WEEK 2026-09-28 · 22 STORIES · SKILLHUB WEEKLY
SkillHub.news Weekly AI
Sep 22 ~ Sep 28, 2026 · Weekly · sources 4
Highlights22 stories · approx.9min read
01OFFICIAL EVENTS14 stories
Kākāpō Party
02RANKING MOVES8 stories
obra/superpowers stars +79,301
01OFFICIAL EVENTS14 stories
官方·Simon Willison2026-09-27Confidence High
Original ↗官方·Simon Willison2026-09-26Confidence High
Original ↗官方·Simon Willison2026-09-26Confidence High
Original ↗官方·Simon Willison2026-09-26Confidence High
Original ↗YouTube·Google Cloud2026-09-25Confidence High
Original ↗YouTube·Salesforce2026-09-25Confidence High
Original ↗YouTube·Cerebras2026-09-25Confidence High
Original ↗YouTube·Salesforce2026-09-25Confidence High
Original ↗YouTube·Google Cloud2026-09-25Confidence High
Original ↗YouTube·Midjourney2026-09-25Confidence High
Original ↗Ecosystem Pulseself-updating
2026-09-03 → 2026-09-10 Star growth：obra/superpowers +79,301★ · mattpocock/skills +71,905★ · multica-ai/andrej-karpathy-skills +59,162★ · anthropics/skills +49,058★ · Shubhamsaboo/awesome-llm-apps +38,266★ · farion1231/cc-switch +36,887★ · nextlevelbuilder/ui-ux-pro-max-skill +35,298★ · Graphify-Labs/graphify +32,474★
※Risk Levelsgray rhino / black swan
Gray rhino · foreseeable
- If OpenAI's pause on 'tool-use-related training/evaluation/inference' extends beyond two weeks, Q4 frontier model release cadence will be forcibly rearranged (Source: IT之家, Sina Finance, Axios/The Paper, 2026-09-27; local 20-seat intelligence unanimously lists it as the first tracking node, medium-high)
- Capital-compute self-cycle: NVIDIA/Amazon/Google are both Anthropic investors and compute suppliers; supplier revenue = investee spending = part of financing funds. Any financing freeze would hit all three at once and pollute compute price signals (Source: Reuters/Gelonghui, 2026-09-27; local 20-seat consensus, medium)
- Credential-side governance is nearly unprotected: dark-web permission black market discounts up to 97%, promises to replace banned accounts, plus tens of thousands of pending boundary-violation incidents at two companies, leaving customers exposed to paying for stolen traffic (Source: IT之家 citing UK Financial Times 2026-09-28, Axios/澎湃 2026-09-27, medium-high)
- Iteration speed exceeds ecosystem absorption capacity: industry average of a new flagship every 6 days, downstream regression and adaptation costs become a hidden tax, and already-accepted versions depreciate rapidly (Source: 网易 2026-09-27; multiple local seats agree, medium)
Black swan · low-probability
- AI security incidents escalate from internal corporate governance to legislative/access review (Australian Senate has summoned OpenAI and Anthropic heads over government system intrusions; this issue relays only 1~2 seats of intelligence, single source pending verification), potentially triggering access lists or procurement restrictions (low-medium)
- Anthropic IPO pricing falls short of expectations → long-term compute orders renegotiated, neocloud financing paradigm hit, compute leasing pricing anchor breaks (low, single source pending verification)
- Domestic compute commitments (e.g., 真武 V900 mass production in 2027Q1) vs actual production capacity gap, or the conflict between DeepSeek financing figures of 7.5 billion USD/300 million USD is confirmed, triggering reversal in valuation and supply-chain expectations (low, single source pending verification)
✦Action Items3
1. Run an immediate health check on API credentials for 'hard quota caps + least privilege + usage anomaly alerts'Sample 30 enterprise accounts for usage anomaly detection (rented credentials will always show off-hours anomalies), and check whether credential governance coverage is >95%; internally implement hard quota caps and least privilege; externally write 'credential transferability/sharing bans' and 'burden of proof for usage spikes' into contracts. Basis: dark-web model-permission black-market discounts of up to 97%, sellers promising 'free replacement credentials if the account is banned' (IT之家 citing the UK's Financial Times 2026-09-28), combined with OpenAI/Anthropic investigating tens of thousands of boundary-violation incidents (Axios/澎湃 2026-09-27).
