Back to Blog
Galen Guan

Daily Briefing · 2026-10-09: Fed Minutes Turn Hawkish, Hormuz Tensions Cool, A-Share Tech Keeps De-Rating

📰 Daily Briefing · 2026-10-09 (Friday)

🌐 Politics · Geopolitics

  • Trump: the US will not attack Iran before the November midterms — Posting on social media Thursday (10-08), Trump said talks with Iran were "productive" and committed to no military action before the November 3 midterms, while the naval blockade stays in place. The White House had earlier asked the Pentagon to prepare strike options, so the statement marks a real de-escalation in market pricing. Context: US-Iran talks had stalled over control of the Strait of Hormuz; after back-channel contacts during September's UN General Assembly, Trump rejected an Iranian proposal. But Iranian attacks on tankers continue — several heavy explosions were reported on shipping lanes south of Hormuz on Thursday, driving oil to its biggest one-day gain in a month. Watch: whether the pre-election restraint holds post-November, and whether attacks derail the recovery of oil flows through the strait (now near pre-war levels). Sina Finance
  • US suspends Microsoft, Adobe, and Infosys from the green-card labor program over fraud allegations — The US government froze PERM green-card labor applications from the three companies as part of its crackdown on foreign-worker programs, leaving thousands of visa-holding employees in limbo. It's the first time tech giants have been named directly in this enforcement wave, a material signal for global talent mobility and offshore hiring. Sina Finance

💰 Finance · Markets

Market snapshot (Oct-08 close / early Oct-09 Beijing time): S&P 500 7765.36 (-0.47%) · Nasdaq 27193.34 (-1.25%) · Dow 51231.64 (+0.10%) · Shanghai Composite 3811.90 (-0.79%) · CSI 300 4310.28 (-1.09%) · USDCNY 6.7031 · US 10Y / DXY / gold: not available

  • Fed September minutes lean hawkish: one more hike this year "may be appropriate," but officials in no rush — The FOMC minutes released Wednesday show all 19 officials backed September's 25bp hike to 3.75%-4% (the first since July 2023), and "most participants judged that an additional tightening this year would likely be appropriate." Still, futures-implied odds of an October hike have fallen from ~70% right after the decision to ~20%, after dovish remarks from Williams and Jefferson. The pivot now hinges on CPI due October 14 — another hot print would swing expectations right back. Worth noting: this is the first hiking cycle under new Chair Kevin Warsh, whose "remove a dose of accommodation" framing sits uneasily alongside the wait-and-see camp; the internal split is wider than the minutes' prose suggests. Long-end Treasuries held gains after a solid 30-year auction. Sina Finance
  • A-share tech keeps de-rating: STAR 50 down nearly 5% on heavy volume, ChiNext hits a fresh stage low — Chinese equities fell Thursday, with the STAR 50 off almost 5%, a CPO leader limit-down at 20% (its first in 18 months), and the ChiNext down more than 3% to another stage low; solid-state battery concepts were the lone bright spot. Shanghai closed at 3811.90 (-0.79%). AI hardware is the epicenter of the unwind. The counterpoint: a wave of strong Q3 pre-announcements (Dongyue Silicon's net profit guided up as much as 190x), so valuation compression in tech and earnings-driven strength elsewhere are running in parallel. CLS
  • PBOC lays out its FX stance systematically: no preset yuan target, guard against destructive overshooting — The People's Bank of China stated the yuan has no preset level, that China has no need or intention to gain trade advantage through depreciation, and that policy focuses on preventing short-term destructive overshooting. A rare systematic statement on FX; with USDCNY around 6.70 it reads as a clear policy floor signal. Watch the daily fixing versus offshore yuan for confirmation. CLS
  • Golden Week property data improves: Shanghai second-hand daily volumes +30% YoY, Shenzhen +45% — Holiday home sales warmed across first-tier cities. Meanwhile memory chips entered a "money-printing moment" (Samsung's quarterly operating profit topping 100 trillion won), but surging memory prices are squeezing downstream — Samsung plans to cut phone output by up to 30% in Q4. Marginal property recovery and the memory supercycle are the two most important China-asset fundamental threads; Samsung's cutbacks show upstream pricing is now feeding into end devices. CLS

