01MODEL RELEASES4 stories
Meituan LongCat-2.5 official version API launched same day: 1.6 trillion parameters, native multimodal, tops multiple subcategories across nine benchmarks
美团中关村在线，2026-11-03Confidence MidProgress · Day 2AI score 88/100
A 1.6 trillion-parameter native multimodal model, ranking first in multiple sub-tasks across nine mainstream benchmarks in three directions—code generation, general agents, and visual agents; the official API launched the same day and strengthens long-horizon task capabilities. The model release was reported yesterday; today's focus is the API opening.
Take · Only the API launch turns the 'parameter narrative' into 'billable calls'; on the first release day there is no call volume or pricing—note the capability side first, and wait for the pricing table on the commercial side.
Kimi K3.1 Pro promises to release full weights by 11-04: 2.8T MoE / 104B active, 1M context
KimiAIHOT，2026-11-04Confidence MidProgress · Day 5AI score 86/100
Promises to release full weights by today: MoE total parameters 2.8T, 104B active (activation ratio about 3.7%), 1M context, multimodal support; Artificial Analysis Intelligence Index score 55 rises to No. 4, up 11 points from K3, with the largest gain in agent tasks.
Take · Today is fulfillment day, and there is only one thing to verify: whether the weights can be downloaded and whether the license allows commercial fine-tuning. An activation ratio of 3.7% means inference is billed by active parameters; parameter inflation no longer proportionally raises inference unit prices.
SenseTime's SenseNova 6.8 Flash official version opens for trial: cross-session long-term memory + editable design source file export
商汤南方都市报（网易转载），2026-11-03Confidence HighNew topicAI score 84/100
The official version opens for trial, introducing cross-session 'long-term memory' and editable design source file export; the official narrative points directly at 'memory gaps' and 'delivery rework' in enterprise business flows.
Take · The selling point shifts from 'smarter' to 'less rework', suggesting buyers already price by hours saved; memory and editable source files are exactly the focus of follow-up negotiations to 'lock in costs'.
Qwen 3.8 27B can interpret model weights locally.
QwenSimon Willison，2026-11-03（官方视频监控台账）Confidence MidNew topicAI score 74/100
The author documents running Qwen 3.8 27B locally, using it to interpret model weights and produce usable results; this is a firsthand test record, with no model card or benchmark scores.
Take · A model that can read weights locally brings the cost of 'checking the upstream black box' down to one laptop; weight auditing no longer has to rely only on vendor self-reports.
02PRODUCT LAUNCHES4 stories
NetflixTechCrunch AI，2026-11-03（官方视频监控台账）Confidence MidNew topicAI score 80/100
Adds an AI Discover tab that recommends content based on users’ current mood. The source does not specify coverage regions, launch timing, or model provider.
Take · Recommendations shift from “what you’ve watched” to “what state you’re in now”; cold start and catalog utilization both benefit; the real effect depends on whether it changes the definition of completion rate.
Original ↗Appledeveloper.apple.com（官方视频监控台账），2026-11-03Confidence MidNew topicAI score 79/100
Apple Developer site announcement: Siri AI opens to all developers; the source does not list capability boundaries, invocation quotas, or billing methods.
Take · System-level voice entry opens, shifting the app voice-entry battle from in-house assistants back to the Apple ecosystem; whoever integrates first gets the default path.
