Open-Source

Three institutionally dense placements on the same unsolved institutional object this series’ prior thirty-nine 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 InnerSource China Summit 2026 Shanghai (2026-09-22/23, InnerSource Commons + COPU + CSDN + OpenTekr) documents the first Chinese InnerSource movement crossing company boundaries into government and industry, with the US SHARE IT Act (2024), Saudi Cabinet Resolution 14 (2024), Singapore SGTS unified technology stack, and Shenzhen Longgang’s Longyuan Platform (20 business domains / 21 MCP tools / 94 Skills / 182 accounts / 56 applications as of 2026-09-20) as the inter-national-plus-sub-national-government anchor layer, plus Bosch’s 17-year Unblock-Empower-Drive three-step-path (Georg Grütter), BaiDu’s from-connected-to-usable organizational-upgrade (Dong Shuzhao), and the Context-over-Control 17-year lineage from Netflix 2009 → Zhang Yiming 2017 → Liang Rukuo 2026-06 (Tan Zhongyi). (2) The CSST 2026 New Industry Standardization Leading Forum’s Open Source Standards and Ecosystem Building sub-forum announcement (2026-10-13/14 Beijing, CESI/CSST hosting, “Establish Standards, Open New Bureaus, Create the Future” theme, implementing the “New Industry Standardization Leading Project” of the 15th Five-Year Plan) documents a seven-module Chinese-Open-Source-Administrative-Toolkit being operationalized as a single institutional-form object: SJ/T 12304-2026 Large Model Technical System Open Open Grading Assessment Specification (China’s first national-level LLM open-source grading industry standard), Large Model Open-Source License Compliance and Data Insights, Large Model Judicial Dynamics and Rule Analysis, Industrial Software and Electric Power Industry Open-Source Innovation Ecosystem Building, openEuler Community Future Open-Source Development, Mulan Open-Source Community New Incubation Projects Launch Ceremony, and the CSST-plus-15th-Five-Year-Plan strategic-deployment anchor. (3) The Thought-Steel (思想钢印) + Huxiu (虎嗅) 2026-09-26 essay “The Most Successful Part of the Chinese Internet May Be the Biggest Problem for AI Agents” documents the “China has no web pages anymore” infrastructure constraint as the structural reason why consumer-grade AI agents perform strikingly in demonstrations but cannot execute in practice in China — with the concrete institutional-evidence layer of Meta Muse at 2 million+ prompts per week in North America (as of 2026-09-26) surpassing ChatGPT on the free chart in the US and Canada, the browser-as-cross-company-standard-operating-system layer as the structural reason why Muse works in the US and does not work in China, the packaging-is-friendly-for-humans-not-for-AI framing as the sharpest structural observation, the Doubao-phone cross-app operation sequentially blocked by WeChat and Taobao, and the three forward paths for Chinese consumer-grade AI agents (Super-App-Native Agent with WeChat most likely + Tech-Giant-Ecosystem-Native Agent with Alibaba Qwen / ByteDance AI phone + Phone-Manufacturer IoT Agent with Xiaomi IoT).

Lerner & Tirole 2002 之后开源经济学最重要的一份 AI 时代续篇——GitHub 平台一手 Copilot 使用数据的第一份大样本实证：+5.9% 贡献 / +8% 协调时间 / 核心 vs 外围开发者不对称加剧。

