01MODEL RELEASES4 stories
GPT-6 fully rolls out globally, replacing GPT-5.6 Sol/Luna
OpenAI腾讯新闻Confidence HighFull rollout
10/9 OpenAI fully rolled out GPT-6 to ChatGPT users worldwide, replacing GPT-5.6 Sol/Luna, available to both free and paid users, with Astra safety technology integrated; market rumors put its annualized revenue at nearly USD 50 billion (Tencent News 2026-10-09). The competitive focus shifts from 'who is strongest' to 'who can get paid'.
Take · The competitive focus shifts from 'who is strongest' to 'who can get paid'; Astra being written into GPT-6 = safety shifts from 'after-the-fact compliance' to a 'release threshold'.
vivo launches 3B on-device multimodal model: 128K context, output >200 token/s
vivo腾讯新闻·蓝鲸Confidence HighNew release
10/10 vivo launches 3B on-device multimodal model (128K, output >200 token/s), official says 'the endgame for on-device is Agent' (Tencent News 2026-10-10).
Take · On-device lowers the cost base another level, moving the effective per-token cost curve down and lowering the break-even call volume threshold for the application layer.
Xiaomi MiMo-V2.5 to be retired: switch to V2.6 on 10/14; old name invalidated on 10/21
小米IT之家Confidence HighMigration window
Xiaomi MiMo-V2.5 to be retired; auto-forward to V2.6 on 10/14; old name invalidated on 10/21; pricing for equivalent intelligence is only 1/20–1/60 overseas (IT之家 2026-10-09).
Take · Offers a ready-made, reviewable sample for the “model retirement—migration window—rollback period”; any pipeline bound to the old model name must be changed this week.
On-device open-source gap-filler: Liquid LFM2.5-2.6B runs locally on phones + JetBrains Mellum 2.1
Liquid AI / JetBrainsIT之家 / LLM ReferenceConfidence MidSame-direction gap-filling
Liquid AI LFM2.5-2.6B runs locally on phones (IT之家 2026-10); JetBrains Mellum 2.1 Thinking (12B total params / 2.5B active, 131K context) (LLM Reference 2026-10-08)—four companies push toward the 2.5B—3B tier in the same week, direction aligned.
Take · Open inference artifacts complete one piece of the localization loop, making on-device 'reproducible + privatizable' the new standard.
02PRODUCT5 stories
ChatGPT adds Intelligent UI: directly generates clickable, computable charts
OpenAITechCrunch 经 AI Pulse 转载Confidence MidInteraction upgrade
10/7—8 ChatGPT adds Intelligent UI, can directly generate clickable, computable charts, no longer only returning plain text (TechCrunch via AI Pulse, 2026-10-08).
Take · Interaction form upgrades from text artifacts to actionable objects, a structural leap at the entry layer.
LangChainAI 官方视频台账Confidence MidProduction-grade
LangChain launches Managed Deep Agents (deploying mission-critical Agents to production) (AI official video log, 2026-10-10).
Take · Agent moves from 'can run' to 'operable and managed'; the managed layer becomes new infrastructure for the ecosystem.
Original ↗DatabricksAI 官方视频台账Confidence MidOfficial Case Studies
Databricks releases case of HelloFresh using Genie Agents to optimize 100% of Meta ad delivery (AI official video ledger, 2026-10-10).
Take · An implementation sample of Agents taking over the full business loop; the data platform side enters a 'measurable output' phase.
Original ↗Natura launches $99 AI Agent smart ring Interface
NaturaZGLG AI NewsConfidence MidSingle-source
Natura launches $99 smart ring Interface, which can summon AI agents (ZGLG AI News, 2026-10-08, single source).
Take · Entry points continue to move off-screen, interaction migrates to the body and local, and distribution-side checklists need a new item: 'on-device reachable surface.'
Google Cloud launches Gemini office agent.
谷歌云中财网Confidence Mid
Google Cloud launches Gemini office agent; institutions say the agent industry chain has entered a period of certainty-driven growth (China Finance Online, 2026-10-09).
Take · Agent shifts from the demo stage to a standard cloud vendor product, marking a turning point in commercialization framing.
03INDUSTRY8 stories
Compute capital divergence: NVIDIA plans to invest up to $100 billion in OpenAI; Firmus IPO bookbuilding stalls.
