J. Ling's Blog

The reason I started this blog is simple: there are always many thoughts running through my mind, and I wanted a place to record them.

I first tried sharing these ideas with people around me, but I soon realized that most were not very interested. It is not because they do not care about me. It is because it is naturally difficult for anyone to truly understand the thoughts of others. When we cannot understand something, we often assume it is not relevant or interesting to us. Real understanding usually happens only when two people are shaped by similar experiences and environments.

Geography and time are two of the biggest factors that create differences between people. Where we live and when we live shape our views, our values, and the way we think. By documenting my thoughts online, I hope to overcome some of these barriers. People in other parts of the world may see my ideas, and people in the future may discover them, even if those around me today cannot relate.

I hope that some of these thoughts will be interesting and perhaps even inspiring to those who read them. Thank you for visiting.

我开设这个博客的原因很简单：我的脑海中总有许多想法不断闪现，我想找一个地方把它们记录下来。

起初，我尝试和身边的人分享这些想法，但很快发现，大多数人并不感兴趣。这并不是因为他们不在乎我，而是因为人们本来就很难真正理解他人的想法。当我们无法理解一件事情时，往往会下意识地觉得它与自己无关，也就失去了兴趣。真正的理解，往往只有在两个人经历过相似的环境和体验时才有可能发生。

地理与时间是造成人与人之间差异的两个重要因素。我们所处的地方和时代塑造了我们的视角、价值观和思维方式。通过把我的想法记录在网上，我希望能跨越这些隔阂。或许在世界的另一端，会有人看到这些文字；或许在未来的某一天，会有人与这些想法产生共鸣，即便今天身边的人未必理解。

我希望，这些思考能对看到它们的人来说有趣，甚至可能带来一些启发。感谢你愿意停下脚步，来到这里。

如果你仔细观察美国的头部 AI 公司，比如 Google、Meta、OpenAI、Groq、英伟达，你会发现一个有趣的现象：核心研究人员中相当大一部分都是华人。当你点开这些研究人员的个人主页或 LinkedIn 时，会发现一个非常明显的规律：很多人本科就读于中国的 985 理工科院校，之后前往美国攻读数学、统计、计算机科学、运筹学或数据科学的博士学位。相比之下，美国本土本科生——无论是白人、印度人还是华人——在这些顶尖 AI 团队中的比例要少得多。

造成这种现象的根本原因之一是，美国本科教育在 STEM（科学、技术、工程、数学）领域的整体要求正在不断降低。在美国本科院校，你会看到大量数学或统计专业的学生连基础的导数、积分都不会算，线性代数的矩阵求逆都做不出来，更不用说对学科的学术意义有深刻理解。但这并不是教育的失败，而是美国教育体系的刻意设计：通过设置相对低的 STEM 门槛，让大量水平一般甚至很弱的学生能够顺利毕业并支付学费，而从这些收入中反哺少数极为顶尖的天才学生。大多数普通学生通过低标准的课程拿到学位，皆大欢喜；而天才们则借助充足的资源，把科研的“上限”推到极高的水平。这也是为什么美国在机器学习、统计模型理论等 AI 核心领域，依然领先全球的根本原因：普通学生的“学费输血”给了顶尖学者充足的科研资源。

与之相对，中国的教育模式则几乎是另一种极端。中国长期奉行平均主义的教育思路：不管你是学业最优秀的，还是成绩垫底的，都必须遵守同样的制度和节奏。对于那些最顶尖的学生来说，学校的规则、班级的繁文缛节、各种不必要的集体活动，往往会打乱他们自主学习的节奏。有些人满腹怨言却不得不接受现实，有些人试图突破却被老师批评或被同学排挤。但即便如此，中国顶尖 985 理工科院校依然保持了相对较高的标准。正因为如此，在头部 985 理工科院校的前 25% 学生，其平均智商与学术能力普遍高于美国 top 30 理工科大学的前 25% 学生，大致相当于美国 top 30 理工科院校中前 5%~10% 的水平。

