01MODEL RELEASES4 stories
Zhipu GLM 5.2 open-sourced: 744 billion total parameters, 40 billion activated, native 1 million tokens context
智谱三亩地周报整理 2026-11-01（单源）Confidence MidNew topic
GLM 5.2 is open-sourced under the MIT license, with 744 billion total parameters in MoE, 40 billion activated (activation ratio about 5.4%), and native 1 million tokens context; it offers four quantization tiers BF16/FP8/Q4_K_M/UD-IQ2, can be configured with vLLM/SGLang, and allows enterprises fully private deployment.
Take · Shift procurement decision-making power from 'vendor API pricing' back to 'customer CIO budgets.' Caveat: parameter details come from a single source, are pending verification against official announcements, and have not been included in any third-party independent benchmark — no leaderboard conclusions can be drawn from this today.
DeepSeek discloses Super-agent architecture: the AI autonomously finds and reproduces defects even when no one @s it
DeepSeek证券时报 2026-11-01Confidence HighNew topic
The Harness team said in a 10-31 post that internal instances have been running long-term: even if no one @s the AI in group chats, the system autonomously identifies defects, reproduces them in the background, and seeks fixes.
Take · This is the first observable form of an 'always-on Agent': shifting from passive response to proactive inspection, with billing and responsibility boundaries changing accordingly. The original text is the team's own account and lacks success rate and false-positive rate readings.
Avartha says its open-weight speech model beats OpenAI/Google/xAI
AvarthaHacker News AI 官方源摘录 2026-11-02New topic
The official blog calls it 'the fastest open-weight frontier speech AI' and says it outperforms OpenAI, Google and xAI in comparisons; the in-depth version labels this as a vendor claim with no third-party evaluation.
Take · Speech-side open-weight competition has just begun; if valid, it would lower unit prices for real-time speech agents; but until evaluation sets and concurrency conditions are disclosed, it can only be treated as a candidate, not a conclusion.
NVIDIA × MIT claim 5-second 2K video generation cut to 22.45 seconds, with Open-H3 3.9× faster
英伟达 × MIT燚燃情报转述（原始出处与日期未标）New topic
The in-depth version notes in one sentence that it cuts 5-second 2K video generation from the original baseline to 22.45 seconds, with Open-H3 3.9× faster; original source and date are not indicated and remain to be verified.
Take · If the readings hold, generation costs for short dramas and ad creatives would fall another notch; but with no comparison baseline, it is not yet included in the capacity ledger.
02PRODUCT LAUNCHES4 stories
Qwen integrated into Apple Intelligence: testing begins with the iOS developer Beta on 11-04; official version to launch in China in December
阿里 × Apple科创板日报 2026-11-01Confidence HighNew topic
Chinese users can use Qwen's image-text understanding and generation in Siri, Writing Tools, and Visual Intelligence; testing with the iOS developer Beta on November 4, and launching for users in China with the official iOS 27.3 release in December.
Take · This is the first time a domestic large model reaches iPhone users through a system-level entry point, moving Agent distribution from inside apps to the system layer. Availability and permission boundaries will depend on the first round of developer testing on 11-04.
Tencent Buddy One/WorkBuddy scale figures: half of the top 500 are using it, 200,000 Skills, 9 billion tasks
腾讯36氪·智能涌现 2026-11-01Confidence MidNew topic
Buddy One connects Tencent Docs/Meeting/ima/cloud drive and interoperates accounts and AI credits; WorkBuddy's open architecture supports enterprise-developed and third-party agents, and the company says 50% of China's top 500 already use it, the open platform has 200,000 Skills, cumulative task submissions exceed 9 billion, and commercialization has no numerical target.
Take · This contrasts with the walled gardens of 155 apps: outside the wall, acting on users' behalf is refused; inside, Agents are invited into workflows. The figures are the vendor's own account and can serve as a scale reference, not as market share.