🚀 Tech · AI

  • OpenAI's annualized revenue approaches $50B — below earlier $70B reports — According to people familiar with the matter, OpenAI's run-rate revenue is on track for roughly $50 billion, up sharply year over year but well under the "nearly $70B" reported last month. The discrepancy stems from investors trying to align methodology with Anthropic, including how cloud-provider-mediated sales are counted. Even on the conservative figure OpenAI is the highest-revenue AI company on earth, but the growth narrative is losing its generous accounting — the argument over methodology is itself a signal of deceleration. Next datapoint: which curve bends first, consumer ChatGPT subscriptions or API. Sina Finance
  • SpaceX buys low-band spectrum for direct-to-cell satellite service — SpaceX is acquiring low-band spectrum to expand mobile communications, a decisive step from Starlink's direct-to-phone demos toward commercial operation. Direct competition for rural coverage with legacy carriers; the combination of spectrum assets and launch capacity is a barrier few (AST SpaceMobile et al.) can match. Sina Finance
  • Apple announces Oct-13 "Homecoming" event in New York, focused on smart home — Apple will hold a product event in New York on October 13, widely expected to showcase its smart-home hardware line (with robotics rumored). Shares rose over 1% on the day. Smart home is Apple's biggest hardware growth story after wearables, and the on-device AI angle is what to watch. Sina Finance
  • Microsoft claims its new flagship laptop beats MacBook on some AI tasks — Microsoft publicly asserted its new flagship laptop outperforms Apple's MacBook on selected AI workloads, a counterattack in the on-device NPU performance narrative. The marketing war is heating up, but independent benchmarks will decide it; for developers, the maturity gap between the two NPU toolchains remains the real selection criterion. Sina Finance

📄 Papers · Research

  • arXiv: decoupling exploration from optimization in RLVR — Decoupling Exploration from Optimization in RLVR (2610.10536). Argues that RL with verifiable rewards on already-trained checkpoints has a structural tension between its core promise (discovering new reasoning strategies) and its optimization objective, and proposes explicitly decoupling exploration from optimization. Directly relevant to anyone training agents — including behavioral-diversity problems in Copilot-style systems: stacking more RLVR rounds does not automatically produce new strategies. arXiv
  • arXiv: scaling laws for robotic latent world models — RoboJEPA: Scaling Robotic Latent World Models (2610.10515). Provides the first systematic estimate of how latent world models scale with model size and data in robotics, filling in the "can world models scale like LLMs" question. Posted the same day as Long-WAM (2610.10528, scaling causal world-action model context) — the two threads (context extension + scaling laws) advancing in the same week suggest the field is entering a systematization phase. arXiv

🧠 Voices

  • Musk vs. Asia's richest man goes public; India's government issues a rebuttal — Musk's apparent public jab at Mukesh Ambani escalated to the point where the Indian government issued a formal statement. Behind the spat lies a structural collision between Musk's Starlink/Tesla interests in India and Ambani's telecom-energy empire — a substantive contest over access to the Indian market. CLS

📝 New Blogs (24h)

  • The DeepSeek site blog shows signs of an Oct-08 update, but the page is a JS-rendered SPA and no title/content could be extracted from static HTML — no reliable content obtained. None of today's target blogs (Lilian Weng, Anthropic Engineering, Cognition, Manus, Baoyu, Qwen, Chip Huyen) published within 24h; latest updates date to 2026-07-04 / 05-25 / 09-25 / 07-24 / 09-02 / 2025-09-23 / 2025-01-16 respectively. Reported honestly: no fresh posts today, no filler. deepseek.com

