01MODEL RELEASES4 stories
谷歌官方博客+DeepMind 频道Confidence HighOfficial first-hand
Google's official blog released Gemini 4 Argon, with DeepMind simultaneously launching multiple demos: 500-page chemistry patent analysis, Chromium ultra-large-scale codebase maintenance, and network attack/defense protection (YouTube·Google DeepMind).
Take · The flagship lands, and the demos point directly at three engineering scenarios—long documents, ultra-large codebases, and security—the day's strongest first-hand source at the model layer.
Original ↗VISTA from Kaiming He's team: visual Harness lets Claude Opus 5.0 pass the ARC-AGI-3 public set with a perfect score
何恺明团队arXiv+36氪/量子位Confidence HighNewcomer
Minimalist visual Harness—using only raw screenshot input + lossless visual memory + model-driven visual self-check—gets Opus 5.0 through all 25 public games in ARC-AGI-3 (a perfect 100 on relative human action efficiency; 57.4% fewer actions than human novices).
Take · The first pure-vision solution to achieve this, with no program synthesis needed—the bottleneck for top models is not reasoning but the perception and memory interface, and the 'base + orchestration' division-of-labor paradigm holds.
Zhipu releases legal vertical model GLM-Legal
智谱z.ai 官方博客+华泰研报Confidence HighNewcomer
Built on GLM-5.3-Flash, it uses 'domain mid-training + reinforcement learning + Agent Harness co-optimization'; Harvey LAB all-task pass rate 19.2% (+12.5pct vs base, first among open-weight models but not overall), LegalBench 86.2, and Enterprise Beta API now open.
Take · Vertical-scenario cost-effectiveness won't be solved by general-purpose large models—differentiated moats for industry models are taking shape.
Fish Audio open-sources directable TTS model Drama 3
Fish Audio深度版Confidence MidOne-sentence summary
Supports word-level emotion control, 14-platform distribution, and Day-0 inference support.
Take · Open-source speech models continue to fill in 'adjustable emotion' capabilities, and options for dubbing workflows keep widening.
02PRODUCT5 stories
Anthropic drops git, builds in-house version control system Scion
Anthropic官方工程博客Confidence MidSingle-source
Engineering blog reveals: coding Agent drove daily merge volume in the monorepo from 2500 to 30000+ in four months; team took 5 months to migrate to in-house version control system Scion.
Take · First top-tier AI company publicly declares 'git is not enough'—as Agents become primary commit producers, every layer of VCS/CI/Code Review may be rebuilt.
ReadMe releases Restless: API auto-iterates with Agent calls
ReadMeHeadsUpAIConfidence MidEarly signal
Reportedly, Restless lets APIs automatically improve based on real Agent calls ('Agent Recovery' error responses, Agent questions automatically converted into documentation improvements).
Take · MCP solves 'how Agents call APIs'; Restless answers 'how APIs change for Agents'—APIs shift from static contracts to adaptive interfaces.
Token hits e-commerce shelves: MiniMax opens a Tmall flagship store, Tmall launches 'AI Space Station' for aggregated top-ups
MiniMax / 天猫每经/BigGoConfidence HighNewcomer
M Plan subscription package hits shelves to ride Double 11 traffic (first order 50% off until 10-14), with Kimi, StepFun and others in talks to join.
Take · This marks a landmark moment for domestic large models' 'Token commoditization + e-commerce channel distribution,' opening an e-commerce peripheral window for price comparison/proxy top-up/model-selection guides.
Salesforce官方Confidence MidVendor claims
Official blog post 'From Radar to Reasoning: Why Enterprise AI Finally Watches Your Business in Real Time' goes live, with a Casey help-agent demo released alongside.
Take · Enterprise AI is shifting from 'post-hoc analysis' to 'real-time operational monitoring,' in step with agents moving from demos to procurement lists.
Original ↗OpenAI官方 News/BlogConfidence MidVendor's own account
OpenAI's official page posts the Higgsfield case study, saying it boosts AI video duration from 1 minute to over 10 minutes.
Take · The duration bottleneck in long-video generation has been directly broken by a vendor case study, expanding the usable radius of creator toolchains.
Original ↗03INDUSTRY11 stories
IMF GFSR: AI-driven market pullback listed among three major systemic risks to global financial stability
IMFReuters/Yahoo 多源Confidence HighOfficial report
IMF 10-13 released the October issue of the Global Financial Stability Report, listing an AI market pullback alongside surging sovereign bond yields and severe cyber incidents as three major near-term systemic risks, and singled out 'AI circular financing' as a main transmission channel; downside scenario modeling: a sharp plunge in AI investment would cause a 1.1% global GDP loss and a 1.8% U.S. output loss by 2028.