Meta Muse hits 1040 万次 downloads in under two months; domestic Today AI, Manus Cue, Rabbit OS3 follow suit
Meta36氪·智能涌现 2026-11-01Confidence MidNew topic
Muse reaches 1040 万次 cumulative downloads in under two months; domestic Today AI, Manus Cue, Rabbit OS3 accelerate their entry
Take · The user-education phase for general-purpose personal assistants is over, and competition is shifting to unit economics: another same-day estimate put the monthly cost per user for general-purpose proactive assistants at at least 1000 元 (single-party estimate, methodology undisclosed), while subscription prices are 20—30 元
Nonce open-sources 'least-privilege Gmail access' for AI agents
NonceHacker News AI 2026-11-02Confidence MidNew topic
Show HN launches Nonce: open-source, least-privilege Gmail access for AI agents
Take · Agent bottlenecks are shifting from model capability to credentials and permissions. The authorization layer already has open-source supply; what’s missing is standards and auditing—another 'declaration vs. execution' gap beyond this week’s robots.txt exit signal
03INDUSTRY7 stories
Google admits for the first time that AI Mode is cannibalizing search ads: Alphabet falls about 15% in one day, with about 5000 亿美元 in market value wiped out
Alphabet／GoogleLatent Space／AINews 2026-10-31（跌幅单源，待财报核对）Confidence MidNew topic
First publicly acknowledges that the generative search interface is eroding core ad revenue; Alphabet fell about 15% in one day and lost about $500 billion in market value, its worst single day since its 2004 listing; ads alongside AI Mode have rolled out globally, but unit revenue remains below that of the mature search business.
Take · The chain is straightforward: AI Mode → longer dwell time → fewer outbound link clicks → lower display inventory → lower unit monetization. If it holds, AI search's revenue structure will shift from advertising to 'subscriptions + Agent revenue sharing,' and the intermediary layers that rely on sharing search traffic revenue will be hurt as well. The decline figure is from a single source and needs to be checked against earnings reports.
Anthropic files + becomes SpaceX's largest customer: $120.5 billion long-term compute contract in exchange for a $1.5–2 trillion valuation
Anthropic／SpaceX中国新闻网／财联社 2026-11-01Confidence HighNew topic
Has filed a prospectus and plans to list on Nasdaq in mid-November; the market expects it to raise up to $100 billion at a $1.5 trillion–$2 trillion valuation; it also committed to spend up to $120.5 billion on compute through 2029 (about $3.1 billion per month starting December), while about 34.5% of SpaceX's revenue last quarter came from this same customer.
Take · Supplier, customer, and investor rolled into one: revenue moves around the closed loop and back to the compute bill. On the way up it is a flywheel; on the way down it is an amplifier—this is the single-point structural risk to watch most closely in this issue.
Inference data volume exceeds training data volume for the first time: 175 trillion tokens per day on average, a thousandfold increase in two years
赛迪研究院央视新闻／IT之家 2026-11-01Confidence HighNew topic
As of June 2026, daily average token calls exceeded 175 trillion, more than 1000 times the level at the start of 2024; supported by more than 500 excellence-level smart factories and more than 1000 scenarios across 20 industries.
Take · AI spending has shifted from a one-off capital item to a recurring operating bill—this is the core evidence that 'the unit of account has changed.' Note on basis: another secondary citation says 'National Data Administration: 500 trillion per day,' a difference of nearly 3 times; this issue does not combine or convert them and cites only the traceable one.
Four vendors unify data-scraping opt-out standard: one declaration applies to multiple companies, with implementation timing differing by six months to a year.
OpenAI／谷歌／微软／AnthropicIT之家 2026-11-01Confidence HighNew topic
Four companies are adopting the unified opt-out signal Anthropic proposed to the IETF and integrated into robots.txt; content providers can make a one-time declaration to multiple AI companies on whether their content may be used for training. Anthropic has adopted it; OpenAI will implement within months; Google and Microsoft next year; Meta supports it but has no timetable.
Take · Corpus property rights now have a machine-readable carrier for the first time. But the first version only addresses declaration, not enforcement—declaration is zero-cost; verifiability is the high cost; what really determines the landscape is who first delivers crawl auditing and settlement channels.
Capital's three-way pricing divergence: OpenAI voluntarily cuts 100 亿, DeepSeek expands to 1055 亿, 可灵 receives 6.45 亿 less.
OpenAI／DeepSeek／可灵FT／证券时报、新浪财经、钱江晚报·潮新闻 2026-11-01Confidence HighNew topic
Three contrasting readings on the same day: OpenAI cut its valuation from 1.4 万亿 to 1.25 万亿 and its financing from 300 亿 to about 200 亿; DeepSeek expanded its financing to 1055 亿元, with a post-money valuation of 6155 亿元 (nearly 30 industrial shareholders); 可灵 actually received 198.02 亿元, 6.45 亿 less than the 204.47 亿 subscription cap, and two originally planned investors withdrew.
