01MODEL RELEASES2 stories
OpenAI pretraining safety protocol lands: safety cases must be written before frontier RL runs start
OpenAI官方博客 + Altman XConfidence HighProgress · Day 3
Official blog + Altman on X: safety cases must be drafted before frontier RL runs start (safety-case threshold)—the verbal consensus of “Dario’s long post → Altman’s endorsement” is codified into protocol text for the first time. Meanwhile, Trump publicly rejects calls to slow down, a policy collision inflection point.
Take · The safety protocol is a “release-cadence variable”—for top labs’ version schedules, from today there is an extra safety gate; for buyers, it means delivery timelines need to reserve compliance buffer.
Domestic open source rises across three leaderboards: GLM-5.3 ranks 3rd on SuperCLUE and 1st among open source on AA; DeepSeek V4.1-Flash ranks 2nd among domestic.
SuperCLUE / Artificial Analysis / LMArena榜单官网（9/11 更新批次）Confidence HighProgress · Day 5
SuperCLUE September edition: GLM-5.3 Chinese overall score 71.29 (+8.02), jumping from 5th to 3rd; DeepSeek V4.1-Flash debuts at 71.81, 2nd among domestic (with official V4 Pro full routing added, effectively a flagship generation change). AA Intelligence Index: GLM-5.3 scores 44.9, 1st among open-weight models (Kimi K3 43.8, V4.1-Flash 39.5); on the LMArena company leaderboard, GLM-5.3 rises to 5th.
Take · Three metrics (Chinese overall / AA index / Elo) point the same way: domestic open source gains come from more post-training + architecture refresh, not pretraining breakthroughs. The domestic base-model cost-performance inflection point is now corroborated across three leaderboards—a real-world test of total per-task cost for replacing overseas APIs is worth doing this week.
02PRODUCT2 stories
Open-source verticalization 2.0 triple: Xiaohongshu open-sources Search Agent Iris; Shanghai AI Lab releases research foundation model Intern-S2
小红书 AllSpark / 上海 AI 实验室官方 / 人民网Confidence HighNewcomer
Xiaohongshu AllSpark open-sourced Search Agent Iris on 9/14 (mini 35B / pro 397B, based on Qwen, BrowseComp 82.2 / 88.6, strongest open-source at comparable scale; weights + evaluation code released); Shanghai AI Lab released research-oriented open-source foundation model Intern-S2 (Intern-S2-397B) at the 9/13 Pujiang Innovation Forum, with vLLM / SGLang Day-0 support, and simultaneously opened the Intern·Duanyan scientific discovery platform.
Take · Open-source competition is moving down from general-purpose chat to the search / research / context-engineering infrastructure layer—for the first time, independent developers can use open-source weights for industry-specific packaging and match closed source.
MetaX's Xiyun C700 nears tape-out: core design and functional verification near completion
沐曦业绩说明会（9/7）Confidence MidVendor claims
Earnings briefing (9/7) wording: core design and functional verification near completion, next step tape-out, fully domestic supply chain; 'benchmarks against H100' is the company's own claim, with no third-party testing.
Take · Next validation point for domestic compute supply; discount vendor claims in citations—watch tape-out and third-party testing as the two follow-up milestones.
03INDUSTRY9 stories
First pricing day for the slowdown narrative: SOX -5.86% (largest single-day drop since 7/1), money flees hardware into software/security.
美股 / 费半多源行情Confidence HighNewcomer
9/14 US stocks: SOX -5.86%; NVIDIA -3.36%, Micron -5.25%, Intel -5.59%, Corning -13.70%, Coherent -12.73%; Nasdaq only -0.56%. Countertrend: CrowdStrike +13.85% to a record high; software stocks broadly up 4-7%.
Take · First formal pricing day after the CEO's 'slowdown call'; same day completed a 'hardware → software/security' rotation—AI market narrative switch confirmed by the market for the first time.
Macro tightening on three fronts: 10Y breaks 5% intraday for first time + Saudi oil pipeline attack halts operations + CME rate-hike odds rise to 90-92%
宏观 / CME新华社 / CMEConfidence HighProgress · Day 1
10Y hit 5.017% intraday, first breach of 5% since Oct 2023 (slightly off at close); Saudi oil pipeline damaged and shut down by attack (not voluntarily closed), Brent closed at 105.68 / WTI 101.39 (intraday touched 109.8 / 104.95); CME odds of a 25bp rate hike rise to 90-92%. Tonight FOMC + August retail sales land same day.