Take · For the first time, AI capex has been modeled by the IMF as a macro variable on par with sovereign debt cycles—macro institutions have formally launched it as a project.
Goldman Sachs three-signal framework: two lights already on; capex growth to peak in 2026Q3, entering validation window
高盛华尔街见闻/鉅亨/ZAKER 多源转引Confidence HighSell-side light turns on
Three falsifiable signals (stock price-earnings divergence, credit repricing for AI borrowers, actual slowdown in capex growth); the first two have been triggered; forecasts hyperscaler capex growth to peak in 2026Q3 (peak +116% YoY) and fall to just over 20% by end-2027. 10 core AI stocks account for about 40% of S&P 500 market cap (above the 2000 dot-com bubble peak), while US equities' order-book depth per $1 billion traded has shrunk about 93% from 2017.
Take · All three key figures—116%/40%/93%—passed multi-source verification; the 'top' has moved from opinion to a falsifiable framework.
Tiered pricing in AI bond market: Tencent's $5.7 billion bond trades tight to Treasuries; CoreWeave's unsecured debt spread breaks 1000bp for first time.
腾讯 / CoreWeaveReuters（腾讯）Confidence MidCoreWeave pending verification
Tencent launches about $5.7 billion dual-currency bond issuance ($1.8 billion dollar bonds + RMB26 billion dim sum bonds; orders exceeded $23 billion; 10-year dollar bond spread about 90bp); meanwhile CoreWeave unsecured bond yield rose to 15.1%, spread reached 1000bp for the first time, and its H100 rental price fell about 12% in one month (single source). 10-year Treasury yield rose to 5.24%; cash/capex coverage at the top five clouds has fallen from 1.79x to 1.06x.
Take · Funding bottlenecks are moving from 'executive verbal warnings' into tiered pricing in the credit market—good names trade close to Treasuries, while heavy leverage breaks 1,000bp.
Anthropic IPO vetting: $2 trillion valuation target, first Q3 disclosure this Wednesday
AnthropicCNBC/华尔街见闻/Reuters 多源Confidence MidBasis correction
Planned IPO valuation target about $2 trillion (Bloomberg range $1.8-2 trillion; if completed, would surpass SpaceX as largest IPO ever), planned listing in November; this Wednesday (10-14) first disclosure of Q3 financial data to institutional investors. Clarification: the rumored 'cumulative loss of 42 billion' is actually a single-year net loss for 2025, of which about 34 billion is non-cash accounting losses from valuation changes in financing instruments, with actual operating loss about $8.06 billion.
Take · S-1 not yet public, but disclosure basis will become a new pricing anchor for the AI sector—the 'OpenAI 50 billion revenue blunder' has shown that a single metric can crash chip stocks.
On-device AI chip profits materialize: 星宸科技 Q3 report pre-announces 663%~698% profit growth
星宸科技上交所业绩预告+上证报Confidence HighProgress
First three quarters revenue expected at 45.38-46.08 亿元 (+109%~113%), net profit attributable to shareholders 15.43-16.13 亿元 (+663%~698%, earnings forecast not final); standalone Q3 non-recurring-adjusted net profit +34%~48% QoQ.
Take · AI boom is spreading from cloud to endpoints—星宸 is a newcomer in on-device/edge AI SoC; risk: Q3 net profit QoQ growth is markedly lower than YoY; watch for profit margins peaking after competition intensifies.
Order diffusion: compute boom spreads from core to supporting segments, visibility to 2027
产业链多家上市公司公告/21财经/中央社Confidence HighAnnouncement endorsement
JPT has cumulative USD 163 million in overseas optical interconnect orders (about RMB 1.11 billion); VeriSilicon signed RMB 15.142 billion in new orders this year (about 90% AI compute-related); Jingzhida has RMB 1.576 billion in semiconductor test equipment orders; AMEC has a backlog over RMB 18 billion, with deliveries scheduled through 1H 2027; Taiwan packaging/testing firm Sigurd will spend NT$2.188 billion to purchase/build a Hsinchu Science Park plant to expand AI chip testing (production in 2027).
Take · Capacity tightness is transmitting from advanced packaging to testing/optical components—the infrastructure side of the boom is broadening across the board.
