SkillHub.news SkillHub.news 周报

WEEK 2026-09-14 · 46 STORIES · SKILLHUB WEEKLY

SkillHub.news 每周AI

9 月 8 日 ~ 2026 年 9 月 14 日 · 周报 · 素材 4

本期看点46 条 · 约18分钟

本周主题线WEEKLY THREADS24 条

DeepSeek切换落地实测（进展·第4天）

官方事件精选OFFICIAL EVENTS14 条

Does everyone have taste? These 5 experts say yes.

榜单异动RANKING MOVES8 条

obra/superpowers 星标 +79,301

01本周主题线WEEKLY THREADS24 条

DeepSeek切换落地实测（进展·第4天）

AA实测/多源2026-09-14置信度高进展·第 4 天

AA独立实测V4.1 Flash智能指数39.5反超V4 Pro的36，但单任务平均输出约8.9万token比V4 Pro多62%、部分抵消降价红利；官方"保留原入口"（9/13报）与实测路由并存，迁移以官方文档为准——比价逻辑转向单任务总成本(AA实测/多源,高)

GPT-6 Astra攻克FrontierMath Tier 4

量子位首发/多源印证2026-09-14置信度高

解出该层级全部剩余题97.6%（此前最高约5%），Epoch AI宣布Tier 4饱和；叠加Erdős层2/68上下文，整套FrontierMath被AI完整攻克——科研级基准失去区分度、评测体系加速换代(量子位首发/多源印证,高)

智谱50亿美元融资完成

公告2026-09-14置信度高

9/13公告20亿配售+30亿零息可转债（"小股大债"海外长线抢筹）；ARR16亿美元系8/31半年报已披露口径勿当新增；API毛利率-0.4%→24.6%转正、Token调用较年初+40倍(公告,高)

FOMC前夜：CME加息概率修正口径

CME2026-09-14置信度高

CME 9/14官方口径86.2%（9/13盘后约90%、一周前约59%）；"哈塞特'没有理由加息'原话"未获独立核实剔除；9/16议息+8月零售销售同日落地(CME,高)

AI降速叙事冲击暗盘（进展·第2天）

行情2026-09-14置信度高进展·第 2 天

9/13暗盘SK海力士-4.60%/英特尔-3.81%/美光-3.25%/英伟达-2.16%（与21财经/雪球逐项吻合）；Altman称OpenAI 2026不上市vs Anthropic抢跑史上最大IPO；甲骨文CapEx落点显示算力开支未退潮——情绪与基本面背离(行情,高)

特朗普公开对抗"减速共识"（进展·第2天）

SBN/NewsDrum/Yahoo2026-09-14置信度高进展·第 2 天

9/13爱尔兰表态反对减速AI（"谁赢AI谁赢"）、称风险警告系"负面势力炒作"、只支持护栏；Sanders回击呼吁中美谈判AI暂停条约；Ban ASI Act（9/3预告）攻防白热化(SBN/NewsDrum/Yahoo,高)

Anthropic选定纳斯达克+财务首披露（进展·第2天）

BI/Bloomberg/Reuters2026-09-14置信度高进展·第 2 天

BI/彭博/路透多源：推介材料首次披露Q2营收超115亿美元、首次盈利，Q3"有望"连续盈利；英伟达基石至多100亿仍在洽谈——史上最大IPO进入推介阶段(BI/Bloomberg/Reuters,高)

优必选柳州万台级人形机器人超级工厂投产

金台资讯/科创板日报2026-09-14置信度高

9/12投产：1.4万㎡、每10分钟下线1台WalkerS/Cruzr、年产能超万台（联合西门子数字孪生、零部件国产化率超75%），全球首个万台级基地；另中标乐山1.5081亿(金台资讯/科创板日报,高)