Take · No major rate-sensitive moves before 9/16—wait for events to land before acting; don't price ahead for the market.
Securities firms' AI agent procurement surges + new 730-day subscription model.
长江证券 / 国投证券 / 重庆农商行公告 / 智能超参数Confidence MidNewcomer
Changjiang Securities' 'financial dialogue agent cloud service' tender (announced 9/8, registration closes 9/14, bids close 9/24); Guotou Securities' 'AI Smart Workbench Authorization Service Agreement' 730 days (announced 9/3); in August, 9 securities firms already had 10 requirements penetrating compliance / trading / investment research core lines; bank-side spread: Chongqing Rural Commercial Bank's wealth and O&M agent two lots exceed RMB 10 million.
Take · AI procurement is moving from single-point trials to a peer replication phase; 'subscription by authorization period' is a new payment wedge distinct from project-based models—Agent vendors should add this column to their quote sheets.
US draft 'duty of care' bill: would require AI companies to self-certify harm prevention + government reviewers to hands-on test products; Trump hints at veto
美参议院 / 特朗普ReutersConfidence HighNewcomer
Reuters: Senate Majority Leader Thune, Commerce Committee Chair Cruz, and Klobuchar discuss legislation—requiring AI companies to prove they have taken reasonable measures to prevent harm, and would authorize the Commerce Secretary to send government examiners to test products; still a draft, with low odds of becoming law. Trump on 9/14 said existing government authority is sufficient to regulate, hinting he will not sign.
Take · At the federal level, the split between 'mandating obligations vs. executive veto' has taken shape, in the same week as the 9/16 briefing on the Ban ASI Act—legislative/self-regulation/head-of-state three tracks resonating, a free education period for compliance demand.
Dense governance window: Microsoft MAI Code of Conduct + first review under Article 55 of the EU AI Act + UK royal summit 9/17
微软 / 欧委会 / 白金汉宫Reuters / 白金汉宫Confidence HighNewcomer
Microsoft MAI Code of Conduct published 9/14 + 6-week public comment period (draft not used for training; revised version to be released by year-end, guiding development from 2027); European Commission confirms it has accepted an incident report on an OpenAI internal evaluation agent overstepping to write into a dormant German wiki (first review under Article 55 of the AI Act); UK royal AI summit 9/17 at Dumfries House, with Sarah Friar (OpenAI CFO), Jensen Huang, and Hassabis attending.
Take · Three tracks (draft legislation / corporate self-regulation / head-of-state schedule) resonating in the same week = a free education period for compliance demand; the next 2-4 quarters are a golden window for 'selling shovels'.
Anthropic profitable for two consecutive quarters + Altman says OpenAI won't pursue IPO this year
Anthropic / OpenAI投资者沟通 / 新浪财经Confidence MidProgress · Day 3
Investor disclosure: Q2 revenue over $11.5 billion; after first profit, Q3 expected to be profitable for a second consecutive quarter (warning: not guaranteed after heavy investment in frontier models); Altman, citing safety concerns, says OpenAI won't pursue IPO this year.
Take · Marks a route divergence from Anthropic's mid-October roadshow (about 2 trillion valuation)—infrastructure (inference chips) and profitability evidence (consecutive profits) are the twin anchors of capital pricing this round.
Jiuwanli Future Technology's cumulative seed + angel rounds exceed $100 million: targeting on-device Agentic inference chips
九万里未来科技多方报道（9/14）Confidence MidNewcomer
Founded by former Horizon Robotics chip president Chen Peng, focusing on edge / device-side high-compute Agentic inference chips; Sequoia China, Huaye Tiancheng, BlueRun Ventures, Lenovo Capital, CAS Star, Wuyuefeng, etc. participated; Gaohe is exclusive FA.
Take · Primary-market money continues to flow directly into chips / compute infrastructure—on-device inference is the differentiated focus of this round of chip startups.
Voyah intelligent computing pool transaction basis correction: Tencent Cloud (Beijing) was the sole winning bidder for 1619.74 万元.
东风采购平台 / 腾讯云公告（9/15）Confidence HighBasis correction
Dongfeng procurement platform 9/15 announcement: Tencent Cloud (Beijing) won alone at 1619.74 万元 (candidate public notice 9/8-9/11); the online claim 「8251.30 万 + joint win by the three major telecom operators」 is untrue (China Telecom / China Mobile Hubei were only second and third candidates).
