01MODEL RELEASES4 stories
OpenAI DevDay overview: GPT-6.1 Sol + 7×24 always-on agent dots, ChatGPT weekly active users at 1.2 billion
OpenAI本机20席情报（18/20 席同报）Confidence HighProgress · Day 2 (already reported on 10-01)
The San Francisco developer event released about 20 items at once: 7×24h online autonomous agent dots; GPT-6.1 Sol, with performance close to Astra and standard token pricing about one-fifth of Astra’s; Agents/Decisions API, Spaces, and Marketplace; Sora 2 and a standalone social app opened for free at the same event; ChatGPT weekly active users reached 1.2 billion.
Take · The four layers—model, agent, API, ecosystem—were filled in on a single day, rewriting the competitive unit from “single call” to “always-on seat”; Sora 2’s exact release date remains uncertain (continuing the pending verification from 10-01), and this issue does not cite it as hard evidence.
Google Gemini 4 Argon opens on a targeted basis under “Project Fairwind,” tops multiple leaderboards on launch day
Google本机20席情报（18/20 席同报）Confidence HighProgress · Day 2 (reported 10-01)
Nearly 10 months since the Gemini 3 series, it focuses on long-workflow deep reasoning, with a single output cap of about 1 million tokens; first opened to government users and trusted cybersecurity parties, topped multiple leaderboards on launch day, and only then rolled out to the public.
Take · Flagship refresh cycles have been compressed to about 10 months, and tiered distribution—trusted parties first, the public gets the next tier down—has become the norm; the half-life of capability gaps is shrinking in tandem.
Meituan launches LongCat-2.5-Preview: 1.6 trillion total parameters, about 48 billion activated, 1 million token context
美团观察者网 via 网易（13/20 席同报）Confidence MidNew topic
1.6 trillion total parameters, about 48 billion parameters activated per inference, supports 1 million Token context and native multimodality; can operate terminals, browsers, GUI, and spreadsheets.
Take · On the domestic side, a rare complete spec of 'trillion parameters + long context + GUI operation'; but parameters and context are both upper-bound values, and same-caliber real-world tests are needed before discussing replacement.
Kling 4.0 opens beta: native 30 seconds per generation, up to 10 keyframes, 4K 10-bit HDR
快手（可灵）36氪 / 大数跨境（14/20 席同报）Confidence MidNew topic
Native generation up to 30 seconds per run, up to 10 keyframes, supports 4K/1080p 10-bit HDR; Flash version first, official version launches in October.
Take · Directly targets the most fragile link in the talking-head/AI video pipeline (character drift and lip-sync mismatch from multi-segment stitching), and is the only action item today that can be immediately scheduled for A/B testing.
02PRODUCT4 stories
Shopify官方·TechCrunch AIConfidence HighNew topic
Official report (headline taken verbatim from the ledger): Shopify launches Canvas, allowing users to build an online store via chat-based conversation.
Take · With 'conversation as a production tool' applied to e-commerce site building—a paid use case—AI shifts from an assistive feature to the transaction entry point itself.
Original ↗OpenAI官方·TechCrunch AIConfidence HighNew topic
Official report (headline taken verbatim from the ledger): ChatGPT can now let users virtually try on clothes.
Take · Image generation lands directly in retail conversion, a rare directly bookable monetization path in consumer AI.
Original ↗亚马逊官方·TechCrunch AIConfidence MidNew topic
Official report (headline taken from the original ledger): Amazon releases its own Jev clone as decision models appear in large numbers online.
Take · Big tech follows the new 'decision model' category: billing granularity shifts from token to 'per-use'; when granularity changes, the gross-margin structure changes.
Original ↗RunwayYouTube·RunwayConfidence MidNew topic
Runway Labs releases Continuum via an official video, describing it as a real-time video operating system (headline taken from the original ledger).
Take · Video generation moves from 'offline output' to a 'real-time operating system,' directly affecting latency metrics for finished output in talking-head/short-video pipelines.
