报告形态 · 五维分类 + 风险 + 行动 + 时间线 ← 回站点
WEEK 2026-10-28 · 46 STORIES · SKILLHUB WEEKLY

SkillHub.news 每周AI

10 月 22 日 ~ 2026 年 10 月 28 日 · 周报 · 素材 4
本期看点46 条 · 约18分钟
01
本周主题线WEEKLY THREADS24 条
OpenAI 30-90%折扣撬走Anthropic大客户
02
官方事件精选OFFICIAL EVENTS14 条
Getting a Head Start with Codex | Mitchell Hashimoto
03
榜单异动RANKING MOVES8 条
obra/superpowers 星标 +79,301
01本周主题线WEEKLY THREADS24 条

OpenAI 30-90%折扣撬走Anthropic大客户

报道台账2026-10-26
Replit获超30%折扣后转移年合同额超1亿美元AI支出;Rippling编码负载OpenAI份额16%→52%(Anthropic 57%→18%);某70亿美元资产银行每席不足7美元获500席;投资者备忘录显示Anthropic Q2→Q3增速放缓——恰逢递表日冲击IPO定价叙事(The Information 10-24/The Decoder, 高)

寒武纪Q3业绩翻倍+非GPU加速器出货首超GPU

报道台账2026-10-26
寒武纪Q3营收35.31亿+104.47%、净利14.04亿+147.74%(前三季营收95.26亿+106.76%);海光+57.58%、沐曦扭亏预告;IDC:2026H1中国加速器出货200.8万颗,非GPU 101.6万颗(+80%)首超GPU,英伟达份额降至34%、华为升至28%(57.2万颗)(三季报/IDC转引, 高/中)

Cerebras 59片晶圆级芯片跑通3T Kimi K3

Cerebras官方博客多源转引2026-10-26置信度 高
4-bit量化分布59片WSE-3T(44GB片上SRAM装下每层权重),推理1000-2000+ tok/s、功耗约1.18MW;非英伟达硅片TB2.1 88.3%/BrowseComp 91.2%逼近GPT-5.6 Sol——前沿推理瓶颈证实为内存与互连而非FLOPs(Cerebras官方博客多源转引, 高)

FTC拟双查OpenAI/Anthropic+Storm-0940实战

含METR,Axios独家,立案文书未公布2026-10-26置信度 中
FTC拟依FTC法第5条向两家前沿厂商发强制取证(含METR,Axios独家,立案文书未公布,中);微软披露Storm-0940用智能体框架ARTEX攻破2250+ M365租户(部分3小时夺特权)、x402协议25漏洞、五大行拦截AI代理登录(微软官方/The Hacker News, 高/部分待核)

TikTok USDS洽购4GW+词元水电费化

娱乐资本论/政采公告2026-10-26置信度 高
USDS(银湖/甲骨文/MGX各15%、字节保留19.9%)接洽算力供应商,目标2030年代初锁定4GW≈甲骨文近5个季度交付量(850-1340亿美元量级,CNBC双源、4GW单源,中);北京台300万单一来源采购火山引擎词元(12个月),259家广电144条招标、融媒数智化重点项目9228万(娱乐资本论/政采公告, 高)

vLLM整柜跑满GB300逼近DeepSeek自研栈

报道台账2026-10-26
P/D分离+Wide-EP将DeepSeek-V4.1-Flash跑满GB300 NVL72:单GPU吞吐95.4万tok/s(与DeepInfer差距<6%)、上下文128K→855K仅+7%延迟——开源推理栈首逼头部闭源自研栈(vLLM官方博客, 中·无第三方复测)

Vals首版Router Index+法律榜倒挂

Vals/AA转引2026-10-26置信度 中
5款模型路由器无一进前十(最好Switchyard 57.8%排13/27,openrouter/auto仅52.2%),Gemini 4 Argon 69.2%居首;Muse Wonder Full法律指数77.86%登顶(Opus 5.5仅第七)+AA Harvey通过率13.06% vs Argon 3.89%碾压——通用榜与垂直场景严重倒挂(Vals/AA转引, 中)

微软开源Agensh+Mercury 3扩散LLM

报道台账2026-10-26
Agensh(MIT)去中心化多智能体框架:去中央编排器、1024编码智能体自组织,ProgramBench五难任务19.31%→28.78%(+49%);Inception Mercury 3扩散LLM输出1068 tok/s(约Luna 8倍),GPQA Diamond 89.23反超Luna 86.36,$0.35/M(MarkTechPost/机器之心, 高/中)

上海银行给AI一人公司放贷:经营数据=信用

上海市政府官网10-252026-10-26置信度 高
静安支行×闹海AI把平台上OPC创作者接单量/交付质量/回款数据纳入授信依据,9月已通过政银担机制发300万创业担保贷——轻资产AI创业主体首获系统性金融配套(上海市政府官网10-25, 高)

