NoaX

Alice Huang: From Chinese Television to a $4.8 Billion AI Chip Business and the Jensen Huang Scandal If this were a spy thriller, people would probably complain that the plot was too unrealistic. A woman reportedly starts her career in Chinese television, moves into private banking and investment management, becomes involved in a Singapore-based AI computing company, and eventually develops a relationship with the CEO of the world's most valuable AI chipmaker. Now she's at the center of an international controversy involving Nvidia, US intelligence, Chinese technology companies, and billions of dollars in advanced AI hardware. Her name is Alice Huang, also known as Huang Le. And her career is arguably one of the most fascinating parts of the entire story. But there's an important distinction to make upfront: while US intelligence officials have reportedly suspected her of working for Chinese intelligence, those allegations have not been established in court. No public espionage charges against her have been reported. So what do we actually know? 1. A surprisingly unusual career path According to a profile attributed to Huang in a Chinese private investment database, she spent approximately four years working in television journalism, including reported positions at Hunan TV and China Central Television. She then reportedly moved into finance. Her résumé lists a position in HSBC's private banking division, followed by a vice president role at CITIC Securities. That's already a pretty unusual transition. From journalism to international banking, then into China's securities industry. However, these early career claims remain difficult to independently verify. The New York Post contacted a Hunan TV executive who said he was unfamiliar with her. The newspaper also couldn't find corresponding journalist profiles on Hunan TV or CCTV's websites. HSBC and CITIC Securities did not publicly confirm the positions. So the résumé is interesting, but it shouldn't be treated as fully verified. 2. The Shanghai investment connection The next part of her reported career takes us into Shanghai's private investment industry. A Chinese fund database lists a fund manager named Huang Le at Shanghai Hexi Investment Management. One fund, Hequ Shengcai Steady Growth, was established on September 20, 2016, with Huang Le and Kang Wei listed as investment managers. This gives us a more concrete financial record, although the name alone does not independently establish every detail of her earlier résumé. What's particularly interesting is Hexi's reported investment connection to 7Road, a Chinese gaming company. Because 7Road eventually became connected to the Singapore company at the center of this entire controversy. Megaspeed. 3. From a Chinese gaming company to Singapore's AI infrastructure industry Megaspeed wasn't originally established as a major AI computing business. Its predecessor, 7Road International, was incorporated in Singapore in 2015 and had connections to the Chinese gaming industry. In March 2023, the company changed its name to Megaspeed. Around the same time, Alice Huang appeared in corporate filings as a managing director. Ownership was transferred to an entity called Huang Le Limited. Then, just eight months later, another major change occurred. In November 2023, ownership moved to a Singapore company called Swiftdata, and Huang had stepped down from the board. That's a remarkably short period for such significant management and ownership changes. Of course, corporate restructuring is not evidence of wrongdoing. But the timeline matters because Megaspeed subsequently became a major participant in the Nvidia AI infrastructure ecosystem. And the scale of its business was extraordinary. 4. The $4.8 billion GPU business This is where the story becomes much bigger than one executive's résumé. According to investigations based on corporate and international trade records, Megaspeed and its affiliated companies acquired billions of dollars worth of Nvidia-related AI equipment. One reported figure puts purchases involving its Malaysian affiliate, Speedmatrix, at more than $4.8 billion over roughly two years. Bloomberg separately analyzed customs records showing Megaspeed-related imports involving approximately 136,000 Nvidia GPUs through November 2025. Think about that. A relatively obscure Singapore-based computing company had become associated with an AI hardware supply chain worth billions of dollars. Its business involved purchasing servers, deploying computing infrastructure, and providing access to GPU capacity. Some of its reported customers were Chinese technology companies. And that placed Megaspeed in a particularly sensitive position. The United States has spent years restricting China's access to advanced AI chips. But providing computing services through overseas data centers raises a more complicated question. If a restricted customer cannot directly purchase certain GPUs, what happens when it accesses those GPUs remotely through a third-party cloud provider? Where does legitimate international cloud computing end, and where might export-control circumvention begin? Those are precisely the kinds of questions US regulators have been investigating. The reported equipment purchases also shouldn't be confused with Alice Huang's personal wealth. The figures refer to business transactions involving companies in the Megaspeed network. 5. A late-night gathering in Taipei Now comes the part that sounds like it belongs in a movie. June 2024. Computex was taking place in Taipei, bringing together executives from Nvidia, hardware manufacturers, cloud providers, and other technology companies. At a late-night industry gathering, Alice Huang was reportedly socializing with several executives. According to Parker Schmitt, CEO of NetThunder, she suggested that she could get Jensen Huang to join them. And shortly afterward, Jensen actually arrived. Black leather jacket, accompanied by two Nvidia executives. They drank whiskey, talked with people at the gathering, and appeared together in photographs later posted on LinkedIn. Nvidia offered an important clarification: Jensen's appearance was already part of the event arrangements, rather than an unexpected visit solely prompted by Alice Huang. Still, the photographs established that they were moving within the same industry network. And this wasn't their only publicly documented encounter. 6. Another appearance in Taiwan In May 2025, Jensen Huang attended another business dinner in Taipei involving AI industry executives and suppliers. As usual, reporters gathered outside to speak with him. He answered questions, interacted with the crowd, and handed out food from the restaurant. In footage broadcast by Taiwan's SET News, Alice Huang could be seen standing behind him. At the time, the footage attracted little unusual attention. Jensen frequently attends dinners with partners and suppliers, especially during his visits to Taiwan. But once the intelligence allegations became public, those earlier photographs and videos were suddenly reexamined. What had previously appeared to be ordinary industry networking became part of a much larger story. 7. October 2026: The intelligence allegations emerge On October 29, The New York Times reported that US intelligence agencies had been investigating Alice Huang's possible connections to Chinese intelligence. Later reporting described an intelligence memorandum alleging that China's Ministry of State Security had recruited her to gather information related to Jensen Huang and Nvidia. The concern reportedly involved Nvidia's business plans, advanced AI chips, and US export restrictions on China. The story became even more serious when reports emerged about a personal relationship between Jensen Huang and Alice Huang. On November 8, Jensen publicly acknowledged that he had been unfaithful to his wife and apologized to his family and Nvidia employees. Nvidia's board announced an independent investigation into reports that its CEO had been targeted by a foreign intelligence effort. The company said it had no reason to believe national security information or Nvidia trade secrets had been compromised. Meanwhile, Singapore's Ministry of Home Affairs issued an explicit clarification. It said Singapore had not investigated Alice Huang over her alleged intelligence ties or her relationship with Jensen Huang. The Singapore police did confirm an ongoing investigation involving Megaspeed, but that is a separate matter from the personal espionage allegations. These distinctions are extremely important. A business relationship is not evidence of espionage. An unusual corporate structure is not proof of illegal activity. And an intelligence allegation, however serious, is not the same as an established finding of guilt. 8. What makes this story so fascinating? If you put the timeline together, the career is remarkable. A reported beginning in Chinese television. A move into private banking and securities. A connection to Shanghai's private investment industry. Corporate links involving Chinese gaming businesses. A management role at a Singapore AI computing company. Billions of dollars in Nvidia-related hardware purchases. Industry gatherings in Taipei. And finally, a controversy involving one of the most influential executives in modern technology. It's a story that brings together industries we normally think of as completely separate. Media, finance, investment, semiconductors, cloud infrastructure, and geopolitics. And I think the most interesting part isn't even the personal scandal. It's how these commercial networks connect. How does someone move between so many industries? How do relationships in investment and finance translate into opportunities in AI infrastructure? How does a relatively unknown company become involved in billions of dollars of advanced computing equipment? And how much do personal networks matter in an industry where access to GPUs can determine who gets to build the next generation of AI? We spend so much time discussing Nvidia's GPU architectures, CUDA ecosystem, and technological advantages. But the global AI industry also runs on corporate relationships, financing arrangements, regulatory decisions, and personal networks that most people never see. Jensen Huang spent decades building Nvidia into one of the most important companies in the world. And now, a personal relationship involving someone from its extended commercial ecosystem has become the subject of an international investigation. The allegations against Alice Huang remain allegations, and many details of her early career are still unverified. But the documented business connections alone reveal something fascinating about the AI industry. Behind the GPUs, data centers, and trillion-dollar valuations is a complicated network of people, capital, companies, and governments. And sometimes, the most consequential stories in technology have very little to do with the technology itself.

