01MODEL RELEASES3 stories
Hunyuan Hy4 official release adds 'computer use' for the first time: hybrid command-line and GUI, API priced the same as preview
腾讯混元少数派《派早报》Confidence HighProgress · Day 2AI score 92/100
Released on 11-04, the Hunyuan Hy4 official release introduces visual understanding and 'computer use' for the first time; it combines command line and GUI, with special training for professional software such as Blender and CAD and long-horizon cross-application tasks; it is already on Tencent Cloud and OpenRouter, API pricing is on par with preview, and weights will be open-sourced soon.
Take · For the first time, a domestic model has put 'operating someone else's software' on its official-release capability list. But the success rate is currently vendor-reported with no third-party testing—before adding it to a procurement list, first self-test once on three real desktop tasks of your own.
GoogleIT之家／官方·Google AI BlogConfidence HighNew topicAI score 90/100
Native 3–15 seconds, 1-second granularity, multi-turn conversational editing and stronger reference adherence, with SynthID built into outputs; 30% cheaper than 1.1, down to $0.07 per second for 720p, and can draft at 360p then upscale to 4K. The same model is now available on ElevenLabs ElevenCreative.
Take · From 'gacha' to 'revision,' the content pipeline can for the first time calculate per-item cost: 15 seconds at 720p = 15 × 0.07 = $1.05 ≈ RMB 7.5 (local calculation), versus about $1.5 per item before the price cut.
Original ↗OpenAI官方·OpenAI NewsConfidence MidNew topicAI score 88/100
OpenAI officially releases GPT‑6 Cyber, positioned as an aligned cybersecurity agent; the same day, TechCrunch asks, "Is one new cybersecurity model enough?"
Take · Security models are starting to put "alignment" into product names—compliance narratives are outweighing capability narratives. For buyers, this means adding an "auditability" column to the selection matrix, not just looking at detection rates.
Original ↗02PRODUCT LAUNCHES5 stories
Kimi Hosted Agents: selling Harness as a cloud service
月之暗面北京商报Confidence HighNew topicAI score 86/100
Moonshot AI launches Kimi Hosted Agents, selling Harness as a cloud service and, on the same day as DeepSeek, turning the Agent runtime environment into a commodity.
Take · Pricing units shift from Token to "seat + runtime environment." Hosting has built-in migration costs; choosing a provider means paying a "foundation tax"—be sure to include an exit clause in hosting-layer contracts.
DeepSeek Harness: 1.54 million DAU, about 92 trillion Token consumed per day, and 35,000 third-party plugins
DeepSeekIT之家Confidence HighNew topicAI score 90/100
As of 11-04: DAU 1.54 million, daily consumption about 92 trillion Tokens, 35,000 third-party plugins, about 60% of active users have installed plugins; officials will build a plugin marketplace and prepare Token incentives.
Take · 92 trillion ÷ 1.54 million ≈ 60 million Tokens/person/day (local division estimate, not original source basis): this scale can only come from long-context multi-turn Agent loops, not chat. It proves 'runtime' is already an independent commodity.
Meta官方·TechCrunch AIConfidence HighNew topicAI score 82/100
TechCrunch reports: US users can now make video calls with their own Muse AI avatars.
Take · On the consumer side, 'video calls with AI' is being made a default entry point, in contrast to Uber's '4 AI orders in 320 million trips': interaction innovation is rising, but transaction authorization has not moved.
Original ↗Anthropic官方·TechCrunch AIConfidence HighNew topicAI score 80/100
TechCrunch reports: Claude adds Personal Finance tool to analyze personal spending.
Take · This is not a capability upgrade but a compliance-oriented entry point: assistants are starting to touch regulated billing data. On the money side, authorization currently only goes as far as 'view,' not 'pay'—exactly complementing Uber's 4 times.
Original ↗Enterprise-facing public messaging sells 'roles,' not 'models.'
Snowflake／Salesforce／GitHub官方视频台账 2026-11-06 节收录Confidence MidNew topicAI score 74/100
Snowflake discussed an AI-first sales and marketing operations model, Salesforce discussed a Slackbot for Forward Deployed Engineers, and GitHub discussed 'humans set direction, agents close the loop'; all three went public on the same day.
Take · All three frame this around headcount, not model parameters: enterprise budget categories are shifting from 'software licenses' to 'labor replacement,' determining whether procurement comes out of IT budgets or HR budgets.
03INDUSTRY9 stories
Oracle cancels roughly $7 billion chip lease with Tencent.
甲骨文／腾讯少数派《派早报》引 FTConfidence MidNew topicAI score 94/100
According to FT, the deal ended after US pressure: originally a five-year lease covering multiple Southeast Asian data centers and about 100,000 advanced AI chips. Single source; no direct confirmation from either side.
Take · Compute has shifted from 'can you afford it' to 'will they provide it.' Political terms in cross-border leases must go into SLAs and training schedules—for the first time, counterparty risk on delivery commitments is as important as price.