AI safety written into central top-level document on 'new quality productive forces'
中共中央、国务院21经济网/财联社Confidence MidNewcomer
The 'Opinions on Developing New Quality Productive Forces' was issued, explicitly stating in the 'AI+' initiative: 'build a technology monitoring, risk early warning, and emergency response system to ensure AI is safe, reliable, and controllable.'
Take · China's main AI governance line officially shifts from 'content compliance' to 'behavioral controllability'; comprehensive legislation is on the agenda.
FTC v. Amazon algorithmic pricing case opens + State Administration for Market Regulation: 'AI cannot be a shield for illegality'
FTC / 市场监管总局法院排期+封面新闻Confidence HighNewcomer
FTC and 17 states' case against Amazon opens 10-13 (phased trial), with core allegations that its AI pricing system 'Project Nessie' predicted rivals would follow price increases to maintain high prices, causing consumers to overpay more than $1 billion; the same day, the State Administration for Market Regulation said, 'algorithms are designed by people, and responsibility always rests with companies.'
Take · China and the U.S. simultaneously shift regulatory focus from 'what models say' to 'what algorithms do'—algorithmic responsibility falls on corporate entities.
Stardust Intelligence completes 10 亿元 financing, valued at 150 亿
星尘智能界面/贝壳财经Confidence HighNewcomer
Shenzhen rope-driven humanoid robot company raised its 4th round in nearly half a year, led by Toukong Donghai (Shenzhen Investment Holdings); valuation rose from RMB 10 billion to RMB 15 billion in 3 months; T1 has already pushed high-precision robots into the RMB 90,000 range.
Take · Embodied AI raised about RMB 93.5 billion in H1, but the top 20 took nearly 60%—a chase-the-leaders model of state capital + leading financial capital has taken shape.
DeepSeek to lead seed round in embodied AI startup Winverse AI
DeepSeek彭博/福布斯中文Confidence MidUnsettled
Tens of millions of USD, post-money valuation of about $400 million (Bloomberg cites people familiar, deal not closed); Meituan Longzhu committed to invest alongside; founder Hu Han is a core author of Swin Transformer and former head of multimodal at Tencent Hunyuan.
Take · Second embodied AI investment in two months (participated in Unitree's IPO strategic placement in August); dual-track “body + brain” layout takes shape.
Kroger / Character.aiHacker News AI 台账Confidence MidOfficial ledger, firsthand
Retail giant Kroger plans to cut about 3,000 corporate jobs over three years while ramping up AI; Character.ai has filed for bankruptcy, reports say, with a Disney bid and platform shutdown involved.
Take · AI replacing white-collar jobs moves from prediction to public plan as a consumer companion app ends in bankruptcy—the sector's hot and cold in one frame.
Original ↗04RESEARCH3 stories
Vision-only Harness passes ARC-AGI-3 public set (research side of the same achievement)
何恺明团队arXiv:2610.02200Confidence HighPaper + open source
Minimalist visual Harness relies only on raw screenshot input + lossless visual memory + model-driven visual inspection; Opus 5.0 cleared all 25 public games, with 57.4% fewer actions than human novices (arXiv:2610.02200 + open-source code).
Take · Limited to a public set of 25 public games—suggesting that the 'perception and memory interface,' not reasoning ability, is the current main bottleneck, with major evaluation implications.
Anthropic warns: small models under 30B may be capable of cross-device autonomous replication by the end of 2026
Anthropic官方博客/ReutersConfidence MidProgress
Official blog discloses experiments—small models can already complete most individual tasks required for autonomous replication; GLM-5.2 and stronger models can achieve end-to-end replication in simplified test environments; if improvement speed holds, by the end of 2026 they may replicate across devices in an IoT botnet style. Caveat: some test steps were artificially simplified.
Take · First technical substantiation of Amodei's verbal warning last month, with a timeline and threshold model (30B parameters)—a new gray rhino moves from prophecy to experiment.
Leaderboard approaches diverge: Agent Arena adds Bash Reliability signal; LiveBench 2.0 focuses on 'anti-score-gaming'.
Agent Arena / LiveBenchAGI HuntConfidence MidSecondhand citation pending verification
After the update, the top five are all closed-source models (Opus 5.5 Max, GPT-6.1 Sol Max, Fable 5.1 Max, Astra Max, Gemini 4 Argon), while the open-source camp as a whole falls out of the top 14; on LiveBench 2.0 (a continuous question-rotation benchmark), Claude Fable 5.1 ranks first with 88.0 and is the most expensive at the top of the leaderboard.