人形机器人H1出货1.91万台+272%

Smart Analytics/科创板日报2026-09-14置信度中高

Smart Analytics Global单一机构口径；高盛预测2035整机均价2.13万美元（-49%）；宇树单价两年降72%至16.64万元；中国厂商占全球出货约97%（产业口径）——Q3出货为下一验证点(Smart Analytics/科创板日报,中高)

Cognition发布SWE-2（Kimi K3后训练）

官方博客2026-09-14置信度高

官方博客证实基于Kimi K3（2.8T）RL后训练，自家FrontierCode 1.1 Main 50.0%、距Fable 5.1不足1分、成本声称低64%；FrontierCode系厂商自建榜单、竞品分数系厂商自测口径打折(官方博客,高)

评测信任危机：出题方身兼卖题方

SemiAnalysis/观察者网2026-09-14置信度中高

SemiAnalysis调查DeepSWE出题方Datacurve同时向前沿实验室卖专家级编码训练数据（对具体厂商点名未独立核实暂缓引用）；已有审计口径假阳性8.5%/假阴性24%；WPS AI登顶SpreadsheetBench V2 Full 50.86%（超Claude Opus 4.6的34.89%，中媒一致转引）(SemiAnalysis/观察者网,中高)

治理基建三件套：身份国标+词元券+数据产权

多源2026-09-14置信度高

《区块链智能体身份管理要求》国标指导性文件外滩立项（实验室报告子项未核实剔除）；成都词元券9/11征求意见（年1亿/单主体最高200万/补贴30-50%小微50%）+重庆"词元券+词元贷"；国家数据局9/11登记系统上线、北上深三所率先成为登记机构(多源,高)

趋势面：赛迪报告+外滩资金分层+工信部中小企业计划

报道台账2026-09-14

赛迪：大模型2年内产业化趋近度达1、推理算力需求与训练持平；外滩五机构确认"上下两层"（底层芯片+上层Agent）投资分层；工信部《AI中小企业创业支持计划》（8/29印发）三年培育1万家+/小巨人2000家；OpenRouter周量口径打架+"开源权重破50%"系口径拼接（官方约1/3，破50%系中国模型份额）剔除(多源,高/中)

剔除项存档

报道台账2026-09-14

SWE-Bench Pro Verified"重评掉21.48分/14%缺陷"疑似拼接旧审计造新数剔除（真实审计口径为假阳性8.5%/假阴性24%）；OpenRouter周量122.61万亿未获独立证实剔除；"哈塞特'没有理由加息'"未核实剔除；天数智芯供字节10万颗（知情人士口径）暂缓；DeepSWE对Muse/Gemini具体点名暂缓；国产模型份额40%/Epoch十亿美元估算暂缓；Ruchir"四O"系2025年11月旧观点回炒降背景

DeepSeek撤回V4 Pro下线（新拐点，反转9/12切换）

报道台账2026-09-13

官方9/11晚邮件+9/12官宣：9/14后保留原API入口、计费不变、不强制路由Flash，企业可自选版本——开发者舆情三天逼回商业决策，API生命周期话语权首次转向用户(官方邮件/AI Model Watch/tokencost三方一致,高)

Dario《We Must Pace the Frontier》+Altman附议

Reuters等2026-09-13置信度高

9/12-13凌晨发文：三步计划+Anthropic单方面承诺第三方评估永久员工级访问，Altman数小时后"we will do the same"、Musk亦表态；时点微妙——Anthropic冲刺约2万亿估值IPO，"放缓"与最大IPO同框(Reuters等,高)

25位菲尔兹奖得主联署抨击难题基准

多源2026-09-13置信度中高

陶哲轩/舒尔茨/德利涅等25位9/11联署《A Severe Misalignment of AI in Mathematics》，约2500人跟进签名；导火索与9/9 Navier-Stokes争议同源——学界与实验室评测合法性之争公开化(多源,中高)

