Back to Blog
Galen Guan

Daily Briefing · 2026-10-04: Weak Jobs Report Crushes Rate-Hike Bets as Microsoft Enters AI4S Race

📰 Daily Briefing · 2026-10-04 (Sunday)

🌐 Politics · Geopolitics

  • Pentagon sends a third aircraft carrier to the Middle East — Per WSJ and other outlets (Oct 1), US officials say the Pentagon will deploy a third aircraft carrier plus more Marine vessels to the Middle East, adding 9,000–10,000 personnel. Shipping attacks near Iranian waters continue while the US blockades Iranian ports. The Middle East remains the swing variable for oil and shipping: Saudi Arabia's East-West pipeline, hit by Iraqi missiles last month, is already back above 80% capacity, and Hormuz transit volumes have surged to near wartime peaks. CLS
  • G7 to release up to 100 million barrels of diesel and crude reserves — French President Macron announced the G7 will release diesel and crude reserves within four months, up to 100 million barrels, coordinated through the IEA. Trump then said the US would not ban diesel exports. Reserve releases plus restored Saudi pipeline capacity are pressuring oil: WTI November fell 1.9% Friday to $91.11/barrel, down 1.4% on the week. CLS
  • Trump: taxpayer-funded self-promotional ads; Clayton eyed as AI czar — Trump reportedly directed the use of taxpayer funds for ads celebrating his presidency; separately he is expected to name former SEC chairman Clayton as the White House's AI lead — a signal worth watching for US AI regulatory direction. Sina · Sina
  • China: PLA Daily reiterates reunification stance — The Oct 3 PLA Daily commentary reaffirmed the reunification line, coinciding with Trump's tariff/base-fee threats toward South Korea, keeping East Asia trade and security uncertainty elevated. CLS

💰 Finance · Markets

Snapshot (Oct 3 close, Sina): S&P 500 7,722.72 (+0.73%) · Nasdaq 27,190.86 (+1.19%) · Dow 51,176.96 (+0.49%) · Shanghai Composite 3,842.19 (+0.31%) · CSI 300 4,357.62 (+0.29%) · USDCNY 6.7050 (US 10Y / DXY / gold not available)

  • September payrolls miss badly; hike bets collapse — US nonfarm payrolls rose just 29K versus expectations of 84–90K (prior 162K). Traders slashed October hike bets: swap contracts no longer price a full 25bp hike this year, and the implied probability of an October move fell from ~70% to about 20%. The Fed hiked just last month — its first in three years — against elevated energy prices and above-target inflation; the jobs slump pushes the committee toward a wait-and-see stance. Wednesday's September minutes are next week's key event for clues on whether the next hike lands in October, December, or later. CLS
  • 10-year Treasury yield touched 5.34%, a 24-year high — Weeks of global bond selling pushed long-end yields to 5.34% this week. Treasury Secretary Bessent called the rise "a global trend, nothing to panic about" — while dismissing AI-risk warnings as alarmism. High yields + high oil (Brent briefly neared $100) + AI capex expectations form a triple test for Q4 equities. Roughly three-quarters of S&P 500 members fell in September, showing how narrow the rally remains. CLS
  • Nvidia record high; Micron earnings lift semis — Friday's most-active list showed Nvidia notching its first record in over four months; Tesla's Q3 deliveries beat. US equities move to 23-hour trading in December. Sina
  • Oil: Saudi 'lifeline' back above 80% throughput — Saudi Aramco's East-West pipeline is moving ~6M bpd (7M capacity), ~4.5M bpd of it exportable, restarted within two weeks of last month's attack. JPMorgan and Goldman estimate Middle East exports have surged to near pre-war levels. CLS
  • China: holiday motorway EV charging hits record — Grid data showed 28.05 GWh of highway charging on National Day's first day, +60.4% YoY, a single-day holiday record; railways carried a record 25.2M passengers on Oct 1. Airfares have dropped ~20% from Oct 2. CLS

🚀 Tech · AI

  • Microsoft unveils Quine, a biology world model — AI4S race heats up — Microsoft Research launched Quine, an experimental AI scientist spanning genomics, proteins, chemistry, RNA, cell states and bio-imaging, wired into a closed loop of hypothesis → model reasoning → wet-lab validation → feedback. In one disclosed case it ranked thousands of compounds for their ability to shift pancreatic cancer cell states; the top candidate produced the largest expected change in validation — barely a weekend from shortlist to candidate. Microsoft is partnering with the Broad Institute. Competing: Anthropic stood up a life-sciences team on Sep 23 (Claude analyzed 200K+ reverse-transcriptase sequences down to 3,500 candidates); OpenAI is also in. AI4S is becoming the next battleground. CLS / Kechuang Board Daily
  • OpenAI safety staffer resigns, calls for 'nuclear-grade' safeguards — An OpenAI safety-team member publicly resigned, urging nuclear-power-grade safeguards for frontier AI. Sina
  • Data-center resistance spreads from the US to Europe and Asia — STL Partners estimates public opposition has put ~$42B of European and ~$77B of US data-center investment at risk (delays/cancellations). The IEA sees global data-center power use nearly doubling from ~485 TWh (2025) to ~950 TWh by 2030, with AI-dedicated capacity tripling. Europe rejected or restricted 70+ projects in Jan–Apr alone, more than all of 2025: Scotland paused approvals for hyperscale sites, Denmark passed emergency law deprioritizing grid access, and Spain proposed localization requirements. CLS
  • TSMC in talks on Musk's Terafab — TSMC is reportedly exploring cooperation with Musk's Terafab megafab project — a potential reshuffle of leading-edge capacity. Sina
  • Apple: tighter Mac privacy for the agent era; iPhone 18 Pro Max AT&T defect confirmed — Apple announced tighter Mac privacy controls, explicitly flagging rising risks from AI agents; it also confirmed an AT&T network defect on the iPhone 18 Pro Max requiring device replacement for affected users. Sina · CLS