我其实挺赞同这个观点的。 最近看了太多关于 AI Conference、Peer Review 和学术评价体系的讨论，越来越觉得，很多时候我们其实没有那么多选择。 我们可以批评现在的科研越来越功利，论文越来越多，Reviewer 越来越不负责任，甚至很多研究已经变成了一场围绕 Deadline 的工业化生产。 这些问题确实存在。 但作为一个普通 PhD，我们既没有能力改变整个学术评价体系，也很难改变导师、学校和招聘市场对 Paper 的要求。 现实就是，你想毕业，想拿到更好的机会，很多时候还是需要论文。 所以我越来越觉得，年轻 PhD 应该学会一种很重要的能力：在理解规则的同时，尽可能高效地完成规则要求的事情。 能用 AI 加速的就用 AI，能自动化的就自动化，能提前规划的就不要拖到 Deadline。 把那些不需要太多创造力的重复工作尽可能压缩，把时间留给真正值得思考的研究问题。 而且我觉得，科研理想和现实主义本来就不应该是对立的。 你可以一边认为现在的 Conference System 有很多问题，一边认真准备下一篇投稿。 可以一边批评 Paper 数量不应该决定一个研究者的价值，一边努力发表足够多的论文，让自己拥有更多选择权。 这不是什么虚伪。 很多时候，只有先让自己在这个系统里生存下来，才有机会去做那些真正想做的事情。 当然，我也不希望一个 PhD 最后只剩下发论文这一件事。 毕竟我们读博，最开始应该还是因为对某些科学问题有真正的好奇心。 我们未必有能力改变科研界的游戏规则，但至少可以努力不让这些规则消耗掉自己全部的时间和热情。 先学会在现实里生存，再尽可能给理想留一点空间。 我觉得这也是一种成熟

