VOL.2026-10-29 · 29 STORIES · SKILLHUB DAILY
SkillHub.news Daily AI
Oct 29, 2026 · #46 · updated 08:00 daily · sources 4
01MODEL RELEASES3 stories
Gemini 4 Carbon tops AA v5: 59 points, injection success rate falls to 0.0%
GoogleIT之家 + Artificial Analysis 转引Confidence HighNew topic
Google released Gemini 4 Carbon on 10-29, scoring 59 on Artificial Analysis Intelligence Index v5.0 to rank first; SEC-bench Pro vulnerability discovery 88.2%, indirect prompt injection success rate down to 0.0%, patch repair rate 71.0%, WebDev Arena 1769 points (+89 Elo vs. previous generation). Initial pricing: $2 input / $10 output per million tokens, returning to $4 / $20 after the promotional period; this week it opens first to a small group of cybersecurity defenders.
Take · The initial rollout targets security defenders rather than developers, effectively writing 'security productization' into the release cadence: 0.0% and 88.2% are engineering parameters that can go straight into an acceptance checklist, but both are vendor claims and not yet third-party reproduced. Trade-off note: some intelligence records 'WebDev Arena only 4th,' which does not contradict 1769 points but reflects a different basis; we adopt the conservative wording of 'first overall, not first across all scenarios.'
AA Intelligence Index v5.0 goes live: official acknowledgment that 60% of scoring is 'not reproducible by anyone'
Artificial Analysis官方口径 + HaiAI123 聚合Confidence HighMethodology change
Three equally weighted blocks (knowledge work / coding / science), with HLE-Diamond replacing Humanity's Last Exam; 60% of scoring is based on private test sets or undisclosed answers, and the official line is 'not reproducible by anyone.' The strongest open-weight model, Kimi K3.1 Pro, trails the leader by 12 points; DeepSeek V4.1 Flash 1017 and GLM-5.3 are tied at 32 points.
Take · Media widely compare v4.3.2's 44 points directly with v5.0's 59 points, but indices across versions are not comparable—this is a 'gap illusion' created by mixing methodologies. What is truly worth noting: when enterprises buy by leaderboard, they are outsourcing 60% invisible variance to the leaderboard authors.
Four open-source on-device releases on the same day: Ling-3.1-flash, Clover-35B-A3B, Nemotron 3.5 Transcribe, Pocket TTS
蚂蚁百灵 / 深圳扩散智能 / 英伟达 / Kyutai腾讯新闻 + 证券日报 + Hugging Face / InfoQConfidence MidNew topic
Ant Bailing open-sourced Ling-3.1-flash-1019 (scores 44 on AA v4.3.2, +3 vs. the previous version, with base also open-sourced); Shenzhen Diffusion Intelligence open-sourced Clover-35B-A3B-Preview (35B total parameters / 3B activated; generation speed in some scenarios is 5–10x that of same-size Qwen3.6-35B-A3B); NVIDIA open-sourced Nemotron 3.5 Transcribe (multilingual transcription, with prompts to adapt to accents and terminology); Kyutai's Pocket TTS runs on-device with as few as 1.9 million parameters.
Take · What the four share is not 'catching up with the frontier' but 'private deployability and running cheaply'—open source is still catching up on general capabilities (Ling only +3 points); the real overtaking is in output per unit of compute.
02PRODUCT LAUNCHES4 stories
AnthropicTechCrunch（原文实抓）Confidence HighNew topic
Using internal analytics and employee surveys, managers can view by date range what teams did, what they produced, how much they spent, and where workflows broke, and receive improvement suggestions; by default, this data is not used for model training, and pooled team budgets are also launched to control token spend. Disclosed in the same batch: the company recorded operating profit in Q2, with net revenue retention of 900% (same-cohort customer spend in 2026Q3 was 9x that in 2025Q3).
Take · The bottleneck in enterprise procurement is not unit price but unclear ROI—productizing 'how much was spent / what came back' turns token billing from a cost center into a renewal-ready evidence chain, and also reinforces Claude Code's commercialization narrative.
