01MODEL RELEASES4 stories
混元 Hy4 released: official says it has entered the top tier of strongest open-source models for the first time; system card published before weights.
腾讯人民网（经网易转载），2026-11-04Confidence HighNew topicAI score 91/100
Released on November 4; official says it has entered the top tier of strongest open-source models for the first time, publishing a system card at the same time, highlighting cyber and biosecurity risks, and promising to open-source the weights soon.
Take · Open-source strategy shifts from 'releasing weights' to 'releasing weights + constraints': issue a safety card before releasing weights, treating auditability as the entry ticket. Before the weights are released, 'top tier' is only the official self-assessment; only third-party retesting counts.
openPangu-2.0 sets retirement date: sales stop 11/6, service ends 11/30; 2.1-Pro is already available as a paid offering in Beta 小艺 Work.
华为IT之家，2026-11-05Confidence HighNew topicAI score 89/100
华为 announces openPangu-2.0-Flash／2.0-Pro sales stop on November 6 and retirement on November 30; 2.1-Pro is already available as a paid offering in Beta 小艺 Work, with parameters undisclosed.
Take · First public case in China of a 'forced retirement of an old model'; only 24 days from sales stop to retirement—the model replacement cycle is now shorter than enterprise system upgrade cycles. Buyers have one more hard clause in negotiations: version lifecycle commitment.
Unreleased Claude Fable 5.5 hands-on test leaks: 2 hours 40 minutes of continuous coding, pure Python generates a 66-second animation
AnthropicIT之家，2026-11-04Confidence MidNew topicAI score 84/100
The developer says it coded continuously for 2 hours 40 minutes, generating a 66-second animation in pure Python (including visuals, soundtrack, and sound effects packaged into a single HTML), without using any image/video generation tools throughout.
Take · The direction matters more than the score: multimodal media may fall back from 'dedicated generative models' to 'code as the medium.' If this holds, the cost structure of video production lines must be recalculated as 'code + rendering,' not 'seconds × unit price.'
Galbot, together with Peking University and others, releases NavFoM: the world's first cross-embodiment full-domain surround-view navigation foundation model
银河通用极客时间，2026-11-05Confidence MidNew topicAI score 80/100
Using TVI Tokens + BATS strategy, the same navigation capabilities can be reused across different robot embodiments—turning 'retrain every time you switch embodiments' into 'one foundation, multiple embodiments.'
Take · The competitive focus in embodied intelligence is shifting from 'single-machine demos' to 'cross-embodiment reuse rate'; procurement criteria accordingly become: how many embodiments one foundation can cover, and whether switching requires recalibration.
02PRODUCT LAUNCHES4 stories
Apple Intelligence China edition opens for testing: trials start 11-04, rolling out with iOS 27.3 in December; Siri and others to integrate Alibaba Qwen
Apple／阿里Apple 官网新闻稿，2026-10-31（11-04 起测试）Confidence HighProgress · Day 2AI score 90/100
Opens for testing starting Nov 4 and rolls out in December with iOS 27.3; the first batch includes real-time translation, writing tools, visual intelligence, photo eraser, and notification summaries; Siri / writing tools / visual intelligence connect to Alibaba Qwen; this release does not include next-generation Siri.
Take · From “planned to start testing on 11-04” (this column reported on 11-02) to the actual start, only the rollout cadence remains as the gap. A default entry point with a billion-level installed base landing can reshape the distribution landscape more than any model benchmark.
OpenAIOpenAI News（官方），2026-11-05Confidence HighNew topicAI score 78/100
OpenAI official announcement: Daybreak Red is now available in the AWS GovCloud (US-East) region.
Take · Government cloud becomes the compliance foothold for frontier capabilities—this is deployment-side rather than capability-side news, but it determines who can procure it: those seeking to enter the public sector should first check whether the compliance zone includes them.
Original ↗Alibaba Cloud Qovest AI investment research workstation opens individual invite-only testing at 99 元/月
阿里云腾讯新闻·AIbase，2026-11-04Confidence MidNew topicAI score 76/100
Combine market quotes, research reports, data analysis, and AI generation into one workbench, opening individual invite-only trials at 99 yuan/month.
Take · Financial AI shifts from selling tools to selling subscriptions, with price as the entry point: 99 yuan sets an anchor for similar products and brings professional investment research down to a price individuals can pay.
OpenAI／DevRevOpenAI News（官方），2026-11-05Confidence HighNew topicAI score 72/100
OpenAI's official page publishes the DevRev case: delivering autonomous self-service with GPT-6.1 Sol.