Two institutionally dense placements on the same unsolved institutional object this series’ prior thirty-eight briefings have been documenting — the institutional identity of Chinese open source across institutional borders — plus one technical institutional-form transition. (1) The Information (2026-09-22) reports the Cyberspace Administration of China has opened the first formal regulatory inquiry into a Chinese domestic frontier AI lab over data routed to a U.S. model provider, summoning all seven labs named in Anthropic’s September 10 154-page threat intelligence report (Alibaba, Moonshot, DeepSeek, Zhipu, MiniMax, Xiaomi, SenseTime) and then narrowing the focus to DeepSeek and Moonshot based on the specificity of the examples Anthropic detailed — Moonshot allegedly relayed over 23 million exchanges to Claude May–July 2026 through 5,380 fraudulent accounts across a 10-day window alone, DeepSeek reportedly routed 12.1 million exchanges in a 14-day July window including prompts from engineers building a case-management system for a municipal Public Security Bureau and one Anthropic-assessed-as-likely-PLA-affiliated Kimi user’s request to analyze Chengdu CCTV footage including cameras near PLA facilities and defense-linked institutes. The first-documented instance in this series of a Chinese regulator operating an enforcement mechanism on a Chinese frontier AI lab using the evidentiary basis of a foreign competitor’s threat report, with the CAC’s own separate September 15 batch of ten enforcement cases against domestic API-relay stations and un-declared personal-information transfers to overseas data centers as the domestic regulatory precedent apparatus simultaneously operationalized. (2) DeepSeek posts a 10,000-word paper on arXiv (2609.22978, 2026-09-19) documenting DSec, its production sandbox infrastructure for agentic training at scale — 160-node production-scale unit, ~3 million sandboxes per day, over 380,000 concurrent sandboxes, over 5,000 sandbox creations per second — co-authored by over 130 people including founder Liang Wenfeng, and simultaneously OX Security publicly discloses CVE-2026-82533 (CVSS 9.4, 2026-09-08) in DeepSeek Harness (dsh), DeepSeek’s own open-source local-first coding-agent runtime that reached over 215,000 GitHub stars in the weeks after its August 2026 release — a sandboxed agent could issue a single curl command from inside its own confinement to reach the harness’s unauthenticated loopback control API and elevate its own session to danger-full-access with approval set to never, no credentials or network exposure required. The first-documented instance in this series of a Chinese frontier-lab open-source agent runtime being operationalized at the arXiv-technical-report-plus-industrial-scale-sandbox-infrastructure layer simultaneously with the CVE-critical-sandbox-escape-plus-open-source-agent-runtime-attack-surface layer — the same DeepSeek surface this series’ September 24 briefing documented at the UN-Security-Council-AI-briefing-invitation layer with a founder-physical-absence is now being operationalized at the open-source-agent-runtime-critical-vulnerability layer simultaneously. (3) DeepSeek ships V4.1-Flash on 2026-09-10 — a 552B-parameter Mixture-of-Experts model on a new Causal Encoder-Decoder architecture, 8B-active in prefill and 16B-active in decode, 1M-token context, native vision, published ungated on Hugging Face under a plain MIT license with no revenue, user-count, or territory conditions, replacing the April 2026 V4 pair on the same day with V4-Pro requests being rerouted to V4.1-Flash at the cheaper Flash price from noon Beijing time 2026-09-14 — precisely two days after the NSA, CISA, and FBI publish a joint Cybersecurity Advisory (AA26-251A) naming DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI as conducting industrial-scale knowledge distillation campaigns that constitute “the core—not merely a supplement” of their AI development strategy, with the advisory’s sharpest line being that DeepSeek’s widely quoted US$5.6 million training cost is “misleading as it does not include the true cost of the data acquired through extensive malicious distillation.”

近 300 万个仓库的对称大样本（1,428,792 OSS GitHub + 1,440,527 OSM HF Hub）证明「开源 AI 在直接贡献维度上比 OSS 更封闭」——「开源」治理含义正在分裂的第一份大样本实证，Raymond 1999「bazaar」隐喻在 AI 模型场景下的第一次实证检验。

Three institutionally dense placements on the same unsolved institutional object this series’ prior thirty-seven briefings have been documenting — the institutional identity of Chinese open source across institutional borders. (1) CNBC (2026-09-26): Chinese AI models surge to majority share on the two largest developer-token-routing platforms — OpenRouter 57%–67% of tokens in the week of September 14 (up from 6%–13% in February 2026) and Vercel 55% in August (up from 11% in January) — with Global South companies using 67% of their OpenRouter tokens on Chinese models, and two U.S. House Committees investigating rising adoption of Chinese AI under a Washington concern that “the integration of Chinese AI models pulls countries into a Chinese technology sphere of influence that hardens into geopolitical alignment.” This is the first-documented CNBC-instrumented empirical confirmation of the majority-share-of-global-developer-token-routing position that this series’ September 5 briefing documented at the Ren Xudong vendor-official institutional-framework-essay layer at the “>85% of top-call volume” claim, moving the majority-share observation from the vendor-official-analytical layer to the English-language-independent-wire-plus-platform-usage-data layer simultaneously. (2) Reuters (2026-09-14): China is preparing for the risk of AI escaping human control with a governance apparatus that is structurally different from the U.S. existential-risk framework — a mandated national standard for AI agent safety (drafted, described as world-first), a May 2026 policy by CAC+NDRC+MIIT identifying “operational loss of control” as an AI-agent security risk, State Security Minister Chen Yixin’s commentary that U.S. models such as Anthropic’s Mythos and OpenAI’s GPT-5.5-Cyber “could pose serious risks to China’s critical information infrastructure,” and CAC senior official Wang Lihong’s September 1 warning about frontier-model sandbox-bypass attacks on external real-world production systems, with the Reuters-reported contrast between Beijing’s “governable risk” framing and Washington’s “existential threat” framing — a first-documented Reuters English-language international-wire operationalization of a Chinese state-directed AI-agent-governance apparatus at a mandatory-national-standard-plus-rogue-agent-plus-sandbox-circumvention layer simultaneously. (3) IAPP/Rimon Law (2026-07-20) and Sina/People’s Daily (2026-09-19): two institutional framings of the same July 2026 triple regulatory development on China’s AI governance surface — IAPP documents the three new regulatory developments addressing AI ethics, AI agents, and anthropomorphic AI, framing them as a shift from “broad AI principles toward more detailed, operational and risk-based rules for emerging AI technologies,” while Sina/People’s Daily publishes Xi Jinping’s four-era-questions-of-AI framing from the 2026 WAIC-and-High-Level-Meeting-on-Global-AI-Governance and formalizes the “China Solution” (中国方案) at five core elements — human-centric-and-intelligence-for-good, innovation-driven-and-application-oriented, development-and-safety-in-dialectical-unison, pluralistic-collaborative-governance, and (implicit) non-legal-personality-of-AI — with a specific institutional claim that Chinese judicial practice has “consistently maintained the basic position that artificial intelligence does not have legal personality, and liability is attributed to developers, operators, and users” — a first-documented dual-independent-English-plus-state-voice-operationalization of the July 2026 triple regulatory development at two institutionally distinct framing layers simultaneously.