资本快照:亿铸10亿/Enact 3000万/Lambda 40亿

综合公开报道2026-10-26置信度 中
亿铸科技(3D DRAM存算一体已流片)超10亿元新融资;Enact Intelligence 3000万美元Pre-seed(LeCun/Mensch个人跟投);Lambda拟IPO前筹至多40亿美元@估值145亿——资金转向已验证流片/收入标的与明星学者早期票(综合公开报道, 中)

暂缓池(未采信)

报道台账2026-10-26置信度 中
"AWS 2026-27产能售罄、积压4960亿美元"(今日非财报日,数字疑为财报预期口径混入,暂缓);"1Password for Agents驱动年营收5亿美元"(单源待核);旧金山法院受理OpenAI智能体未授权访问诉讼(中文编译信源,待核)

长期跟踪区拐点更新

报道台账2026-10-26
Anthropic IPO:今日10-26为S-1递交窗口+OpenAI折扣战直接冲击定价(投资者预期回落至1.5万亿区间);OpenAI 300亿融资签约窗口下周;两者为下期首要跟踪点

合规卡600万颗排队:产能只够5-10%

报道台账2026-10-24
中企(字节/阿里/腾讯等)订购超600万颗RTX Pro 5500约350亿美元(单价约1.14万美元),英伟达2027全球产能上限约300万颗、商业交付2027Q1,现阶段仅满足5-10%——需求2倍于全球产能,算力稀缺制度化(路透10-20首发/DW/TrendForce多源,高)

AI债市全面走阔+SpaceX 400亿举债+Meta承诺8250亿

彭博/FT/SEC文件2026-10-24置信度 高
全球IG利差129bp创18个月最宽(CDX IG 64bp),年内AI债融资超4500亿美元(Janus Henderson:明年或近1万亿);SpaceX拟400亿美元举债购卡(100亿贷款+300亿IG债,Apollo牵头,BBB——系10-07旧闻降为背景增量);Meta租赁与合同承诺7000亿→8250亿美元,明言「赶在竞争对手占满融资渠道之前」(彭博/FT/SEC文件,高)

Anthropic IPO倒计时+OpenAI 300亿签约(双王定价)

报道台账2026-10-24
Anthropic拟10-26公开递交S-1、11-9路演、次周定价,拟募至多1000亿美元@至多2万亿、算力承诺升至7500亿(WSJ),纽约市审计长致信呼吁暂缓,投资者估值预期回落至1.5万亿区间;OpenAI 300亿融资(阿联酋资本+Coatue领投)传下周签约@投前1.25万亿,定位2027 IPO前过桥(IFR/CNBC,高/中高)

Kimi K3.1 Pro发布+OpenRouter中国71%

财联社/OpenRouter2026-10-24置信度 高
3T参数/100万上下文,TB4.0 65.7%超GPT-6 Astra 58.2%、SWE-Marathon 51.7%超Opus 5.5(46.3%),另有Highspeed高速版,权重最迟11-4开放;OpenRouter当周中国模型承载71% tokens(年初仅7%)、开源权重77%、DeepSeek请求份额24.4%居首(财联社/OpenRouter,高)

国办AI就业友好型意见18项 vs 巴克莱裁5000人

新华社/Guardian2026-10-24置信度 高
国办发〔2026〕29号(10-23发布,成文10-16)6方面18项:失业缓冲期(表述见「加大转型转岗就业支持」)、算法就业歧视治理、岗位替代风险评估、AI新职业开发(数据标注/AI训练师)、中小企业「小快轻准」轻量化升级;巴克莱裁5000人归因AI(利润+12%);Gartner 2002名CIO仅13%见实质回报(新华社/Guardian,高)

规则双线:得州第10州起诉OpenAI+训练禁直连实时互联网

The Hill/官方博客/央视2026-10-24置信度 高
Paxton 10-23起诉OpenAI+Altman(消费者保护法+SCOPE未成年人法,寻求禁未成年使用、新模型第三方评估批准,每违规罚1万美元,此前8州+佛州=第10州);OpenAI官宣前沿模型训练禁直连实时互联网、改缓存网页沙箱(红队发现VM逃逸/侧信道,起因9-20 DNS漏洞短暂联网,10-1已恢复RL训练);商务部解读:开源出口「即报即出」、API/SaaS/端侧不算技术出口、DeepSeek/MiniMax已履行登记(The Hill/官方博客/央视,高)

谷歌TPU首进Meta生产环境+产业交付线

报道台账2026-10-24
PyTorch原生运行时免改写跑TPU推理,Meta数据中心实际部署——英伟达垄断首被生产级撬动;Vera Rubin NVL72批量交付(每百万token推理成本前代1/45);辉羲智能年内融资超25亿元(机器人芯片出货全球第二约1/5);宇树G1累计下线1-2万台(王兴兴:普及需年产千万台);Cybercab车队169→319辆、11月进达拉斯、无监督FSD推迟2027Q2;中国移动宁夏中卫37.1亿EPC(186MW,与10-11交换机集采为不同项目)(Tom's Hardware/澎湃/财报会/北极星,高/中高)

