本文以北京一名药企经理 Stacey Tang 在 2022 年 11 月（北京新冠封控高峰期）接到「天才班」甄选通知为引子，描述中国覆盖面极大的理工人才分流体系：每年约 100,000 名天赋突出的青少年被选入重点高中的「实验／竞赛」班，通常在 16–18 岁进行高强度训练，目标是数学、物理、化学、生物与计算机等国际竞赛并争取升学捷径。中国以国家力量扩大 STEM 供给：据新华社，中国每年培养约 5,000,000 名 STEM 专业毕业生，而美国约 500,000 名；同时，改革让部分竞赛优胜者可绕过 gaokao，进入「985 工程」的 39 所顶尖高校，但每年最终能直接录取者仅约 3%。在投诉压力下，政策于 2025 年底收紧：仅允许全国竞赛获奖者中的前 10% 直接获得清华与北大的录取资格。

文章将上述人才管线与 AI 竞争力直接连结，指出多家挑战美国技术优势的企业（如 ByteDance、Taobao、PDD、Meituan、Cambricon、DeepSeek、Alibaba 的 Qwen 等）核心人才常出自该体系。以 DeepSeek 为例：2024 年，21 岁的 Wang Zihan 参与开发 V2，作为后续 R1 的基础；在当时「中国因美国出口管制而落后 1–2 年、主要在复制 OpenAI 与 Meta」的主流叙事下，DeepSeek 却以更少的先进晶片训练出世界级推理模型，并公开整个研发流程、提供下载，且团队「超过 100 人」几乎全为本土培养、多人有国际竞赛奖牌。文中亦提到 Jensen Huang 在 2024 年 5 月称中国 AI 研究者「世界级」，并观察到 Anthropic、OpenAI、Google DeepMind 充斥来自中国的研究者；同时，Wang Zihan 转赴美国 Northwestern University 攻读博士，表示美国理工博士中「约一半」为中国学生，但签证不确定性正提高回流意愿。

历史脉络上，国家对科学教育的自上而下动员可追溯至二战后；1958 年「大跃进」强化了「科学＝国力」的公共叙事，而以 International Science Olympiads 为代表的竞赛制度（数学奥赛始于 1959 年）成为选拔与资源配置的核心。中国 1985 年首次派 2 名学生参加在芬兰 Joutsa 举办的国际数学奥赛并获 1 枚铜牌；1986 年派出 6 人完整队伍赴华沙并获 3 枚铜牌；到 2025 年，中国在各奥赛共派 23 名选手，其中 22 人夺金，显示长期的「以赛选才」累积效应。大学端的代表是清华「Yao Class」：Andrew Yao 于 2004 年离开 Princeton 回国创办，每年约 30 人；校方报告称其 2019 届 27 人中 24 人有金牌、3 人为省级 gaokao 第一。案例包括 Lou Tiancheng（Pony.ai CTO；公司 IPO 后估值 US$6.9bn），以及 Dai Wenyuan（2014 年创办 Fourth Paradigm；称中国已有超过 1,000 个已注册生成式 AI 模型）。文章并以趋势收束：计算机竞赛热度超过传统理科，AI 被列为 2017 年「国家增长战略」，仅 2018 年就新增 35 个名称含「AI」的高中与大学专班，人才供给的规模化正在加速外溢到产业竞争力。

The article opens with a Beijing pharmaceutical manager, Stacey Tang, receiving a “genius class” recruitment call in November 2022, during peak Covid-19 lockdowns, and uses it to explain China’s unusually large, state-driven STEM talent pipeline. An estimated 100,000 high-aptitude teenagers are selected each year into science-focused “experiment/competition” classes in top high schools, typically for intensive training at ages 16–18 aimed at international contests in maths, physics, chemistry, biology, and computer science. The scale is reinforced by mass higher-education output: state media Xinhua puts China at roughly 5,000,000 STEM graduates per year versus about 500,000 in the US. Admissions reforms let some competition winners bypass the gaokao and enter elite “985 Project” universities (39 members), but only about 3% of participants secure direct admission annually; policy tightened at the end of 2025 so only the top 10% of national competition prize winners qualify for direct entry to Tsinghua and Peking University.

It then links that pipeline to AI performance, arguing that many leaders and key engineers at firms challenging US tech dominance—such as ByteDance, Taobao, PDD, Meituan, Cambricon, DeepSeek, and Alibaba’s Qwen—are products of the system. DeepSeek is framed as a case study: in 2024, Wang Zihan began an internship at age 21 and worked on V2, a precursor to the foundation behind R1, while the prevailing view in Silicon Valley and Washington held that US export controls had left China 1–2 years behind and largely copying OpenAI and Meta. Instead, DeepSeek produced a world-class reasoning model using significantly fewer advanced chips, made its development process public, and offered R1 for download; its team of more than 100 was described as almost entirely homegrown and heavy with international-competition medalists. The article adds external validation and mobility pressures: Jensen Huang said in May 2024 that Chinese AI researchers across Anthropic, OpenAI, and Google DeepMind are “world-class,” while Wang later left for a PhD at Northwestern University, noting that Chinese students make up about half of US science-major PhDs and that visa uncertainty is pushing more to consider returning.

Historically, China’s top-down emphasis on science education is traced to the post‑war period and intensified after 1958, with Olympiad-style contests (the first in 1959) becoming a durable selection mechanism. China’s results are presented as a quantitative marker: in 1985 it sent 2 students to the International Mathematical Olympiad in Joutsa, Finland and won 1 bronze; in 1986 it sent a full team of 6 to Warsaw and won 3 bronzes; by 2025, across Olympiads, it sent 23 contestants and brought home 22 golds. At university level, Tsinghua’s “Yao Class,” founded when Andrew Yao left Princeton in 2004, admits about 30 students per year; a school report said its 2019 cohort had 27 students, including 24 gold medalists and 3 provincial gaokao top-scorers. The article highlights spillovers into firms and scale claims—Lou Tiancheng’s Pony.ai valued at US$6.9bn after its IPO, Dai Wenyuan founding Fourth Paradigm in 2014 and pointing to more than 1,000 registered generative AI models—and notes that computer science has surged, AI was named a key national growth strategy in 2017, and 35 new “AI”-named special classes were created in 2018 alone.