Take · No uniform conclusion that 'capital has cooled' is drawn; instead, it is judged as tiered pricing by auditable cash flow: one has long-term contracts as a backstop, one lacks cash-flow backing, and one is a marginal subscription retreat. The three evidence chains are independent of each other, so they are merged into one item rather than each telling its own story.
Anthropic prospectus financial metrics: 85% gross margin, $1 billion Q2 operating profit, 48% of revenue via cloud resale
Anthropic明见、坤舆转述 S-1 2026-11-01（未见原件）Confidence MidNew topic
Secondhand accounts say its gross margin is 85%, Q2 single-quarter operating profit is USD 1 billion, and 48% of revenue is resold through cloud vendors.
Take · If the 85% gross margin is true, it is selling a software premium rather than compute, conflicting with the popular narrative that 'AI is a heavy-asset business.' Note: the original S-1 was not seen; all three figures are secondhand.
2026 Wuzhen Summit opens today: 27 sub-forums; the theme changes from 'interconnectivity' to 'open source, openness, co-building and sharing'.
世界互联网大会新华网／IT之家／中国日报网 2026-11-02Confidence HighNew topic
The event runs November 2–5 under the theme “Open Source, Openness, Co-building and Sharing,” with 27 sub-forums; key topics are the global AI open-source ecosystem, agent governance, and computing infrastructure. The “Light of the Internet” expo is held concurrently.
Take · The yardstick shifts from “whose model is strongest” to “whose ecosystem others can use.” Verifiable outputs are downloadable weights, accessible protocols, and compute quotas that can be applied for—during the event, watch whether these three have implementation documents.
04RESEARCH2 stories
“Broken on Arrival”: a spot check of 305 GGUF quantization tags finds the registry generally lacks functional test records.
HF Daily PapersHF Daily Papers 2026-10-25（坤舆情报转述）Confidence MidNew topic
The paper empirically tests 305 GGUF quantization tags in Ollama’s official library and finds the registry generally has no functional test records—“download and use” does not hold.
Take · Today’s only hard signal on the benchmark side: it turns “quantization level” from a marketing parameter into an artifact attribute requiring acceptance. The direct implication for private delivery is that quantization packages must include smoke tests and quality inflection points, not just report parameter counts.
Supply-chain poisoning path: tampering with core infrastructure is enough to backdoor AI models.
EclypsiumEclypsium 研究／Cybersecurity Dive 2026-11-02Confidence MidNew topic
Cybersecurity Dive reports Eclypsium research: attackers can poison AI models by tampering with core infrastructure.
Take · Move model security down from the weights and prompt layer to the hardware and supply chain layer. For private deployment delivery teams, acceptance checklists need hardened-component and device-integrity items—this is one of the black swans in this issue's risk table.
05PLAYBOOK6 stories
Private deployment delivery package: VRAM configuration sheet + tokens/s and quality inflection point stress tests for each tier + acceptance checklist
尚吉深度版 Top3 行动建议 2026-11-02Confidence MidNew topic
For government and enterprise customers whose 'data cannot leave the premises,' package GLM 5.2's four quantization tiers into a deployment package: VRAM configuration sheet, tokens/s and quality inflection point stress test reports for each tier, and acceptance checklist. The path is to first build 1 benchmark customer (tiers runnable on a single machine with no more than 8 GPUs) → solidify into a three-piece SKU set → convert to annual O&M subscription (upgrades, quantization recalibration, expansion review).
Take · The weights have just been released and customers are evaluating; whoever first produces verifiable stress-test numbers sets the pricing line. The time window is 6—8 weeks from now.
Narrow-scenario Agent: build a single point inside the walled garden; don't sell an all-purpose assistant
霆钧／惊蛰／艮山深度版·商机维度 2026-11-02Confidence MidNew topic
Target cross-App but low-adversarial entry points such as expense reconciliation, inspection work orders, and recruiting follow-up, using the customer’s own allowlisted accounts and pricing by work hours saved. Path: a 3-day verifiable narrow-scope pilot → reach a 95%+ success rate before discussing expansion → replicate horizontally within the same industry → wait for the first super App to open delegated operations, then scale up.
Take · The 155-item rejection list is still expanding, and the unit economics of the general route are negative (measured estimates put monthly cost per user at at least 1000 yuan vs. a 20—30 yuan subscription); narrow entry points must survive first. Time window: 6—12 months.