GPT-5.6 Sol挖出PHP十年高危漏洞+生态安全层

OpenAI/Wiz/GitHub/澎湃2026-10-24置信度 高
两个未授权RCE「PHP2Shell」CVSS 9.2波及PHP 5.4-8.5,Wiz独立复现、官方发4个安全更新——AI找洞从演示变交付;MCP参考实现v2.0.0+Tasks and Events工作组设立(Agents WG降级兴趣组);Skills捐AAIF生态数据(100万Skills/2.8亿安装,厂商自报);美DHS要求联邦机构清退DeepSeek(OpenAI/Wiz/GitHub/澎湃,高)

数据三线:token单价-41%用量+50%/ChatGPT广告83国/Muse千万装机

报道台账2026-10-24
Ramp企业AI指数:有效token单价3月$1.15→9月$0.68(-41%)、用量环比+50%创纪录、企业侧Anthropic 51% vs OpenAI 44.5%、开源/中国企业采用仅6.4%(补课空间);ChatGPT广告新增26国至83国、广告主超10万(8月仅2万,月增约5倍);Muse Wonder超1000万活跃装机/约250万DAU(Sensor Tower)(Ramp多源转引/中高)

榜单口径战:AA v5 TB Science首测+TB版本混淆警示

报道台账2026-10-24
v5新增Terminal-Bench Science首测(Astra 63.3%/Opus 5.5 61.9%)+私有防污染集,v4.3.2发布后不可比;DeepSeek 1017 checkpoint登HF agents-last-exam 35.4+deep-swe 77双榜;第三方对比:同一模型TB2.1约76-88% vs TB4.0严格口径仅3-28%——引用高分先核基准版本号(AlphaSignal/dev.to/BenchLeader,中高/中)

资本快照:DeepSeek交割细节/Rivian收购世界模型/微亿智造挂牌

彭博/TechCrunch/Crunchbase2026-10-24置信度 高
DeepSeek交割投前约700亿美元(宁德时代拟增资100亿翻番,新进君联/石溪/吉利/中芯聚源/横店);Rivian收购Third Dimension(前Tesla/Wayve世界模型团队,扩建伦敦AI中心);Flapping Airplanes 10亿美元A轮(Lux+Index领投,成立不足1年);Higgsfield获英伟达+TCV追加(8月B轮4亿/估值54亿扩展);微亿智造10-28港股挂牌「工业具身智能第一股」(彭博/TechCrunch/Crunchbase,高)

趋势:实验室中层密集改口AGI+灭绝警告升级

报道台账2026-10-24置信度 中
前DeepMind Alex Turner:10年内灭绝概率1/3、超级智能12-18个月;DeepMind安全负责人Irving呼吁现在停止进步(先单边后中美协议);Hinton称AI主观体验与人类「一样多」;Anthropic RL负责人Douglas称AGI一两年内(与Hassabis 2-3年同向);Muse Wonder定价$0.2/M vs $200/月(单源待核)(NDTV/agihunt/Dealroom,中高/中)
02官方事件精选OFFICIAL EVENTS14 条

Genie 4: A new generation of world simulation

YouTube·Google DeepMind2026-10-28置信度 高
原文 ↗

Introducing Genie 4

YouTube·Google Developers2026-10-28置信度 高
原文 ↗

Developing with advanced agent workflows in Looker

YouTube·Google Cloud2026-10-28置信度 高
原文 ↗

BigQuery to Databricks: A Strategic Framework

官方·Databricks2026-10-28置信度 高
原文 ↗
03榜单异动RANKING MOVES8 条

obra/superpowers 星标 +79,301

GitHub2026-09-03 → 2026-09-10
原文 ↗

mattpocock/skills 星标 +71,905

GitHub2026-09-03 → 2026-09-10
原文 ↗

anthropics/skills 星标 +49,058

GitHub2026-09-03 → 2026-09-10
原文 ↗

farion1231/cc-switch 星标 +36,887

GitHub2026-09-03 → 2026-09-10
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
  • 世界模型「可演示不可交付」:Genie 4 仍是 research preview,官方承认长时漂移、音频准确性与画面内文字三项未解,影视/游戏采购若按成片标准签约,返工与验收争议将在数周内集中爆发(中高,50-70%)
  • AI slop 侵蚀科研与内容信任:2025 年约 57% 论文出现大模型痕迹、《Nature Communications》超 12% 评审全 AI 生成,而每 40 篇仅 1 篇披露——评价体系与采购验收若仍以「文本产出量」计量,劣币出清会更慢(中高,40-60%)
  • 资本向少数标的集中:100 亿美元单笔高至 10 亿美元的晚期基金 + 世界模型侧连续并购,二线 AI 公司在 2027 年前被迫卖身或转债权融资,估值分化加剧(中,40-60%)
  • 治理折价显性化:Anthropic 的信托 + 创始人 50.1% 结构让普通股东几乎无表决权,被动资金纳入指数时须重估风险溢价,同类「公益公司」标的将集体面对估值折让(中,35-55%)
黑天鹅 · 低概率大冲击
  • 密码学应急:若 EHT 否证成立且算法外溢到现有密码参数,依赖密钥长度的系统需集体评估;OpenAI 称已检查、认为不构成破解,但学术界尚未独立复核(低,<25%)
  • 仿真-现实差距事故:世界模型与仿真开始进入自动驾驶与机器人测试(Rivian 收购 Third Dimension、亚马逊洽谈 Decart、AMD 收购 World Labs),一旦以仿真通过替代实车验证,安全事件的责任划分将无先例可循(低,<20%)
  • AI 生成证据污染司法与监管:IEEE 实验证明无专业背景者可用商用模型造出足以进入评审流程的假论文,同类手法平移到法律、医疗与财报场景会直接冲击「文档即证据」的默认假设(低,<20%)
✦行动建议3
1. 把「团队共享上下文」做成一周可交付的小产品