看了一位台湾半导体分析师 Andrew Lu 对黄仁勋事件的分析，感觉这个角度还挺值得琢磨的。 他的观点是，这次风波真正可能影响的，远不止黄仁勋的私人生活，而是美国过去一直没能彻底解决的三个问题。 第一，削弱黄仁勋在华盛顿的游说影响力。 过去几年，黄仁勋一直主张不要彻底切断 NVIDIA 对中国市场的芯片出口。 他的逻辑也很简单：如果美国公司主动退出中国市场，最后可能只是把市场份额拱手让给华为等中国企业。 但现在不一样了。 一旦他被卷入涉嫌中国情报活动的争议，以后再公开支持放宽对华芯片出口限制，政治成本就会高得多。 原本是一场正常的商业政策争论，现在很容易被上升到国家安全层面。 第二，重新收紧 H200 对华出口。 这个时间点也确实很微妙。 就在相关风波持续发酵的时候，《纽约时报》又调查了部分获准处理 NVIDIA H200 芯片的中国企业。 记者实地走访发现，有些公司的注册地址竟然是空办公室，背后股东还与美国制裁名单上的企业存在关联。 这就让美国商务部之前的出口审批再次受到质疑。 如果这些问题继续发酵，原本已经有所松动的 H200 出口政策，可能再次面临收紧压力。 第三，让 NVIDIA 接受更严格的国家安全审查。 现在 NVIDIA 董事会已经启动独立调查，美国国会也要求黄仁勋就相关问题进行闭门说明。 对于一家掌握全球关键 AI 算力供应的公司来说，这件事可能意味着未来在客户审查、芯片流向、出口合规等问题上，要面对更严格的监督。 我觉得这位分析师最有意思的判断，是他认为这三件事并不是孤立的。 黄仁勋的个人争议、H200 出口漏洞，以及美国对中国 AI 算力发展的限制，最后可能汇集成同一场政治博弈。 当然，这套分析目前还只是对政治动机的推测。几件事情恰好在相近时间发生，并不能证明背后有人统一策划，更不能认定黄仁勋已经泄露了什么机密。 但它确实让我想到一个问题。 过去我们一直认为 NVIDIA 最大的竞争优势来自 CUDA、GPU 架构，以及它在 AI Infrastructure 上建立的生态。 可当一家公司重要到足以影响国家的技术竞争格局时，它的商业决策就很难再完全由市场决定。 黄仁勋可以决定 NVIDIA 如何竞争，但他未必能够决定 NVIDIA 被允许参与怎样的竞争。 AI 发展到今天，技术、资本和地缘政治之间的边界，已经越来越模糊了。

Jensen on Monday be like

Strongly agree. Jensen made a serious mistake, and he should take responsibility for it. Any national security concerns deserve a thorough investigation. But we shouldn't confuse accountability with erasing someone's contributions. Jensen has spent decades helping build the computing infrastructure that made modern AI possible. You can acknowledge his mistakes while recognizing how extraordinarily important his leadership has been. Holding someone accountable and recognizing their value are not mutually exclusive.