Content authorization credential: upgrade from “I declared an opt-out” to “I can prove who failed to comply”
乾机／巽风／明见深度版·商机维度 2026-11-02Confidence MidNew topic
For content organizations and corporate sites, build “robots.txt opt-out declaration + crawl logs + content fingerprints + licensing price list” into evidence-ready, reconcilable quarterly monitoring. Path: free one-time authorization blind-spot audit → quarterly monitoring subscription (UA/fingerprint reconciliation) → dispute evidence and negotiation support, while tracking the IETF call for comments.
Take · Robot standards only govern “declaration,” not “execution”; the execution layer is the gap. Standards implementation lags by six months to a year, and this window is content providers’ only bargaining period. Time window: 4—6 weeks.
Provision capacity at p95, not at the average: training looks at peak FLOPS; inference looks at the arrival process.
乾机／厚坤深度版·前瞻维度 2026-11-02Confidence MidNew topic
Inference surpassing training (daily average 175 trillion tokens) and Wuzhen's 100,000-card engineering point in the same direction: training looks at peak FLOPS, inference at the request arrival process (Poisson arrivals plus daily cycle), and capacity must be provisioned to p95 rather than the mean.
Take · Otherwise there will be queues by day and idle capacity by night—this is the first arithmetic question of the operations-bill era, and the item most likely to be asked by customers in a private-deployment delivery package.
Handling conflicting measurement bases: 175 trillion and 500 trillion are not added together or converted; only the traceable figure is accepted.
本期口径说明深度版·数据维度 2026-11-02Confidence MidProgress
Both under the same June 2026 basis: 赛迪 gives a daily average of 175 trillion tokens, while another secondhand citation says National Data Administration daily Token calls exceed 500 trillion, a nearly 3x difference; neither statistical scope is public. This issue adopts 赛迪—the releasing institution and release date are traceable, and it is today's multi-source item on the same topic; 500 trillion is recorded as pending verification.
Take · This is the report's reproducible convention: conflicting readings are listed side by side with the reasons for selection; no weighted averaging, no unit conversion, so readers can verify back to the source themselves.
Always-on Agent meets per-seat pricing: UI value sinks, data pipeline and permissions value rises
惊蛰深度版·前瞻维度 2026-11-02Confidence MidNew topic
Agents don’t need UI: interface value declines while data pipelines and permissions gain value, loosening the underlying logic of SaaS per-seat billing.
Take · Mutually causal with “inference over training”: pricing units shift from person×month to calls×tasks; quotes must accommodate both seat-based and metered bases, or they won’t reconcile at renewal.
Ecosystem Pulseself-updating
News600Skills72OpenHub68Reddit252026-09-02 → 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. Private-deployment delivery package (GLM 5.2, four-tier quantization)For government and enterprise customers whose “data cannot leave the premises,” sell deployment packages rather than models: GPU memory configuration sheet + stress-test report on tokens/s and quality inflection point by tier + acceptance checklist. Path: first make 1 benchmark customer (tier runnable on a single machine within 8 GPUs) → solidify into a three-piece SKU set → convert to annual operations subscription (upgrades, quantization recalibration, expansion review).
window Starting now, 6—8 weeks: weights have just been released and customers are evaluating; whoever first produces verifiable stress-test figures sets the pricing line. 2. Narrow-scenario Agent (single point within the wall)For entry points like expense reimbursement reconciliation / inspection work orders / recruitment follow-up—“cross-App but low-adversarial”—use customers’ own whitelisted accounts, quote by labor hours saved, and do not sell an all-purpose assistant. Path: a 3-day verifiable narrow-scope pilot → reach 95%+ success rate before discussing expansion → replicate horizontally within the same industry → wait for the first super App to open delegated operations, then scale.
window 6—12 months: the 155-product list is still expanding; unit economics of the general-purpose route are negative (single-party estimate: monthly cost at least 1000 yuan vs 20—30 yuan subscription); niche players survive first 3. Content licensing credentials and monitoringFor content institutions/corporate sites, build 'robots.txt opt-out declaration + crawl logs + content fingerprints + licensing price list,' upgrading 'I declared opt-out' to 'I can prove who didn't comply.' Path: free one-time licensing blind-spot check → quarterly monitoring subscription (UA/fingerprint reconciliation) → dispute evidence and negotiation support; simultaneously follow the IETF call for comments.
window Next 4—6 weeks: standards implementation timelines differ by six months to a year; this window is the content side's only bargaining period