Kim 等 (2026, arXiv 2609.05395, EMNLP Industry) 提出 KOPA-Bench——145 个真实韩国公开 API 多步任务基准，用 EDGE 数据合成框架把 9B 开源 LLM 拉到接近 27B 未微调模型水平。这是『数据主权法规反向驱动开源供给』的第一次实证：制度环境层（Williamson L2）绕过治理机制直接压 L4 技术栈，开源从审美偏好升级为合规义务——但开源的能力供给缺口尚未填上。
Mitic 等 (2026, arXiv 2609.01275) 首次把前沿 LLM 系统定义为『宪法性制度』，用 23 个前沿 LLM 原型 × 1649 名美国参与者的匹配审计，实证发现前沿 AI 的隐宪只覆盖公众价值需求的 ~2%——37% 的用户『宪法无处可归』。这是阿西莫格鲁『包容性 vs 汲取性制度』在 AI 价值层最锋利的实证：开源从『审美偏好』升级为『包容性制度的必要条件』，闭源主权不是代码，是宪法起草权。
人类治理强自主行动者从来不是靠监控其思维过程，而是靠关键行动上的独立见证——Salfeld-Nebgen 将这一制度模式计算化，为 AI Agent 时代的开源治理提供语法。
GitHub 上 12,110 个 .cursorrules 文件的实证快照——开源治理从人类可读的 CONTRIBUTING.md 跃迁为机器可执行的 AI Agent 规则，定义权悄然易主。
US-China schedule first official AI talks under Trump; Bessent warns of sanctions over AI IP ’theft’; China formalizes WAICO with 29 founding members; SPC overhauls chip-design IP with punitive damages; controversial open-source licensing rulings continue to shape China’s legal landscape.
Qiushi Issue 15 frames open-weight AI as a boon for the world; The Economist analyzes China’s superior AI efficiency; Foreign Policy says US is betting the house on AI race; TechCrunch/AISI reports open-weight models catching frontier; Washington Post says Washington mobilizing against Chinese AI threat.
Taiwan’s hardware giants launch TAIONE foundation to claim open-source AI governance; Chinese AI IPO rush accelerates with DeepSeek and Moonshot; US lead over China in AI disappears; the social costs of AI deployment in China come into focus; DeepSeek pauses $71B funding round after founder’s viral comments; China creates world’s first AI agent regulatory framework.
State media consolidates open-weight AI success narrative; openEuler Hong Kong User Group launches with 22 members; West debates slowing AI as China builds parallel governance; the open-weight dilemma emerges.
Kimi K3 open-source model challenges closed-source AI dominance; OpenAtom ‘Park Tour’ Hong Kong Station launches openEuler’s first overseas regional community.