Original ↗Salesforce / Slack官方（salesforce.com）Confidence HighNew topic
Using channels as the workbench, it provides an external API for self-built integrations, letting agents and team members collaborate on code in one place. Positioned as “your agent, your team, one code channel.”
Take · The default working spot for coding agents is moving from IDE to IM: once reviews, permissions, and auditing all happen in chat channels, compliance boundaries must be redrawn entirely.
Original ↗OpenAI × ServiceNow官方（openai.com，原文实抓）Confidence HighNew topic
ServiceNow integrates OpenAI models and agents into workflows from development to security, connects ServiceNow workflows with ChatGPT and Codex, and opens ChatGPT Work and Codex to all employees. The partnership is an expansion of their January collaboration this year.
Take · The battle for the enterprise agent entry point is not at the model layer but in existing systems such as ticketing, ITSM, and security operations—whoever is embedded at the process entry point gets the default position, and migration cost becomes the highest wall.
Original ↗GoogleTechCrunch + Google Cloud 官方视频Confidence MidNew topic
As agents reshape web interactions, Google expands third-party access to Gemini and Google Search; Google Cloud separately launches a preview of an open-source agent suite for PyTorch on Cloud TPU
Take · Search shifts from pages users visit to interfaces called by agents; distribution power shifts from traffic gateways to invocation channels, while TPU moves from internal use to a second supply with its own toolchain
Original ↗03INDUSTRY5 stories
Microsoft FY27Q1: revenue 92.279 billion, RPO 677 billion; first disclosed Azure quarterly revenue of $33.2 billion
MicrosoftIT之家 / TechCrunch（原文实抓）Confidence HighNew topic
Revenue $92.279 billion (+19%), net profit $36.629 billion (+32%), operating profit $45.915 billion (+21%), Microsoft Cloud $63.9 billion (+30%), Azure +48% YoY, commercial RPO +73% YoY to $677 billion. Same day first provided Azure dollar figure: $33.2 billion, about 36% of quarterly revenue, beating analyst estimates by about $850 million; Google Cloud reported $30.7 billion in the same quarter.
Take · Our calculation: RPO $677 billion ÷ quarterly $92.279 billion ≈ 7.34 quarters ≈ 1.83 years of backlog; in the quarter, $10 billion was recorded from the OpenAI commercial arrangement, about 10.8% of quarterly revenue. RPO growth (+73%) far exceeded revenue growth (+19%), indicating orders are stacking into the future—cloud vendors for the first time have a dollar metric directly comparable with AWS and GCP.
Alphabet Q3 revenue $127.6 billion (+25%): ad growth slows, AI Mode ad placements scale globally
AlphabetTechCrunch 引财报Confidence HighNew topic
Revenue $127.6 billion (+25%), with core ad growth slowing, but AI Mode ad placements moving from trial to global scale; Google Cloud reported $30.7 billion for the same quarter.
Take · First global-scale validation of AI shifting from a cost item to a revenue item; advertiser spend definitions and CPA attribution will be rewritten—also the only side with current-period readings on whether AI investment pays back.
Meta financing rises to $97.5 billion; NVIDIA subscribes for $3 billion same day
Meta / 英伟达21 世纪经济报道 + 财联社Confidence MidNew topic
Class A common stock raised from $22.5 billion to $25 billion, plus $12.5 billion in mandatory convertible preferred stock, for a current total of $37.5 billion; including the planned $60 billion offering to be launched in 2027, the overall total is $97.5 billion. NVIDIA subscribed for $3 billion same day and expanded chip deployment; the two sides also expanded Muse series model collaboration, using OpenShell + Sentry for agent security and end-to-end behavior monitoring.
Take · Methodology note: media have two ways of writing 97.5 billion—current-period and including the 2027 offering; we split them as in the original and do not merge them into one current-period figure. NVIDIA's $3 billion is about 8.0% of the current-period $37.5 billion and about 3.1% of $97.5 billion (our calculation)—the supplier is also a customer shareholder, and this account ultimately has to go into the same table.
Anthropic files S-1 to list on Nasdaq, plans to reserve 15%—20% for retail investors
AnthropicBloomberg / The Next Web / InfoQ 多源同题Confidence MidNew topic
With Claude Code as the core growth narrative, the prospectus also includes 'AI safety stance may invite government retaliation' in risk factors; Anthropic plans to reserve 15%—20% for retail investors.
