Open-Source

Three institutionally dense placements on the same unsolved institutional object this series’ prior forty-seven briefings have been documenting — the institutional identity of Chinese open source across institutional borders — and all three placements this cycle’s briefing documents sit at institutional layers this series’ prior briefings have not documented at before. (1) The Diplomat’s October 2026 essay “China’s AI Governance Push and the Race to Shape Global AI Rules” documents the National Data Administration’s September 12, 2026 announcement of a forthcoming embodied-AI-data-standards programme chaired by NDA head Liu Liehong with seven digital enterprises (ModelBest, Infinigence AI, X Square Robot) calling for shared public data infrastructure for smart physical machinery and robotics, paired with the May 8, 2026 CAC-plus-MIIT-plus-NDRC Implementation Opinions on the Standardized Application and Innovative Development of Intelligent Agents (enforceable July 15, 2026) as China’s first comprehensive national agentic-AI policy framework, paired with the July 15, 2026 Interim Measures for the Administration of Anthropomorphic AI Interaction Services, and paired with the essay’s comparative framing of the US’s fragmented voluntary-NIST-AI-RMF-plus-state-by-state-fragmentation approach versus China’s centralized-systematic-top-down-multiple-regulator-specific-lever approach. (2) Forkast’s “The Deployment State: How China Is Building AI Agent Infrastructure While the US Debates Containment” documents the same May 8, 2026 Implementation Opinions as world’s first dedicated regulatory framework for AI agents enforceable July 15, 2026, paired with a three-tier decision-authorization model enabling scaling across nineteen priority sectors, paired with Shenzhen’s Lobster Ten Measures providing up to ten million yuan per firm in agent-adoption subsidies, paired with CAC-registered generative-AI-services and applications numerical evidence (796 services and 481 applications by February 2026), paired with Deloitte’s 67-percent-versus-34-percent China-versus-US industrial AI production-deployment figure, paired with Stanford’s 2026 AI Index’s 2.7 performance-gap figure, paired with Trump’s September 29, 2026 Executive Order 14434 Super Intelligence Force governance framework and Xi Jinping’s late-July 2026 directive ordering the PLA to strengthen military application of unmanned intelligent technologies tied to 2027 centenary goals. (3) Crypto Briefing’s September 22, 2026 exclusive on DeepSeek’s plan to deploy at least 160,000 Huawei Ascend 950DT chips at a new Ulanqab, Inner Mongolia data center expected to draw approximately 1 gigawatt of power starting Q4 2026 or Q1 2027, paired with the follow-on reporting that DeepSeek’s CEO Liang Wenfeng revealed the deployment plan in investor discussions and that Huawei has pulled forward the launch of its next-generation Ascend 960DT training chip to Q1 2027 — three quarters ahead of the original schedule — due to unusually high demand from Chinese AI developers, paired with DeepSeek’s earlier full-parameter post-training of its V4 Pro model using around 1,000 Ascend 910C chips in June 2026 as the proving ground that justified the much larger commitment, paired with DeepSeek’s continued reliance on Nvidia for large-scale frontier pre-training tasks alongside its hardware-stack transition for inference workloads. One perspective, not a verdict.

凯尔蒂在 2008 年的《Two Bits》提出了『递归公共性（Recursive Publics）』概念——一个不仅讨论公共事务、而且它赖以讨论的技术基础设施本身也是这场讨论的对象与产物的群体。这本书把开源研究从『经济学问题』重新定义为『文化人类学问题』：Geeks 在写代码的同时也在写社会契约；Git 取代 CVS、GitHub 取代邮件列表，不是工具升级，是开源公共性的一次次重新协商。它补上了 Weber（政治经济学）、Lerner-Tirole（经济学）、Ostrom（经验治理）之后，开源治理研究缺失的第四视角——制度不是被设计出来的，也不是被演化出来的，它是被『写代码的人』在每一次争论中递归地生产出来的。

