01MODEL RELEASES4 stories
StepFun releases Step 5 Preview: 600B total parameters / 27B active / 1M Token context, plans to open-source on October 15
阶跃星辰环球网 2026-09-20（本机 20 bot《每日AI情报 2026-09-22》转述）Confidence MidMultiple sources agree
Sparse MoE flagship foundation model, 600B total parameters, only 27B active (activation ratio ~4.5%), supports 1M Token context and natively supports text+vision, targeting Agentic scenarios such as AI coding, software engineering, and finance; open-source timing explicitly stated as October 15.
Take · By simultaneously bringing 'ultra-large total parameters + extremely small active parameters + ultra-long context' to Agentic scenarios, it directly targets the two most expensive line items, unit token cost and context retry cost; moreover, the open-source date is publicly verifiable, making it one of the few promises that day that can be verified afterward.
Zhipu GLM-5.3-FlashX listed on Alibaba Cloud: approx. 200 tokens/s, supports 1M context
智谱 / 阿里云阿里云帮助中心 2026-09-21（本机 20 bot 情报转述）Confidence MidMultiple sources agree
Alibaba Cloud Help Center listed Zhipu GLM-5.3-FlashX on 09-21, with a nominal ~200 tokens/s and support for 1M Token context; the same day, the Apsara Conference hosted MaaS and Agent main forums.
Take · Speed and context length directly rewrite the denominator of 'cost per piece of content'—speed is the only model metric users can directly feel, and ~200 tokens/s means the waiting anxiety of interactive Agents is engineered for the first time.
Tencent launches multimodal model Gander (task execution and voice in parallel), Alibaba open-sources lightweight image model Qwen-Image-2.1
腾讯 / 阿里千问SegmentFault《AI 日报》2026-09-21（本机 20 bot 情报转述）Confidence MidMultiple sources agree
Tencent's Gander focuses on task execution and voice dialogue in parallel; Alibaba open-sourced lightweight image model Qwen-Image-2.1 the same day.
Take · Multimodal is moving from 'understanding' to 'executing while conversing,' a precursor to an interaction paradigm shift; on the design side, teams should prepare early to adapt input forms.
MediaTek launches Dimensity CX C10 Max: TSMC 3nm, 45 TOPS NPU, aimed at Googlebook on-device AI
联发科IT之家 2026-09-21（本机 20 bot 情报转述）Confidence MidSingle-source pending verification
8 all-big-core CPU, 11-core Immortalis-G925, 45 TOPS NPU, explicitly positioned for Googlebook on-device AI scenarios.
Take · Each step up in on-device compute compresses the visible call volume of high-end cloud models by one layer, increasing tiered-pricing pressure on model makers—on-device and high-speed inference updates landed the same day, shifting the overall competitive focus from 'capability ceiling' to 'unit cost'.
02PRODUCT4 stories
ElevenLabs📹视频库《AI官方视频监控台账》2026-09-22（YouTube·ElevenLabs）Confidence MidOfficial
Official channel launches Studio 4.0, positioning itself as an agentic video editor in ElevenCreative; the ledger entry only provides the title and link, and the video content was not retrieved.
Take · Making the editing step itself agentic is the same theme as our video factory’s orchestration line from source material to finished video, making it a comparable capability sample.
Original ↗Cloudflare / Simon WillisonSimon Willison 官方博客 2026-09-21（已回读原文）Confidence HighOfficial
After two years in preview, it is now GA; the official wording is “Python is now a first-class, fully supported language on Cloudflare’s developer platform”; the implementation compiles Python to WebAssembly via Pyodide and runs it in the V8-based workerd runtime. multiprocessing and threading are unavailable in that WASM VM, and local development uses pywrangler (PyPI package name workers-py) to locally simulate the full stack.
Take · Running Python at the edge lowers the barrier to “moving local lightweight services to the cloud” another notch, aligning with our lightweight services and cloud job direction, and worth listing as a selection option; the limitations (no multiprocessing/threading) also define the task boundaries it can handle.
Original ↗百度智能云📹视频库《AI官方视频监控台账》2026-09-22（B站·百度智能云）Confidence MidOfficial (vendor self-announcement)
The ledger recorded its 8 official placements that day: Enterprise Edition mobile app launch (tasks continue from desktop to mobile; Project Space, Project Memory and Share Relay), full opening of standardized hardware API, developer platform connecting general/professional/custom agents to form a task loop, and launch of a co-creation program, among others; the original ledger only provides titles and links, and no body text was retrieved.