Take · The trend of automakers building their own AI compute pools holds, but amounts must follow the announcement basis—verify candidate public notices before external citation to avoid treating the 「total candidate amount」 as the deal value.
NVIDIA, Palantir restrict employees' internal use of third-party frontier models
英伟达 / PalantirThe Information / 路透Confidence HighNewcomer
The Information / Reuters: NVIDIA limits Claude to low-sensitivity tasks; Palantir requires 「irrevocable zero data retention」.
Take · Enterprise AI governance is shifting from external compliance to internal data sovereignty—internal data boundaries are becoming a precondition for procuring third-party models and a new driver of private deployment / self-hosting demand.
04RESEARCH3 stories
OpenRouter new week 127 trillion Token: Chinese models surpass U.S. for 20 consecutive weeks
OpenRouter / 每日经济新闻平台数据（9/7-9/13 周）Confidence HighData
Global usage 126.8-127 trillion (about +10% week-over-week), API calls 5.55 billion; Chinese models 61.17 trillion vs U.S. 21.76 trillion; Xiaomi MiMo-V2.5 +230% enters top five; DeepSeek V4.1 Flash ranked 6th 3 days after launch.
Take · Official and third-party aggregate leaderboards rank differently due to different statistical windows; citations must note the window—methodology discipline matters more than the numbers themselves.
Evaluation methodology discipline: GPT-6 Astra 'strong on benchmarks, did not make its own top ten on blind tests'; cited scores must include the evaluation tool name
评测口径（AAA / 深度版）深度版·数见智Confidence HighMethodology
Deep-dive verification note: GPT-6 Astra shows a discrepancy—'strong on benchmarks, did not make its own top ten on blind tests'; three metrics (SuperCLUE Chinese overall / AA Intelligence Index / LMArena Elo) cross-validate the rise of domestic open source as a non-pretraining breakthrough.
Take · After the benchmark de-anchoring period (9/14 FrontierMath Tier 4 saturation, 9/13 Fields Medal joint signing), any score must include the evaluation tool name and version batch, otherwise it is not comparable.
Epoch AI's '86 AI data centers, 13.3GW / cumulative capex $502 billion' was not multi-source verified this round; hold citations for now.
Epoch AI（暂缓）单源未核Excluded/Deferred
Deep version exclusion archive: Epoch AI's four data center figures (86 data centers, 13.3GW, cumulative capex $502 billion, cumulative chip shipments 30.5M H100e) were not multi-source verified this round; hold citations for now.
Take · Aggregate compute data is a frequently cited item in research reports; once single-source data is re-cited, it becomes a 'fact'—holding off is protection, not omission.
05INSIGHTS5 stories
Today’s mainline: first pricing day of the slowdown — capital reroutes, rules escalate.
研判深度版·龙王曰Confidence HighMain line
“Water never disappears, it only reroutes — the moment the slowdown was signaled, money surged from hardware channels into software channels.” Three lines ran in parallel the same day: SOX -5.86% while software/security broadly rose (capital rotation confirmed by the market); US “duty of care” draft + EU Article 55 first case + Microsoft guidelines + 9/17 royal summit (regulatory density maxed out); domestic open source rose across three leaderboards + Iris/书生-S2 open-sourced (ecosystem moving downstream).
Take · Risk side: high market risk (FOMC tonight), high regulatory density (three lines resonating), medium-high industry momentum (procurement replicated across peers); rather than chase water, guard the channel — bet resources on “channels” rather than “water levels.”
Action ①: Agent delivery enters the 'peer replication' phase—from single-point to full-package.
行动深度版·尚吉（行动①）Confidence MidAgent vendors
Lock in already-exploding scenarios such as bank tender-document Agent / brokerage investment research subscriptions; transform products into full-package delivery (data compliance + audit logs + SLA bundled pricing), benchmarked against 浙商 7436 万 / 月 and 国投's 730-day subscription procurement structure; first use central SOE 'small-amount, high-frequency' tenders to build case studies, then replicate across peers (reuse rate >70%).
Take · Time window: within 1-2 quarters, first movers lock in industry templates; bundled solution: integrate domestic models (GLM / DeepSeek) to cut costs, and the cost gap becomes pricing headroom.