Take · The boundary of self-service is pushed from answering tickets to getting the job done itself; such cases are the most practical proxy for Agent commercialization: look at how many person-hours they replace, not at the demo.
Original ↗03INDUSTRY9 stories
Decision models become a new category: Jev end-to-end 70–500 ms, valuation 200 million→$7.5 billion in 24 days.
TypeSafeAI36氪，2026-11-04Confidence MidNew topicAI score 88/100
Jev outputs only yes/no, scores, and probabilities, with end-to-end latency of 70–500 ms and 20–200× faster than peers; valuation rose from $200 million to $7.5 billion in 24 days (about 37.5×); OpenAI followed with Decisions API within two weeks, AWS/Microsoft/Cloudflare entered, and more than ten similar products appeared within a month.
Take · Using free text to carry one bit is pure waste: the information gain of the vast majority of business decisions is concentrated in a few binary judgments. OpenAI's follow-up within two weeks shows the model-layer moat ≈0; the real assets are the annotation distribution near the decision threshold and workflow embedding.
Kuaishou-backed Lingchuan Technology raises RMB 1 billion Series B: fully domestic 3D stacked chip has returned from fabrication, SL200 has sold nearly 100,000 units
凌川科技36氪·硬氪，2026-11-04Confidence HighNew topicAI score 82/100
Yizhuang Guotou Yitang Shenghe Fund, Qiming Venture Partners, and Zhaohui Capital jointly led the round; its self-developed DiPU deterministic-latency architecture + 3D near-memory MCM targets cost reduction for AIGV video inference, with volume shipments next year; SL200 has sold nearly 100,000 units, covering 99.7% of Kuaishou live-stream transcoding.
Take · For the first time, the cost curve of video inference has empirical evidence of 'dedicated chips + proven scenarios,' rather than PPT TOPS figures. For buyers, this is upstream of the new pricing unit, 'cost per frame.'
Arm FY2027 Q2: AGI CPU orders raised from $1 billion to over $2 billion; server share from 15% to over 50%
ArmIT之家，2026-11-05Confidence HighNew topicAI score 84/100
Revenue $1.443 billion, +27% YoY; technology licensing revenue $773 million, +50% YoY; data center royalties doubled; AGI CPU order value raised from about $1 billion to over $2 billion, with first commercial deliveries set for FY2027 Q4; GAAP net income $130 million, -45% YoY (mainly due to a $46.6 million fair value loss on equity investments), non-GAAP net income $566 million, +36% YoY.
Take · Two directions in the same earnings report: GAAP falling, non-GAAP rising; the order book and royalty structure are the leading indicators. Compute has already been locked in advance through FY2027, meaning 'buying it' is scarcer than 'buying it cheap.'
Anthropic reportedly to complete IPO in the week of November 16: raising $100 billion at a $1.5–2 trillion valuation
Anthropic黑马早报引外媒，2026-11-04（单源·官方未确认）Progress · Day 3AI score 86/100
Local media, citing informed sources, report that it plans to complete the IPO in the week of November 16, raising $100 billion at a $1.5–2 trillion valuation; it previously disclosed revenue of $33.242 billion for the first three quarters but losses widened.
Take · Based on a rough $1.5–2 trillion valuation, PS is about 45–60x (not annualized; about 34–45x after annualization), implying it would need 'revenue to grow severalfold in two years and losses to narrow' to hold up. Single-source and unconfirmed; this column only records the basis and makes no valuation inferences.
US midterm elections: AI data centers 'fully rejected' in 45 ballot measures
行业Tom's Hardware，2026-11-04（经 Hacker News AI 收录）Confidence MidNew topicAI score 78/100
Reports say 45 ballot measures related to AI data centers were overwhelmingly rejected by voters.
Take · Community opposition over siting and electricity use is quantified in ballot form for the first time. Cloud providers' capex timelines now have another variable: not whether they have enough money, but whether the land gets approved.
Uber completed only 4 trips fulfilled by AI Agents last quarter
UberThe Verge，2026-11-04（经 Hacker News AI 收录）Confidence MidNew topicAI score 74/100
Uber disclosed that only 4 trips were ordered by AI Agents last quarter.
Take · The most honest counter-evidence data point for Agent commercialization: between a demo and fulfillment lies an entire operations system (payments, risk control, exception handling). "It works" and "it scales" are two different things.