关于design。我还可以说一嘴。 我自己是非常喜欢design 各种motion 和 动画效果。我个人对审美有很高的要求，但是我不是一个设计师，我只是工科出身的学生，所以我们要诚实承认 不擅长设计，但是不擅长设计不等于，不具备好的对美的感知和taste，所以我们要平时尽可能多搜集好看的design，然后丢给agent 让他提取总结。 前一段很火的个人知识库，我用不太上，但是我随手mark good design的习惯，让我有一个 属于我自己的design库。 哪怕我不是designer，我也能design出好的产品。我认为design本身是非常溢价的行业，而且是有非常好的designer，但是大部分designer是非常普通的，我的这套打法，可以对齐大部分中上水平的designer。

我来告诉你 你不需要看99%的ai 教程 就两个rule： 1. 把自己的想法录下来，或者写下来，假设你是在和一个你最要好的朋友分享自己要做的东西，这东西 你可以说的非常繁琐。但是必须是你自己的意思。整体方向可以含糊，但是你在乎的点要非常细节 2. 和 gpt / claude 讨论，说自己不擅长这个领域（如果你不擅长开发），希望让他一步一步指导你，构建你想做的产品，把整个plan 清晰 先列一遍，然后再反复和agent 讨论。逐步实现，如果遇到你不知道名词，自己先研究一下，然后或者 让agent 去开源社区找替代方案 or 优秀的解法，迁移过来，然后让agent按推荐的来。 我的这个方法，可以解决90%的中小型SaaS开发问题，对于design，则是让他设计非常多的artifacts，你去找好的设计，让他去模仿，同时设计很多版本，然后你来挑选。

在读 PhD 的过程中，我目前对于一些词语已经有生理性的反感： Novelty 不足 Missing Baseline Insufficient Ablation Lack of Theoretical Justification Limited Generalization Reviewer 2 Minor Revision 再补一个实验 这个很快就能做完 下个 Deadline 赶得上 你这个 Idea 可以发 Nature 导师想跟你 Quick Chat 一下 Funding 快没了 毕业的事情我们下次再讨论

黄仁勋公开承认婚外情，更敏感的是，涉事女性还被美国情报部门怀疑与中国情报机构有关。 看了几篇报道，发现这件事其实早就有迹可循。 美国政府此前就已经收到相关情报，特朗普也曾被告知此事。但直到最近，媒体才开始密集披露更多细节，如今 NVIDIA 董事会也正式启动了独立调查。 最让我感慨的，其实是这个时间点。 美国中期选举刚刚结束，市场正在重新评估 AI 基础设施的政策风险，科技股也经历了剧烈回撤。 偏偏在这种时候，全球 AI 算力产业最核心公司的 CEO，又陷入了一场涉及国家安全的争议。 而且接下来就是深圳 APEC 峰会，美中之间的芯片出口和 AI 竞争本来就是非常敏感的议题。 当然，新闻集中曝光不代表背后一定存在政治操作。 资本市场有一个很有意思的特点，上涨的时候，大家似乎总能找到新的理由相信未来。而一旦市场开始下跌，过去被忽略的风险就会一个接一个浮出水面。 现在关键的问题，已经不只是黄仁勋的私人生活。 如果最终只是个人道德问题，影响更多可能集中在声誉和公司治理层面。但如果调查发现确实存在情报渗透或敏感信息泄露，那对 NVIDIA 的管理层、美国芯片政策乃至整个 AI 产业的影响，可能就完全不同了。 目前还没有证据能够证明事情已经发展到那一步。 只是最近这一连串事件，让我越来越清楚地意识到： 我们习惯用 Scaling Law、GPU 和模型能力理解 AI 产业，却经常低估政治、国家安全和人的因素，究竟能在多大程度上改变一家科技公司的命运。 技术可以持续进步，但资本市场对未来的信心，可能比我们想象得脆弱得多