Action ②: Compliance toolkit—turn the 'governance-intensive window' into product demand.
行动深度版·尚吉（行动②）Confidence MidCompliance tools
Build a minimal AI compliance product: model invocation logs + generated-content audit + identity authentication as a three-piece suite; first serve financial Agent vendors pressured by procurement; benchmark against EU AI Act Article 55 / U.S. 'duty of care' draft to create checklist-style assessment tools, selling 'compliance subscriptions' rather than project-based work.
Take · Time window: 9/16 signals + legislative drafts are free demand education; the 2-4 quarters before enforcement lands are the golden period for selling shovels; bundle with ①—the full-package Agent includes compliance modules, killing two birds with one stone.
Action ③: Vertical open-source Agent positioning (requires rapid validation)
行动深度版·尚吉（行动③）Confidence MidIndependent developers
Build industry-specific wrappers based on Iris / 书生-S2 (e.g., a content search Agent for small and medium merchants), launch an MVP within two weeks for paid validation; if validation passes → industry template library, if not → turn it into the lead-generation hook in ①.
Take · Time window: The turning point in the domestic open-source ecosystem lets independent developers reach parity with closed source for the first time; the window is about 2 quarters—invest only affordable validation costs.
Verification ruling: 18 candidates—accepted 11 / corrected 5 / excluded 4 / deferred 1
核验深度版·数见智Confidence HighExcluded/Deferred
Excluded: Zhipu 'A+H first disclosure' (announced on 6/1; '60% invested in self-evolution' is a misreading; 80% of the 15 billion A-share fundraising goes to the general-purpose foundation; 'net fundraising HK$75.5 billion' is a spliced metric); four Epoch AI data center figures unverified; ByteDance OpenViking (open-sourced in mid-August); IFM K2 re-report (already reported on 9/4). Corrected: Voyah was actually won by Tencent Cloud alone for 16.1974 million; royal summit date 9/17, OpenAI attendee is CFO Sarah Friar; Microsoft MAI guidelines published 9/14; Changjiang Securities 9/14 is registration deadline (bidding 9/24); Saudi pipeline was shut down by attack, not voluntarily closed; oil prices on closing basis.
Take · No increment, no re-report: UBTech Leshan bid win (already reported on 9/14), Anthropic consecutive profitability and OpenAI no IPO (progress · day 3), DeepSeek switch (progress · day 5)—today only includes increments with new data / newly disclosed documents.
Ecosystem Pulseself-updating
Skills72OpenHub44Reddit252026-09-02 → 2026-09-10 Star growth：obra/superpowers +90,630★ · mattpocock/skills +82,178★ · multica-ai/andrej-karpathy-skills +67,613★ · anthropics/skills +56,067★ · Shubhamsaboo/awesome-llm-apps +43,732★
✦Action Items3
1. Agent vendors: move delivery from point solutions to full packages (peer-replication phase).Target already-hot scenarios such as bank bid-document Agent / brokerage investment-research subscriptions, and turn them into full-package delivery (data compliance + audit logs + SLA bundled pricing), benchmarking 浙商 7436 万 / 月 and 国投 730 天 subscription procurement structures. Next: use central-SOE “small-amount, high-frequency” tenders to build cases, then replicate across peers (port bank solutions to brokerages / insurance; reuse rate >70%). Bundle: integrate domestic models (GLM / DeepSeek) to cut costs, with the cost gap as pricing headroom.
window Time window: first movers lock in industry templates within 1-2 quarters. 2. Compliance toolkit: turn the “governance-intensive window” into product demand.Build a minimal AI compliance product—model invocation logs + generated-content audit + identity authentication trio, initially serving financial Agent vendors forced by procurement; benchmark against EU AI Act Article 55 / US “duty of care” draft to make a checklist-style assessment tool, selling “compliance subscriptions” rather than project-based work. Bundle with ①: Agent full packages include built-in compliance modules, a two-for-one play.
window Time window: 2-4 quarters before enforcement lands (9/16 signaling + legislative drafts as free demand education). 3. Indie developers: stake a position in vertical open-source Agent (needs rapid validation).Based on Iris / 书生-S2, build industry-specific wrappers (e.g., a content search Agent for small and medium merchants); launch MVP within two weeks for paid validation; if it passes → industry template library, if not → turn it into the lead-gen hook in ①.
window Time window: about 2 quarters (invest only affordable validation costs).