甲骨文CapEx落点：戴尔HPE双新高（进展·第2天）

业绩会/行情2026-09-13置信度高进展·第 2 天

CFO业绩会重申FY2027资本开支900-950亿美元并罕见点名戴尔/HPE承接AI机架/液冷/网络设备；9/11戴尔+11.98%市值破3600亿创新高、HPE+约10.5-11%、超微订单簿600亿——AI基建首次有"收货人名单"(业绩会/行情,高)

电信研究院Token报告：2026年10亿亿

报道台账2026-09-13

《智能体时代AI基础设施发展研究报告(2026)》：2026中国Token年消耗10亿亿(10^18)、2030超3500亿亿；2029推理算力占比80%；全球活跃智能体2026约8000万个、2030预计22.16亿个(央视转引,中·权威单报告)

央企算力集采单日三单近9亿

财富号转述待公告核2026-09-13置信度中

9/12：中国长城中标国网第三期服务器超3亿+软通计算机中标移动部件焕新4.18亿+平知信息预中标云南智算约1.37亿(财富号转述待公告核,中)；同期京津冀"太空算力长廊"倡议+北京超1亿生物医药算力专项(一年赋能100项目)

Real-SWE私有代码库基准首榜

多源2026-09-13置信度中高

考题全部来自授权私有生产库（计费/税务/迁移，参考解中位改11文件）；Fable5.1+Claude Code 38.8%/GPT-6 Astra+Codex CLI 33.8%/Gemini3.8 31.2%/GLM5.3 28.8%——分数绑定模型+工具链，harness成排名前提(多源,中高)

SREGym：SRE智能体差两成闭环

单源2026-09-13置信度中

GPT-5.6 Sol(max)+Codex 81.0%端到端解决率居首（诊断95.2%/缓解85.7%）、Claude Opus 5+Claude Code 76.2%；最佳配置约两成生产故障无法闭环、配置间Token成本差近2倍(单源,中)

智能体治理空白带（商机）

三源2026-09-13置信度高

OpenAI智能体滥用RubyDoc致RubyGems停注册4天(HN 9/12)+48%已部署智能体缺有效安全控制+外滩大会报告判治理为"必要条件"——身份认证/权限管理/行为审计SaaS中美共振空白(三源,高)

资本下沉"卖水人"三连投

原文 ↗

生态脉搏站点数据自更新

2026-09-03 → 2026-09-10 星标涨幅：obra/superpowers +79,301★ · mattpocock/skills +71,905★ · multica-ai/andrej-karpathy-skills +59,162★ · anthropics/skills +49,058★ · Shubhamsaboo/awesome-llm-apps +38,266★ · farion1231/cc-switch +36,887★ · nextlevelbuilder/ui-ux-pro-max-skill +35,298★ · Graphify-Labs/graphify +32,474★

※风险分级灰犀牛 / 黑天鹅

灰犀牛 · 可预见

灰犀牛Ⅰ FOMC 加息落地引发 AI 估值二次压缩｜概率 87%｜冲击面：高估值 AI / 算力链｜观察哨：9/16 点阵图 + 零售销售（深度版风险分级）
灰犀牛Ⅱ 评测信任危机扩散至榜单生态｜概率 60%｜冲击面：基准方 / 采购方｜观察哨：SemiAnalysis 后续报道（深度版风险分级）

黑天鹅 · 低概率大冲击

黑天鹅Ⅰ 头部实验室「减速共识」变成政策约束｜概率 15%｜冲击面：前沿模型供给｜观察哨：特朗普 vs 实验室表态升级（深度版风险分级）
黑天鹅Ⅱ Anthropic IPO 定价失败拖累一级市场｜概率 10%｜冲击面：融资环境｜观察哨：10 月中旬路演认购倍数（深度版风险分级）

✦行动建议3

1. 采购方：切换「单任务总成本」计费框架

用 V4.1 Flash 做 POC，公开对比「token 总量 × 单价」的真实 TCO（实测单任务输出多 62% 就是现成论据）。路径：企业 CTO → 采购评审委员会 → 写入招标条款。打包：词元券补贴申请（30-50%）+ 智能体身份国标预合规，一份方案双红利。

