope for peaceful relations between China and the world. Our concerns are specifically with the risks to humanity posed by any powerful authoritarian political systems with access to frontier AI systems. Anthropic supports international AI safety dialogue with AI experts in China, when possible. The world has a vested interest in safe AI, regardless of where it is developed and deployed. There are a range of risks that could emerge from frontier AI systems requiring engagement between the US and China. Efforts that identify shared challenges and advance ideas to prepare for and mitigate these risks are in our shared interests. The prospects for productive engagement are best when the US maintains a large capabilities advantage. Responsibly building a lead in developing and deploying the most advanced AI augments our ability to influence AI safety in China and elsewhere. Mythos Preview, a model that we released to select partners as part of Project Glasswing in April, signals the arrival of an acceleration period that makes policy action even more urgent. With access to the model, Firefox was able to fix more security bugs last month than it had in all of 2025, and almost 20 times more than its monthly average security bug fixes in 2025. In response to the model, one PRC cybersecurity analyst wrote that China is “still sharpening our swords while the other side has suddenly mounted a fully automatic Gatling gun.” Frontier AI capabilities will quickly approach the “ country of geniuses in a datacenter ” portrayal of transformative AI. This acceleration will be driven by the logic of scaling laws , in which model performance improves predictably with increases in computing power and data inputs, and by AI itself increasingly being used to accelerate the development of new models. There is a high likelihood that we will look back on 2026 as the breakaway opportunity for American AI. American labs have the most advanced AI models, a large lead in both the quantity and quality of the advanced AI chips required to push the frontier, and a colossal capital advantage from revenues and financing to back the necessary investments to achieve it. PRC labs have real strengths: world-class, innovative talent, abundant and cheap energy, and plenty of data. All are requirements for developing frontier intelligence. But they simply do not have sufficient domestic compute to compete, nor do they have

the revenues and capital to fund it. The US and China are engaged in a competition for strategic advantage in frontier technologies like AI. Statements from both Beijing and Washington reflect that view. Calling that competition a “race” can give the false impression that there is a finish line, after which one side will conclusively secure victory. Rather, the competition will be an ongoing contest for advantage, in which either democracies or authoritarian regimes successfully position themselves to shape the values, rules, and norms of an AI-enabled future. This competition is playing out on four fronts: Intelligence is the most important of the four fronts. We anticipate that frontier model capabilities will drive the most consequential changes for geopolitical competition. Model capabilities are also a primary driver of market adoption and global distribution. But intelligence alone is not sufficient. If the CCP integrates near-frontier AI systems quicker and more effectively into China’s economy and the CCP security apparatus, and drives global adoption of subsidized, low-cost AI, then it could secure advantages over democracies that overcome an intelligence deficit. Beijing’s AI+ Initiative and its focus on “embodied intelligence” accordingly put high priority on policies that advance the integration of frontier intelligence into their economy and state apparatuses. The Trump administration’s AI Action Plan , and its focus on “ promoting the export of the American AI technology stack ,” also speaks to the strategic advantage of driving global adoption. While we won’t focus on it in this essay, we believe resilience will be an important front of AI competition. Being able to sustain stability, cohesion, and good policymaking in this period will be a critical advantage, and a vulnerability for those who cannot. Compute—the advanced semiconductors needed to train and deploy frontier AI—is an essential input on each front of the competition described above. The race for global AI leadership is in large part a race for compute. For more than a decade, model capability has scaled with compute, and the majority of performance gains in AI capabilities have historically come from simply using more of it. Moreover, compute is needed to serve customers’ use of AI (also known as “inference” capacity), not just to train new models. Compute will be critical both for training the most intelligent models and for deploying them in commercial and national security spheres. Access to top talent, copious amounts of data, and critical algorithmic advances all matter to the race for intelligence—but each of those inputs is irrelevant if the compute is insufficient. Democracies are winning the competition for compute leadership today. While some worry that export controls could accelerate the CCP’s own efforts to develop an advanced chip supply chain, little evidence suggests that China’s indigenization efforts will challenge US and allied leadership in advanced compute technology. Beijing has invested enormous resources into China’s chip sector, with major industrial policy initiatives like the Made in China 2025 strategy and the China Integrated Circuit Industry Investment Fund launched years before the imposition of export controls. Despite this state-backed investment, PRC AI labs and chipmakers remain stymied by US and allied export controls on advanced chips and chipmaking equipment. As a result, the compute gap appears to be widening. An