英伟达 / FirmusCNNBC / 新浪财经经36氪转载Confidence HighPrimary market hot, secondary market cold
NVIDIA plans to invest up to $100 billion in OpenAI to build data centers and separately reach a $6.3 billion cloud deal (CoreWeave +5%); on the same day, NVIDIA-backed Australian data center company Firmus's IPO ended bookbuilding but still could not set a price, and shareholder Maas Group was at one point -30% (CNNBC / Sina Finance via 36Kr, 2026-10-08).
Take · Capital is forming an internal loop between compute and models; this must be read alongside Firmus's setback—the first alarm bell at the peak of the boom.
TSMC September sales NT$511.86 billion, +54.6% YoY
台积电财联社Confidence HighHard data
September sales NT$511.86 billion; first three quarters NT$3.89873 trillion, +41.1% YoY, with AI demand continuing to drive growth (Cailian Press, 2026-10-09).
Take · The day's only self-explanatory monthly hard data, providing the hardest physical quantity for the entire compute narrative that can be falsified next month.
Application-layer valuation dual anchors rise: Manus valued at about $4 billion, Databricks at $43 billion
Manus（蝴蝶效应）/ Databricks36氪 / 华尔街见闻Confidence HighDual valuation anchors
Manus parent Butterfly Effect completes over $500 million funding at about $4 billion valuation (led by Boyu/IDG, roughly 2x Meta's acquisition price, 36Kr, 2026-10-09); Databricks valued at $43 billion in new round, called the hottest AI IPO candidate (Wallstreetcn).
Take · Application-layer Agents get a platform-level pricing anchor for the first time, with valuations rising in tandem with the data infrastructure layer.
L4 White Rhino completes $100 million Series C
白犀牛36氪Confidence MidSingle-source
On the capital side of embodied AI, L4 White Rhino completes $100 million Series C (led by Hidden Hill Capital, 36Kr, 2026-10-08).
Take · Capital goes into end-to-end and physical AI, as embodied AI becomes an independent capital-side track.
Amazon cuts jobs again across multiple teams, affecting India and the UK.
亚马逊新浪财经Confidence MidSingle-source
Amazon layoffs hit multiple teams again, affecting India and UK; employees flood a nearly 37,000-person Slack channel to press for severance (Sina Finance, 2026-10, single source).
Take · Big Tech cost cuts and compute capex advance in parallel, with organizational structures reshuffled along AI investment direction.
Meta FAIR releases 8-billion-parameter robot world model RoboJEPA
Meta FAIRAI Daily NewsstandConfidence MidSingle-source
Trained on 15022 hours of video, focused on robot representation learning (AI Daily Newsstand, 2026-10-09, single source).
Take · World models are becoming a new foundation for embodied intelligence, bringing with them the risk that 'skill poisoning can penetrate to the physical layer.'
Six AI giants sign 'Joint Commitment on Frontier Responsibility' + Democrats unveil AI safety framework
规则 · 治理Nemo 社区 Edge AI Daily / AI UnderstandingConfidence MidExistential commitments
On 9/29, Trump and the CEOs of Anthropic, OpenAI, Google, Meta, xAI, and Nvidia signed a commitment to strengthen safety oversight of frontier models (the document misspelled United States as 'Unites States'); Democrats unveiled an AI safety and responsibility framework for the midterm elections (2026-10-03).
Take · Governance enters an 'election-cycle-driven' phase; signings are existential and non-binding—do not rely on them long term.
AnthropicTechCrunch AIConfidence MidReal incident
TechCrunch reports a real law enforcement incident triggered by a model false alarm (TechCrunch AI, 2026-10-09).
Take · Agent overreach/false reporting has moved from speculation to real cases, direct evidence of missing permission and audit rules.
Original ↗04RESEARCH3 stories
Skill Constellations: The 'No Registry' Gap in the Agent Skill Supply Chain
HuggingFace 论文HF Daily PapersConfidence MidSingle paper
The paper notes that SKILL.md and scripts spread through repo-to-repo copying, with no registry, no versioning, and no provenance, yet execute with user privileges—an untraceable software supply chain (HF Daily Papers, 2026-10-08, single paper).
Take · Whoever first offers 'skill source registry + version locking + signed distribution' will control the first gate for enterprise Agent deployment; the first skill poisoning could penetrate to production and the physical layer.
On-device models' 'output speed/benchmark score' metrics are highly inconsistent and cannot be directly compared.
