Back to Blog
Galen Guan

Daily Briefing · 2026-11-06: AI Selloff Day Two — Nasdaq Plunges 5.5% Into Correction

📰 Daily Briefing · 2026-11-06 (Friday)

🌐 Politics · Geopolitics

  • DF-61 ICBM now operational with the Rocket Force — On the 66th anniversary of the DF-1's first launch, state media confirmed the new DF-61 land-based intercontinental ballistic missile has entered active service and is conducting regular field training: mobile maneuvers, position changes, and live launch drills. The missile debuted at the September 2025 military parade; the pace from public debut to operational readiness is far faster than the traditional ICBM cycle. Analyst Shao Yongling reads the speed as evidence the missile was fully developed, tested, and reliability-proven before its debut, and that it integrates with existing command, communications, and support chains. China's land-based nuclear retaliatory force gains mobility and penetration capability. CLS
  • China's commerce ministry: open to buying more US farm goods — but only if reciprocal tariff cuts land first — At a regular briefing, spokesperson Huang Ling said China is open to expanding agricultural purchases from the US, preconditioned on the agreed "30-billion-for-30-billion" reciprocal tariff-reduction arrangement being implemented. She also pushed back on the EU's new €2 handling fee on low-value imports (effective Nov 1), urging Brussels to weigh the impact on cross-border e-commerce. Farm purchases have historically been China's opening concession in trade talks; tying them to tariff reciprocity means Beijing is linking any purchase commitment to concrete US tariff action. CLS

💰 Finance · Markets

Market snapshot (Sina Finance, 2026-11-06 05:34–06:01 CST = Thursday's US close): S&P 500 7353.73 (-4.36%) · Nasdaq 24,315.00 (-5.52%) · Dow 51,523.06 (-3.44%) · Shanghai Composite 3689.91 (+0.10%) · CSI 300 4113.41 (+0.29%) · USDCNY 6.6680 · Brent / US 10Y / DXY / spot gold: not retrieved (gold rebounded above $4,100 per a single CLS report)

  • AI selloff, day two: Nasdaq -5.5%, officially in correction — US tech sold off en masse Thursday: Nasdaq -5.52%, S&P 500 -4.36%, Dow -3.44%; the Philadelphia Semiconductor Index fell as much as 9% intraday to 9,331. Nvidia and Broadcom dropped for a second straight session, with data-center and cloud-supply-chain names down in sympathy. Three compounding drivers: ① mounting doubts about returns on massive AI capex; ② high financing costs slowing new data-center construction; ③ post-election regulatory risk — with Democrats sweeping Congress, UBS estimates ~30% of planned 2028 US data-center capacity could face construction pauses. Tuesday's election rally evaporated in one session. CLS
  • The storage supercycle meets skepticism: 80% margins still sold off — SK Hynix, Kioxia, and Samsung just posted record quarterly profits (Hynix and Kioxia at ~80% Q3 operating margins) — and their stocks cratered anyway: Seagate -8.13%, SanDisk -7.91%, Micron -7.00%, WD -6.77% in the US; SK Hynix -9.46%, Samsung -7.57%, Kioxia -13.14% in Asia; A-share CXMT -3.54% with ¥2.49B net outflow. The core issue is fading pricing momentum: TrendForce pegs commodity DRAM contract-price growth at 58–63% in Q2 2026, narrowing to 13–18% in Q3 and 10–15% in Q4. The market's real question is how long price hikes can last from an all-time-high margin base, and whether the cycle extends into 2027. The election was a trigger, not the cause. CLS
  • A-shares rally on for a second day, diverging sharply from Wall Street — The ChiNext rose nearly 1% Thursday with over 4,800 stocks up; compute-leasing names surged and solar rallied, while coal and oil lagged. Shanghai Composite +0.10%, CSI 300 +0.29%. The AI selloff isn't infecting Chinese equities — money is rotating into "picks-and-shovels" branches like compute leasing instead. Whether the divergence holds depends on tonight's US session and northbound flows. CLS market wrap
  • Ackman: Trump could pull off "the greatest real estate deal ever" — Bill Ackman argued Trump could release Fannie Mae and Freddie Mac from government conservatorship, creating $500B in value. Note this aligns with his long-standing long position in the GSEs — single-source commentary for now. Sina Finance

🚀 Tech · AI

  • Compute's narrative shift at CIIE: from specs to deployable systems — Samsung, SK Hynix, Intel, and AMD shared the stage at the China International Import Expo, all pointing the same direction: competition is moving from chip specs to system-level deliverability — cooling, power, full racks. Memory vendors scrambling to prove system value amid the margin-peak debate dovetails with the storage selloff above. CLS
  • Qualcomm's outlook disappoints; analysts turn lukewarm — Qualcomm issued a below-consensus profit outlook, leaving analysts unenthusiastic: soft phone SoC demand persists while PC and automotive revenue isn't yet big enough to pick up the slack. The AI-phone story isn't showing up in Qualcomm's numbers. Sina Finance
  • AI agents haven't changed how people book travel, says Expedia CEO — Agentic AI has yet to alter actual booking behavior, per Expedia's chief executive — consumer habit change is lagging industry hype. When the flagship vertical for booking agents pours cold water on its own category, agent startups should take note. Sina Finance

📄 Papers · Research

(72h window; arXiv cs.AI Thursday batch — no new batch on Fridays)

