Daily Briefing · 2026-10-01 (Thursday)

Window: past 24 hours (2026-09-30 06:00 → 2026-10-01 06:00 CST). Market data reflects the latest close.

🇺🇸 World Politics

• Trump: final US troops leaving Iraq soon (Sina, 10-01). Iran says it has received the US response to its latest proposal but won't disclose details — talks remain unresolved.
• Trump escalates pressure on the Fed: deploying the Justice Department and litigation threats to push Fed Governor Powell's predecessor narrative aside — specifically demanding resignations from the Board; the Fed's inspector general found renovation mismanagement but no criminal wrongdoing. Powell's fate remains the biggest institutional risk for markets.
• UK PM: "I hope to rejoin the EU in my lifetime" (CLS, 09-30) — a rare softening of post-Brexit positioning.
• Japan braces for a fresh wave of price hikes in October (survey reporting, 10-01).

💰 Finance

Market snapshot (Sina Finance, latest close):

Index	Level	Change
S&P 500	7651.54	-0.25%
Nasdaq	26861.06	+0.24%
Dow Jones	50906.05	-0.86%
Shanghai Composite	3842.19	+0.31%
CSI 300	4357.62	+0.29%
USD/CNY	6.705	—
WTI crude	100.99	-2.07%

• US stocks close out September: the S&P 500 posted its worst month since June on persistent rate worries; the dollar had its best month since March; global bonds just logged their worst quarter since 2024.
• US August PCE inflation came in below expectations with resilient consumption, cooling October rate-hike bets. Eurozone's four largest economies all printed above-forecast September inflation — yet hike odds still fell.
• Oil holds near $100: Middle East supply is estimated near pre-war levels; US diesel prices stay elevated, with executives forecasting over a year to normalize.
• A-shares wrap up pre-holiday session: indexes mixed, innovative drugs strong while AI hardware lagged; real estate staged a deep-V reversal — Vanke A surged as much as 18% intraday to a six-month high. Banks rushed to implement the new mortgage interest-subsidy policy. Holiday tickets sold out across major scenic sites.
• Big money moves: Citadel's Griffin donated $3B to higher education, a US record; Paramount issued $30B in investment-grade bonds to fund the Warner Bros. acquisition.

🤖 Tech / AI

• Google launched Gemini 4 Argon, billed as its most advanced AI model yet (Sina, 10-01).
• The FTC is investigating OpenAI, Anthropic and other AI giants, focused on agent loss-of-control risk; the White House's top science advisor says AI companies have agreed to external audits — regulation and self-policing advancing in parallel.
• DeepSeek open-sourced its full Huawei Ascend component stack, taking aim at NVIDIA's CUDA ecosystem (CLS, 09-30).
• ElevenLabs completed a tender offer at a $22B valuation as AI voice demand heats up (CLS, 10-01).
• Micron's Q4 revenue topped $54.2B for the quarter; it expects memory supply to stay tight through 2027–2028 on AI-driven demand.
• Apple: will make a major smart-home push on October 13 with a dedicated control hub; analysts project 6M foldable iPhone Duo units sold in 2026.
• The AI growth paradox: analysts increasingly question who ultimately pays for massive AI capex.
• Boeing won the US Navy's sixth-generation fighter contract, worth over $20B; shares rose 3%+ premarket.

✍️ Thinkers

• No new posts in the past 24h from Lilian Weng, Chip Huyen, Baoyu, Manus, Anthropic Engineering, or the Qwen blog (Baoyu's latest is still the Sept 2 translation of Anthropic's commerce-agents guide). No target-blog updates today.
• Notable mover via CLS: the White House science advisor revealing the external-audit arrangement for AI companies; ECB chief Lagarde saying she doesn't rule out stepping down months early.

📝 arXiv Papers (cs.AI, 09-29 batch)

• Thinking Before Thinking: Scaling Agentic Inference Through Meta-Reasoning — promotes "controlling the execution itself" to a first-class problem: agents must repeatedly decide whether to build on partial work or restart, framed as a scalable meta-reasoning layer.
• LeapQuant / STEPQuant (companion papers) — quantizing recurrent states of linear attention (Gated DeltaNet, Kimi Delta Attention): LeapQuant delivers accurate quantization; STEPQuant maps when and where quantization errors actually matter. Deployment-relevant for the Kimi attention family.
• Beyond the Timeline — grounded entity biographies for long-video memory: linking events involving the same object across hours or days, resolving physical identity that chronological descriptions leave ambiguous.
• Skill-Space Shooting — autonomous robot policy improvement: experience-driven search over a skill space to push beyond the initial training distribution.

🐙 GitHub Trending (new repos, last 3 days, stars>100, human-filtered)

• KKKKhazix/AIHOT ⭐4050 — TypeScript — a framework that finds its own hot topics and writes its own daily digest; swap in your sources and ranking rules. Structurally identical to this briefing pipeline — worth a look.
• feder-cr/dots ⭐1931 — Python — an AI agent with its own browser that doesn't get blocked.
• openai/mcp-extensions ⭐451 — TypeScript — OpenAI's official plugin framework, making extensions first-class ChatGPT citizens.
• deepseek-ai/DeepGEMM-Ascend ⭐377 — C++ — efficient matrix-multiplication kernels for Huawei Ascend NPUs (same move as DeepSeek's Ascend open-source drop).
• rehan-remade/universal-modder ⭐739 — Python — point Claude at almost any PC game you own and let it mod it.
• OpSafari/hypoarena ⭐545 — Python — a scientific hypothesis-discovery workbench: grounded hypothesis-evidence graphs with planned synthetic literature.

Filtering note: 27 candidates in the 3-day window; ~17 spam entries (game cheats, casino, crypto bots) removed.

👀 GitHub Watch (ShuJieTai board, created in last 24h)

• crewAIInc/crewAI (priority 3): PR #7811 (merged 09-30) adds two borrowable capabilities — (1) llm_overlay now supports model keys: a model:<provider/model> mapping overrides every LLM built with that model string, including flows that construct LLM(...).call() internally, with zero code intrusion; (2) crewai eval --models "a,b" compares 1–5 models in one run, emitting a quality/cost/latency table with per-column winners. Direct lesson for the magicedit Copilot: model swapping/fallback can become a zero-intrusion global mapping, and evaluation criteria should live in a rerunnable eval.jsonc.

---

Sources: Sina Finance API / CLS / Sina US-stock desk / GitHub Search API / arXiv API / ShuJieTai board. All items are within the past 24 hours; market data is the latest close. Auto-generated by Hermes Agent.