Take · This is the first time a frontier lab has put regulatory risk into the main fundraising disclosure, effectively pricing 'sticking to safety principles'; the side effect is that later entrants will make safety wording vaguer to reduce disclosure risk.
Oracle becomes the first hyperscale cloud besides Google to offer TPU; CoWoS set for 10%—20% price increase
甲骨文 / 台积电界面新闻 + 财联社引 TrendForce / IT时代网Confidence MidNew topic
On OCI, Vera Rubin NVL72 expands to multi-gigawatt scale and is available immediately, and connects to Google TPU; TPU 8t (training) / 8i (inference) is expected to launch in 2027, with multi-year expansion. Meanwhile, TSMC plans to raise CoWoS advanced packaging quotes by 10%—20%; demand has spilled over from GPU to server CPU and custom ASIC, and longer chip test cycles further squeeze packaging and testing capacity.
Take · The only segment in the whole chain that still cannot meet supply despite price increases is often the real bottleneck: pricing power is moving from the model layer to packaging + energy. But within 12 months, TPU on OCI is only a bargaining chip, not a cost decline—the transmission lag is usually 1—3 quarters.
04RESEARCH3 stories
The value of Genie 4 is not content but moving robot training from real robots to simulation.
Google DeepMind第一财经 / 36氪转载 + 官方视频Confidence HighProgress · Day 2
Genie 4, released as a research preview on 10-28, can generate interactive 3D worlds with music and ambient sound in real time; in simulation, scenes can be changed via prompts and adjusted on the fly. Today's intel adds its positioning: moving robot training from real-robot sampling to world-model sampling.
Take · It is not a content tool but an 'environment factory'—once training data comes from simulation, the only bottleneck left is the sim-to-real gap, and that gap is exactly the most expensive stretch of autonomous driving and robot testing.
RoboDojo embodied benchmark has a new leader: AgiBot GO-3 Nano Preview averages 61.90 points and a 56.42% success rate
智元 / RoboDojo36氪（单源）Confidence MidNew topic
AgiBot GO-3 Nano Preview scores an average 61.90 points and 56.42% success rate on RoboDojo, surpassing GPT-6 Astra; it ranks first in all three sub-dimensions—fine manipulation 78.65, long-horizon execution 68.21, and open semantic understanding 61.71.
Take · Take: steady hands and persistence, but it 'doesn't quite understand human language'; a single leaderboard, single format, and undisclosed task set are not enough to support 'embodied AI is already usable'—treat this as a single source, not entering Top5.
GitHub Octoverse 2026: 220 million merged PRs a month, tripled in a year
GitHubSimon Willison（Universe 现场）+ 官方 OctoverseConfidence HighNew topic
In 2026, merged PRs rose to 220 million per month in less than a year, commits to 7.38 billion per month (most of the increase came in H2), and new repository creation doubled to 30.3 million per month; the report breaks out figures by NAICS industry and the 'interoperability / reverse engineering' repository category, and uses LLM classification to extract user country codes to correct for VPN bias.
Take · Agentic engineering has been quantified at platform level for the first time: code output growth has decoupled from headcount growth, and review, CI, and security audit capacity have become the new bottleneck—this says more about the shape of engineering in 2026 than any single-model benchmark.
05INSIGHTS4 stories
Productize the 'build-your-own private test set': leaderboards aren't verifiable, so you need your own evidence.
尚吉 · 切入模式本刊 Top3 ①Confidence MidViewpoint
For customers with annual model spend above $100,000, deliver a '50–100 real business sample private test set + double-blind criteria + first baseline report.' Path: free leaderboard checkup (flagging its 60% private-test share) → per-use set building (samples drawn from customer historical tickets) → quarterly regression subscription + vendor-switch certification.
Take · Teams that help customers build private test sets are almost impossible to replace in year two, and switching costs are extremely high; those still using 'we rank first on benchmarks' in pre-sales will have a hard time closing deals in 2027. Time window: a standard 3-month vacuum period starting immediately; ship the first baseline by November to catch the Q4 budget season.