Original ↗DatabricksDatabricks 官方，2026-11-03（官方视频监控台账）Confidence MidNew topicAI score 78/100
Databricks official blog: Claude Platform can now report observability data on Databricks MLflow natively via OpenTelemetry
Take · The bar for putting agents into production shifts from "can it run" to "can it be traced"; only with unified trace standards can evaluation and billing instrumentation for the model-switchable layer share a common foundation
Original ↗GitHubGitHub 官方（官方视频监控台账），2026-11-03Confidence MidNew topicAI score 72/100
Official GitHub video introduces a new way to use Copilot in JetBrains IDEs; this item relies only on the title and official release record, and the video was not reviewed
Take · Distribution on the IDE side matters more than model capability in determining daily usage; JetBrains users are precisely the existing base that Copilot and Cursor are competing for
Original ↗03INDUSTRY8 stories
Anthropic returns to operating loss in Q3, first discloses approximately USD 8954 hundred million ($895.4 billion) compute procurement commitment
Anthropic财联社／华尔街见闻，2026-11-03Confidence HighProgress · Day 2AI score 95/100
Files amended prospectus with SEC: Q3 revenue nearly USD 170 hundred million (about $17 billion, about 11x YoY), net loss widened 92% YoY, and for the first time includes nearly USD 9000 hundred million ($900 billion) compute procurement commitment in public disclosure. At quarterly revenue, about 52.6 quarters; at annualized 680 hundred million (68 billion), about 13.2 years.
Take · Commitments are a rigid floor, revenue a flexible ceiling: from the moment of signing, 'growth below expectations' turns from earning less to repaying debt. The watch point is whether the commitment/annualized revenue multiple is still rising.
Tech giants collectively take on debt to invest in AI: Alphabet launches its second euro bond of the year, Meta plans to issue at least USD 200 hundred million ($20 billion) in multi-currency debt
美国科技巨头金融界（网易转载），2026-11-03Confidence HighNew topicAI score 92/100
Alphabet launches its second euro bond of the year; Meta considers at least USD 200 hundred million ($20 billion) in multi-currency high-grade debt; Oracle about USD 200 hundred million ($20 billion) in equity and debt financing. Fidelity International estimates related issuance pipeline at EUR 400–600 hundred million (€40–60 billion), a third-party estimate, not an issuance announcement.
Take · Financing shifts from equity to debt; sector beta shifts from 'compute supply' to 'interest rates': when long-end rates rise, the cash-flow narrative will be questioned before model capability.
AI inference company Fal raises $1.3 billion, valuation doubles to $16 billion
FalIT时代网（网易转载），2026-11-04Confidence MidNew topicAI score 86/100
Bessemer led, with NVIDIA, Sound Ventures and T. Rowe Price participating; the E and F rounds total $1.65 billion, doubling the valuation from $8 billion in April this year (about 7 months). The original text gives no ARR or gross margin, so PS cannot be calculated.
Take · Capital is shifting from the training layer to the inference layer; the same money is seeking outlets in both equity and debt, indicating that risk appetite has diverged.
AMD FY2026 Q3 revenue reaches a record quarterly high of $13.377 billion; data center revenue up +97% YoY
AMDIT之家，2026-11-04Confidence HighNew topicAI score 85/100
Lisa Su confirmed that 2nm Zen 6 sixth-generation EPYC Venice has begun revenue shipments; Q4 guidance is about $17 billion, with the first OEM systems to launch.
Take · The second pole on the supply side is starting to show real financials, so procurement price-comparison weight should be raised; with the cost floor and financing costs rising at the same time, the middle layer is the most squeezed.
NVIDIA sets a 49% compute exposure cap for Chinese customers
NVIDIA腾讯新闻（转 FT Technology News），2026-11-03Confidence MidNew topicAI score 88/100
Requires data center customers to sell no more than 49% of a single data center's total compute to China; US policy commentators argue the reasonable threshold should be zero. Single-source paraphrase; all reports mark it “pending second cross-check.”
Take · A ratio-based rather than list-based system: auditable, spot-checkable, and punishable, shifting from “whether to sell” to “how to rent.” Based on 49%, reaching 100 units of capacity requires at least 3 IDCs; if any one is blocked, about 1/3 of capacity is lost.
Hong Kong SFC’s first AI survey of large institutions: 30% already use agentic AI; plans dedicated guidance for agentic AI
香港证监会中国证券网·上海证券报，2026-11-02Confidence HighProgress · Day 2AI score 84/100
Covers 20 large financial institutions + 78 Hong Kong-linked licensed institutions, with involved transactions accounting for more than 75% of reported securities trading; 30% already use agentic AI, four major regulatory directions have been identified, and dedicated guidance for agentic AI is planned.
