01MODEL RELEASES4 stories
DeepSeek discloses DeepInfer production parameters: 155T tokens processed daily, 96% cache hit rate
DeepSeek官方+多源Confidence HighProgress · Day 2
Discloses the production-grade inference engine and token codec components supporting V4.1 Flash: 155T tokens processed daily per cluster, 96% input cache hit rate, 400ms first token, 230 tokens/s generation (机器之心, 10-22).
Take · The previous issue reported open-source status and throughput multiples; this issue gives four verifiable engineering parameters; “82% gross margin” and “81-day payback” are community estimates back-calculated from these parameters, the company has never disclosed gross margin, and any citation must be labeled as an estimate.
OpenAI opens access to the Jev-competitor decision model and simultaneously launches the Decisions API
OpenAI官方+外媒Confidence High
Decision models are now broadly available, alongside a Decisions API that makes applications more responsive (TechCrunch AI, OpenAI official, 10-22).
Take · Same day as 蚂蚁 LingDT, the move points to specialization in the decision domain: general models handle understanding, decision models handle convergence, and procurement is beginning tiered model selection by task domain.
蚂蚁数科 LingDT-Decision-1.0: 4B/27B dual tiers specialize in structured decision-making
蚂蚁数科厂商口径·量子位Confidence MidVendor-selected benchmarks
Officials say its composite score on 23 open-source benchmarks surpasses Microsoft-Decision-1 and others, with token consumption down over 90% vs general large models and time efficiency improved about 10x (量子位, 10-22).
Take · The 23 benchmarks were vendor-selected; weighting and task distribution were not disclosed. Before procurement, use your own 20~50 real decision cases for a single-blind test instead of ordering directly based on composite scores.
Flint adds benchmark readings: SWE-bench Verified 74.2%, Terminal-Bench 2.1 at 54.9%
Recursive自报读数Confidence MidPreviously reported · readings added
39.8B MoE, activates only 2.5B per token, Apache 2.0, runs on MacBook Pro (机器之心 / 量子位, 10-22).
Take · The previous issue covered the 'runs on a 48GB Mac' usability angle; this issue gives the score: the first public reading of an on-device coding agent entering the 'usable' range.
02PRODUCT5 stories
Google ACE gives each agent a hardware-isolated micro-VM
Google读懂AI时代日报引谷歌官方Confidence Mid
Agent Compute Engine uses Titanium chips for hardware isolation and automatically pauses when idle; with the open-source K8s runtime Agent Substrate, it achieves 10x single-machine density and <500ms pause/resume for 200,000 instances, and is already used for Gemini (10-23).
Take · Isolation and density for agent runtimes are now platform components—enterprise procurement of always-on agents now has infrastructure to discuss, and the security boundary has moved from the prompt layer down to the virtualization layer.
Canva releases its first image generation model
Canva外媒Confidence Mid
Design tools shift from 'integrating third-party models' to in-house image generation (TechCrunch AI, 10-22).
Take · The image-generation race gains another distributor with 200 million-level design users: beyond capability, distribution becomes the second variable in this race.
Uber rolls out ChatGPT Work to 30,000 employees company-wide
Uber官方+外媒Confidence High
Enterprise side moves from 'partial-team pilots' to company-wide rollout (TechCrunch AI, OpenAI official blog, 10-22).
Take · This week's most concrete procurement-side signal: not a launch event, but public disclosure of seat counts; only after company-wide rollout will renewal and usage curves reveal true retention.
1Password lets AI agents log in for you without seeing passwords
1Password外媒Confidence High
Issues agents 'usable credential capabilities' rather than 'the credentials themselves' (TechCrunch AI, 10-22).
Take · Directly targets the credential-management gap in the agent era; a prerequisite for agent-operated scenarios and a control point for the blast radius of unauthorized-access incidents.
Spotify's AI bet: becoming 'the music service agents will choose'
Spotify外媒Confidence Mid
Content platforms begin designing interfaces and authorization structures for "agent invocation" (TechCrunch AI, 10-22).
Take · The distribution entry point shifts from "users opening the App" to "agents calling APIs"; copyright and authorization structures must be rewritten—a transformation all content platforms will face.
03INDUSTRY7 stories
OpenAI plans to raise up to $30 billion at a $1.25 trillion valuation, has confidentially filed for an IPO.
OpenAICNBC 单源转载Confidence MidSingle-source pending verification
UAE investors and Coatue lead; target is to sign "by next weekend"; positioned as a bridge arrangement before a potential 2027 IPO; the company has confidentially filed for an IPO (CNBC, 10-23).