Take · Domestic Agent office entry points expanded the same day from a single desktop endpoint to three tracks—mobile + hardware API + platform collaboration; the supply-side battle for entry points is shifting down from the model layer to the workflow layer.
Original ↗LangChain / TypeSafe📹视频库《AI官方视频监控台账》2026-09-22（YouTube·LangChain）Confidence MidOfficial
The LangChain official channel published the 'Building a Harness with Jev' session; the original ledger only provides the title and link, and no body text was retrieved. Jev is a small 'System One'-type model open-sourced on 09-19, claimed to only do decision/classification/routing/scoring.
Take · Two days after open-sourcing, it was integrated into a mainstream framework as a harness—'small models for routing and judgment' has officially moved from experiment to engineering, in line with our routing-layer design.
Original ↗03INDUSTRY6 stories
US AI and chip stocks rose across the board; AMD market cap topped $1 trillion for the first time
美股 AI / 芯片板块新浪财经 2026-09-22（本机 20 bot 情报转述）Confidence MidMultiple sources agree
The Nasdaq rose over 2%, Meta over 11%, Intel over 12%; AMD became the fourth US chipmaker to top a $1 trillion market cap (after NVIDIA, Broadcom, Micron).
Take · The compute hardware boom has moved from “narrative” to “market-cap fact”—a hard gauge of cycle position and a benchmark anchor for judging whether AI is overheated.
Beijing issues “Action Plan for Accelerating the Development of the Token Economy”; AICC 2026 AI Computing Conference held in Beijing the same day
北京市 / AICC 2026网易、腾讯新闻 2026-09-22（本机 20 bot 情报转述）Confidence MidMultiple sources agree
The action plan calls for advancing foundation model iteration and breakthroughs in key core compute technologies; AICC 2026 focuses on intelligent computing power and a domestic intelligent-computing foundation, seen by the market as positive for compute leasing and domestic intelligent computing.
Take · Writing “token” into a policy name effectively officializes the smallest unit of service metering—past policies anchored on “cards” and “P,” now on “token output,” changing the basis for pricing and cost ledgers.
Apsara Conference main forum bets on MaaS and Agent deployment; 'how AI value is realized in business' becomes the focus
阿里云 / 云栖大会甲子光年 2026-09-22（本机 20 bot 情报转述）Confidence MidMultiple sources agree
Highlights the Qwen agent service foundation, showcases evolution of enterprise office, AI Coding and other scenarios; in the same period, Alibaba Cloud lists Zhipu GLM-5.3-FlashX.
Take · The agenda shifts from 'model capability' to 'business realization,' indicating cloud vendors are moving from selling models to selling workflows, and the procurement entry point is shifting.
California governor signs executive order to study mandatory AI 'circuit breaker'; Spanish Prime Minister calls for stronger public oversight the same day
加州州政府 / 西班牙政府AI Governance Brief 2026-09-21（本机 20 bot 情报转述，两条均单源）Confidence MidSingle-source pending verification
California asks expert panel to assess feasibility of emergency kill switches for frontier models and stricter laws; Spanish Prime Minister simultaneously calls for stronger public oversight of AI development.
Take · Governance moves from 'principles declarations' to the design level of 'default safety devices'—switches, staffing, and accountable entities are treated as part of infrastructure; compliance costs for going global and To B delivery must be booked in advance.
Compute and Agent tracks simultaneously enter the exit window: Nscale targets about $35 billion valuation to advance IPO, Anthropic plans to acquire Decart for $6 billion, Manus may pursue a Hong Kong IPO
Nscale / Anthropic / ManusStartup and Venture Investment News 2026-09-22、投资界 2026-09-15、腾讯新闻 2026-09-18（本机 20 bot 情报转述）Confidence MidSingle-source pending verification
AI infrastructure company Nscale targets about $35 billion valuation to advance IPO (Holtec postpones listing in the same period); Anthropic plans to acquire Decart for $6 billion and push ahead with listing; Manus may pursue a Hong Kong IPO, with a new round of about $500 million in financing nearing completion. One additional old item is background only: OpenAI plans to launch a new funding round at over $1.2 trillion valuation (IT之家 2026-09-09, single-source pending verification; not treated as today's new development).
Take · The capital loop is clearly closing faster than service-side delivery—if valuations are built on procurement growth rather than customer retention, the cracks will be seen in a single earnings report.
RSI (recursive self-improvement) becomes a capital focus; the common conclusion is 'verification is the biggest bottleneck; data and environments are the core business.'