如果用一个形象的比喻，美国的 AI 人才更像 CPU：数量不多，但单核算力极强；而中国的 AI 人才更像 GPU：单个计算核心的性能略逊于美国，但并行数量巨大。AI 的发展既需要 CPU 的高性能，也需要 GPU 的高并发。并不是所有运算都能靠 GPU 完成，但也没有足够的时间让 CPU 慢慢计算完一切。因此，中国和美国在 AI 领域一直存在某种天然的合作关系。通过 CPT、OPT 和 H1B 签证，大量中国的顶尖理工人才进入美国的大型科技公司工作，形成了 AI 人才的“CPU + GPU”双重优势。这也是为什么今天美国依然在 AI 综合实力上领先全球。

然而，这种格局未必能一直持续。如今在高被引 AI 论文中，中文名字的比例正在不断上升，这说明“GPU”的贡献越来越大。如果美国的政策日趋封闭，移民环境越来越民粹化，是否还能长期留住这些中国 AI 研究者？这是一个悬而未决的问题。我相信，如果中国能够提供更好的科研环境与更具吸引力的经济机会，大量如今在美国工作的中国 AI 人才会愿意回国发展。至于原因，我会在后续的文章中展开更深入的讨论。

If you look closely at leading U.S. AI companies such as Google, Meta, OpenAI, Groq, and NVIDIA, you will notice an interesting pattern: a substantial share of their core research staff are Chinese. When you open these researchers’ personal webpages or LinkedIn profiles, another pattern emerges. Many completed undergraduate studies at China’s elite 985 engineering and science universities, then went to the United States for a PhD in mathematics, statistics, computer science, operations research, or data science. By comparison, U.S.-trained undergraduates, whether white, Indian, or Chinese, are far less represented on these top-tier AI teams.

One root cause is that undergraduate STEM standards in the United States have steadily declined. At American universities, you can find many math or statistics majors who cannot compute derivatives or integrals, who struggle to invert a matrix, and who have little grasp of the deeper academic meaning of these subjects. Yet this is not necessarily a failure. It reflects a deliberate design: set relatively low thresholds so that large numbers of average or even weak students can graduate and pay tuition, then use that revenue to support a small cohort of exceptional talents. Most students pass low-bar courses, receive their degrees, and feel satisfied, while the exceptional few, backed by abundant resources, push the research frontier to very high levels. This, I argue, is why the United States still leads in core areas of AI such as machine learning and statistical theory. The tuition paid by the many subsidizes the work of the few.

China’s model is almost the opposite. For a long time, the education system has emphasized egalitarian uniformity. Whether you are top of the class or near the bottom, you are expected to follow the same rules and pace. For the very best students, school regulations, classroom bureaucracy, and compulsory activities often disrupt independent study. Some resign themselves to it, while others try to opt out and end up criticized by teachers or ostracized by peers. Even so, the top 985 engineering programs still maintain relatively high standards. As a result, the top 25 percent of students at these institutions typically show higher average cognitive and academic ability than the top 25 percent at U.S. top-30 engineering schools, and are roughly comparable to the top 5 to 10 percent at those U.S. schools.

Here is a metaphor. U.S. AI talent is like CPUs, few in number but extremely strong per core. Chinese AI talent is like GPUs, slightly weaker per unit but available in very large parallel quantities. AI needs both high single-thread performance and high throughput. Not every computation can be done on GPUs, yet there is not enough time for CPUs to do everything alone. This is why the two countries have formed a natural partnership in AI. Through CPT, OPT, and H-1B programs, many top Chinese engineers join major U.S. tech companies, creating a combined CPU plus GPU advantage. This is also why the United States still leads in overall AI strength today.

That balance may not last. The share of Chinese names among highly cited AI papers has been rising, which suggests the contribution from the GPU side is growing. If U.S. policy becomes more restrictive and the immigration environment more populist, it may struggle to retain these Chinese AI researchers over the long term. I believe that, if China can offer a better research ecosystem and stronger economic opportunities, many Chinese researchers now working in the United States will choose to return. I will explore the specific reasons in future posts.