analysis of Huawei and NVIDIA’s roadmaps found that Huawei will produce just 4% of NVIDIA’s aggregate compute in 2026 in total processing performance, and 2% in 2027. Moreover, NVIDIA represents only part of the US and allied compute ecosystem, with Google and Amazon ramping up production of their own chips (TPUs and Trainium, respectively) to meet demand from American frontier AI labs and their customers. Further exacerbating their compute shortfalls, China has made little progress in many of the most technologically complex segments of the semiconductor supply chain. Without access to extreme ultraviolet (EUV) technology, and even more so if policymakers can close loopholes on deep ultraviolet (DUV) technology and servicing and maintenance thereof, China’s chipmakers will remain unable to manufacture chips in sufficient quantity or quality to challenge US compute leadership. China’s inability to manufacture high-bandwidth memory at scale further exacerbates this gap. If the US strengthens its restrictions on the CCP’s ability to access US compute, one study estimates that America will have access to roughly 11 times more compute than China’s AI sector. There are two main reasons for the compute lead. The first is the incredible innovation of companies like NVIDIA, AMD, Micron, TSMC, Samsung, ASML, and others across democracies like Japan, South Korea, Taiwan, the Netherlands, and the US, who together have built the unique technologies in the world’s most advanced semiconductors. Today’s AI achievements would not be possible without the feats of engineering and decades of sustained R&D investments that c

AI在浏览器虚拟机中用汇编与eBPF技术生成分形图形

开发者Yossi Eliaz创建了一个名为wolfram-fb0的创新项目，该演示展示了AI如何直接在浏览器中的一个真实Linux虚拟机内，编写x86_64汇编语言及eBPF程序，用于生成沃尔夫勒姆（Wolfram）规则、曼德博集合等分形图形，并将图像直接输出到帧缓冲设备/dev/fb0。该项目特别强调了为何必须使用真实的虚拟机而非容器：因为eBPF追踪、访问物理帧缓冲设备以及嵌套虚拟化等功能均需真实的内核和硬件访问权限。 整个构建过程由AI智能体驱动，包括五个阶段：在虚拟机中预置开发环境；AI使用本地Gemma模型编写纯汇编代码；自动化验证与优化；通过QEMU启动并将AI编写的程序加载到真实帧缓冲；最后通过公网URL实时分享图像流和内核事件追踪流。项目已开源并采用MIT许可，其在线平台为新用户提供了免费额度，使得任何人都能一键复现并观察AI代码与系统内核交互生成分形的完整过程。 #AI #编程 #虚拟化 #分形 #开源 #图形计算 #eBPF #技术前沿

Bursting the AI Bubble: Fed Could Take Away the "Who Could Have Known?" Defense

7 2 3 Share The collapse of both the 90s tech bubble and the 00s housing bubble had a devastating impact on the lives of tens of millions of workers. And with the collapse of the housing bubble, millions also lost their homes and their lives’ savings. When something causes so much damage, it would be nice to see the people responsible pay some price. In the case of the 1990s bubble, there were some instances of fraudulent accounting where the perps did get nailed. Enron and Worldcom are two that stand out, where the people most responsible did get prosecuted and face time in prison. Thanks for reading! Subscribe for free to receive new posts and support my work. This was less the case with the housing bubble. Hundreds of billions of dollars of fraudulent mortgages were packaged into securiti