申请博士时，不同家庭背景的人，面对的根本不是同一套选择题。 有些人，首先考虑的是 Cambridge、Oxford，哪个导师更好，哪个方向更有意思。 但对于另一些人来说，第一件事可能是看 Funding。 有没有全奖？Stipend 够不够生活？需不需要 TA？读博期间能不能负担房租，家里有没有能力承担突发开销？ 甚至有时候，一个人再喜欢某个研究方向，没有 Funding，也只能放弃。 我觉得这件事最令人感慨的地方，是大家可能都在讨论科研理想，但对风险的理解完全不同。 对于家庭条件优越的学生，读博几年没有收入上的明显增长，可能只是一种人生探索。 但对于没有经济支持的人，这几年可能意味着推迟经济独立，放弃其他职业机会，还要承担未来就业的不确定性。 而且这种差异会进一步影响科研本身。 一个不需要担心生活费的 PhD，可能更有余裕去探索高风险、长期没有成果的科学问题。 另一个学生却可能需要尽快发 Paper、毕业、找到工作，因为他根本承担不起太长时间的不确定性。 所以有时候我们看到两个研究者最后取得了不同的成绩，很容易归结为天赋、努力或者科研能力。 但我们往往看不见，他们从一开始能够承受的失败次数，就完全不同。 我并不觉得家庭富裕有什么问题。 只是我越来越觉得，很多看起来属于个人选择的事情，背后其实早就存在非常不同的经济约束。 有些人可以自由选择自己最热爱的研究。 另一些人必须先确认，自己有没有经济条件去热爱它。 而我希望，至少在科研教育里，一个真正有潜力的年轻人，不应该因为家庭无法承担读博成本，就失去探索科学的机会

Nature 最近报道了一件挺有意思的事。 一家叫 XSci 的平台开始给审稿人发钱了。 5 天内完成 peer review，拿 $200。10 天内完成，拿 $100。钱由希望获得快速评审的论文作者支付。 我觉得这件事最大的意义，倒不是博士生终于可以靠审稿赚点零花钱了。 而是学术界一个长期被忽略的矛盾，终于开始有人尝试用市场机制解决。 你想想，学术出版其实是一个非常奇怪的商业体系。 研究者花几年时间做实验、写论文，通常不拿稿费。 审稿人花自己的时间免费审稿。 编辑和出版社负责组织整个流程，最后论文可能还需要作者支付几千美元的 APC。 整个系统最核心的知识生产和质量判断，很大程度上依赖研究者的劳动。 而且，我们一边要求审稿人严谨负责，一边又希望他们在繁忙的科研工作之外，无偿、快速地完成这些任务。 这本身就是一种激励错配。 但我觉得，付费审稿真正有意思的地方，是它让一个以前很难讨论的问题浮出水面： 我们到底是在为审稿人的时间付费，还是在为审稿判断的质量付费？ 两者完全不是一回事。 如果按速度付钱，审稿人自然有动力更快完成工作。但快不意味着好，甚至可能催生更加模板化、浅层化的审稿意见。 更重要的是，当论文作者成为直接付费的一方，平台又如何保证审稿人的独立性？ 我其实支持给审稿人合理报酬，但前提是建立相应的质量控制与利益冲突机制。 尤其在 AI 时代，LLM 已经可以快速生成一份看起来相当专业的 review。 未来真正稀缺的，可能不再是写出几页审稿意见的能力，而是一个有领域判断力的人，愿意为自己的学术判断承担责任。 AI 可以让审稿意见越来越便宜，但不应该让可信的学术判断也随之贬值。 说到底，同行评审真正值得付费的，从来不是那几页文字，而是文字背后多年积累的专业判断。

One of the most important skills everyone should learn is sales. Not because everyone needs to become a salesperson, but because most people have no idea what their skills are actually worth. You might be an amazing engineer, researcher, or designer. Your employer pays you a salary, so naturally you assume your skills are valuable. But here's the thing: being good at something and getting someone to pay for it are two completely different skills. Your employer has already figured out the business model for you. They've found the customers, identified the demand, and built a system that turns your work into revenue. You just happen to be one part of that system. The moment you step outside it, you realize something uncomfortable: nobody pays you simply because you're talented. They pay because you solve a problem they care about. And sometimes that problem isn't even technical. People pay for convenience, confidence, peace of mind, entertainment, or simply feeling understood. That's why I think everyone should try selling something at least once. Talk to real customers. Ask what frustrates them. Find out what they've already tried and why they're willing to spend money fixing it. You'll probably learn more about the market in ten customer conversations than in months of building something nobody asked for. Sales isn't just about convincing people to buy. It's how you discover what people actually value. And until you understand that, you don't really have a business. You just have a skill you're hoping someone will pay for.