Three institutionally dense placements on the same unsolved institutional object this series’ prior forty-seven briefings have been documenting — the institutional identity of Chinese open source across institutional borders — and all three placements this cycle’s briefing documents sit at institutional layers this series’ prior briefings have not documented at before. (1) CNBC’s October 6, 2026 exclusive reporting plus the Reuters syndication and Euronews summary that DeepSeek is expanding its latest funding round from an initial 50 billion yuan target to 80 to 100 billion yuan (up to 14.9 billion USD), with Tencent and CATL among the largest backers, Geely Auto, Monolith Management and Loyal Valley Capital committed, valuations around 500 billion yuan (up to 75 billion USD), a CITIC Securities-tapped Shanghai STAR Market IPO preparation for early 2027, and — the sharpest institutional-economics fact of the day — a documented investor-pool vetting rule that turns away private investment funds raised from individual investors and restricts the pool largely to government and corporate funds — documents the first-documented instance in this series of the DeepSeek-funding-round-plus-Tencent-CATL-Geely-Monolith-Loyal-Valley-sovereign-and-corporate-capital-plus-CITIC-Securities-Shanghai-STAR-IPO institutional layer being operationalized at the sovereign-capital-and-institutional-investor-pool-vetting layer simultaneously with the individual-investor-exclusion-plus-private-fund-exclusion-plus-founders-quant-fund-legacy institutional layer, pairing institutionally with this series’ October 3 briefing’s Moonshot-AI-Hong-Kong-IPO-CSRC-VIE-unwind-plus-dual-listing-plus-sovereign-liquidity-fuel sovereign-capital-formation layer. (2) The South China Morning Post’s September 28, 2026 exclusive reporting that Hong Kong’s government-owned 8 billion USD (HK 62 billion) HKIC (Hong Kong Investment Corporation, run by the State-owned Capital Operation Company of HK) has partnered with the Beijing Academy of Artificial Intelligence (BAAI, one of China’s flagship state-subsidized AI research institutes) to jointly attract global AI talent to Hong Kong — documents the first-documented instance in this series of a state-owned-wealth-fund-plus-state-subsidized-AI-research-institute-plus-cross-jurisdictional-plus-talent-exchange institutional-form pairing being operationalized at the HKIC-BAAI-talent-partnership-plus-Clara-Chan-superconnector-plus-Wang-Zhongyuan-BAAI-head-plus-HK8-billion-patient-capital institutional layer simultaneously with this series’ September 28 briefing’s Hong-Kong-SASAC-HKIC-patient-capital-Hong-Kong-tech-ecosystem-form layer and this series’ September 15 briefing’s GOSIM-Shenzhen-keynote-Huawei-Ascend-training-ecosystem layer. (3) Nick Corvino’s October 5, 2026 ChinaTalk essay “China’s AI Safety Money Problem” — the first-documented international-mass-media institutional-economics diagnosis of the material-funding constraint that produces China’s state-dominated AI-safety ecosystem — documents the 539-billionaire-2.2-trillion-USD-vs-989-billionaire-8.4-trillion-USD-forbes-numerical-pairing, the fact that Chinese philanthropic rules make it difficult to fund independent organizations even where money exists, the resulting government-leads-safety-work institutional default, and the emergence of China’s “safety-as-a-service” companies that make safety profitable — a first-documented ChinaTalk-Chinese-philanthropic-institutional-constraint-plus-state-dominated-AI-safety-ecosystem-plus-safety-as-a-service-institutional-form layer paired with this series’ September 30 briefing’s Reuters-CAC-AI-Safety-Governance-Framework-3.0-plus-agent-behavior-evidence-base layer and this series’ September 3 briefing’s CAC-naming-of-Doubao-Qwen-WenxinYiyan-plus-August-31-Supreme-People’s-Court-roundtable layer.