先量三个漏损指标(周重复解题次数、洞察复用率、卡点时长),再落地最小闭环:共享提示词库 + 团队级 agent 工作区 + 每次会话沉淀为可检索条目。路径:内部先用一周跑出基线 → 对外包装成「AI 上下文审计」按团队数收费 → 进阶做跨系统上下文层(对接工单、IM、知识库)。判据:四周内周重复解题次数下降过半。

时间窗 即刻起 4—6 周(企业 Q4 预算窗口)
2. 给现有 agent 上「预算三档 + MCP 白名单」

在一周内把成本与越权从事故变成配置:token 预算 80% 告警/90% 降级/100% 阻断,MCP 服务白名单,单 agent 与整组 kill switch,子代理按角色标注。可自研也可直接自托管 flightdeck(Apache 2.0,进程内 sensor 不改代码),但需自行过一遍代码与抓包确认默认不落提示词。判据:一次压测里能把某条越权 MCP 调用拦在出进程前。

时间窗 1—2 周(先止血,再谈治理平台)
3. 抢「AI 预演」这一格的交付位置

世界模型现在能给的只有分镜与创意预演,成片仍要走既有管线。路径:先用现成工具(Veo/GWM Worlds 类)把「一句话→分镜预演」做成 48 小时交付包,同时按官方承认的三条缺口写验收表(两分钟不漂移、音频与环境匹配、文字可读);等 Genie 4 向 AI Ultra 与 Gemini API 早期伙伴开放时,第一批就有可对比的交付样本。判据:拿到 3 个愿意付费的预演单。

时间窗 3—6 个月(开放窗口前抢跑)
⏱未来 3 个月时间线概率
时间窗事件概率依据
未来数周Genie 4 从 trusted testers 扩到 Google AI Ultra 订阅者(Project Genie)与 Gemini API/AI Studio 早期伙伴高(85%)TechCrunch 引 Google 简报的官方时点
未来 3 个月Anthropic 公布发行定价区间并挂牌(S-1 已交,TechCrunch 称正接近 IPO)高(75%)SEC 424B 文件与交易所公告为准
未来 3 个月世界模型侧整合落地:AMD 收购 World Labs、亚马逊收购 Decart、Rivian 收购 Third Dimension 的最终条款披露中(55%)TechCrunch 报道方向已定,交易完成待监管与官宣
未来 3 个月计算机科学界对 EHT 否证的公开评审结论(工作组/会议产出)中(45%)Ryan Williams 已发公开信征集反馈,尚无评审日程
未来 3 个月Oracle AI Workspace 在 ChatGPT 内开放形态与计价口径披露(推断,官方未给时点)中(40%)推断:官方仅称 Oracle 在开发,未披露时间与定价
素材来源:报道台账(滚动查重):2026-10-22 ~ 2026-10-28 共 41 条主题记录 · AI 官方视频监控台账:7 天 · 607 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 2026-10-28 日报
SkillHub.news — 全球 Skill 榜单
WEEK 2026-10-28 · 46 STORIES · SKILLHUB WEEKLY
SkillHub.news Weekly AI
Oct 22 ~ Oct 28, 2026 · Weekly · sources 4
Highlights46 stories · approx.18min read
01
WEEKLY THREADS24 stories
OpenAI's 30-90% discounts poach Anthropic's major customers
02
OFFICIAL EVENTS14 stories
Getting a Head Start with Codex | Mitchell Hashimoto
03
RANKING MOVES8 stories
obra/superpowers stars +79,301
01WEEKLY THREADS24 stories

OpenAI's 30-90% discounts poach Anthropic's major customers

报道台账2026-10-26
After securing over 30% discounts, Replit shifted more than $100 million in annual contract value of AI spending; for Rippling's coding workloads, OpenAI's share went 16%→52% (Anthropic 57%→18%); a bank with $7 billion in assets got 500 seats at under $7 per seat; an investor memo shows Anthropic's Q2→Q3 growth slowed—coinciding with filing day and hitting the IPO pricing narrative (The Information 10-24/The Decoder, High)