es and sold around the world. There was little effort to determine criminal culpability, as the Obama Justice Department seemed to have decided not to look into the mess. But apart from the people who might have literally committed crimes, both bubbles were driven by people who were criminally stupid. I’m thinking of the people who get paid very high salaries to manage pensions, endowments, or other large pots of money, who apparently thought that the record high price-to-earnings ratios of the dotcom era made sense. In the next decade, they weren’t bothered by the unprecedented departure of house sale prices from rents of the housing bubble. It would be reasonable to think that people, some of whom were paid millions of dollars a year, would be able to see things that, certainly in retrospect, seemed very obvious. And to some of us, seemed very obvious even before the collapse. But very few of these people faced any consequence. To be clear, I’m not talking about jail time; I just mean that their careers should have suffered. After all, those lower down the pay ladder are held responsible for the quality of their work. The dishwasher that breaks a lot of dishes or the custodian who leaves a dirty toilet gets fired. Shouldn’t the investment manager who loses 20%, 30%, or 40% of the value of their portfolio also be sent packing? What kept most of these highly paid failures in their jobs was the “who could have known?” defense. This just meant that all of them could point to peers who made equally stupid calls in their investment decisions. If everyone on Wall Street thought was a $100 billion company, can you blame your investment manager for failing to recognize it was on the edge of bankruptcy? This is where the Fed can play a useful role. Around 200 economists work for the Federal Reserve Board in Washington, and roughly 200 more work for the 12 district banks around the country. The new Fed chair, Kevin Warsh, could assign some of the Fed economists to assess whether the current valuation of the stock market is consistent with the Fed’s projections for the future growth of GDP and profits. Unless their arithmetic is very different than the stuff the rest of us use, they will have to conclude that stock valuations are not consistent, unless today’s crop of stockholders expect very low future returns. That seems unlikely, but that is the alternative to saying that the market is in a bubble. This can be very useful in deflating the bubble because it will force every investment fund manager to deal with the argument. With the 90s tech bubble and the 00s housing bubble, the investment managers could get away with saying they didn’t pay attention to the small number of naysayers. They can’t get away with saying that they didn’t pay attention to the research that was being cited by the chair of the Federal Reserve Board. If they have an answer to it, fine. Maybe they will claim that the Fed is hugely underestimating future growth. That’s always possible, but a rather strong claim. Alternatively, maybe they would say that the Fed is missing a massive shift from wages to profits, going far beyond what we have already seen. Again, this is possible, but they would be painting a very dark picture of the world that does not seem to be widely shared. In any case, since the collapse of the AI bubble will have enormous consequences for financial markets and the economy, and these consequences will only become more severe as the bubble grows

further, it should be the Fed’s responsibility to try to rein it in. Unfortunately, since the new chair likely views his main responsibility as keeping Donald Trump happy, we shouldn’t anticipate that he would go this route. But this is one of the options that is on his table if he chooses to use it. Thanks for reading! Subscribe for free to receive new posts and support my work.

AI驱动知识工作即时化

AI在工作场所带来的最深刻影响，并非简单的自动化，而是对认知准备周期的极致压缩。传统知识工作依赖大量“认知库存”——包括研究笔记、构思草稿、背景阅读等预备材料，以缓冲从构思到产出的时间差。而如今AI工具能在数小时内生成高质量初稿，将原本需要数天甚至数周的准备过程急剧压缩。 这种变革正将知识工作推向“即时生产”模式：管理者推后决策、截止期限不断收紧、准备缓冲逐渐消失。正如制造业中的准时生产系统在提升效率的同时剥离了供应链韧性，AI在优化知识工作流程时，也可能系统性消解那些曾被视为“冗余”的认知缓冲——包括深度思考、错误校验和独立判断的时间。当人类角色从内容生成者转变为高压下的AI输出审核者时，工作的认知结构已发生根本转变。 更深层的风险在于，组织可能将人类认知逐步异化为按需调取的“即时资源”，在最小库存、最小冗余、高度压缩周期的模式下运行。这虽然创造了效率杠杆，却也构建了新的结构性脆弱。我们或许终将意识到，知识工作中那些曾被视为浪费的“冗余”——充分准备、思维沉淀与背景积累——恰恰是组织应对不确定性的关键韧性所在。 #AI应用 #职场变革 #组织管理 #知识工作 #技术风险 #即时生产 #认知科学