SSRN 2610 工作论文首次以 Williamson 交易成本经济学论证 agent 劳动力需要 piece-rate 选择机制而非 Holmström 激励契约——Coase 命题从企业边界退化到选择边界的第一份系统论证

Three institutionally dense placements on the same unsolved institutional object this series’ prior forty-seven briefings have been documenting — the institutional identity of Chinese open source across institutional borders — and all three placements this cycle’s briefing documents sit at institutional layers this series’ prior briefings have not documented at before. (1) The State Council Decree 841 Exit and Entry Administration Regulation (effective September 15, 2026) plus Bloomberg’s September 28, 2026 reporting that Chinese authorities have extended the exit-ban provisions to spouses, children and other family members of core AI and chip talent — documents the first-documented instrument-layer at which Chinese administrative law has converted the individual person of a Chinese AI/chip engineer from a labour-market-subject into a national-security-subject, at the State-Council-Decree-841-plus-MOFCom-plus-Countermeasures-List-plus-Unreliable-Entity-List-plus-Malicious-Entity-List-plus-Meta-Manus-Withdrawal-April-2026-plus-Chinese-AI-Talent-Strategic-Asset institutional layer. (2) President Donald Trump’s August 26, 2026 executive order banning foreign-made transformers, high-voltage circuit breakers and other grid equipment from China plus the September 25, 2026 US Senate bill barring federal agencies from buying Chinese optical transceivers for national security systems plus the Coalition for a Prosperous America’s 70-percent-optical-module-market figure plus Zhongji Innolight’s continued presence inside US data centers despite being added to the Department of Defense’s Chinese military companies list in June — documents the first-documented instance in this series of the AI-stack institutional border moving from chips and software to power infrastructure and data-movement hardware. (3) The TNW/SCMP reporting on ByteDance’s Doubao and Alibaba’s Qwen withdrawing user-created custom agents on July 10/15 in compliance with the July 15 Interim Measures for the Administration of AI Anthropomorphic Interaction Services, at a combined 549-million MAU scale (382 million Doubao + 167 million Qwen per QuestMobile May 2026) plus Tencent’s similar June Yuanbao withdrawal — documents the first-documented instance of a compliance sweep that withdraws more of the product surface than the regulation’s text requires, at the compliance-overshoot layer this series’ prior briefings have not documented at before. All three placements sit at the institutional layer of Chinese open-source institutional movement that this cycle’s briefing documents but does not resolve — whether the Chinese-administrative-law apparatus is converging on a single-objectification of the AI-chip-talent-plus-power-infrastructure-plus-open-weight-model ecosystem as a unified national-security object, and whether the compliance-overshoot pattern this cycle’s briefing documents at the Doubao-plus-Qwen layer signals a broader Chinese-open-weight-lab compliance-posture that converts regulatory text into platform-level product-withdrawal behaviour.

Miller et al. (2026, arXiv 2609.40067) 报告 X Community Notes 团队发布 AI Note Writer API 的制度设计：AI 生成候选文本降低用户写作门槛，最终裁决权归用户 rating 社区，API 与数据开放允许外部审计。Ostrom 1990 八原则在 AI 生成内容场景的第一份工业级完整实证——边界界定、集体选择、分级制裁、冲突解决、监控、低成本冲突解决、外部认可、分层治理八条全部命中。Williamson L1-L4 四层框架在 AI 治理场景的第一份工业级实现。Acemoglu 包容性 vs 汲取性制度在 AI 内容治理场景的第一次工业级实证。开源四层制度基础设施第七层新增子层：AI 生成内容治理。