Uber: 4 orders placed via AI assistant among 3.2 亿 trips, penetration rate 1.25×10⁻⁸
Uber少数派《派早报》引 11-04 财报电话会Confidence HighProgress · Day 2AI score 95/100
In 2026Q3, ride-hailing and food delivery totaled 3.2 亿 orders; only 4 ride-hailing orders were placed via AI assistant (about 500 food-delivery orders in Halloween week) → 4 ÷ 3.2 亿 = 1.25×10⁻⁸ (local calculation, actual division).
Take · The point is not 4 but 3.2 亿—the whole market is giving numerators; this is one of the rare first-hand datasets that also gives the denominator. The 500 food-delivery orders, converted on the same basis, are about 2×10⁻⁵; the two differ by three orders of magnitude: subsidies and coupons can tolerate failure, but ride-hailing has zero fault tolerance.
NRTA sets tone: 'support live-action dramas, regulate AI dramas'
广电总局网易·政知新媒体Confidence HighNew topicAI score 88/100
11-05 in Beijing: differentiated stance on live-action micro-dramas and AI micro-dramas clarified, setting the tone for AI film and TV content.
Take · It is not a ban, but positioning AI dramas as an 'assistive tool' rather than 'replacement production': part of the cost advantage will be consumed by filing, review, and labeling; ROI for capacity-driven teams must be recalculated.
Draft 'Ban Artificial Superintelligence Act': up to 20 years in prison for individuals; companies may be forcibly dissolved
美国民主党新浪财经（快科技）Confidence MidNew topicAI score 84/100
The draft calls for a total ban on ASI R&D and deployment, defining ASI as 'reaching or surpassing human cognition across broad domains.' Single source; draft stage.
Take · The real risk is in the definition: no benchmark, no threshold, no measuring entity; the burden of proof would be reversed—the challenged party must first prove it is 'not ASI.'
Dropbox／Tigris Data官方·TechCrunch AIConfidence HighNew topicAI score 82/100
TechCrunch reports: Dropbox acquires Tigris Data for $350 million to strengthen AI infrastructure.
Take · Application-layer companies are starting to buy 'data foundations': money is going to distributed object storage rather than models. AI's marginal cost is materializing on the storage side, not the inference side.
Original ↗Chengdu Ruixinxing raises tens of millions of yuan: human-brain-like embodied model; replacing parts does not require re-collecting data
成都睿芯行36氪·硬氪Confidence MidNew topicAI score 80/100
An embodied model based on human-brain-like mechanisms can perform industrial picking and grasping without collecting training data for new parts; it has been tested at Zhejiang Zhongli, Haier, and BOE factories, with investors including Sinovation Ventures, Zhongli Shares, and Today International.
Take · “Training-free” eliminates the most expensive hidden cost for industrial customers—data collection and line-stop calibration. The selling point is not accuracy, but line-change time.
Three AI chip startups have completed 10 亿元-level financing in quick succession.
AI 芯片新锐（含墨芯）网易·全球半导体观察Confidence MidNew topicAI score 78/100
NetEase · Global Semiconductor Observer reports: three AI chip startups have completed 10 亿元-level financing in quick succession, including Moffett AI.
Take · This forms a “supply cutoff—substitution” hedge with Oracle’s same-day order cuts: money is moving toward the inference side and the domestic substitution chain, not toward peak training compute.
Anthropic signs its first news content licensing deal: partnering with AP for real-time election-night results.
Anthropic／美联社官方视频台账 2026-11-06 节收录Confidence MidNew topicAI score 76/100
Anthropic signs deal with Associated Press to show AP results live on election night with attribution; the company had previously argued that scraping was fair use.
Take · From 'scraped' to 'signed': content costs shift from variable costs to fixed expense items, rewriting the unit economics of every RAG product.
Google adjusts K-12 education AI policy: off by default; school administrators must actively enable it.
Google少数派《派早报》Confidence MidNew topicAI score 74/100
Starting in 2027, if schools ban AI entirely, students will not even be able to use Google Search.
Take · Compliance no longer relies on after-the-fact audits but changes the defaults: the cheapest and hardest-to-roll-back form of regulation, shifting costs to procurement.
04RESEARCH5 stories
Microsoft Research官方·MSR BlogConfidence MidNew topicAI score 84/100
Microsoft Research blog introduces PDD Kit, using formal verification to check Agent-generated full-stack web applications.
Take · The first engineering answer to 'how to prove Agent outputs are correct.' For buyers, this is a prerequisite for advancing Agent delivery from 'trial' to 'acceptance.'
Original ↗Structure-guided RNA foundation model published in Nature
中国科大中国科大新闻网Confidence MidNew topicAI score 78/100
The structure-guided RNA foundation model from Zhou Shaohua and Xiong Peng's team at USTC has been published in Nature.
Take · AI for Science is one of the few cumulative tracks unaffected by traffic cycles. The value of such results is measured by 'reusable experimental criteria,' not by leaderboard rankings.
Apple ML官方·Apple MLConfidence MidNew topicAI score 70/100
Apple Machine Learning Research blog post: “Faster Retrieval for Better, More Efficient RAG.”