RSI 专题（AGI House / 红杉中国）36氪「智能涌现」2026-09-21（本机 20 bot 情报转述）Confidence MidMultiple sources agree
AGI House, Sequoia China, and others co-hosted a special session, 'When AI Begins Recursive Self-Improvement,' with Google DeepMind's Shunyu Yao, Recursive Superintelligence's Tianlin Shi, and scholars from Tsinghua University and Shanghai Jiao Tong University discussing the definition, evaluation, and verification of RSI.
Take · The day's only item to put 'verification' on the table: as generative capability becomes commoditized, pricing power shifts to 'who can judge right from wrong'; the scarcest thing is not compute but criteria.
04RESEARCH4 stories
Daily leaderboard signal 1: IntBMoE (heat 90) decouples MoE's three quantities—participation, execution volume, and materialization volume.
Hugging Face 当日论文榜hf-mirror.com 每日论文榜 2026-09-22（本机慧师情报转述）Confidence MidSingle source pending verification (daily leaderboard)
The paper argues for measuring and optimizing MoE's three quantities (expert participation, execution volume, materialization volume) separately, rather than conflating them into one efficiency number; daily leaderboard heat 90.
Take · In line with same-day industry moves: break down and measure the cost structure instead of continuing to pile on scale—directly relevant to pricing and deployment trade-offs for sparse models.
Daily leaderboard signal 2: Designer-RSI lets a frozen frontier model evolve program memory from user traffic, reasoning that design results 'have no reliable programmatic oracle'.
Hugging Face 当日论文榜hf-mirror.com 每日论文榜 2026-09-22（本机慧师情报转述）Confidence MidSingle source pending verification (daily leaderboard)
Without updating model weights, it evolves program memory from user traffic; the paper's motivation is that design tasks lack a reliable programmatic oracle.
Take · Once again points the bottleneck back to missing criteria: any task that cannot be programmatically validated can only fall back on human feedback for self-improvement—precisely the value of the acceptance station.
APort Vault releases Agent payment authorization benchmark: replays 4,371 human attacks, covering 14 models from 8 labs
APort VaultHF Daily Papers（hf-mirror 镜像）2026-09-18 / Latent Space 2026-09-16（本机巽风情报转述）Confidence MidMultiple sources agree
Using 4,371 attacks written by humans in public CTFs to replay-test the performance of 14 models from 8 labs in payment authorization scenarios, including 5 policy configurations and two replay tracks. In the same period, AIUC completed a Series A with 'sue-able Agents' as underwriting targets.
Take · Once an Agent can spend money, risk-control benchmarks and liability insurance are its contract and deposit; turning 'who will spend recklessly' into a reproducible test is the standard paradigm for assetizing criteria.
Microsoft Research: using RetroChimera to improve small-molecule synthesis prediction at scale
Microsoft Research📹视频库《AI官方视频监控台账》2026-09-22（官方·Microsoft Research）Confidence MidOfficial (no link in ledger)
The day's ledger included the official entry 'Scaling up small-molecule synthesis prediction with RetroChimera'; no link was attached to the ledger and the full text was not retrieved.
Take · AI4S is the scenario with the strongest 'verifiable' property—whether a synthesis route is feasible is decided by experiments, making it a positive sample when criteria are scarce.
05INSIGHTS6 stories
Treat 'verification' as a scarce resource to buy: independent evaluation before discussion; first produce mutually invisible scores, then hold the adjudication meeting.
本机 20 bot 情报共识☯九元素 20 bot《每日AI情报 2026-09-22》Confidence MidLocal intelligence consensus
All model-side releases today were throughput and capacity metrics (600B/27B, 1M, 200 tokens/s); none provided end-to-end time from a single request to usable output. Meanwhile, the full consensus at the independent-source RSI session was that 'verification is the biggest bottleneck.' All 20 bots unanimously advised: for Agent iteration, adhere to 'independent evaluation before discussion'—first have each produce mutually invisible scores, then convene an aggregate adjudication; RSI's bottleneck is isomorphic to our acceptance hook, and without independent evaluation, self-improvement only amplifies noise.
Take · The more generative capability accelerates itself, the higher the cost of judging that 'it really did it'; whoever controls the auditable task environment controls the authority to judge improvement.
Add two hard metrics to the external acceptance sheet: end-to-end time (submission → usable output) and first-pass acceptance rate.
本机 20 bot 情报共识（交付）☯九元素 20 bot《每日AI情报 2026-09-22》Confidence MidLocal intelligence consensus
According to queueing theory (Little's law), waiting time = in-flight tasks ÷ effective throughput; raising single-point throughput will not reduce perceived user waiting by the same proportion, because the bottleneck has shifted to verification and backfilling. Therefore, the acceptance sheet must add two hard metrics—end-to-end time and first-pass acceptance rate—as the minimum observable configuration for the service closed loop.