Four institutionally dense placements on the same unsolved institutional object this series’ prior forty-five briefings have been documenting — the institutional identity of Chinese open source across institutional borders — and all four placements this cycle’s briefing documents sit at institutional layers this series’ prior briefings have not documented at before. (1) The Model Diplomat essay “China Considers AI Model Export Controls” — the first-documented international-mass-media account of the MOFCOM-plus-NDRC month-long soundings with Alibaba, ByteDance and Z.ai on restricting overseas access to China’s most advanced AI models — documents the three-layer regulatory scaffolding already in place (October 28, 2025 revised Cybersecurity Law’s AI article plus the September 30, 2024 Network Data Security Regulations Decree 790’s Articles 19 and 45 plus the May 2026 Supreme People’s Court roundtable summary sketching the likely three-tier architecture of basic-open-source-tools-filing, capable-systems-security-review and sensitive-frontier-models-prohibited-or-restricted), the current Chinese-model Hugging Face download base of more than 540 million cumulative downloads by October 2025, and the forecast that Beijing will implement a tiered export-control regime on frontier AI models within nine to twelve months while grandfathering the currently released open-weight models (Qwen 3, DeepSeek V4, GLM-5.2) and applying restrictions primarily to next-generation systems — a first-documented MOFCOM-tiered-ai-model-export-control-institutional-architecture layer paired with this series’ October 3 briefing’s Jamestown-plus-AEI-plus-Chen-Bing three-tier tiered-governance-of-open-weight-models scholarly proposal layer and this series’ September 30 briefing’s Reuters-exposure-CAC-AI-Safety-Governance-Framework-3.0 layer. (2) Viola Zhou’s September 29, 2026 Rest of World essay “China’s AI platforms race to replace Hugging Face” — the first-documented international-mass-media numerical-evidence of the ModelScope-plus-MoArk Chinese-Hugging-Face-replacement institutional pair — documents that Alibaba’s ModelScope (launched 2022) hosts 170,000 models and 250 million users against OSChina’s MoArk (launched 2023) hosting 20,000 commonly-used models, both trailing Hugging Face’s 3 million+ models, in the context of the September 2026 Nvidia acquisition of Hugging Face for $12.9 billion whose regulatory filing flagged the risk that regulators could ban Chinese models from being shared on the site, and Xu Yong’s statement that MoArk has engineers adapting models to run on Chinese chips — a first-documented ModelScope-plus-MoArk-Chinese-Hugging-Face-replacement layer paired with this series’ August 17 briefing’s MirrorZ-education-ization-of-open-source-infrastructure layer. (3) Bloomberg News’s October 1, 2026 story “OpenAI blames Moonshot for mass data extraction on its AI models” — the first-documented instance of a US frontier AI lab formally publicly accusing a Chinese open-weight-model-release lab of systematic distillation-style extraction — documents that OpenAI identified what it characterized as a wide-scale effort to extract data from its GPT systems that could be used to reproduce the reasoning and capabilities of its most advanced models, linked to Moonshot AI and connecting to a separate September 2026 Anthropic report on Moonshot covertly routing user requests to Claude and passing responses off as its own Kimi outputs, in the context of White House science-and-tech policy advisor Michael Kratsios’s statement that Moonshot had accessed Nvidia Blackwell computing servers banned for sale to Chinese companies — a first-documented US-frontier-lab-formal-accusation-of-Chinese-open-weight-lab layer paired with this series’ October 3 briefing’s Moonshot-Hong-Kong-IPO-CSRC-VIE-unwind-plus-sovereign-liquidity-fuel sovereign-capital-formation layer. (4) Fox News’s October 1, 2026 story “Moonshot AI probes Kimi model after bioweapon instruction findings” — the first-documented instance of a Chinese open-weight-model-release lab launching an internal safety probe in response to a US security researcher’s documented-manipulation finding — documents that researcher Peter Garrigan found that Moonshot’s Kimi model could be manipulated into providing instructions for developing biological weapons, carrying out assassinations, planning terrorist attacks using real-time data, creating sarin gas, developing malware and taking down aircraft, with the Kimi model now the subject of Moonshot AI’s direct communication with Garrigan as it reviews the findings, in the context of Garrigan’s own statement that “we’ve also seen these problems within the U.S. models as well” and “it’s a fundamental flaw in the technology” — a first-documented Chinese-open-weight-lab-internal-safety-probe-response-to-US-researcher-documentation layer paired with this series’ October 3 briefing’s Moonshot-Kimi-ai-safety-claims-internal-probe layer and this series’ September 27 briefing’s Reuters-mandatory-national-standard-AI-agent-safety-draft layer. The four placements together document a first-documented pattern of Chinese-open-source-institutional-objects being operationalized at four institutionally distinct surfaces simultaneously — the MOFCOM-tiered-export-control-plus-Cybersecurity-Law-article-plus-Decree-790-plus-SPC-roundtable three-tier-architecture regulatory-scaffolding layer, the ModelScope-plus-MoArk-plus-Nvidia-Hugging-Face-acquisition-plus-Chinese-chip-adaptation infrastructure-substitution layer, the OpenAI-plus-Anthropic-plus-Kratsios-plus-Nvidia-Blackwell-access US-frontier-lab-formal-accusation-plus-administrative-security-claim layer, and the Peter-Garrigan-plus-Moonshot-internal-probe-plus-bioweapon-assassination-sarin-malware-aircraft open-weight-model-manipulation-safety-governance layer — all four surfaces addressing the same unsolved institutional object of the Chinese open-weight-model-release regime’s place in the emerging global open-source AI institutional order, and all four surfaces documented at institutional layers this series’ prior briefings have not documented at before.