Original ↗03INDUSTRY8 stories
Tencent to lease Oracle's Southeast Asia data center for about $7 billion, involving about 100,000 AI chips
腾讯 × 甲骨文IT之家 引金融时报（19/20 席同报）Confidence HighNew topic
Five-year lease, valued at about $7 billion, with about 30% upfront payment, including placement of about 100,000 advanced AI chips; it is Oracle's largest overseas lease agreement.
Take · Compute acquisition shifts from 'buying cards' to 'renting land,' turning capex into opex and leaving five-year uncertainty to lessors—the gap in compute 'carriers' lies in sites and power.
NVIDIA adds $150 billion buyback (total authorization $235 billion) and is in talks to backstop AI chip loans
英伟达新浪财经 / 财联社（18/20 席同报）Confidence HighNew topic
Adds $150 billion stock buyback, total authorization reaches $235 billion (largest in U.S. history); same day, in talks with insurers to backstop AI chip loan risk; Jensen Huang says chips should be seen as an 'investable asset class.'
Take · Financial engineering, not a capacity signal: the buyback amplifies earnings per share; physical depreciation does not slow due to financial packaging; equipment risk moves from manufacturers' balance sheets to the financial system.
Google launches satellite carrying in-house TPU, Project Suncatcher enters live testing
谷歌 × SpaceXThe Verge via 网易（7/20 席同报）Confidence HighNew topic
Uses Falcon 9 to send a satellite carrying an in-house TPU into low Earth orbit, focusing on verifying the chip's usability under radiation, shock, and severe temperature swings.
Take · Extending compute buildout to the most extreme form of “site” (orbit) is a first-principles response to terrestrial power bottlenecks, forming the other end of the same proposition as Tencent leasing sites.
AMD to acquire World Labs for about $8.2 billion in all-stock deal; Fei-Fei Li to serve as EVP and Chief Scientist
AMD红星资本局 via 36氪（17/20 席同报）Confidence HighProgress · Day 2 (10-01 already reported)
All-stock transaction expected to close by the end of 2026, pending regulatory approval; Fei-Fei Li will report directly to Lisa Su.
Take · Compute vendors are shifting from “selling compute” to “buying capability,” directly absorbing world model/spatial intelligence teams; the all-stock consideration ties the seller to the buyer’s long-term interests.
Anthropic prospectus: FY2025 revenue $4.59 billion (+1088%), operating loss over $8 billion
Anthropic财联社 / 新浪财经（19/20 席同报）Confidence MidProgress · Day 2 (already reported on 10-01)
Nearly a quarter of revenue comes from two customers, and most major customers have not signed long-term contracts; separately, it is purchasing Akamai cloud infrastructure for $11.6 billion over 7 years (about 15/20 outlets reported the same), with purchase commitments not absolutely rigid and tied to delivery and availability metrics.
Take · The revenue side is cancellable while the spending side is non-cancellable, a classic maturity mismatch; source correction: the source's '$518 billion in compute investment over the next year' is a multi-year cumulative obligation and should not be cited as a single-year budget.
OpenAI plans to raise at least $30 billion at a $1.4 trillion valuation; SoftBank's $30 billion is fully in place
OpenAI界面新闻 / IT之家（16/20 席同报）Confidence MidProgress · Day 2 (already reported on 10-01)
SoftBank completed the final tranche via the Vision Fund and is expected to hold about 13%; a March round this year had secured $122 billion in committed capital at a $852 billion valuation.
Take · The self-reinforcing loop of 'financing—compute—valuation' has taken shape and is the main anchor for AI asset prices in this cycle, directly determining upstream and downstream costs and talent flows.
Twin governance tracks on the same day: California's Anti-Robot Boss Act takes effect, and research heads at four labs jointly call for regulating 'research automation'
监管（加州）IT之家 / 新浪财经（17/20 席同报）Confidence HighNew topic
California requires human oversight and disclosure when automated systems are used for termination and disciplinary decisions; on the same day, research heads at OpenAI, Anthropic, Meta, and Microsoft co-authored an article saying research automation could trigger an 'intelligence explosion' and calling on policymakers to investigate companies' degree of research automation.