MetaTechCrunch，2026-11-04Confidence MidNew topicAI score 72/100
TechCrunch reports that Meta has completed its acquisition of AI coding startup Conductor.
Take · The final piece for big companies to complete AI coding remains “context and workflows,” not models. For small and medium-sized teams, the implication is that the exit path for vertical coding tools is narrowing.
Original ↗Apple cuts internal AI usage budget to reduce spending
AppleBloomberg，2026-11-04（经 Hacker News AI 收录）Confidence MidNew topicAI score 73/100
Bloomberg reports that Apple has lowered its internal AI usage budget as part of broader spending cuts.
Take · This comes on the same day as “the launch of the China edition of AI intelligence,” showing that leading companies are shifting AI spending from “universal quotas” to “allocation by output”—internal usage management itself has become a product requirement.
NDRC proposes “accelerating foundational and comprehensive AI legislation”: security runs through all stages of R&D, application, and governance.
国家发改委京报网／北京日报，2026-11-03Confidence HighNew topicAI score 76/100
The November 3 “15th Five-Year Plan” press conference proposed that security run through all stages of R&D, application, and governance, that compliance boundaries be clarified, and that attention be paid to the impact on employment structure.
Take · Moving from “special regulations” to “comprehensive legislation” means compliance shifts from a project-based arrangement to a recurring cost; companies’ self-inspection ledgers must upgrade from “whether there is a filing” to “whether they can self-prove the full process.”
04RESEARCH4 stories
OpenAIIT之家，2026-11-05（官方页：A thousand fewer paths to root）Confidence HighNew topicAI score 95/100
A five-stage pipeline scanned 5034 targets and about 2.8 million lines of code in 30 hours: 2308 deduplicated candidates → 1724 reproduced (74.7%) → 948 that allow local privilege escalation, API cost $1.2134 million (about $1279 per vulnerability); within 10 days, only 61 were fixed upstream (6.4%) and 43 received staged patches.
Take · Discovery is industrialized, but remediation still depends on human scheduling—cost per vulnerability fell from $3095 to $1279 (down about 58.7%), total output was 5.7x while total spending was only 2.4x. The pipeline’s most valuable output is not the number 948, but turning “can it all be fixed” into a priceable capacity problem.
Original ↗New study tests whether AI models “know what they are thinking”
Quanta MagazineQuanta Magazine，2026-11-04（经 Hacker News AI 收录）Confidence MidNew topicAI score 70/100
Quanta reports on a new study testing whether models can report their own internal states (i.e., “know their own thoughts”). The ledger only records the title and link; the body details were not reviewed, and this item cites no specific numbers or conclusions.
Take · If interpretability can only be measured externally and models cannot explain themselves, then all compliance designs for 'making the model self-certify' must be rewritten. Verification point: whether the experiment distinguishes 'reporting' from 'post-hoc fabrication'.
Can 'reverse game theory' rein in AI Agents: a public discussion of the mechanism design route
MIT Technology ReviewMIT Technology Review，2026-11-03（经 Hacker News AI 收录）Confidence MidNew topicAI score 68/100
MIT Tech Review published an article discussing the use of reverse game theory (reverse game theory/mechanism design) to constrain AI Agent behavior. The ledger only includes the title and link; body details were not reviewed.
Take · Moving the alignment problem from 'training models' to 'setting rules' aligns with this week's 'pricing by outcome' main theme: executable incentive structures are easier to implement than post-hoc audits.
OSFOSF 预印本，2026-11-05 台账收录New topicAI score 62/100
A PDF preprint titled AI as a Multi-Vector Civilizational Failure was published on OSF. The ledger only includes the title and link; the full text was not read.
Take · The “multi-vector failure” framework is worth borrowing: split risk by vector (supply, energy, cognition, governance); this is easier to turn into monitorable metrics than a single “existential risk” narrative.
Original ↗05PLAYBOOK5 stories
Security budgets switch pricing units: buy outcomes by the “hole,” then sell “patch throughput” separately
观点据 IT之家 2026-11-05 与官方页推得Confidence MidViewpointAI score 85/100
Human audits are priced per person-day and SaaS per seat; AI has now set a third price—1279 USD per exploitable privilege-escalation hole. If buying outcomes by the “hole” holds, security budget line items and vendor structure will both need rewriting.
Take · The biggest gap now is not scanning but “patch throughput”: only 61 of 948 holes were fixed in ten days. Selling weekly fix counts / staged patch counts / upstream regression rates is closer to buyer pain points than selling another scanner.