Take · If validated, 'anti-contamination + continuous question rotation' may become standard for next-generation evaluations—leaderboard credibility defense and methodological fragmentation in the same frame.
05INSIGHTS7 stories
Three documents complete a 'triangulation' of the AI capital cycle: IMF project initiation, Goldman Sachs lights up, capital verification at Anthropic's door
前瞻深度版研判Confidence MidToday’s main thread
IMF GFSR adds an AI correction to three major systemic risks and calls out AI circular financing (macro institution project); Goldman Sachs's Molavi three-signal framework has two lights on, with the third pointing to 2026Q3 (sell-side warning); Anthropic enters investor disclosure with a $2 trillion valuation target plus a $42 billion single-year loss (corporate verification).
Take · A top doesn't equal a crash, but it means the second-half decisive factor switches from 'who raises the most' to 'who turns cash flow positive'—watch Anthropic's listing, CoreWeave spreads, and Q4 cloud vendors' capex remarks.
Payment is really happening on shelves and purchase orders: Token on Tmall, AI digital therapeutics enter top-tier hospitals.
前瞻深度版研判Confidence MidPaid fulfillment
MiniMax opens a Tmall flagship store and lists Token subscription packages; Tmall launches 'AI Space Station' aggregated top-up center; Beijing Anzhen Hospital buys 'AI cognitive assessment and intervention training' services for 970.885 ten-thousand yuan (Beijing Zhijingling Technology).
Take · Paid services shift from 'selling compute' to 'selling outcomes'—application-layer cash flow is the most solid bottom support in this cycle.
Agent procurement moves down to vocational schools: an 'ant market' of 30-100 万 per deal.
前瞻深度版研判Confidence MidDownmarket channel expansion
Changzhou Textile and Garment Vocational College's 'AI Agent Platform' won a 46.8 万 bid, and Huaibei Industrial School opened bids for a 54.3 万 AI trainer training room.
Take · Thousands of vocational colleges nationwide are launching projects in batches under the 'AI general education courses + agent practical training' policy; short decision chains—a window for small and midsize integrators.
Textbook case of split leaderboard methodology: Muse Wonder AA 56 points, No. 2 vs 31.7 points in real-task tests.
前瞻深度版研判Confidence MidDefinitional split
On AA Intelligence Index v4.3.2, Wonder Full scored 56, beating GPT-6 Astra (53); but in Pawel Huryn's real-world test with 105 bugs planted in two real repositories, it scored only 31.7 (Astra 45.0). A VulcanBench sample says it uses 72.4 万 effective tokens per task, 6.5x that of GPT-6.1 Sol (blogger's small sample, not a benchmark).
Take · When citing any index, cost methodology and harness differences are now must-read footnotes—real traffic is disproving leaderboard premiums.
OpenRouter weekly volume hits a new high of 172 trillion; DeepSeek V4.1 Flash tops the week at 40.5T
前瞻深度版研判Confidence HighTraffic votes
GLM 5.3 Flash 12.7T is 2nd, Step 5 Preview 3rd; V4.1 Flash tops mainly due to migration of existing V4 Pro calls + low cache costs suited to Agent scenarios. Third-party Dirac methodology (single-day snapshot, OpenRouter channel only): open-weight models account for 66.9% of call volume, DeepSeek single-day share 35.4%.
Take · Real traffic (money's vote) is replacing leaderboards (institutions' vote) as the basis for selection—the selection market's pricing logic is shifting axes.
Linus uses AI to write kernel patches: coding agents go from toy to authoritative endorsement
前瞻深度版研判Confidence MidTool personification
Report: Linus Torvalds has begun using AI to write Linux kernel patches; meanwhile Anthropic develops in-house Scion to replace git, and LiquidAI discusses real frontier issues for on-device Agents.
Take · The most conservative engineering authority turns to AI coding, a landmark endorsement that 'coding agents have grown up'—tools shift from assistant to primary force.
Regulatory focus shifts from 'what models say' to 'what algorithms do'
前瞻深度版研判Confidence MidGovernance shift
A central top-level document explicitly states 'ensure AI is safe, reliable, and controllable'; FTC v. Amazon 'Project Nessie' algorithmic pricing case goes to trial; the State Administration for Market Regulation says 'algorithms are designed by humans; responsibility always lies with the enterprise.'
Take · China and the U.S. simultaneously push governance from content compliance to behavioral controllability—opening a policy window for compliance-as-a-product.