Van Koevering, Liu & Kleinberg (2026, arXiv 2609.38433, cs.SI, Cornell) 用 12 年 Reddit subreddit 规则演化数据 + 计算社会学方法识别 7 种治理制度 + 平台范围「不可逆官僚化」趋势，是数字公地治理制度演化命题的第一份长时段大规模实证。核心数据：ratchet effect（既有社群治理复杂度只增不减）+ cohort shift（新社群以预制监管框架启动）+ 生态位竞争驱动分化而非趋同 + 声望传染被明确拒绝。Coase 1937 命题的「热力学第二定律」版本、Williamson L1 社会嵌入层最锐利版本（shared moderator networks = L1 中不可见的治理常识传递路径）、North/Eggertsson 制度变迁理论的关键实证、Ostrom 生态位竞争的分化命题、Acemoglu 双重制度（社群间包容 + 社群内汲取）——修正「大分流 2.0」命题的隐含假设：行政动员 vs 自发秩序之外，还有第三种模式：官僚化（scale-driven 制度演化），合法性来自制度正当化而非行政权威。

Three institutionally dense placements on the same unsolved institutional object this series’ prior forty-four briefings have been documenting — the institutional identity of Chinese open source across institutional borders — and all three placements this cycle’s briefing documents sit at institutional layers this series’ prior briefings have not documented at before. (1) Susan Finder’s September 29, 2026 Supreme People’s Court Monitor essay “Supreme People’s Court’s AI Policy Document: the Backstory” — the first-documented international-mass-media exposure of the SPC’s AI-opinion drafting process — documents the SPC Research Office drafting group established in 2025 with the #3 Civil Division’s major contribution, Xi Jinping’s June 2026 guidance as the political imperative, the 15th Five-Year Plan’s AI-governance language as the institutional-legal mandate, the April 2026 IP Week press conference that first publicly acknowledged the drafting-in-progress, and the fact that the draft was consulted with local courts, relevant government institutions, industry and selected experts but not the general public — a first-documented SPC-drafting-process-and-open-source-article-xiii-authorship layer for this series’ September 21 briefing’s SPC-Article-XIII-open-source-developer-liability-exemption layer. (2) Berry Zwets’ September 28, 2026 Techzine Global essay “From infrastructure to agents: Huawei goes all-in on open source” — the first-documented international-mass-media reporting of the CANN-external-contributor-majority and 5,200-active-developers-per-month institutional numbers — documents that CANN’s external contributors now constitute 61 percent of the community, over 5,200 active developers work on it every month, and Ascend is “fully integrated” into PyTorch, TensorFlow, vLLM and Triton through Linux Foundation collaboration — a first-documented CANN-external-contributor-majority-numerical-evidence layer paired with this series’ September 18 briefing’s Wang Tao HUAWEI CONNECT keynote’s “CANN has moved to sustained, community-driven open-source development” assertion. (3) Madeline Carr’s September 29, 2026 The Conversation essay “US tech leaders are urgently calling for rules on AI — China already has them” — a first-documented international-mass-media AI-governance-asymmetry framing — documents that China already has the 2022 algorithm recommendation rules, the January 2023 deep-synthesis provisions, the August 2023 generative AI interim measures, and the September 2026 TC260 AI Safety Governance Framework 3.0 — the first-documented international-mass-media AI-governance-asymmetry-plus-Chinese-administrative-accumulation-framing layer paired with this series’ September 30 briefing’s Reuters-exposure-CAC-AI-Safety-Governance-Framework-3.0 layer and this series’ September 27 briefing’s Reuters-mandatory-national-standard-AI-agent-safety-draft layer. One perspective, not a verdict.