Cambricon Q3 results double; non-GPU accelerator shipments surpass GPU for first time

报道台账2026-10-26
Cambricon Q3 revenue 35.31亿, +104.47%, net profit 14.04亿, +147.74% (first three quarters revenue 95.26亿, +106.76%); Hygon +57.58%, MetaX preannounces turnaround; IDC: 2026H1 China accelerator shipments 200.8万颗, non-GPU 101.6万颗 (+80%), surpassing GPU for first time; NVIDIA share down to 34%, Huawei up to 28% (57.2万颗) (Q3 report/IDC cited, high/medium)

Cerebras runs 3T Kimi K3 on 59 wafer-scale chips

Cerebras官方博客多源转引2026-10-26Confidence High
4-bit quantization distributed across 59 WSE-3T chips (44GB on-chip SRAM holds each layer's weights), inference 1000-2000+ tok/s, power about 1.18MW; on non-NVIDIA silicon, TB2.1 88.3%/BrowseComp 91.2% nears GPT-5.6 Sol—frontier inference bottleneck confirmed as memory and interconnect rather than FLOPs (Cerebras official blog via multiple sources, High)

FTC plans dual probes into OpenAI/Anthropic + Storm-0940 real-world operations

含METR,Axios独家,立案文书未公布2026-10-26Confidence Mid
FTC plans to issue compulsory evidence demands to two frontier vendors under Section 5 of the FTC Act (including METR; Axios exclusive; case filing not disclosed; Medium); Microsoft discloses Storm-0940 used agent framework ARTEX to breach 2250+ M365 tenants (some gained privileges in 3 hours), 25 vulnerabilities in the x402 protocol, and five major banks blocked AI agent logins (Microsoft official/The Hacker News, High/partly pending verification)

TikTok USDS in talks to buy 4GW+; tokens becoming utility-like

娱乐资本论/政采公告2026-10-26Confidence High
USDS (Silver Lake/Oracle/MGX 15% each, ByteDance retains 19.9%) in talks with compute suppliers, targeting 4GW by the early 2030s ≈ Oracle's deliveries over nearly 5 quarters ($85–134 billion scale; CNBC double-source, 4GW single-source, medium confidence); Beijing TV 3 million single-source procurement of Volcano Engine tokens (12 months), 259 broadcasters 144 tenders, key converged-media digital-intelligence projects 92.28 million (娱乐资本论/government procurement notices, high confidence)

vLLM saturates a full GB300 rack, approaching DeepSeek's in-house stack

报道台账2026-10-26
P/D separation + Wide-EP saturates GB300 NVL72 with DeepSeek-V4.1-Flash: per-GPU throughput 954k tok/s (gap vs DeepInfer <6%), context 128K→855K adds only +7% latency—the first open-source inference stack to close in on leading closed-source in-house stacks (vLLM official blog, medium confidence · no third-party replication)

Vals' first Router Index + legal leaderboard inversion

Vals/AA转引2026-10-26Confidence Mid
None of the 5 model routers made the top 10 (best Switchyard 57.8% ranks 13/27; openrouter/auto only 52.2%), Gemini 4 Argon tops at 69.2%; Muse Wonder Full tops legal index at 77.86% (Opus 5.5 only seventh); AA Harvey pass rate 13.06% vs Argon 3.89% crushes — severe inversion between general leaderboard and vertical scenarios (Vals/AA cited, Medium)

Microsoft open-sources Agensh + Mercury 3 diffusion LLM

报道台账2026-10-26
Agensh (MIT) decentralized multi-agent framework: no central orchestrator, 1024 coding agents self-organize; ProgramBench five hard tasks 19.31%→28.78% (+49%); Inception Mercury 3 diffusion LLM outputs 1068 tok/s (about 8x Luna), GPQA Diamond 89.23 overtakes Luna's 86.36, $0.35/M (MarkTechPost/机器之心, High/Medium)

上海银行 lends to AI one-person companies: operating data = credit

上海市政府官网10-252026-10-26Confidence High
静安支行 × 闹海AI includes platform OPC creators' order volume/delivery quality/payment collection data in credit-granting criteria; in 9月 it issued 300万 in entrepreneurship guarantee loans via the government-bank-guarantee mechanism—asset-light AI entrepreneurial entities receive systematic financial support for the first time (Shanghai Municipal Government official website, 10-25, High)

Capital snapshot: 亿铸 10亿 / Enact 3000万 / Lambda 40亿

综合公开报道2026-10-26Confidence Mid
亿铸科技 (3D DRAM compute-in-memory has taped out) raises over 10亿元 in new funding; Enact Intelligence 3000万美元 Pre-seed (personal follow-on by LeCun/Mensch); Lambda plans to raise up to 40亿美元 pre-IPO @ valuation 145亿—capital shifts to targets with verified tape-out/revenue and early bets on star scholars (synthesized public reports, Medium)

Deferred pool (not adopted)