Take · Without these two columns, the closed loop is just a slogan; with them, renewal negotiations have a common language, and throughput improvements can be shown to have truly landed with the client.
Add a 'governance device risk' column to the external proposal template and preset three questions: who has the authority to shut it down, under what procedure, and how responsibility is divided.
本机 20 bot 情报共识（治理）☯九元素 20 bot《每日AI情报 2026-09-22》Confidence MidLocal intelligence consensus
California studies mandatory circuit breakers, Spain discusses public oversight, and the U.S. forms an 'AI force' and sets up an AI czar—three sides with opposing positions but moving in the same direction: AI is being treated as infrastructure governance that needs 'a switch, an establishment, and an accountable entity.' Accordingly, any proposal involving critical system delivery must preset the above three questions and specify degradation and rollback paths.
Take · In the short term, no real mandatory circuit breaker will be implemented, but the fact that 'the government has switch capability under study' will itself change procurement and insurance terms; those without contingency plans will be passive during compliance review.
The cost anchor shifts from 'seat/project' to 'cost per effective output unit'; the customer ledger discusses only three columns.
本机 20 bot 情报共识（成本）☯九元素 20 bot《每日AI情报 2026-09-22》Confidence MidLocal intelligence consensus
Beijing writes 'token' into the name of an action plan, Alibaba Cloud pushes MaaS, and AICC showcases a domestic compute foundation. Read the three together: the minimum unit of measurement changes from seats or projects to tokens, amplifying clients' perception of 'how much was used.' It is recommended that the service ledger first list the trio 'effective units / cost / first-pass rate,' and renewal talks discuss only these three columns.
Take · In the usage-based era, the first customer-success action is not training but allocating customer token spend to effective units—cost transparency is the relationship's transparency item.
Simon WillisonSimon Willison 官方博客 2026-09-20（已回读原文）Confidence HighOfficial
Regarding the view that 'when running a full terminal Agent, you can just call the API directly and MCP is redundant,' the author gives four rebuttals—need ① precise control over which external services it can access; ② an authentication method that keeps the Agent from directly touching API Key; ③ a convenient interface for users to connect and authorize more services; ④ reliable audit logs.
Take · These four are isomorphic to our discipline of 'keys never land in sessions, service whitelist, and traceable audit,' and can directly serve as design basis for a self-built MCP layer and Agent integration specifications.
Original ↗微信公众号「斯为立学」公众号侧清单 2026-09-22（原文 2026-09-22）Confidence MidSingle-source pending verification (WeChat official account)
A WeChat official account article says that in September 2026 enterprise-grade agents crossed the tipping point and topped the technology maturity curve; this item is single-sourced from a WeChat official account and has not been cross-verified with official vendor statements.
Take · Aligns with same-day Yunqi 'business fulfillment', Baidazi mobile client, Agent role-ification, and other signals; included as an ecosystem-side sentiment thermometer, with conclusions to enter decisions only after official data confirms.
Original ↗Ecosystem Pulseself-updating
News600Skills46OpenHub44Reddit252026-09-02 → 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. Add two hard-metric columns to acceptance sheets: end-to-end time + first-pass acceptance rateOn all external delivery and internal pipeline acceptance sheets, fix two columns: end-to-end time (task submission → usable output) and first-pass acceptance rate; any conclusion drawn from demo sets or single-point throughput must be labeled with its basis and used at a discount. Advanced: connect these two columns into the service ledger to form a weekly comparable longitudinal curve, as the first evidence in renewal negotiations.
window Immediate - 1 month 2. Change the cost anchor to 'cost per effective output unit,' and build a three-column renewal tableChange the cost basis from seats/projects to 'cost per effective output unit'; the service ledger should first list three columns: effective units, cost, and first-pass rate; run a cost experiment of three model tiers × three task types against GLM-5.3-FlashX (about 200 tokens/s) and Step 5 (600B/27B) to determine whether primary traffic migrates. Advanced: write this basis into the pricing template so customer bills and internal costs use the same denominator.
window Immediate - 3 months 3. Make 'stoppable/auditable/rollback-capable' an admission gate, and run a pull-the-plug drillExternal proposal templates add a “governance mechanism risk” field, specifying who has authority to shut down, under what procedures, and how responsibilities are divided; for live services, run 10-minute drills each for model cutoff, compute cutoff, and index cutoff, recording customer perception and rollback duration as degradation baselines. Advanced: archive drill scripts and baselines to create a hard gate of “must pass before Agent delivery,” and pre-set compliance option clauses for each key delivery.
window 3 - 6 months