Take · RAG optimization is shifting from the “generation side” to the “retrieval side”: at the same answer quality, cost differences mainly come from the recall layer, not decoding parameters.
Original ↗eunomia.dev官方·Hacker News AIConfidence MidNew topicAI score 72/100
The project page introduces AgentSight, which uses eBPF to profile and monitor AI Agents across the whole system.
Take · Agent observability is shifting from “application logs” down to the “kernel side”: forensics on unauthorized access and data exfiltration must ultimately land at this layer to serve as evidence.
Original ↗arXiv官方·Hacker News AIConfidence MidNew topicAI score 68/100
arXiv paper: “Joules-to-GDP: How Much Economic Value Do AI Systems Produce per Unit of Energy?”
Take · Switching the compute accounting from “dollars” to “joules” is an external validation of the same methodological line as “effective bandwidth per watt × available power.”
Original ↗05INSIGHTS5 stories
The three dimensions cannot be added: transaction side, supply side, and unit-price side.
口径提醒深度版日报·数据维度Confidence HighViewpointAI score 86/100
Today’s three key figures belong to different dimensions: 4 order placements (transaction side), 92 trillion Token (supply side), and 0.07 USD/second (unit-price side).
Take · The most common methodological error across the 18 intelligence items is putting “92 trillion Token” and “4 order placements” into the same chart to calculate an average. The three cannot be added together, nor can they cross-validate penetration rate—separate them into columns before writing them into the quarterly review.
Revocable small-amount authorization: start with scenarios where “a 30-yuan mistake can be refunded”.
商机深度版日报·商机维度Confidence MidViewpointAI score 84/100
For scenarios with extremely low payout caps, such as price comparison, rebooking prefill, expense reimbursement aggregation, and meeting-minutes backfill: first provide a suggestion form (do not execute directly) → one-click confirmation → revocable afterward.
Take · Uber’s 4 instances show the bottleneck lies in authorization and failure fallback, not model capability. Use “completion rate without human intervention” as the KPI and prohibit call volume/impressions—otherwise teams will sacrifice revocability to chase volume.
Plugin slot-grab window: about 3 months
商机深度版日报·商机维度Confidence MidViewpointAI score 80/100
35,000 plugins, a 60% installation rate, and official Token incentives in preparation—before the plugin marketplace officially launches is the only window to grab a slot.
Take · After launch, positions are locked in by call volume; to get in later, you must buy placement based on call volume. The time-window judgment is an inference; the two figures 35,000 and 60% are substantiated in the original text.
Video costs move to two-tier quotes: drafts per second, final versions per video.
成本深度版日报·商机维度Confidence MidViewpointAI score 78/100
Recalculate material costs at 0.07 USD/second: repeatedly test drafts at 360p, upgrade the final version once to 1080p/4K, and change retry costs from 'per item' to 'per second'.
Take · Split one high-cost generation into multiple low-cost iterations. Add three items to the quote sheet: per-second unit price, resolution tier, and whether retries are billed—without the third, all quotes are floating.
The more prosperous the plugin ecosystem, the higher the returns from supply chain attacks.
风险深度版日报·生态维度Confidence MidViewpointAI score 76/100
Anthropic's official VS Code plugin was poisoned on Cursor for nearly 12 hours, affecting up to about 5,000 people (reported 2026-11-03; via link, full text not read).
Take · In line with '60% install base + plugin incentives': the wider the install base and the larger the per-call incentives, the higher the expected payoff from poisoning. Those who installed it should first check extensions and credentials; uninstalling the extension may not be enough.
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. Revocable proxy order-placement packageEntry point: only do 'low-payout, rollbackable' scenarios—price comparison, rebooking pre-fill, expense consolidation, meeting-notes backfill—with the one-liner 'If it's wrong, 30 块 can be refunded; we won't make you pay for AI.' Path: first provide a suggested order (do not execute directly) → one-click confirmation → can be withdrawn afterward; use 'completion rate without human intervention' as KPI, and prohibit call volume/impressions.
window Starting now for 3 months: Uber cold data has just come out, and no one in the same track is doing authorization interaction; the first mover defines the defaults. 2. Build plugins, not an entry pointApproach: package capabilities as plugins and list them in the DeepSeek/Kimi marketplaces, trading others’ distribution for time; building your own Harness is the most expensive route. Path: pick one quantifiable pain point for launch → target 'next-day retention calls' rather than total call volume → build a failure-sample library (cloud vendors won’t do it for you).
window About a 3-month land-grab window (before the plugin marketplace officially launches); once the Token incentive details are out, the outcome will be clear. 3. Video production line switches to two-tier pricingApproach: recalculate in-house footage costs at $0.07/second—draft at 360p for repeated testing, upgrade the final cut once to 1080p/4K, and shift retry costs from 'per item' to 'per second.' Path: lock script and storyboard first → refine pacing at low resolution → do high-res final rendering only for proven hits; add three items to the quote sheet: 'per-second unit price + resolution tier + whether retries are billed.'
window Run one real production test within two weeks; if OpenAI/ByteDance follows with price cuts, recalculate the entire cost model immediately.