⭐ GitHub Trending

(GitHub Search API, 3-day window created:>2026-10-05, stars>100; 61 results, top ten shown; site-source repos like playgta5.com filtered out as low-value per usual practice)

  • openai/math (Lean · ⭐12001) — OpenAI's open-source math formalization repo; created Oct-06 and past 12k stars in three days. An official OpenAI repo tied to Lean/formal proofs — most likely the companion to a math-reasoning capability release. A strong AI-for-Math signal; the description is still empty, so go read the repo directly. repo
  • mhtsec/ARTEX (Go · ⭐678) — Autonomous AI penetration-testing system; champion of Baidu's "agent+" security challenge. A complete engineering implementation of a security agent (planning + tool calls + attack/defense knowledge) — a high-quality reference for agent orchestration and red-team automation. repo
  • pingdotgg/ts-rust (Rust · ⭐682) — An experimental Rust port of the TypeScript 7 compiler (tsc), by Theo. After the official Go port, the community now has a Rust contender — the compiler-performance arms race is in full swing; watch whether it sets new type-checking throughput baselines. repo
  • storytold/wordcraft + cadcraft (Rust · ⭐759/689) — The same team clean-room rewriting Microsoft Word and an AutoCAD-style app in pure Rust. The "rewrite everything in Rust" movement expands from editors to two traditional software giants; engineering completeness worth tracking. repo
  • nullmoth/nvidia-macos-driver (Rust · ⭐955) — An open-source Metal driver for NVIDIA GeForce RTX cards on macOS 15 Sequoia (Intel/OpenCore hackintosh). Years after Apple dropped NVIDIA support, the community built its own Metal driver — a must-have for hackintosh + local LLM inference users. repo
  • alchaincyf/huashu-art-motion (JavaScript · ⭐2404) — An art-animation skill: 35 art styles and 9 narration grammars to bring paintings to life in code. A Chinese-developed style-transfer + animation pipeline for classic artworks; 2,400 stars in three days — a breakout hit. repo

🔭 GitHub Watch · Watchlist

(11 of the 26 watched repos had commits in the past 24h; 2 new [Watch] tasks appeared on the ShuJieTai board, with the "no actionable improvements" digest excluded; both included below)

  • langchain-ai/langgraph: ToolNode must not swallow interrupts and resume-validation errors (PR #9232 merged) — Fixes two error-handling flaws in the prebuilt ToolNode: (1) a pydantic ValidationError thrown inside a tool's interrupt(response_schema=...) was misclassified as a model argument error and fed back to the model, so human-approval interrupts silently vanished; (2) with wrap_tool_call plus a non-default handle_tool_errors, GraphInterrupt was wrapped into an error ToolMessage and the interrupt never paused. Adds a langgraph.errors.is_invalid_resume(error) discriminator. This hits the Copilot tool-execution layer squarely: if you do human review or schema-validation resumes inside tools/sub-agents, the same bug blames the model for validation errors and keeps the run going as if the interrupt were a mere error message — a silent failure of the whole approval chain. Action items: upgrade langgraph and run Copilot's approval/interrupt self-checks; for a hand-rolled dispatch path, port the is_invalid_resume logic. Priority 3 — schedule this week. repo
  • Comfy-Org/ComfyUI: MTP/CUDAGraph offload fix (PR #16877) as a lesson for video-generation VRAM strategy — Fixes Qwen3.5 MTP embedding offload: unconditional CUDAGraph capture would bake dynamically-migrated weights or offload cast buffers into the graph and compute on garbage data; the fix routes everything through the model-prefetcher (graphs auto-invalidated on offload), splits embeddings into a separate graph, and falls back to eager CPU embeddings when offloading. The offload + CUDAGraph combination under VRAM pressure produces silent garbage inference — degraded output with no errors, the hardest class of video-quality bugs to diagnose. Action items: audit magicedit's common workflow templates for video-model offload combined with CUDAGraph; add an "invalidate/disable CUDA graph on offload" rule to Copilot's VRAM-tuning advice. Priority 2 — track. repo

References