报道台账2026-10-26Confidence Mid
"AWS 2026-27 capacity sold out, backlog 4960亿美元" (today is not an earnings-report day; the figure is suspected to be mixed in from earnings-expectation basis; deferred); "1Password for Agents drives annual revenue of 5亿美元" (single-source, pending verification); San Francisco court accepts OpenAI agent unauthorized-access lawsuit (Chinese compiled source, pending verification)

Long-term tracking section: inflection-point update

报道台账2026-10-26
Anthropic IPO: today 10-26 is the S-1 filing window + OpenAI discount war directly hits pricing (investor expectations fall back to 1.5万亿 range); OpenAI 300亿 funding signing window next week; both are primary tracking points for the next issue

6 million compliant cards in queue: capacity only enough for 5-10%

报道台账2026-10-24
Chinese firms (ByteDance/Alibaba/Tencent etc.) ordered over 6 million RTX Pro 5500 for about $35 billion (unit price about $11,400); NVIDIA 2027 global capacity cap about 3 million units, commercial delivery 2027Q1, currently only satisfying 5-10%—demand is 2x global capacity, compute scarcity becoming institutionalized (Reuters 10-20 first report/DW/TrendForce multi-source, High)

AI bond market spreads widen across the board + SpaceX $40 billion debt raise + Meta $825 billion commitments

彭博/FT/SEC文件2026-10-24Confidence High
Global IG spreads 129bp, widest in 18 months (CDX IG 64bp); AI debt financing exceeds $450 billion year-to-date (Janus Henderson: could be near $1 trillion next year); SpaceX plans $40 billion debt raise to buy cards ($10 billion loan + $30 billion IG bonds, Apollo-led, BBB—old news from 10-07 downgraded to background increment); Meta lease and contract commitments $700 billion→$825 billion, explicitly saying before competitors fill up financing channels (Bloomberg/FT/SEC filing, High)

Anthropic IPO countdown + OpenAI $30 billion signing (dual-king pricing)

报道台账2026-10-24
Anthropic plans to publicly file S-1 on 10-26, roadshow on 11-9, price the following week; aiming to raise up to $100 billion at up to $2 trillion, compute commitment raised to $750 billion (WSJ); New York City comptroller sends letter calling for delay; investor valuation expectations fall back to $1.5 trillion range; OpenAI $30 billion financing (UAE capital + Coatue lead) reportedly to sign next week at $1.25 trillion pre-money, positioned as a bridge before 2027 IPO (IFR/CNBC, High/Medium-High)

Kimi K3.1 Pro released + OpenRouter China 71%

财联社/OpenRouter2026-10-24Confidence High
3T parameters/1M context, TB4.0 65.7% beats GPT-6 Astra 58.2%, SWE-Marathon 51.7% beats Opus 5.5 (46.3%), plus Highspeed fast version, weights to open by 11-4 at latest; on OpenRouter that week Chinese models carried 71% of tokens (only 7% at start of year), open-source weights 77%, DeepSeek request share 24.4% ranked first (Cailian Press/OpenRouter, High)

State Council General Office AI employment-friendly guidelines: 18 items vs Barclays cutting 5000 jobs

新华社/Guardian2026-10-24Confidence High
State Council General Office Document [2026] No. 29 (issued 10-23, dated 10-16), 6 areas and 18 items: unemployment buffer period (wording under “increase support for transition and job-transfer employment”), algorithmic employment discrimination governance, job-displacement risk assessment, AI new-profession development (data annotation/AI trainer), and “small, fast, light, precise” lightweight upgrading for SMEs; Barclays’ 5000 layoffs attributed to AI (profit +12%); Gartner 2002 CIOs: only 13% see substantive returns (Xinhua/Guardian, high).

Two-track rules: Texas becomes 10th state to sue OpenAI; training banned from direct connection to the real-time internet

The Hill/官方博客/央视2026-10-24Confidence High
Paxton sued OpenAI + Altman on 10-23 (consumer protection law + SCOPE minors law; seeking to ban minors from use and require third-party evaluation and approval for new models, USD 10,000 per violation; with 8 prior states plus Florida, making it the 10th state); OpenAI announced that frontier model training is banned from direct connection to the real-time internet and switched to cached-web sandboxes (red team found VM escape/side channel; caused by a 9-20 DNS vulnerability that briefly connected to the internet; RL training resumed on 10-1); Ministry of Commerce interpretation: open-source exports “report and export immediately”; API/SaaS/device-side do not count as technology exports; DeepSeek/MiniMax have completed registration (The Hill/official blog/CCTV, high).