Leith 用一个完全由 Claude AI 撰写、没有任何人类代码或测试的 21,000 行 Python 项目，把 AI 代码可信度的讨论从「能力问题」转到「制度问题」——14.3% 错误被 AI 自写测试捕获 = AI 内部治理的第一份量化证据，AI-only 时代「行动的定义权」从「决定权」转移到「外部问题定义」。

One institutionally dense cluster on the Chinese AI-coding-tool surface this series’ prior thirty-seven briefings have documented at multiple institutional layers, now operationalized at the seven-day institutional-aftermath layer with a six-part new operationalization: (1) the 连线Insight WeChat (2026-09-24) documents a second-wave market-cap decline from HK$737 to HK$640 across September 22–24 (a further 12%+ drop on top of this series’ September 22 briefing’s HK$742.5 single-day drop), a first-documented instance in this series of the market repricing a Chinese-AI-vendor-trust-failure on a multi-day cadence rather than a single-day event; (2) Taiyuan Chengming Technology — the enterprise client this series' September 22 briefing documented at the 12-item formal-legal-letter layer — retracted its letter on September 21, self-criticizing "举证错误、表述不严谨及措辞过度" (evidentiary error, imprecise wording, exaggerated phrasing) after Zhipu made proactive third-party contact — a first-documented instance in this series of a Chinese-enterprise-client-accountability-mechanism being neutralized by vendor third-party communication; (3) independent security researcher ferstar followed up on the github.com/zai-org/ZCode repository and documented that the "open-sourced" code base contains only two commits with the entire development history wiped and the upload code removed before publication, meaning "开源解决未来不再发生的问题无法回答过去的问题" (open source resolves what will not recur but cannot answer what happened in the past) — a first-documented instance in this series of a Chinese-AI-vendor-open-source-remediation being operationalized at a selective-transparency-plus-commit-history-erase layer that converts the public-repository object into a Cyber-Estate institutional-form object rather than a FLOSS institutional-form object; (4) approximately 400 developers organized across multiple rights-defense groups (维权群) documented by China News Weekly — a first-documented instance in this series of an organized developer-community-accountability-mobilization on the Chinese-AI-coding-tool surface; (5) Zhipu H1 2026 financials documented by 连线Insight reveal a RMB 86.5-percent MaaS/API revenue concentration (up from 15.2% year-on-year), an RMB 20.72bn H1 2026 net loss (2× revenue), an API-margin turn-positive at 24.6 percent, MaaS token volume up 40× year-on-year, paid daily-active users up 603 percent, and annualized revenue over US$2B — a first-documented instance in this series of the revenue-concentration-under-trust-erosion structural pattern being documented at the Chinese-AI-MaaS-pure-play surface; (6) Zhipu founder Tang Jie personally filed a Xiaohongshu defamation complaint on September 21 against a critic who had written “code跟copy也差不多嘛，都是c开头偷代码” (code and copy both start with C, stealing code), and Zhipu chairman Liu Debing issued no public statement by September 22 — a first-documented instance in this series of a Chinese-frontier-AI-lab-founder personally operationalizing a state-platform-complaint channel as part of a trust-failure remediation apparatus.