Take · Regulatory lever shifts from 'what content is generated' to 'who makes the decision'; compliance shifts from a cost item to an entry requirement. The joint statement is a single article from a single source; conclusions are downgraded as 'mechanism hypothesis, lacking causal evidence'.
OpenAI accuses Moonshot AI of systematic adversarial distillation: over 15,000 accounts, blocked on July 28
OpenAI × 月之暗面IT之家（引 OpenAI 声明）Single-source pending verification
OpenAI statement says it has faced systematic adversarial distillation since July; over 15,000 accounts were used to extract model capabilities, and all were fully blocked on July 28 (no response from the other party, single source).
Take · Model asset 'ownership confirmation' appears for the first time in large-scale attack-defense form; what can be distilled is the output distribution, while what cannot be moved away is the validated evaluation criteria and failure list.
04RESEARCH4 stories
何恺明团队官方·量子位Confidence MidNew topic
The report says the work uses general visual data to learn and transfer to the ARC abstract reasoning challenge (title taken from the original ledger text).
Take · Abstract reasoning is shifting from specialized task training to general visual pretraining transfer, a new example of the 'less labeling, more generalization' route.
Original ↗Apple ML官方·Apple MLConfidence MidNew topic
Apple Machine Learning Research: letting models internalize self-generated feedback as an improvement signal (title taken from the original ledger text).
Take · Echoing the same-day 'four labs jointly call for regulating research automation,' the self-improvement route is both a capability-ceiling indicator and a compliance-risk indicator.
Original ↗Prompted Identity Degrades Cooperation in Multi-Agent LLM Systems: Identity labels weaken multi-agent collaboration
Hugging Face Daily Papers本机20席情报（巽风/玄枢 引 HF Daily Papers，2026-09-28 批）Confidence MidListed batch has been marked.
Experiments show that once agents know each other's model family, the group splits into cliques and cooperation drops significantly (2026-09-28 listed batch, not same-day news).
Take · Direct experimental support for multi-agent collaboration rules: fewer identity labels, expose only roles and responsibilities—not a matter of politeness but of efficiency.
Rubric Rewards from Item Response Theory: synthesize rubrics into scalar rewards; see also DAGent deep research agent
Hugging Face Daily Papers本机20席情报（巽风 引 HF Daily Papers，2026-09-28 批）Confidence MidListed batch annotated.
Two other transferable takeaways from the same-day list: synthesize rubrics (rubric) into scalar rewards—i.e., 'a rubric is not a document; it is an incentive function'; DAGent proposes a DAG-style deep research agent (both in the 2026-09-28 listed batch).
Take · A skewed scorecard makes models optimize in the wrong direction; the deep research agent's DAG orchestration aligns with our multi-step task-chain design.
05INSIGHTS6 stories
Bottleneck shifts from 'chips' to the 'power—site—orbit' trio.
观点（本机20席情报·乾机／厚坤／燚燃）20 席情报Confidence Mid
Map today's fragments into one picture: chips (NVIDIA buyback, AMD acquisition) → data centers (Tencent's $7 billion lease) → grid interconnection (Musk's at least 15GW shortfall in 2027, chip demand growing 40%-50% annually vs power supply outside China growing 10%-20%); Google sends TPU to space to seek growth.
Take · The shortage is not something scheduling can fix; it is a physical quantity. Compute rental prices will become the pricing anchor before model capabilities do. Falsifiable check: if PPA contract prices broadly rise and transformer delivery lead times lengthen over the next two quarters, the regime shift is confirmed.
“Chips = an investable asset class” is financial engineering, not a capacity signal.
观点（本机20席情报·厚坤／巽风）20 席情报Confidence Mid
$235 billion buyback authorization + insurance backstop for chip loans essentially shifts equipment risk from manufacturers’ balance sheets into the financial system; buybacks amplify EPS, not capacity.