Take · Regulatory focus shifts from “generated content” to “agent behavior,” with accountability to be recorded in ledgers—this will be the first written market-access requirements for agentic AI.
ByteDance's Seedance overseas revenue accounts for 50%, with overseas revenue up nearly 5x over the past six months
字节跳动36氪，2026-11-03Confidence MidNew topicAI score 76/100
Tencent ranks among its top three China-region clients, with real-economy industry clients making up more than one quarter
Take · Repeat purchases and overseas expansion say more about asset quality than parameters; the video-generation price war has already reached overseas markets
Tencent Maps announces shift from 'App-first' to 'Agent-first'; WeChat opens three gray-release tracks the same day
腾讯微盛AI（微信公众号），2026-11-02New topicAI score 62/100
WeChat official account relay: Tencent Maps switches to 'Agent-first'; WeChat opens three gray-release tracks the same day. Single official-account source, not cross-checked against official statements
Take · Super apps handing the entry point to Agents means replacing 'open the App' with 'say one sentence'; the gray-release scope and eligibility await official confirmation
04RESEARCH3 stories
arXivHacker News AI（官方视频监控台账），2026-11-03New topicAI score 60/100
The ledger includes the arXiv paper The Little Book of Generative AI Foundations: An Intuitive Mathematical Primer, recommended via HN. This entry is based only on the title and link; the full text was not read.
Take · The value of introductory materials lies in reducing internal team alignment costs; using it as a textbook requires checking the version and errata first.
Original ↗2026 AgentSkills Technology and Security White Paper: claims the ecosystem already has 900,000 Skills
AgentSkills墨玫人工智能（微信公众号），2026-11-03New topicAI score 58/100
A WeChat public account's interpretation of the report claims the Agent Skill ecosystem has reached 900,000 Skills; the figure is the author's self-report, with no third-party source seen.
Take · While the number of Skills has surged, security is treated in a separate chapter—the Cursor extension tampering incident that same day is an example of the same risk.
分子发现Hacker News AI（官方视频监控台账），2026-11-03New topicAI score 52/100
HN indexed a social media post claiming AI agents outperform human-built tools on molecular discovery tasks; no preprint or controlled experiment disclosed.
Take · The conclusion must wait for reproducible experimental records; for now it can only be a story lead and should not enter any rankings.
Original ↗05PLAYBOOK6 stories
The buyer window for the 'model-switchable layer' is about 2–3 quarters.
模型可切换层霆钧（深度版商机栏），2026-11-04Confidence MidViewpointAI score 82/100
A thin abstraction layer + N backends (open-source weights / domestic APIs / overseas APIs), unified prompts, unified evaluation set, unified billing instrumentation. The first step is to do only one thing: run the current business on 3 backends and record three numbers: per-run cost, failure rate, and integration hours.
Take · Upstream must chase volume to pay interest, so the buyer window is open for the next 2–3 quarters; once IPO pricing is set, bargaining power returns to sellers.
Four components of an agent authorization ledger: who authorizes, quota caps, revocation path, and audit trail
代理授权台账明见／艮山（深度版商机栏），2026-11-04Confidence MidViewpointAI score 80/100
30% of institutions already use agentic AI, but most lack incident ledgers; without a ledger, they cannot self-prove compliance. The ledger should be standardized and directly mapped to the first round of compliance checks under Hong Kong SFC guidelines.
Take · Getting in place 1–2 quarters early at the consultation-draft stage is much cheaper than retrofitting after formal requirements land.
Switch the ROI metric from “accuracy” to “rework rate × person-hours × unit price”
ROI 口径深度版 Top3 行动建议，2026-11-04Confidence MidViewpointAI score 78/100
Test only one process at a time, report “hours per person per week consumed by rework,” not efficiency-improvement percentages; an Agent without a control group does not go into production (no control group = no significance, only a demo).
Take · Percentages do not enter finance’s books; person-hours do; a rework-rate baseline can be produced within two weeks.