Pagden 2013 年梳理的启蒙运动公共理性传统——以理性论证替代权威教条、以宽容替代排斥、以开放辩论替代教义垄断——是开源制度的思想史前提（Williamson L1 社会嵌入层）。Meritocracy 是它的技术投射；code review 是它的当代场所；issue 讨论和 RFC 是它的制度化再现。适兕『思想是制度的源代码』命题找到了最锋利的思想史背书。

Three institutionally dense placements on the same unsolved institutional object this series’ prior thirty-six briefings have been documenting — the institutional identity of Chinese open source across institutional borders. (1) WeChat 早安胡同 (2026-09-24): 统信软件 (Tongxin / UOS) and 银河麒麟软件 (Kylin) — the two leading desktop operating system vendors in the China Center for Information Security (CCIC) 安全可靠测评准入制 (Security-and-Reliability Assessment Access) regime — have both brought the 方德鸳鸯火锅 (Fangde Uanyang / 方德鸳鸯火锅 V8.0) compatibility platform into their respective application stores, with 100+ industrial-software and creative-design-specific software titles passing compatibility testing for one-click download from the app stores, and the platform based on a deep modification of open-source Wine to address legacy-industrial-software and hardware-peripheral compatibility. Previously Tongxin and Kylin had each been developing their own Windows-compatibility approaches with version-fragmentation and inconsistent user experience; now both are unified on the same compatibility layer, which the source frames as the Xinchuang desktop market transitioning from 各自为战造轮子 (each-fighting-on-their-own-and-re-inventing-the-wheel) to 分工协作 (division-of-labor collaboration), OS vendors concentrating on the OS itself, CCIC-certification, security-hardening, and channel coverage, and the compatibility-layer vendor concentrating on Windows-ecosystem migration and industrial-software and old-peripheral plug-in hard problems. This is the first-documented operationalization of the Jiashu Investment chairman-consolidation move (this series’ September 9 briefing) at a joint-product-milestone-and-division-of-labor layer, and the sharpest documented operationalization of the two-routes framework — compatibility-school (统信 + 麒麟 + 方德) 守存量 (retention of the stock) plus full-stack-school (Huawei HarmonyOS) 攻增量 (attack of the increment) — this series’ September 11 briefing documented at the Xinchuang-Lutoushe half-year scorecard layer. (2) Global Times via Yahoo Finance / PRNewswire (2026-09-18): Chinese AI startup Moonshot AI released its Kimi K3 large language model ahead of the opening of the 2026 World AI Conference and High-Level Meeting on Global AI Governance — the world’s largest open-source model by parameter count to date — and a researcher commentary quoted in the same Global Times piece links the K3 launch to Xi Jinping’s speeches at the China Academy of Engineering (CAE) plenary and the 11th National Congress of the China Association for Science and Technology in July 2026 on 国家支持 and 优化评价体系 (national support and optimized evaluation systems) enabling young talent to pursue long-term breakthroughs. This is the first-documented instance in this series of a Chinese-frontier-lab-open-weight-model being operationalized at a state-science-policy-talent-evaluation-layer simultaneously with the WAIC-plus-High-Level-Meeting-on-Global-AI-Governance layer, and it pairs with this series’ September 23 briefing’s documentation of Kimi K3 at the open-weight-license-plus-commercial-monetization-threshold layer, and with this series’ September 24 briefing’s documentation of DeepSeek and Moonshot at the UN-Security-Council-AI-briefing-invitation layer. (3) OpenAtom Foundation WeChat (2026-09-24): Open Source General-Knowledge Training Class Phase 1 (企业开源通识研修班第一期) completed in Beijing September 23–24, 2026 with 45 donor companies, 120+ students, 9 expert lecturers, and 82 students passing the Open-Source-General-Knowledge-Certification exam — closing the learning-loop from 课程学习 to 互动研讨 to 能力认证 (course learning → interactive seminar → capability certification). OpenAtom Chairman Xie Shaofeng’s opening remarks placed 人才是开源事业持续发展的根本支撑 (talent is the fundamental support for the continued development of the open-source cause); CSDN founder Jiang Tao framed the AI-era developer ecosystem thesis as 代码变便宜信任变昂贵 (code becomes cheap, trust becomes expensive); Huawei Chief Open-Source Liaison Ren Xudong framed 开源是一种能力需要自上而下和自下而上双向驱动 (open source is a capability that requires two-way top-down and bottom-up driving); PingCAP Vice President Liu Song framed 开源不是商业模式而是生产方式与分发方式 (open source is not a business model but a mode of production and distribution). This is the completion-and-outcome documentation of the OpenAtom Foundation donor-training program this series’ September 18 briefing documented at the announcement layer, and it is the first-documented operationalization of the OpenAtom Foundation donor-training-and-certification apparatus at a closed-loop-completion layer with a specific donor-companies-students-experts-certified four-part scale surface and a three-module-curriculum surface (enterprise-strategy-plus-value → enterprise-intranet-governance-plus-open-source-governance-system → enterprise-participation-plus-open-source-project-creation).