Google TPU first enters Meta production environment + industrial delivery line

报道台账2026-10-24
PyTorch native runtime runs TPU inference without code rewriting, with actual deployment in Meta data centers—NVIDIA's monopoly first pried open at production level; Vera Rubin NVL72 delivered in volume (inference cost per million tokens at 1/45 of the previous generation); HuiXi Intelligence raised over RMB 2.5 billion this year (robot chip shipments second globally at about 1/5); Unitree G1 cumulative production reached 10,000-20,000 units (Wang Xingxing: mass adoption requires annual output of 10 million units); Cybercab fleet 169→319 vehicles, entering Dallas in November, unsupervised FSD delayed to 2027Q2; China Mobile's RMB 3.71 billion EPC in Zhongwei, Ningxia (186MW, a different project from the 10-11 switch centralized procurement) (Tom's Hardware/The Paper/earnings call/Polaris, high/medium-high)

GPT-5.6 Sol uncovers a ten-year-old high-severity PHP vulnerability + ecosystem security layer

OpenAI/Wiz/GitHub/澎湃2026-10-24Confidence High
Two unauthenticated RCEs 'PHP2Shell' with CVSS 9.2 affect PHP 5.4-8.5; Wiz independently reproduced, and the vendor released 4 security updates—AI vulnerability discovery moves from demo to delivery; MCP reference implementation v2.0.0 + Tasks and Events working group established (Agents WG downgraded to interest group); Skills donates AAIF ecosystem data (1 million Skills/280 million installs, vendor self-reported); US DHS requires federal agencies to remove DeepSeek (OpenAI/Wiz/GitHub/澎湃, high)

Three data points: token unit price -41%, usage +50% / ChatGPT ads in 83 countries / Muse 10 million installs

报道台账2026-10-24
Ramp Enterprise AI Index: effective token unit price $1.15 in March → $0.68 in September (-41%), usage +50% MoM, a record; enterprise-side Anthropic 51% vs OpenAI 44.5%; open-source/Chinese enterprise adoption only 6.4% (room to catch up); ChatGPT ads added 26 countries to reach 83, advertisers over 100,000 (only 20,000 in August, about 5x monthly growth); Muse Wonder over 10 million active installs / about 2.5 million DAU (Sensor Tower) (Ramp multi-source citation/medium-high)

Leaderboard methodology battle: AA v5 TB Science first test + TB version confusion warning

报道台账2026-10-24
v5 adds first Terminal-Bench Science test (Astra 63.3%/Opus 5.5 61.9%) + private anti-contamination set; not comparable after v4.3.2 release; DeepSeek 1017 checkpoint lands on both HF agents-last-exam 35.4 and deep-swe 77 leaderboards; third-party comparison: same model about 76-88% on TB2.1 vs only 3-28% under TB4.0 strict criteria—check benchmark version number before citing high scores (AlphaSignal/dev.to/BenchLeader, medium-high/medium)

Capital snapshot: DeepSeek closing details / Rivian acquires world model / 微亿智造 listing

彭博/TechCrunch/Crunchbase2026-10-24Confidence High
DeepSeek deal pre-money about $70 billion (CATL plans to add 10 billion to double; new investors 君联/石溪/吉利/中芯聚源/横店); Rivian acquires Third Dimension (former Tesla/Wayve world-model team; expanding London AI center); Flapping Airplanes $1 billion Series A (led by Lux + Index; less than 1 year old); Higgsfield gets follow-on from NVIDIA + TCV (August Series B $400 million / $5.4 billion valuation extension); 微亿智造 to list in Hong Kong on 10-28 as the 'first industrial embodied intelligence stock' (Bloomberg/TechCrunch/Crunchbase, high)

Trend: lab mid-level staff shift stance en masse on AGI; extinction warnings escalate

报道台账2026-10-24Confidence Mid
Former DeepMind Alex Turner: 1/3 extinction probability within 10 years, superintelligence in 12-18 months; DeepMind safety lead Irving calls for stopping progress now (unilateral first, then US-China agreement); Hinton says AI subjective experience is 'as much' as humans; Anthropic RL lead Douglas says AGI within one or two years (same direction as Hassabis's 2-3 years); Muse Wonder pricing $0.2/M vs $200/month (single source pending verification) (NDTV/agihunt/Dealroom, medium-high/medium)
02OFFICIAL EVENTS14 stories