AI 深入专业领域

https://careersafter.ai/

随着大模型技术成熟，人工智能正从通用助手快速渗透到具体职业场景。OpenAI 在 ChatGPT 中上线个人理财功能，通过 Plaid 连接逾 1.2 万家金融机构，为美国 Pro 用户整合消费、账单与净资产数据，使其能基于自身账户进行预算咨询。同期，Anthropic 为 Claude 添加了 20 余个法律领域连接器与 12 个实践插件，覆盖合同审查、法律研究、电子取证及文档管理等流程。微软亦在 Word 中推出基于 Frontier 平台的法律代理，可依据内部剧本审核合同并标记条款。此外，企业服务智能体 Agentforce Sales 已能处理从潜客挖掘到报价生成的全套销售任务。这些进展表明，AI 不再局限于生成文本，而是直接接入专业软件和金融账户，执行高价值、情境化的具体任务，深刻改变法律、金融和销售等行业的协作模式。 #AI应用 #法律科技 #金融科技 #职业未来 #大模型落地 #OpenAI #Anthropic #微软

马斯克百亿美元布局Coding Agent，强调其对AI模型训练的战略必要性

马斯克与AI公司Anthropic关系持续紧张，xAI的Cursor账号曾被限制使用Claude模型，促使马斯克寻求自主突破。随后，SpaceX宣布与编程工具Cursor达成战略合作，投资高达100亿美元，旨在获取高质量训练数据。此举背后是AI行业的共识：coding agent产品能提供宝贵的过程监督信号和实时强化学习数据，是模型厂商开发强大编程模型的关键路径。Cursor通过其产品展示了数据飞轮的优势，例如Composer模型利用实时RL从用户交互中迭代优化。此合作凸显了在竞争激烈的AI领域，拥有自研coding agent对数据获取和模型优化的深远影响。 #马斯克 #CodingAgent #AI #模型训练 #SpaceX #Cursor #强化学习 #编程模型 #科技新闻

AI编程智能体成功率提升关键：Anthropic与OpenAI实验证明Harness工程是核心

最近，Anthropic与OpenAI不约而同地通过实验指出，提升AI编程智能体成功率的关键并非更换更强大的模型，而在于为其配备完善的“Harness”工程基础设施。Anthropic的实验显示，同一个模型在裸跑时仅花费9美元但几乎无效，而加入完整的Harness验证循环后，虽然成本增至200美元，但任务完成率从极低水平飙升至100%。OpenAI的百万行代码库实验也得出类似结论，仅通过在仓库根目录添加一个配置文件，就能显著改善智能体表现。 Harness是一套围绕AI智能体构建的工程体系，包含指令、工具、环境、状态与反馈五个子系统，旨在解决智能体常见的过早宣告完成、上下文焦虑导致质量下降以及跨会话失忆等失败模式。业界共识是，模型能力决定了性能上限，而Harness决定了能实际利用到多少上限。目前，除Anthropic和OpenAI外，DeepSeek也开始招聘Harness工程师。建议开发者通过添加项目约定文件、设定安全权限、固化开发环境、维护任务状态文件以及定义可机器验证的完成标准这五个步骤，为智能体安装这套基础设施，以提升开发效率和代码质量。 #AI #编程智能体 #Harness工程 #Anthropic #OpenAI #DeepSeek #软件开发 #AI编程

前调查记者白兆东在泰国失踪半年，疑遭跨国绑架

据维权网消息，曾供职于《财经》杂志的知名调查记者白兆东，自2025年11月27日在泰国发出最后信息后失联，至今已超过半年，音讯全无。报道指出，其失踪并非源于最初怀疑的当地纠纷，而是与中国当局和泰国方面实施的跨境行动有关。白兆东因长期从事深度报道，揭露中国官场与商界的权钱交易黑幕而广受关注。 报道进一步指出，近年来在泰国寻求庇护的中国异见人士和维权者，持续面临来自北京方面的严重安全威胁。相关行动常利用中泰紧密关系，通过所谓的“联合办案”或非正规程序进行。文中列举了多起过往类似案例，包括2019年记者邢鉴、2016年媒体人李新、2015年香港书商桂民海以及获联合国难民身份仍被遣返的姜野飞等人，均曾据称在泰国遭中方人员控制并带回国内。