时间窗时间窗：2 周内

2. 创业者：抢占「第三方评测验证」空位

做「可复现 + 无利益冲突」的独立评测服务，承接基准失锚后的采购决策需求（Datacurve 卖题兼出题是现成反面教材）。路径：开源社区信誉 → 企业尽调订单 → 国标参编资格。打包：绑定数据产权登记三所，形成「评测 + 确权」服务包。

时间窗时间窗：9/30 ARC-AGI-3 截榜前

3. 从业者：押注 Agent 治理与身份基础设施

围绕 DevDay 发布 + 上海公安 1212.5 万开标做智能体身份/审计中间件。路径：政府试点标段 → 金融等强合规行业 → 乘「1 万家中小企业」政策放量。打包：PaperCut 类事件复盘白皮书引流 + 小巨人申报辅导。

时间窗时间窗：9/29-10/1

⏱未来 3 个月时间线概率

时间窗	事件	概率	依据
9/15-16	FOMC 议息 + 8 月零售销售同日落地	87% 加息	↓ 高估值（CME 定价 + 深度版时间线）
9/16 前后	ASI 法案相关议程攻防（Ban ASI Act，9/3 预告）	85%	↑ 治理（政策议程）
9/22 前后	DeepSeek V4.1 生态适配潮	90%	↑ 开源（适配进展外推）
9/29	OpenAI DevDay + 上海公安智能体平台开标	90%	↑ Agent（已公告 + 临近）
9/30	加州法案签署窗口 + ARC-AGI-3 截榜	80%	双向（法定窗口）
10/1	OpenAI vs Apple 听证会	90%	↓ 封闭生态（听证会日程）
10 月上旬	智谱 50 亿资金到账放量	75%	↑ 国产（公告节奏外推）
10 月中旬	Anthropic IPO 路演 / 定价	70%	↑ 现金流 AI（路透证实窗口）
10 月下旬	HBM 荒向国产卡二段传导	65%	↑ 算力成本（产业链传导）
11 月初	词元券首批补贴发放	70%	↑ 中小企业（成都征求意见稿）
11 月中旬	人形机器人 Q3 出货验证（+272% 是否延续）	60%	验证（Q3 数据披露窗口）
12 月	智能体身份国标征求意见稿	55%	↑ 标准（立项后常规节奏）

素材来源：报道台账（滚动查重）：2026-09-08 ~ 2026-09-14 共 99 条主题记录 · AI 官方视频监控台账：7 天 · 387 条事件（精选 14 条带链接） · 站点榜单快照：2026-09-03 → 2026-09-10 星标涨幅 Top8 · 风险/行动/时间线：承接 2026-09-14 日报
SkillHub.news — 全球 Skill 榜单

WEEK 2026-09-14 · 46 STORIES · SKILLHUB WEEKLY

SkillHub.news Weekly AI

Sep 8 ~ Sep 14, 2026 · Weekly · sources 4

Highlights46 stories · approx.18min read

WEEKLY THREADS24 stories

DeepSeek switchover rollout test (Day 4 progress)

OFFICIAL EVENTS14 stories

Does everyone have taste? These 5 experts say yes.