大学越来越擅长衡量知识的生产，却越来越难以衡量知识的传承。 一篇论文有多少引用、发表在哪个期刊、带来了多少经费，这些都有明确的指标。 但一个教授是否改变了学生的思考方式，是否培养出了能够独立提出问题的研究者，是否让一个原本平庸的学生发现了自己的潜力，这些几乎无法被量化。 于是我们看到一种很奇怪的现象： 一个教授可以因为持续发表高水平论文而获得晋升，却未必需要证明自己真正培养出了优秀的学生。 而一位愿意投入大量时间指导学生、认真教学的教授，其付出的机会成本反而可能体现在更少的论文和经费上。 这不是说科研不重要，而是大学正在面临一个典型的 Goodhart's Law 问题：当一个指标成为目标，它就逐渐失去了衡量真实价值的能力。 更讽刺的是，大学本应是最懂得长期价值的机构，却越来越依赖短期可量化的产出来分配资源。 一所大学真正的学术遗产，不仅是它发表过多少论文，更是它培养出来的人，在离开这所大学之后，还能创造多少知识

今天看到一个很有意思的概念：硅谷短期主义（Silicon Valley Short-termism）。 逻辑很简单。 按照一些 AI 乐观主义者和末日论者的设想，20 年后，人类要么不再需要工作，要么连人类文明能否延续都成了问题。 如果未来真的如此不可预测，那我们今天奉行的长期主义，还有多少意义？ 花十年积累某项专业技能，可能五年后就被 AI 彻底重构。为了几十年后的职业发展牺牲当下，可能连职业本身都会消失。 当然，这种预测未必正确。但它确实提出了一个值得思考的问题： 当技术变化的速度超过个人规划的时间尺度，长期主义本身是不是也需要重新定义？ 我一定程度也觉得，与其把人生押注在一个遥远且不确定的未来，不如不断缩短行动与反馈的周期。 有想法就去验证，有机会就去尝试，有价值就尽快创造。不必等所有条件成熟，也不必为了一个十年后的理想版本，持续牺牲当下的可能性。 这并不意味着放弃积累，而是不再执着于某一条固定的积累路径。 毕竟，长期主义真正重要的，从来不是坚持同一个计划十年，而是持续保有在任何时代重新开始的能力。 每天都奉行短期主义，某种程度上，反而是一种更适合 AI 时代的长期主义。

对 PhD 最大的误解，是以为博士训练的终点是掌握某个领域的知识。 其实不是。 本科和硕士更多是在学习如何解决已经被定义的问题，而 PhD 真正困难的部分，是你必须学会在没有标准答案、甚至不知道问题是否成立的情况下，独立定义问题、设计实验并建立证据。 但这里存在一个很微妙的矛盾。 学术界一方面要求博士生成为 independent researcher，另一方面，博士生的研究方向、经费、算力、合作资源乃至毕业时间，往往都深度依赖导师。 我们要求一个人在高度依赖的组织结构中，训练出完全独立的研究能力。 所以我认为，衡量 PhD 训练是否成功，不能只看他能不能独立完成一篇 paper。 更应该看：当导师不再提供题目、资源和研究路线时，他是否还知道什么问题值得研究，以及如何找到实现它的路径。 学术独立性不是你能独自完成多少工作，而是你能否独立决定什么工作值得完成。 一个博士生可以非常擅长执行研究，却始终没有真正学会做研究

一个在经济、工业和民用科技领域受到诸多限制的国家，却能够组织起足以威胁全球大型金融平台的网络攻击力量。 这恰恰说明，一个国家的综合技术水平，并不决定它在某个特定领域能够达到的能力上限。 传统科技竞争需要庞大的工业基础、资本积累和供应链体系。但网络攻防不同。 攻击者不需要建立一个比对手更先进的金融系统，只需要找到对方系统中一个足够关键的薄弱环节。 防守方需要保护数万名员工、复杂的权限体系、持续变化的软件依赖和整个基础设施。 进攻方只需要成功一次。 而朝鲜的特殊之处在于，它可以把网络攻击作为一项长期的国家战略，集中培养人才、积累经验，并持续寻找高价值目标。 这不是简单的技术创新，而是对现代数字经济脆弱性的一种系统性利用。 更值得警惕的是 AI Agent 的出现。 过去，入侵一个系统之后，攻击者仍然需要大量人工操作。而未来，当 Agent 被授予越来越多的代码、生产环境和资产操作权限时，一次权限失控的影响范围可能进一步扩大。 AI 正在降低创造经济价值的成本，也可能在降低破坏经济价值的成本。 朝鲜这个案例真正提醒我们的，是现代社会的经济安全，越来越取决于那些最难被完全消除的系统性薄弱环节。 一个国家不需要拥有最先进的科技体系，也可能对另一个高度数字化的经济体造成巨大的损失。技术越发达，系统越复杂，局部脆弱性所能产生的全球性影响就越值得警惕