口径提醒深度版研判Confidence MidMethodology
The measurement metrics for on-device models are highly inconsistent and mostly single-source; the right approach is to first establish comparable metrics (cost per task, P95 latency, and accuracy, in three columns) and then compare numbers.
Take · In a price war, the easiest pitfall is using readings with different definitions for cross-comparison; conclusions will be systematically distorted.
OpenAI reportedly released 722 math papers at once and solved 90 of 500 open problems.
海外侧传闻Latent SpaceConfidence MidPending verification
Latent Space reports OpenAI released 722 math papers in a single batch and solved 90 of 500 open problems (Latent Space, 2026-10-07).
Take · The scale is significant, but it is a single institution’s self-assessment and pending verification; reviewers noted that only 10 of 719 papers disclosed chain-of-thought and 42% were not formalized—mass output ≠ mass correctness.
05INSIGHTS5 stories
Today’s main theme: the evaluation lens shifts to “unit economics”.
前瞻深度版研判Confidence Mid
GPT-6’s full rollout, on-device small models cutting prices to 1/20–1/60 of overseas levels, and TSMC corroborating with order readings—all three on the same day point to “who has the lowest unit cost and who actually gets paid”; leaderboard influence is being diluted by cash-flow metrics.
Take · The model layer’s excess-profit window is narrowing, with value shifting toward “distribution and trust close to use cases”.
The compute chain’s “primary hot, secondary cold” divergence must be read in tandem.
前瞻深度版研判Confidence Mid
Upstream is leveraging up via 10-billion-scale circular financing, while downstream ability to pay has not kept up; Firmus’s failed price inquiry is the first crack and must be read alongside NVIDIA’s $100 billion investment commitment.
Take · Once the financing structure is first disrupted by external markets, full-chain resonance transmits in reverse.
Seven-day A/B migration for on-device/domestic models
技巧深度版建议Confidence Mid
Take 5 high-frequency real tasks (one each for coding/summarization/extraction/translation/tool calling), with Group A on cloud flagship and Group B on on-device 3B/domestic low-cost; lock metrics to three columns: unit task cost, P95 latency, and accuracy.
Take · Xiaomi 10/14 switchover and 10/21 old-name invalidation are hard deadlines; recalculate ROI each time cost drops one tier.
Agent skill provenance and on-device security health check
技巧深度版建议Confidence Mid
First establish a minimum viable 'three elements of skill source registration'—source repository + committer identity + content hash; internally, change external skills' default status from 'allowed' to 'forbidden', tightening into a private whitelist.
Take · 6-12 month gap (HF papers have flagged it, standards have not yet appeared); once the first third-party registration scheme appears, interface standards must be aligned.
Unit economics scorecard + non-price retention
技巧深度版建议Confidence Mid
Switch the internal scorecard from 'flagship benchmark scores' to 'per-token output / RMB cost per completed qualified task,' and add a 'non-price retention metric' to customer success.
Take · During the price-war window, list 'has independent cash flow' as a higher-weighted field than 'has big-company backing.'
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 Items3
1. Seven-day A/B migration for on-device/domestic modelsTake 5 high-frequency real tasks (one each for coding/summarization/extraction/translation/tool calling); Group A uses cloud flagship, Group B uses on-device 3B/domestic low-cost, with metrics locked to three columns: 'unit task cost, P95 latency, accuracy.' Each task needs ≥30 samples for significance → criteria written in advance (if Group B cost ≤ 1/5 of Group A and accuracy gap ≤5pp, judge 'migratable') → after migration, write 'model shutdown—migration window—rollback period' into a three-clause hard-term template.
window 7 days starting immediately 2. Agent skill provenance and on-device security checkupFirst establish a minimum viable 'three elements of skill source registration'—source repository + submitter identity + content hash; internally change the default status of external skills from 'allowed' to 'prohibited,' consolidating into a private whitelist. Free provenance checklist for lead generation → per-use skill audit (including version locking and permission scope) → SaaS-based continuous monitoring; make 'reproducible' a contract acceptance item.
window 6-12 months 3. Unit economics scorecard + non-price retentionChange the internal scorecard from 'flagship benchmark scores' to 'output per unit token / RMB cost per completed qualified task,' and add a 'non-price retention metric' to Customer Success (how many renewal reasons are unrelated to price). Recalculate the 'AI unit cost table' once this week, and go through last year's 'won't do because too expensive' list item by item → put a 'same-intelligence cost comparison' chart on the customer-facing first screen.
window Starting immediately for 1 month