Compute split into two ledgers: physical budget + compliance budget, plus a column for 'single-IDC dependency'.
算力两本账深度版 Top3 行动建议，2026-11-04Confidence MidViewpointAI score 76/100
Back-calculate the compliant available compute ceiling under the 49% basis and prepare a '30-day supply cut-off' fallback plan; add a dependency column to the procurement sheet and list anything above 49% as a hard risk.
Take · It is still undecided whether 49% is a self-regulatory or regulatory basis; when the implementation rules are released, everyone will move together, and whoever runs the survey first gains the first-mover advantage.
The integer arithmetic of 49%: a single IDC's China cap is 49%, and filling 100 units of capacity requires at least 3 IDCs.
算术坎流推算（深度版数据栏），2026-11-04Confidence MidViewpointAI score 70/100
49+49+2: if any one is blocked, about 1/3 of capacity is lost directly; the number of channels goes from 1 to 3, and intermediary layers and compliance costs rise together.
Take · Only by translating the policy ratio into procurement quantities can you know how many contracts to sign and how many links to prepare.
Embodied AI acceptance metric: number of accepted scenarios ÷ claimed skill library size
验收口径厚坤／总协调（深度版商机栏），2026-11-04ViewpointAI score 66/100
Recommend adding one more question: 'How many customers invoke the same skill?'; projects with a ratio below 10% will see collection problems cluster next year.
Take · Capital is priced quarterly and production lines annually; when financing density exceeds delivery density, the gap ultimately surfaces as payment collection.
Ecosystem Pulseself-updating
News600Skills39OpenHub68Reddit252026-09-02 → 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. First build the 'model-switchable layer,' then discuss price cuts.Entry approach: a thin abstraction layer + N backends (open-source weights / domestic API / overseas API), with unified prompt, unified evaluation set, and unified billing instrumentation; the first step does only one thing—run the current business on 3 backends and record three numbers: per-run cost, failure rate, and integration hours. Advanced path: ① get the three backends running side by side → ② targeted fill-ins based on failure samples (open source sets the baseline, closed source covers the long tail) → ③ negotiating leverage: write 'price ceiling clause + termination right for supply interruption' into a long-term API contract → ④ externally output as a 'model routing' capability.
window 2–3 quarters (from today through 2027-02); basis: Anthropic's 13.2-year commitment and debtization of giants mean upstream must secure volume in the next 2–3 quarters. 2. Compute power is split into two ledgers: physical budget + compliance budget, with an added column for 'single-IDC dependency'.Entry approach: this week, pick one real business line, calculate compute exposure clearly (shares of self-owned / leased / cross-border), back-calculate the 'compliant available compute ceiling' under the 49% threshold, and prepare a '30-day supply cutoff' degradation plan; add a dependency column to the procurement table, flagging anything above 49% as a hard risk. Advanced path: ① dependency census → ② multi-point diversification (non-leading-edge inference cards, rack-level complete machines, third-party IDC) → ③ raise the weighting of AMD / rack solution price comparisons → ④ preset 'exposure adjustment clauses' and 'price cap clauses' in contracts.
window Complete the census within this month and re-sign the clauses within 1 quarter; whether 49% is currently a self-regulatory or regulatory threshold is undecided—those who move first gain first-mover advantage. 3. Change the ROI metric from 'accuracy' to 'rework rate × labor hours × unit price'Entry approach: test only one process at a time; report 'hours per person per week consumed by rework', not the efficiency-improvement percentage—the latter cannot enter finance's books; Agents without a control group do not go into production. Advanced path: ① baseline measurement of single-process rework rate → ② add a memory layer (cross-session / cross-delivery memory retention rate) → ③ write 'source files editable, exportable, migratable' into acceptance criteria → ④ agent authorization ledger (who authorizes / quota cap / revocation path / audit trail) as a standard component, directly aligned with the first round of compliance checks under Hong Kong SFC guidelines.
window Rework-rate baseline data produced within 2 weeks; agent authorization ledger formed within 1 month (1–2 quarters ahead of formal requirements is the best positioning window)