博士学位真正应该证明的，不是你有能力为一个研究体系持续生产成果，而是你已经具备了不依赖这个体系，也能创造新知识的能力。 PhD 训练里最值得警惕的，不是沦为 paper machine，而是逐渐失去定义问题的能力。 很多博士生一开始是带着科学好奇心进入学术界的，几年以后，却开始习惯性地用审稿人的偏好决定研究方向，用 benchmark 的提升判断研究价值，用论文数量衡量自己的成长。 最可怕的不是这些指标本身，而是你逐渐把它们内化成了自己的价值观。 但我也不认同另一种极端：认为不追求发表、只做理论和方法论，就天然比做工程和应用更高尚。 真正重要的工作，从来不由研究形式决定。有些理论论文十年后无人问津，有些看似工程性的突破，却彻底改变了整个领域的研究范式。 科研最重要的自由，不是摆脱评价体系，而是不让评价体系替你决定什么问题值得研究。 我认为，PhD 最应该积累的资产，不是论文数量，而是三种能力：独立定义问题的能力、把想法变成可验证结果的能力，以及离开某个导师、实验室或学术体系之后，依然能够继续研究的能力。 如果一个博士读了五年，发表了十几篇论文，却始终不知道离开导师之后自己想研究什么，那我很难说这是一种成功。 反过来，如果一个人只发表了两三篇论文，却建立了独立的研究判断、扎实的技术能力和长期探索问题的勇气，我反而觉得他的 PhD 训练可能更加成功

学术圈里有一种被严重低估的能力：懂得什么时候不展示自己的聪明。 刚开始做研究时，总觉得学术能力需要被看见。读了多少论文、提出了什么新想法、发现了别人方法里的什么漏洞，都恨不得在每一次组会和讨论中证明自己。 但后来逐渐发现，学术圈并不是一个纯粹按照智力和贡献分配认可的地方。它同样存在竞争、利益、资源分配，以及非常复杂的人际关系。 你指出一个人的错误，未必能让他感谢你；你公开展示自己的成果，也未必能换来真诚的认可。有时候，过早暴露自己的判断和野心，反而会让合作变得更加困难。 我觉得一个成熟的研究者，需要学会区分两件事：追求真理，和证明自己比别人更接近真理。 前者是科研，后者很多时候只是自我满足。 真正有能力的人，不需要在每一次讨论中赢得争论，也不需要让所有人知道自己有多聪明。 该坚持学术原则的时候坚持，该承认不知道的时候承认。对别人的工作保持尊重，对自己的成果保持克制。 但这不意味着一味示弱，更不意味着面对学术不端和不公平时保持沉默。 学术成熟的标志，不是学会隐藏自己的能力，而是不再需要通过展示能力来确认自己的价值。 毕竟，PhD 最终要训练的，是独立发现问题和创造知识的能力，而不是在每一个房间里成为最聪明的那个人

我觉得这里有一个很值得思考的问题：一个社会的幸福感，究竟来自更高的效率，还是更低的生存压力？ 东亚社会习惯把努力、竞争和责任感视为美德，但我们很少讨论，持续追求效率的代价最终由谁承担。 不过，丹麦式的松弛未必是幸福的原因，也可能是社会保障和制度信任所带来的结果。 真正令人羡慕的，不是一个人可以不努力，而是一个人即使不那么努力，也不必为此付出灾难性的代价。 当然，个人可以选择松弛，但公共系统不能把低效率的成本转嫁给所有人。

A friend of mine thought paying $200/month for AI subscriptions was too expensive. So he spent over $20,000 on two Mac Studios to run open-source models locally. After days of setting everything up, he finally got everything running. Only to realize the models weren't even as good as the $200/month ones. He tried to save on subscriptions and ended up prepaying for almost a decade of them. And the best part? He now has a second job maintaining his own AI infrastructure. Bro tried to escape the subscription economy and accidentally became his own cloud provider.