Genie 4: A new generation of world simulation

YouTube·Google DeepMind2026-10-28Confidence High
Original ↗

Introducing Genie 4

YouTube·Google Developers2026-10-28Confidence High
Original ↗

Developing with advanced agent workflows in Looker

YouTube·Google Cloud2026-10-28Confidence High
Original ↗

BigQuery to Databricks: A Strategic Framework

官方·Databricks2026-10-28Confidence High
Original ↗
03RANKING MOVES8 stories

obra/superpowers stars +79,301

GitHub2026-09-03 → 2026-09-10
Original ↗

mattpocock/skills stars +71,905

GitHub2026-09-03 → 2026-09-10
Original ↗

anthropics/skills stars +49,058

GitHub2026-09-03 → 2026-09-10
Original ↗

farion1231/cc-switch stars +36,887

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
  • World models ‘demoable but undeliverable’: Genie 4 remains a research preview; officials acknowledge three unresolved issues—long-duration drift, audio accuracy, and in-frame text; if film/TV and game procurement contracts to finished-film standards, rework and acceptance disputes will erupt within weeks (medium-high, 50-70%)
  • AI slop erodes research and content trust: in 2025 about 57% of papers show large-model traces, over 12% of Nature Communications reviews are fully AI-generated, yet only 1 of every 40 discloses it—if evaluation systems and procurement acceptance still measure by ‘text output volume,’ clearing out bad actors will be slower (medium-high, 40-60%)
  • Capital concentrates into a few targets: a USD 10 billion late-stage fund with single checks up to USD 1 billion + serial M&A on the world-model side; second-tier AI companies will be forced to sell themselves or shift to debt financing before 2027, intensifying valuation divergence (medium, 40-60%)
  • Governance discount becomes explicit: Anthropic’s trust + founder 50.1% structure leaves ordinary shareholders with almost no voting rights; when passive funds include it in indices, they must reassess the risk premium, and similar ‘public benefit corporation’ targets will collectively face valuation discounts (medium, 35-55%)
Black swan · low-probability
  • Cryptography emergency: If the EHT falsification holds and the algorithm spills over into existing cryptographic parameters, systems relying on key length need a collective assessment; OpenAI says it has checked and does not consider it a break, but academia has not yet independently reviewed it (low, <25%)
  • Simulation-reality gap incidents: World models and simulation are starting to enter autonomous driving and robotics testing (Rivian acquires Third Dimension, Amazon in talks over Decart, AMD acquires World Labs); once passing simulation replaces real-vehicle validation, liability allocation for safety incidents will have no precedent to follow (low, <20%)
  • AI-generated evidence contaminates justice and regulation: IEEE experiments show that people without professional backgrounds can use commercial models to create fake papers sufficient to enter the review process; applying the same methods to legal, medical, and financial-reporting scenarios would directly challenge the default assumption that 'documents are evidence' (low, <20%)
✦Action Items3
1. Turn "team shared context" into a small product that can ship in a week.

First measure three leakage metrics (weekly repeated problem-solving count, insight reuse rate, blocker duration), then implement the minimal closed loop: shared prompt library + team-level agent workspace + each session captured as searchable entries. Path: internally run for one week to establish a baseline → package externally as 'AI context audit' and charge per team → advance to a cross-system context layer (integrated with ticketing, IM, knowledge base). Criterion: weekly repeated problem-solving count falls by more than half within four weeks.

window Starting immediately, 4—6 weeks (enterprise Q4 budget window)
2. Give existing agents 'three budget tiers + MCP allowlist'

Within one week, turn cost and unauthorized access from incidents into configuration: token budget alerts at 80% / downgrade at 90% / block at 100%, MCP service allowlist, kill switch for individual agents and the whole group, and sub-agents labeled by role. Can build in-house or directly self-host flightdeck (Apache 2.0, in-process sensor without code changes), but must review the code and capture packets yourself to confirm prompts are not logged by default. Criterion: in one stress test, block an unauthorized MCP call before it leaves the process.

window 1—2 weeks (first stop the bleeding, then discuss a governance platform)
3. Claim the delivery slot for 'AI previsualization'

World models can currently offer only storyboards and creative previsualization; final cuts still go through the existing pipeline. Path: first use off-the-shelf tools (Veo/GWM Worlds-type) to turn 'one sentence → storyboard previsualization' into a 48-hour delivery package, while writing an acceptance checklist for the three officially acknowledged gaps (no drift for two minutes, audio and environment matching, readable text); when Genie 4 opens to AI Ultra and Gemini API early partners, the first batch will have comparable delivery samples. Criterion: secure 3 paying previsualization orders.

window 3—6 months (get ahead before the opening window)
⏱Next 3 MonthsProb.
WindowEventProb.Basis
未来数周Genie 4 expands from trusted testers to Google AI Ultra subscribers (Project Genie) and early Gemini API/AI Studio partnersHigh (85%)TechCrunch cites official timing from Google briefing
未来 3 个月Anthropic announces issue price range and listing (S-1 filed; TechCrunch says nearing IPO)High (75%)Subject to SEC 424B filing and exchange announcement
未来 3 个月World-model consolidation lands: final terms disclosed for AMD's acquisition of World Labs, Amazon's acquisition of Decart, and Rivian's acquisition of Third DimensionMedium (55%)TechCrunch reports direction is set; deal completion awaits regulatory and official announcement
未来 3 个月Computer science community's public review conclusion on the EHT refutation (working group/conference output)Medium (45%)Ryan Williams has issued an open letter soliciting feedback; no review schedule yet
未来 3 个月Oracle AI Workspace’s availability format and pricing basis inside ChatGPT disclosed (inferred; no official timing given)Medium (40%)Inferred: officials only say Oracle is in development; no timeline or pricing disclosed
Sources:Coverage ledger (rolling deduplication):2026-10-22 ~ 2026-10-28, 41 topic records total · AI Official Video Monitoring Log:7 days · 607 events (14 selected with links) · Site leaderboard snapshot:2026-09-03 → 2026-09-10 Top8 Star Growth · Risk/Action/Timeline:Continues from the 2026-10-28 daily report
SkillHub.news — Global Skill rankings