window Immediate - 2 weeks 2. Revise Agent procurement acceptance clauses into three measurable parameters① Number of out-of-bounds alerts per 10,000 tool calls; ② median time from alert to human intervention (public timeline: 15 min alert / 3 min human intervention / 2.5 hours to terminate training, can serve as baseline); ③ average cost per incident. First run a 7-day shadow run with the existing Agent (log only, no intervention) to understand the alert distribution, then write the three numbers into the contract. Basis: unanimous recommendation from vault AI Lab's 20-seat Daily AI Intelligence 2026-09-28; timeline data points from Sina Finance 2026-09-27.
window 1-3 months 3. Products with voice share >30% should recalculate unit costs this week and pick one 'judged-dead scenario' to run a 7-day A/B.Qwen voice line TTS approx. -70%, Realtime approx. -85%, ASR approx. -95%: recalculate unit costs and LTV/CAC, and pick scenarios such as 'long-term meeting transcription' and '7×24 voice customer service' that were previously judged dead due to cost for small-traffic controlled experiments; the only criterion is whether per-conversation unit gross margin turns positive. At the same time, following a 'multi-armed bandit' approach, fix 2–3 candidate models for long-term parallel small-traffic testing, and use business metrics rather than leaderboards to decide whether to keep or drop them. Basis: vault 20-seat intelligence 2026-09-28; Sina Finance/36Kr 2026-09-24~26.
window 3-6 months ⏱Next 3 MonthsProb.
| Window | Event | Prob. | Basis |
|---|
| 一周内 | OpenAI's pause on "tool-use-related training/evaluation/inference": whether it is lifted, and on what conditions (whether quantifiable tool authorization and sandbox isolation standards are provided) | Medium-high (65%) | The local 20-seat intel unanimously lists it as the No. 1 tracking node (sources: IT之家, Sina Finance 2026-09-27); suspension duration is the only hard metric that can falsify a "PR gesture," and 2 weeks is the key observation window. |
| 2 周内 | Whether Anthropic's IPO pricing and the final terms of NVIDIA's equity investment are finalized (a USD 10 billion-level cornerstone investment) | Medium (50%) | Once pricing is finalized, it will mark the entire compute-capital closed loop to market and become a valuation anchor for subsequent AI targets (sources: Reuters/Gelonghui 2026-09-27, single source pending verification). |
| 9/30 | Quarter-end compute order fulfillment criteria (acceptance and collections, not signed contract value) | Medium (55%) | Quarter-end settlement convention; commitment tables expose risk earlier than revenue tables, with a 6–18 month lag between contract value and revenue recognition (local multi-seat consensus view) |
| 30 天内 | Are application-side ROI-flip cases emerging after voice pipeline price cuts (ASR -95%/TTS -70%); horizontal spread of Agent billing models (will Muse and resident agent "o" follow usage-based pricing) | Medium (50%) | The segment with the largest price cuts = the most homogenized segment; billing-unit migration typically precedes product-form migration by half a year, and first movers take pricing power (vault 20-seat intel 2026-09-28) |
| Q4 | Capacity validation for 2027Q1 mass-production commitments including 真武 V900; actual usability of on-device personal agents (Qualcomm itself says it will only be stable in 2027—2028) | Medium (50%) | Promises and capacity must be priced separately: long-dated options do not constitute today's combat power; vendors' self-reported time windows are the testimony most worth cross-checking (Sources: Kaiyuan Securities research report 2026-09-27, 36Kr 2026-09-27) |