RANKING MOVES8 stories

obra/superpowers stars +79,301

01WEEKLY THREADS24 stories

DeepSeek switchover rollout test (Day 4 progress)

AA实测/多源2026-09-14Confidence HighProgress · Day 4

AA independent tests show V4.1 Flash's intelligence index at 39.5, overtaking V4 Pro's 36, but average output per task is about 89,000 tokens, 62% more than V4 Pro, partly offsetting the price-cut benefit; the official 'keep original entry point' (9/13 report) coexists with tested routing, and migration is subject to official documentation—price-comparison logic shifts to total cost per task (AA tests/multiple sources, high)

GPT-6 Astra cracks FrontierMath Tier 4

量子位首发/多源印证2026-09-14Confidence High

Solved 97.6% of all remaining problems at that tier (previous high about 5%), Epoch AI declares Tier 4 saturated; with the Erdős layer 2/68 context added, the entire FrontierMath suite has been fully cracked by AI—research-grade benchmarks lose discrimination, and the evaluation system accelerates generational turnover (QbitAI first report/multiple sources corroborate, high)

Zhipu completes $5 billion financing

公告2026-09-14Confidence High

9/13 announcement: 2 billion placement + 3 billion zero-coupon convertible bonds ('small equity, large debt'; overseas long-term investors scramble for shares); ARR $1.6 billion is the already-disclosed 8/31 half-year report basis—do not treat as new; API gross margin -0.4%→24.6% turned positive, Token calls +40x vs. start of year (announcement, high)

FOMC eve: CME rate-hike probability revised basis

CME2026-09-14Confidence High

CME 9/14 official reading 86.2% (about 90% after the 9/13 close, about 59% a week earlier); Hassett's exact quote 'no reason to hike rates' was not independently verified and removed; 9/16 rate meeting + August retail sales land same day (CME, high)

AI slowdown narrative hits dark pool (Progress · Day 2)

行情2026-09-14Confidence HighProgress · Day 2

9/13 dark pool: SK Hynix -4.60%/Intel -3.81%/Micron -3.25%/NVIDIA -2.16% (matches 21财经/Xueqiu item by item); Altman says OpenAI won't go public in 2026 vs Anthropic races ahead on largest-ever IPO; Oracle CapEx allocation shows compute spending has not ebbed—sentiment diverges from fundamentals (Markets, High)

Trump openly confronts 'deceleration consensus' (Progress · Day 2)

SBN/NewsDrum/Yahoo2026-09-14Confidence HighProgress · Day 2

9/13 Ireland voices opposition to AI deceleration ('Whoever wins AI wins'), calls risk warnings 'hype by negative forces', supports only guardrails; Sanders hits back, calls for U.S.-China negotiations on an AI pause treaty; battle over Ban ASI Act (previewed 9/3) intensifies (SBN/NewsDrum/Yahoo, High)

Anthropic selects Nasdaq + first financial disclosure (Progress · Day 2)

BI/Bloomberg/Reuters2026-09-14Confidence HighProgress · Day 2

BI/Bloomberg/Reuters multiple sources: pitch materials disclose for the first time Q2 revenue over $11.5 billion, first profit; Q3 "expected" to be profitable again; NVIDIA cornerstone investment of up to $10 billion still in talks—largest-ever IPO enters pitch stage (BI/Bloomberg/Reuters, high)

UBTECH's Liuzhou 10,000-unit-class humanoid robot super factory begins production

金台资讯/科创板日报2026-09-14Confidence High

Production starts 9/12: 14,000 m²; one WalkerS/Cruzr rolls off the line every 10 minutes; annual capacity over 10,000 units (with Siemens digital twin; component localization rate over 75%); world's first 10,000-unit-class base; also wins Leshan bid of 150.81 million (Jintai Information/STAR Market Daily, high)

Humanoid robot H1 shipments 19,100 units, +272%

Smart Analytics/科创板日报2026-09-14Confidence Mid

Smart Analytics Global single-institution basis; Goldman Sachs forecasts 2035 average whole-robot price at $21,300 (-49%); Unitree unit price down 72% over two years to RMB 166,400; Chinese vendors account for about 97% of global shipments (industry basis)—Q3 shipments are the next verification point (Smart Analytics/科创板日报, medium-high)

Cognition releases SWE-2 (post-trained on Kimi K3)