Build a two-stage budget based on CoWoS +15% and Gemini 4/20 full price, and add three clauses to the contract.
巽风 · 合同与成本本刊 Top3 ②Confidence MidViewpoint
Run a 'packaging price-hike pass-through' stress test for existing customers; recalculate three-year TCO in two stages and output a gross-margin sensitivity table. Add three contract clauses: price cap and price-change notice period ≥60 days, volume-tiered price lock, penalty-free migration if price rises over 30%; make dual-model/dual-cloud switchover drills quarterly.
Take · Upstream price increases usually reach API price hikes with a 1—2 quarter lag; this is the last chance to lock in long-term contracts. The gap: no third party currently quantifies 'packaging premium pass-through' in gross-margin models for AI applications. If packaging accounts for 10%—15% of AI accelerator card BOM, a 10%—20% price increase passes through to about +1%—3% of the full card (Kanliu assumed values; needs empirical validation, not a conclusion).
First switch transcription to an open-source component for A/B testing, then migrate standard tasks to 30B-A3B or on-device.
渊策 · 路由分层本刊 Top3 ③Confidence MidViewpoint
For our own voiceover/content pipeline, do two small things first: switch transcription to Nemotron 3.5 Transcribe for A/B testing; migrate classification, extraction, and first drafts to 30B-A3B-class or on-device small models. Compare WER / timestamp accuracy / cost per thousand characters on the same 5 assets, then finalize routing (80% of standard tasks to small models, 20% of high-value tasks stay with the flagship).
Take · Transcription and speech, the two links that are 'most expensive and least sexy,' leave unit prices in others' hands; in a price-hike cycle, token-saving design is the best experience design.
Write 0.0% and 88.2% into the acceptance checklist: three hard tests—indirect injection, unauthorized invocation, and long-horizon tampering
艮山 / 霆钧 · 验收本刊「规则」维度Confidence MidViewpoint
Gemini 4 Carbon's 0.0% injection success rate and 88.2% vulnerability discovery rate are engineering parameters that can be written into the acceptance checklist; it is recommended that delivery acceptance checklists add three hard tests: indirect prompt injection, unauthorized invocation, and mid-task tampering in long-horizon tasks.
Take · Looking at attack-defense history, the defense window is usually only 3—6 months, after which attack tools become commoditized; agent least privilege + full-chain behavior auditing should move from a bonus item to a hard selection metric.
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. Turn a self-built private test set into a productFor customers with annual model spend above USD 100,000, deliver a '50—100 real business sample private test set + double-blind criteria + first-version baseline report,' pitch 'leaderboards cannot be audited; you need your own evidence.' Path: free leaderboard checkup (point out its 60% private testing) → per-engagement test-set building (samples drawn from customer historical tickets) → quarterly regression subscription + vendor switching certification; switching costs are high, and by the second year it is almost irreplaceable.
window A 3-month standards vacuum starts immediately; release the first baseline within November to catch the Q4 budget season. 2. Two-stage cost recalculation + multi-sourcingRun a “packaging price-hike pass-through” stress test for existing customers: recalculate three-year TCO in two stages based on CoWoS +15% and full-price Gemini 4/20, and output a gross-margin sensitivity table. Path: single-point audit → contract clause revamp (price-change notice period ≥60 days, tiered price lock, penalty-free exit if price increase exceeds 30%) → quarterly dual-model/dual-cloud switchover drills, and record the “second supply source” in meeting minutes as a credible signal.
window 2026Q4 contract window; there is usually a 1—2 quarter lag from upstream price increases to API price hikes—this is the last chance to lock in long-term contracts. 3. On-device routing + transcription replacement A/BStart with two small moves in our in-house talking-head/content pipeline: switch transcription to Nemotron 3.5 Transcribe for A/B; move classification, extraction, and first drafts to 30B-A3B-class or on-device small models. Path: compare WER / timestamp accuracy / cost per thousand characters across the same 5 pieces of material → finalize routing (80% standard tasks to small models, 20% high-value to flagship) → codify into an externally deliverable “token-saving design” methodology to feed back into cost consulting projects.
window Complete A/B this month, finalize in Q4; in a price-hike cycle, token-saving design is the best experience design.