Take · An industry-wide primary-market pricing anchor: if the anchor shifts, long-term compute contracts, talent costs, and competitor valuations will all be repriced. CNBC single-source only, not officially confirmed, and differs from the 1.4 trillion figure reported on 10-21; this edition uses 1.25 trillion as the basis.
Amazon's "dark GPU": $5 billion in GB300 stranded in transit warehouses due to power shortage.
亚马逊财联社·内部邮件Confidence MidSingle-source internal email
Internal emails show it launched an "early rack acceptance" plan and plans about $3 billion per month in additional hardware procurement in the future, but multiple AWS sites have insufficient power capacity, leaving many GPUs stranded in transit warehouses unable to enter production (Cailian Press, 10-22).
Take · Stuck in warehouses but still depreciating; the bottleneck has shifted from chips to power and grid connection. On the same day, Blackstone's president said AI demand is growing 'parabolically'—the pace of supply-side delivery is the real constraint on valuations.
China's first fully AI-generated theatrical feature film, 'Sanxingdui: Future Past,' opens today
博纳影业 / 生数科技多源Confidence High
Produced by Bona Film Group and granted public screening approval; developed over two years, generated more than 1.2 million pre-production images, runtime about 100 minutes; Vidu by Shengshu Technology provided lead technical support (Dudong AI Era Daily / toolin.ai, 10-23).
Take · AI content is receiving its first public vote via box office and word of mouth; tonight, build a comparison table of 'opening-day box office + Douban curve + social media buzzwords,' and label 'AI-generated' separately as a variable.
Unitree launches humanoid robot H2-X 'Kung Fu Legend': official claims maximum battery life of about 8 hours
宇树36氪 / 金十数据Confidence Mid
Wang Xingxing said around the same time that 'aligning model inputs and outputs with real robots' is the most central challenge for embodied intelligence to pass the 'ChatGPT moment' (10-22).
Take · Founder's self-described shortcomings are more informative than parameters: open-world embodied AI remains stuck on alignment, while near-term deliverables are still domains with standard answers.
The General Offices of the CPC Central Committee and the State Council require full implementation of the 'AI+' initiative, with eight departments simultaneously deploying end-to-end implementation.
中办、国办 / 工信部等八部门金十数据Confidence High
Covering R&D, design simulation, pilot-scale validation, and production operations (金十数据, 10-22).
Take · Policy has shifted from 'encouragement' to 'full implementation'; procurement budgets and acceptance criteria will be implemented before technological maturity—this is the most certain source of orders this year.
ConnectWise was breached; attackers reportedly surveilled thousands of companies.
ConnectWise外媒Confidence Mid
IT service providers become a springboard for supply chain attacks (TechCrunch AI, 10-22).
Take · The MSP ecosystem is the most fragile link in the AI era; the same day, Google Threat Intelligence released a vulnerability discovery/exploitation trends report, with AI accelerating both offense and defense.
Anthropic donated Agent Skills to the Linux Foundation's AAIF, under the same umbrella as MCP.
Linux 基金会 AAIF官方+多源Confidence High
70+ agent platforms have integrated it since December 2025; after the donation, existing skills remain usable and the license is unchanged (ClaudeDevs / Simon Willison, 10-23).
Take · The protocol and skills layers are neutral; no single vendor can lock in customers, and cross-platform portability of skill assets is rising. But before governance details (seats, proposal rights, voting weights) are published, it is inadvisable to make selection commitments based on 'already neutral.'
04RESEARCH3 stories
Microsoft's MAI-DxO makes the cover of Science: 83.2% accuracy on difficult cases.
微软The Verge 引论文·《Science》封面Confidence High
On 300+ NEJM clinical pathology cases + 300+ emergency cases, diagnostic accuracy was nearly twice that of human doctors, with lower cost and fewer required tests; the code is open-sourced (2026-10-23).
Take · The only peer-reviewed 'AI beats experts' evidence in this issue with a controlled sample size; however, performance within a single disease spectrum and a single hospital system does not equal cross-institution reproducibility—until replicated, treat it only as a directional signal.
Flint-interactive: a language for generating interactive data visualizations
微软研究院官方博客Confidence Mid
Used to generate interactive data visualizations (Microsoft Research, 2026-10-23).
Take · Same name as Flint, the open-source model from Yuandong Tian's Recursive, but a different thing; do not conflate when citing: one is a visualization language, the other a programming model.
arXivHacker News AI 条目Title only; full text not read
Paper ID arXiv 2610.27251, surfaced that day via a Hacker News AI item (10-23); full text not read locally; conclusions and sample size pending verification.