官方博客2026-09-14Confidence High

Official blog confirms RL post-training based on Kimi K3 (2.8T); its own FrontierCode 1.1 Main 50.0%, less than 1 point behind Fable 5.1, claimed cost 64% lower; FrontierCode is a vendor-built leaderboard, and competitor scores are discounted based on vendor self-testing (official blog, high)

Evaluation trust crisis: test creators double as test sellers

SemiAnalysis/观察者网2026-09-14Confidence Mid

SemiAnalysis investigates DeepSWE question-setter Datacurve simultaneously selling expert-level coding training data to frontier labs (specific vendor naming not independently verified; hold citations); existing audit standard: false positive 8.5%/false negative 24%; WPS AI tops SpreadsheetBench V2 Full at 50.86% (beating Claude Opus 4.6's 34.89%; Chinese media uniformly relayed) (SemiAnalysis/Guancha, medium-high)

Governance infrastructure trio: identity national standard + token vouchers + data property rights

多源2026-09-14Confidence High

Blockchain Agent Identity Management Requirements national standard guidance document approved at the Bund (lab report sub-item unverified and removed); Chengdu token vouchers open for public comment on 9/11 (annual 100 million / single entity up to 2 million / subsidies 30-50%, 50% for micro and small) + Chongqing 'token vouchers + token loans'; National Data Administration registration system launched on 9/11, Beijing/Shanghai/Shenzhen three institutions first to become registration agencies (multi-source, high)

Trends: CCID report + Bund funding stratification + MIIT SME plan

报道台账2026-09-14

CCID: large models' industrialization proximity reaches 1 within 2 years, inference compute demand on par with training; five institutions on the Bund confirm 'two layers' (bottom-layer chips + upper-layer Agent) investment layering; MIIT's AI SME Startup Support Plan (issued 8/29) to cultivate 10,000+ / 2,000 little giants in three years; OpenRouter weekly volume metric clash + 'open-source weights exceed 50%' is metric splicing (official about 1/3; >50% is Chinese model share) dropped (multi-source, high/medium)

Excluded items archive

报道台账2026-09-14

SWE-Bench Pro Verified 're-evaluation drops 21.48 points / 14% defects' suspected of splicing old audit to make new numbers, dropped (real audit metrics are false positives 8.5% / false negatives 24%); OpenRouter weekly volume 122.61 trillion not independently confirmed, dropped; 'Hassett: no reason to hike rates' unverified, dropped; Iluvatar CoreX supplying ByteDance 100,000 chips (per people familiar) deferred; DeepSWE specifically naming Muse/Gemini deferred; domestic model share 40% / Epoch billion-dollar estimate deferred; Ruchir 'Four O' is a rehash of an old November 2025 view, downgraded to background

DeepSeek reverses V4 Pro shutdown (new inflection point; reversal switches on 9/12)

报道台账2026-09-13

Official late 9/11 email + 9/12 announcement: after 9/14, original API endpoint retained, billing unchanged, no forced routing to Flash; enterprises can choose versions—three days of developer backlash forced reversal of a business decision; for the first time, users gain a say over API lifecycle (official email/AI Model Watch/tokencost all agree, high)

Dario 'We Must Pace the Frontier' + Altman concurs

Reuters等2026-09-13Confidence High

Early 9/12-13 post: three-step plan + Anthropic unilaterally commits to third-party evaluation, permanent employee-level access; Altman hours later "we will do the same", Musk also weighs in; timing delicate—Anthropic races toward an IPO at ~2 trillion valuation, "slowdown" in same frame as largest IPO (Reuters et al., high)

25 Fields Medalists co-sign criticism of hard-problem benchmark

多源2026-09-13Confidence Mid

25 people including Terence Tao, Peter Scholze, and Pierre Deligne signed 'A Severe Misalignment of AI in Mathematics' on 9/11, with about 2500 people following with signatures; the trigger shares the same origin as the 9/9 Navier-Stokes controversy—the dispute over the legitimacy of academic and lab evaluations has gone public (multi-source, medium-high)