Build a “model migration ledger” for each customer: four columns are enough
观点据 IT之家 2026-11-05 openPangu 下线口径推得Confidence MidViewpointAI score 80/100
Four columns: what they’re using / next candidate / switch criteria / rollback point. What you’re selling is “version changes without downtime,” not tokens.
Take · openPangu went from end-of-sale to retirement in only 24 days, 混元 switched versions the same week, and Apple is rolling out in phases—model replacement cycles are now shorter than enterprise transformation cycles, turning a ledger from a good habit into a necessity.
Decision-layer bypass retrofit: strip enumerable judgments out of the generation pipeline.
观点据 36氪 2026-11-04 品类型态推得Confidence MidViewpointAI score 79/100
Only one thing: strip enumerable decisions—whether to block, recommend, or ban—out of the LLM generation pipeline and route them through discriminative models or rules; start with the two numbers: latency and unit price.
Take · Latency falls from seconds to hundreds of milliseconds, and per-call cost drops by one to two orders of magnitude. Once OpenAI Decisions API announces pricing, the only window left at this layer is channel and workflow lock-in—do not follow the crowd on valuation; follow on cost.
Earnings basis: when GAAP and non-GAAP diverge, look first at the order book and royalty structure.
观点据 IT之家 2026-11-05 Arm 财报口径推得Confidence MidViewpointAI score 74/100
In the same Arm earnings report: GAAP net income $130 million, -45% YoY (mainly due to a $46.6 million fair value loss on equity investments), non-GAAP net income $566 million, +36% YoY; technology licensing +50%, data center royalties doubled.
Take · Both directions holding at once is not an accounting trick but a difference in valuation anchors: look at GAAP for the current period, and at orders and royalties to judge business momentum. Miss this layer and you will mistake 'fair value fluctuations' for 'weakening demand'.
Treat 'remediation throughput' as the first industry metric to watch.
观点据 IT之家 2026-11-05 数据维度复算Confidence HighViewpointAI score 82/100
Discovery cost fell about 58.7% in four months (3095→1279 USD/vulnerability), while remediation rate stayed at 6.4% (61/948 in 10 days).
Take · If it remains below 15% after two weeks, open-source infrastructure will enter long-term degraded operation. This is the first public accounting of 'stronger capability, heavier liability'—worth tracking weekly more than any benchmark score.
Ecosystem Pulseself-updating
News600Skills46OpenHub68Reddit252026-09-02 → 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. AI security priced on a dual basis: 'per vulnerability + patch throughput'.Entry model: sell outcomes, not scanning—price by 'verified exploitable vulnerabilities' (benchmarked to the public anchor of 1279 USD/vulnerability), and sell 'patch throughput SLA' separately; this is the biggest current gap (only 61 of 948 vulnerabilities fixed in ten days). Advanced path: ① run a free comparable static audit of your own codebase to establish a baseline → ② per-vulnerability-priced verification report → ③ monthly patch-throughput managed service (weekly fixes / staged patches / upstream regression rate).
window From now through the next 6 months: price anchors have just been made public, and buyers have not yet formed price-comparison habits; note the offense-defense asymmetry—attackers are cheaper on the same pipeline, and the report must clearly define the boundaries of the 74.7% reproduction rate. 2. Model migration ledger (version lifecycle engineering)Entry model: build a four-column ledger for each customer—what they are using / next candidate / switch criteria / rollback point; what you sell is 'version changes without downtime,' not tokens. Advanced path: ① first use the openPangu-2.0 retirement (sales stop on 11/6, retirement on 11/30) as the first batch of migration drill samples → ② standardize the regression test case package → ③ add contract clauses for 'version lifecycle commitment + migration period ≥6 months + portable weights'.
window The two weeks starting November 6 are a demand thermometer: see how many customers are forced into emergency migration; after this wave, the terms will become an industry default. 3. Decision-layer bypass retrofit (generation → discrimination)Entry model: do only one thing—strip enumerable judgments such as 'block or not, push or not, ban or not' out of the large-model generation chain and route them through a discriminative model or rules; start by letting latency and unit price speak. Advanced path: ① select a chain with more than 100,000 daily calls for A/B → ② output the error rate and converted manual review volume → ③ package it as 'per thousand decisions' pricing (rather than per seat).
window 1–2 quarters: once OpenAI Decisions API announces pricing, the window for this layer will only be channel and workflow binding; don't chase valuation, chase cost.