  • FrontierPhysics: benchmarking agents for end-to-end physics research (arXiv 2611.06910) — Existing physics benchmarks test fixed questions, isolated subtasks, or paper reproduction from given solutions; FrontierPhysics evaluates agents on full research cycles including research-quality checks. Together with BEAKER, same-day drops signal agent evaluation moving from "answering questions" to "doing science." The evaluation design is structurally reusable for assessing autonomous Copilot workflows. arXiv
  • BEAKER: an expert-curated benchmark for embodied brains in self-driving chemical labs (arXiv 2611.06911) — Self-driving chemical laboratories require machines to perceive lab scenes, reason about experimental state, and act under physical, procedural, and safety constraints. BEAKER, curated by domain experts, measures exactly that — embodied-agent evaluation extending beyond software domains (browsers, terminals) into physical experimentation. arXiv
  • Perceptive Humanoid Motion from Simulation (arXiv 2611.06908) — The bottleneck for generalist humanoid autonomy is data: mapping pixels plus free-form language to low-level joint commands. This work proposes generating perceptive humanoid motion data from simulation, attacking that bottleneck directly. arXiv

🧠 Thinkers · Voices

  • No signal-bearing statements in the past 24h from watchlist figures (Karpathy, LeCun, Altman, Jensen Huang, et al.) were retrievable today: the social-listening tool is unavailable in the cron environment and the blog channels are quiet. The only single-source voice captured was Ackman on the GSEs (filed under Finance).

📝 New Blogs (24h)

  • No updates in the past 24h from any target blog (Lilian Weng / Anthropic Engineering / Cognition / Manus / Baoyu / DeepSeek / Qwen / Chip Huyen). Lilian Weng's latest remains July 4's "Harness Engineering for Self-Improvement"; Baoyu's latest is from Oct 20. HuggingFace Blog fetch timed out and couldn't be verified; kexue.fm remains blocked (~110 bytes).

⭐ GitHub Trending

(Search API, 3-day window stars:>100: 192 hits; top 10 by stars manually filtered — 7 of 10 were auto-uploaders, OSINT people-search, wallpaper packs, and similar low-substance projects, excluded)

  • KKKKhazix/agent-autoloop (TypeScript · ⭐3,217 · created 11-03) — Lets CLI agents pick up work in the background and see it through, only pinging you with a decision card when something genuinely needs your call. Addresses the human-in-the-loop interruption problem for long agent tasks: background autonomy plus critical-decision escalation. 3.2K stars in three days says the pain is real. The interaction model is directly relevant to托管 long-running Copilot tasks. repo
  • andrew-babo/muse4ai (TypeScript · ⭐805 · created 11-03) — Lets your AI agent generate images and videos via a local Windows REST bridge to Muse. Wraps a commercial video-generation model as a local REST service callable by agents — squarely in the AI-video space, worth studying for its bridge-layer interface design. repo
  • brightsongwrath/personal-data-deletor-osint (JavaScript · ⭐649 · created 11-03) — Sends data-deletion requests to 970 data brokers. Free, local, no account, no server. A practical privacy utility; fully local operation is the differentiator. Modest technical depth but fast growth. repo

📡 New-Term Radar

  • No new terms captured today (the radar ledger docs/local/radar-terms.jsonl was not found, so the radar degraded to trending-page scanning only; no new-term faces in the top 10). Ledger status: pending rebuild.

🔭 GitHub Watch · Watchlist

(ShuJieTai board, AI Video Watch project: 3 new tasks in the past 24h; "no actionable improvements" digest excluded)

  • hypit-ai/hypit: evaluate @hypit/bach declarative video-generation DSL as a Copilot video-orchestration reference — hypit v0.3.2 (11-05) promotes @hypit/bach to a default dependency. Bach 1.0 Pro is a declarative video-graph DSL: SVML Surface tags like TextVideo/FrameVideo/ReferenceVideo, with first/last-frame interpolation, reference image/video/audio injection, and seed/resolution/aspect-ratio parameters; elements produce .video artifacts for downstream reference. Why: Copilot currently generates multi-segment video through prompt conventions, with no validated declarative intermediate representation. Bach's tag graph (frame references, durations, constraints like "only the final frame needs 4K") is a mature consistency solution. See packages/bach's surface.ts + validation.ts. Priority 3 (adapt). repo
  • stanfordnlp/dspy: borrow the tool-result media degradation pattern (content blocks + supported_tool_result_media) — dspy PR #10737 (11-05): media like dspy.Image in tool messages is no longer JSON-escaped into text but expanded into content blocks; when the backend doesn't support media in tool results, a placeholder is left and the media attaches to the following user message; LMs declare capability via the supported_tool_result_media routing policy. Why: when Copilot tools (canvas screenshots, video frames) return media to a backend that can't accept it, we need this same degradation-plus-capability-declaration mechanism. Priority 3 (adapt). repo
  • Comfy-Org/ComfyUI: track the comfy-agent thin-host contract pattern (priority 2) — ComfyUI PR #17454 (11-05) adds app/agent_host.py: discovers and loads comfy-agent packages via the comfyui.agent entry-point v1 protocol, registers agent routes on the host aiohttp app, and broadcasts status via feature flag + websocket agent_status. Agent load/start failures degrade only the agent itself — never blocking ComfyUI startup — with failure reasons written to state rather than silently swallowed. Why: a template for extending Copilot with new tools/sub-agents — a thin host contract plus status broadcasting, with fault isolation to the extension itself. Priority 2 (track). repo

References