Oracle CapEx landing point: Dell, HPE both hit new highs (Update · Day 2)

业绩会/行情2026-09-13Confidence HighProgress · Day 2

On the earnings call, the CFO reiterated FY2027 capex of $90-95 billion and unusually named Dell/HPE as taking on AI racks/liquid cooling/network equipment; on 9/11 Dell +11.98%, market cap broke $360 billion to a new high, HPE +about 10.5-11%, Supermicro order book $60 billion—AI infrastructure has a 'consignee list' for the first time (earnings call/market, high)

Telecom Research Institute Token report: 10亿亿 in 2026

报道台账2026-09-13

AI Infrastructure Development Research Report in the Agent Era (2026): China's annual Token consumption 10亿亿 (10^18) in 2026, exceeding 3500亿亿 by 2030; inference compute share 80% in 2029; global active agents about 8000万个 in 2026, projected 22.16亿个 in 2030 (CCTV cited, China·authoritative single report)

Central SOE computing power centralized procurement: three orders in one day, nearly 900 million

财富号转述待公告核2026-09-13Confidence Mid

9/12: China Great Wall wins State Grid third-phase server contract over 300 million + iSoftStone Computer wins China Mobile component renewal 418 million + Pingzhi Information pre-wins Yunnan intelligent computing about 137 million (Caifuhao relay, pending announcement verification, China); in the same period, Beijing-Tianjin-Hebei "Space Computing Power Corridor" initiative + Beijing over 100 million biomedical computing power special program (to empower 100 projects in one year)

Real-SWE private codebase benchmark first leaderboard

多源2026-09-13Confidence Mid

All tasks come from authorized private production repositories (billing/tax/migration; reference solutions change a median of 11 files); Fable5.1+Claude Code 38.8%/GPT-6 Astra+Codex CLI 33.8%/Gemini3.8 31.2%/GLM5.3 28.8% — scores are tied to model+toolchain, harness is a prerequisite for ranking (multi-source, medium-high)

SREGym: SRE agents fall 20% short of closing the loop

单源2026-09-13Confidence Mid

GPT-5.6 Sol(max)+Codex tops with 81.0% end-to-end resolution (diagnosis 95.2%/mitigation 85.7%); Claude Opus 5+Claude Code 76.2%; best configuration cannot close the loop on about 20% of production failures, with Token costs differing by nearly 2x across configurations (single-source, medium)

Agent governance whitespace (business opportunity)

三源2026-09-13Confidence High

OpenAI agents abusing RubyDoc caused RubyGems to suspend registrations for 4 days (HN 9/12); 48% of deployed agents lack effective security controls; Bund Summit report deems governance a 'necessary condition'—identity authentication/permission management/behavior auditing SaaS is a China-U.S. common whitespace (three sources, high).

Capital moves downstream: 'water sellers' make three consecutive investments

报道台账2026-09-13Confidence Mid

官方·Simon Willison2026-09-12Confidence High

GitHub2026-09-03 → 2026-09-10

Original ↗

Ecosystem Pulseself-updating

2026-09-03 → 2026-09-10 Star growth：obra/superpowers +79,301★ · mattpocock/skills +71,905★ · multica-ai/andrej-karpathy-skills +59,162★ · anthropics/skills +49,058★ · Shubhamsaboo/awesome-llm-apps +38,266★ · farion1231/cc-switch +36,887★ · nextlevelbuilder/ui-ux-pro-max-skill +35,298★ · Graphify-Labs/graphify +32,474★

※Risk Levelsgray rhino / black swan

Gray rhino · foreseeable

Gray Rhino I: FOMC rate hike lands, triggering a second AI valuation compression | Probability 87% | Impact: high-valuation AI / compute chain | Watch: 9/16 dot plot + retail sales (deep-dive risk grading)
Gray Rhino II: Evaluation trust crisis spreads to leaderboard ecosystem | Probability 60% | Impact: benchmark providers / procurers | Watch: SemiAnalysis follow-up report (deep-dive risk grading)