Take · Aligned with the day's 'scientific commons' issue: boundary questions after AI collects public research corpora are worth adding to the topic pool, but no conclusions should be cited before the full text is read.
Original ↗05INSIGHT6 stories
Agent-level cost observability remains a gap: the three bills are the easiest to get.
坎流判断·推断Confidence Mid
Today, no company can answer 'how much this agent spent this month and at which step' (Kanliu, 10-23 judgment).
Take · The three bills—vCPU·秒 + token + tool calls—are the easiest entry point for enterprise procurement: first provide the bill, then negotiate price.
Putting cache hit rate into the SLA is the most direct move after DeepInfer's parameters were disclosed.
巽风建议Confidence Mid
Refund pro rata when the hit rate falls below the threshold, and negotiate tiered pricing (Xunfeng, 10-23 suggestion).
Take · At official list prices (1 元/M for misses, 0.02 元/M for hits), blended input cost at a 96% hit rate is about 0.06 元/M—'gross margin' is determined by hit rate × cluster utilization, not model capability.
Bottleneck shift: from 'chips' to 'power and grid connection'
惊蛰、乾机推断Confidence Mid
Chip lead times of 12~18 months allow capacity expansion, while substations and grid-connection permits take 3~5 years and cannot be accelerated (10-23).
Take · The next inflection point in compute pricing is more likely to come from power delivery; the observation point should shift from 'GPU shipments' to 'grid-connected capacity + idle rate'.
Billing units shift from token to 'agent·machine hours'
巽风推断Confidence Mid
Once ACE makes the anchor 'how many agent·hours are occupied,' token price wars have less explanatory power for total cost (10-23).
Take · Bargaining focus shifts to pause strategy, cold-start duration, and density commitments—negotiation terms must change along with the billing unit.
Verifiable domains are harvested first: the criterion is domain verifiability, not demo-video completeness
慧师规律Confidence Mid
Converged domains with standard answers—diagnosis, code, and math—have crossed the threshold (MAI-DxO, Flint); open-world embodied intelligence remains stuck on alignment (慧师, 10-23).
Take · The most reliable pattern this week: first land in domains with standard answers, then discuss general capabilities.
Open source does not mean saving money: revenue-trigger clauses put compliance friction back on the books
巽风推断Confidence Mid
Meta Muse Spark Open (816B total parameters / 42B activated) has limited commercial terms: if revenue exceeds $20 million for 12 consecutive months and MaaS is offered externally, a separate agreement is required (至顶网, 10-22).
Take · Before productization, the three definitions—group basis / entity basis, whether private deployment is included, and rolling 12 months vs fiscal year—must be written into an internal explanatory document; otherwise, dispute costs are ultimately borne by the product side.
Ecosystem Pulseself-updating
News600Skills25OpenHub65Reddit252026-09-02 → 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. Calculate “cost per task” for our own scenarios by cache hit rate tier.First build a one-week cost baseline (token × hit rate × utilization), price high-hit scenarios (code completion, customer service, tickets) separately; route simple structured tasks to 4B-class specialized models and keep complex judgments on flagship models; negotiate tiered pricing and a “cache hit rate written into SLA” clause externally, and align depreciation and power amortization bases with finance.
window Starting immediately, 2~4 weeks (DeepInfer has made cost parameters public; customers will first ask “what is your hit rate?”) 2. Put agent assets in a neutral layer and add an agent-level three-item bill.Turn internal methodologies, workstation cards, and QC criteria into portable skill packages compliant with AAIF standards, not tied to a single platform SDK; pick 1 high-frequency scenario for a cross-platform migration test (whether the same skill behaves consistently across three runtimes) → build a regression test set → use “portable + measurable” as the enterprise procurement entry point; before launching persistent agents, set rollback boundaries and action whitelists.
window 3~6 months (ACE is still in preview and AAIF governance details have not been announced; a low-cost, long-payback option) 3. Build the market validation table for “Sanxingdui: Future Past” tonight.Build a monitoring table for “opening-day box office + Douban curve + social media hot words,” mark “AI-generated” as a separate variable, and compare with live-action films in the same release window; use three reading points—3-day / 7-day / 14-day—to distinguish “novelty-driven opening day” from “word-of-mouth long tail”; feed conclusions back to the in-house content line—talk less about generation capability and more about delivery stability (calculate variance using the worst minute).
window 2 weeks from now (the first weekend will set the tone; industry marketing messaging will shift accordingly within two weeks)