Take · Any financing, leasing, or collateralization based on “chips = assets” must calculate discount rates by physical depreciation cycles, not valuation multiples; as depreciation cycles overlap with lease expirations, around 2027 there may be a “compute stock clearing.”
Procurement metrics shift: from 'price per million tokens' to 'unit task cost,' with 'model replaceability' as a first-class engineering goal
观点（本机20席情报·坎流／玄枢）20 席情报Confidence Mid
Gemini 4 Argon / GPT-6.1 Sol / Claude 5.5 — three generations of flagships clash in the same month, with refresh cycles compressed to about 10 months; multiple price cuts in the same month, while unit prices and call volumes inflate in tandem
Take · The half-life of generational advantage is only 1-2 quarters, and the payoff window for betting on single-point capability gaps is extremely short; decoupling evaluation/process from the model is the only structural investment that can hedge against the accelerating pace
The real bottleneck for always-on Agents isn't the model; it's whether memory and knowledge can be inherited and retired
观点（本机20席情报·坤舆／明见）20 席情报Confidence Mid
dots' 7×24 and Doubao 'Spell' are both turning 'one-question-one-answer' into 'long-term work on your behalf'; whether this compounds depends on whether the knowledge it accumulates is retrievable, inheritable, and retireable
Take · Agent artifacts that aren't retired turn from assets into liabilities; any production team should launch an 'artifact deduplication and archival gate' this quarter, or maintenance costs will backfire in six months
Compliance shifts from a 'variable' to a 'fixed item': making 'automated decision disclosure + human review audit trails' a standard component
观点（本机20席情报·巽风／悦泽）20 席情报Confidence Mid
California's 'Anti-Robot Boss' bill requires human oversight and disclosure for automated decisions; the regulatory lever is shifting from 'model capability' to 'whether someone signs off on the decision process'
Take · Teams that pass review in one go have lower long-term customer acquisition costs; it is recommended to add a 'manual review checkpoint list' now (topic elimination, hit re-rendering, ingestion/discard decisions, etc.).
Distillation channels narrow + identity masking research: moats must shift to data, workflows, and distribution
观点（本机20席情报·乾机／厚坤）20 席情报Confidence Mid
OpenAI's crackdown on large-scale distillation (over 15,000 accounts) and the HF paper 'Prompted Identity Degrades Cooperation' mirror each other; distillation saves startup costs but not verification costs.
Take · Don't build your business on 'capability gaps you get for free'; add the trio of 'verifiable usage + anomalous-call audit trails + stop-on-exceed' to external partnership contracts—real-time enforceability beats after-the-fact accountability.
Ecosystem Pulseself-updating
News600Skills72OpenHub44Reddit252026-09-02 → 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 Items4
1. Lock in 'power and site' as asset clausesFor this quarter's large compute projects, review long-term power and rack contracts as asset clauses (recommend locking in for 3+ years), and change the hard procurement metric from 'price per card' to 'intelligent output per watt, tokens/W'; at equal performance, low-power models go directly into the approved list.
window 6-12 months 2. Engineering 'model replaceability': decouple evaluation/workflows from modelsRe-account in-use model calls by 'actual cost per completed task (including retries and tool calls)'; renegotiate procurement unit prices quarterly instead of annual locks; change external pricing from hourly/per-use to per-task.
window 1-2 quarters 3. Turn the 'compliance delivery package' into a standard componentComplete the trio of 'automated decision disclosure + human review audit trail + data flow diagram'; first produce a 'human review checkpoint list' (topic elimination, viral hit re-rendering, ingestion/rejected draft determination); add 'verifiable usage + abnormal call audit trail + stop upon exceeding limit' to external partnership contracts.
window 2-3 quarters 4. On the video side, immediately run Kling 4.0 cross-shot consistency A/BStart A/B as soon as the Flash version opens, tracking only 'cross-shot consistency' and 'per-item cost' on two axes, freezing all other variables; first see which more affects final video usability rate.
window This week