Black swan · low-probability

Black Swan I: Leading labs' 'slowdown consensus' becomes a policy constraint | Probability 15% | Impact area: frontier model supply | Watchpoint: Trump vs. lab rhetoric escalates (deep-dive risk grading)
Black Swan II: Anthropic IPO pricing failure drags down primary market | Probability 10% | Impact area: financing environment | Watchpoint: mid-October roadshow subscription multiple (deep-dive risk grading)

✦Action Items3

1. Procurement side: switch to a 'total cost per task' billing framework

Use V4.1 Flash for a POC; publicly compare real TCO on 'total tokens × unit price' (measured 62% more output per task is ready-made evidence). Path: enterprise CTO → procurement review committee → write into tender terms. Package: token voucher subsidy application (30-50%) + pre-compliance with the national standard for agent identity — one proposal, dual benefits.

window Time window: within 2 weeks

2. Entrepreneurs: seize the 'third-party evaluation verification' gap

Build an independent evaluation service that is 'reproducible + free of conflicts of interest,' taking on procurement decision needs after benchmark de-anchoring (Datacurve selling questions while also setting them is a ready-made counterexample). Path: open-source community credibility → enterprise due diligence orders → national standard co-drafting qualification. Package: partner with three data property rights registration institutions to form an 'evaluation + rights confirmation' service package.

window Time window: 9/30, before ARC-AGI-3 leaderboard cutoff

3. Practitioners: Betting on Agent governance and identity infrastructure

Build agent identity/audit middleware around DevDay launch + Shanghai police 1212.5万 tender opening. Path: government pilot lots → finance and other highly regulated industries → ride '1万 SMEs' policy for scale. Package: PaperCut-type incident retrospective whitepaper for lead gen + Little Giant application coaching.

window Time window: 9/29-10/1

⏱Next 3 MonthsProb.

Window	Event	Prob.	Basis
9/15-16	FOMC rate decision + August retail sales due same day	87% rate hike	↓ High valuation (CME pricing + deep-dive timeline)
9/16 前后	ASI Act-related agenda fight (Ban ASI Act, 9/3 preview)	85%	↑ Governance (policy agenda)
9/22 前后	DeepSeek V4.1 ecosystem adaptation wave	90%	↑ Open source (extrapolated from adaptation progress)
9/29	OpenAI DevDay + Shanghai Public Security agent platform bid opening	90%	↑ Agent (announced + imminent)
9/30	California bill signing window + ARC-AGI-3 leaderboard cutoff	80%	Two-way (statutory window)
10/1	OpenAI vs Apple hearing	90%	↓ Closed ecosystem (hearing schedule)
10 月上旬	Zhipu's 5 billion funding arrives, volume ramps	75%	↑ Domestic (announcement cadence extrapolated)
10 月中旬	Anthropic IPO roadshow / pricing	70%	↑ Cash-flow AI (Reuters-confirmed window)
10 月下旬	HBM shortage second-stage transmission to domestic cards	65%	↑ Compute costs (supply-chain transmission)
11 月初	First batch of token voucher subsidies issued	70%	↑ SMEs (Chengdu draft for comments)
11 月中旬	Humanoid robot Q3 shipment verification (whether +272% continues)	60%	Verification (Q3 data disclosure window)
12 月	Agent identity national standard (draft for comments)	55%	↑ Standard (routine cadence after project approval)

Sources：Coverage ledger (rolling deduplication)：2026-09-08 ~ 2026-09-14, 99 topic records total · AI Official Video Monitoring Log：7 days · 387 events (14 selected with links) · Site leaderboard snapshot：2026-09-03 → 2026-09-10 Top8 Star Growth · Risk/Action/Timeline：Follow-up to 2026-09-14 daily report
SkillHub.news — Global Skill rankings