📄 Papers · Research

(72h window; arXiv on weekend hiatus — latest batch Oct 1)

  • One Basis to Animate Them All: Gaussian Blendshape Distillation for Real-Time Avatars (Oct 1) — Distills the animation of pretrained 3D-Gaussian avatars into a shared blendshape basis, bypassing costly neural inference for real-time driving. Relevant to avatar/video generation; the "distill and reuse pretrained models" pattern is worth watching alongside backend evaluations. arXiv
  • KaliBench: Cybersecurity Tool Use on Kali Linux (Oct 1) — A fine-grained benchmark for LLM cybersecurity tool-calling, using runtime-verifiable rewards instead of human grading. Ties into the agent tool-calling evaluation thread. arXiv
  • Reconstruct, Practice, Go Real: Guided Self-Improvement for Embodied Agents (Oct 1) — A self-improvement loop for embodied agents: reconstruct skills, practice in simulation, transfer to hardware — cutting manual reward design. The "practice safely in sim, then ship" structure mirrors iterative-refinement agent loops. arXiv
  • ScholarCatalyst: Retrieving Papers That Inspire New Research (Oct 1) — Benchmarks retrieval of papers that spark new research; scientists remain far ahead of AI at sensing which buried idea matters. arXiv
  • SILSA: Sliding-Window Slice Latents for 3D Generation (Oct 1) — Replaces voxel latents + multi-stage pipelines with sliding-window slice latents for topology-preserving, high-resolution 3D generation. arXiv

🧠 Voices

  • Bessent (US Treasury Secretary): AI-risk warnings are 'alarmism' — Bessent publicly dismissed AI-risk warnings, called himself a Musk "superfan," said the bond selloff is "a global trend, nothing to panic about," and predicted no Iranian oil at sea this week. One official shaping both the AI narrative and the bond narrative — plus a proposed AI czar — makes US AI-policy turf fights worth tracking. Sina
  • Musk: Terafab courts TSMC — Talks between TSMC and Musk's Terafab could redraw the leading-edge capacity map (see Tech). Sina

📝 New Blogs (24h)

(Sunday: none of the target overseas or Chinese blogs published within 24h — Lilian Weng last posted Jul 4, Bao Yu Sep 2; Anthropic Engineering, Cognition, Chip Huyen, Qwen, Shanghai AI Lab and Jiqizhixin show nothing after Oct 3; Manus's latest are two Oct 1 posts (Video Editor, Game Developer). kexue.fm remains blocked; HuggingFace Blog unreachable on this network. Reported as-is, no padding.)

⭐ GitHub Trending

  • facebookincubator/muse-gadget-sdk (C · ⭐819 · created 10-02) — Meta's open-source SDK for building Muse hardware gadgets. A rare hardware SDK open-sourcing from Meta — worth watching for ecosystem intent. repo
  • sganggs/Stronghold-Protocol (JavaScript · ⭐330 · created 10-02) — Unofficial fan remake of Arknights' "Stronghold Protocol: Covenant" as a browser auto-chess tower defense, solo or 1–4 player co-op. Game remake; kept only for trending rank, no engineering takeaways. repo
  • QingYunA/answer-me-with-html (JavaScript · ⭐290 · created 10-02) — An agent skill that answers hard questions with a single readable HTML page. "Answer as page" is a neat skill-design idea, aligned with agent-skill packaging trends. repo
  • whirlchat/whirl (TypeScript · ⭐240 · created 10-02) — AI chat app aggregating top models with real memory, living documents, and custom tools. Its memory/documents implementation is worth a look. repo
  • youcci/playport (Kotlin · ⭐235 · created 10-01) — Wireless CarPlay in the browser: a server-side receiver turning any LAN screen into a head unit. repo
  • StayLameBro/backburner (Python · ⭐215 · created 10-01) — Use an iPhone to help a Mac run a 27B model: faster prompt reading and more context over USB-C. A fun take on heterogeneous-device co-inference. repo
  • BootLoops-ai/bootloops (Python · ⭐204 · created 10-01) — Certified computational tools for exact, high-precision physics and quantum calculation. repo

(Also culled as minor/spam: x4gpanell/SuperJinX (no description), blendi-remade/agentcraft (no description), mizorewww/x_gift_bot (X gift CLI).)

🔭 GitHub Watch · Watchlist

  • langchain-ai/langgraph: upgrade fixes checkpoint-fork replay-pollution bug (board task 10-03 20:03 UTC) — Upstream fixed replay pollution in checkpoint-fork scenarios, i.e. correctness of history replay after graph forking. Any multi-step orchestration on LangGraph risks reading dirty state on forked paths; evaluate the upgrade and run regression. repo
  • crewAIInc/crewAI: evaluate Flow's experimental Job lifecycle and Runner for long-running orchestration (board task 10-03 20:03 UTC) — crewAI added a Job lifecycle and Runner abstraction to Flow, aimed at resumable long tasks. If your long pipelines need checkpoint-resume and failure re-entry, the Runner abstraction is a useful reference; medium priority. repo

(2 new [Watch] board tasks within 24h today.)

References