newsGalen Guan

Daily Briefing · 2026-07-23

📰 Daily Briefing · 2026-07-23 Thursday

🌐 Politics & Geopolitics

  • US-Iran conflict escalates: Trump threatens to bomb Iran's bridges and power plants — Trump stated that if Iran launches weapons attacks on ships in the Strait of Hormuz, the US will destroy Iran's bridges and power plants, including those in Tehran. The US also threatened to bomb Iranian civilian infrastructure. ↳ Context: This is the most intense US-Iran rhetorical escalation in recent weeks. Oil markets surged in response. While Secretary of State Rubio indicated willingness for diplomatic resolution, the rhetoric now targets direct infrastructure strikes. Hormuz shipping safety is the key watchpoint. · CLS

  • US military threats against Iranian civilian facilities push oil prices higher on July 22 — US military threat language has expanded from "nuclear facilities" to "civilian infrastructure." Brent crude reached $122.73/barrel. ↳ Context: Oil prices are now pricing in "supply disruption" scenarios rather than mere sanctions expectations, with energy sector and shipping insurance costs likely to rise in tandem. · Sina Finance

  • Middle East tensions suppress risk assets; all three major US indices close lower — Mutual threats reignited rate hike fears. Dow -0.01% to 52,218.58, Nasdaq -0.57%, S&P 500 -0.14%. SpaceX fell 6.7% to new lows; NVIDIA bucked the trend with +2.3%. ↳ Context: Geopolitical conflict suppresses risk appetite while rising oil prices push inflation expectations higher — the "stagflation narrative" is resurfacing. Watch for the Fed's next rate signals. · CLS

💰 Finance & Markets

Market Snapshot: S&P 500 7,498.96 (-0.14%) · Nasdaq 25,690.90 (-0.57%) · Dow 52,218.58 (-0.01%) · Shanghai 3,867.03 (+0.07%) · CSI 300 4,717.24 (-0.46%) · USD/CNY 6.7639 · Brent Crude 122.73 (-0.37%)

  • Alphabet Q2 revenue hits $119.8B, up 24% YoY — Google's parent company delivered strong earnings across the board, beating expectations. Search ads and cloud business drove twin-engine growth. ↳ Context: Google remains resilient in macro headwinds — AI-powered search monetization is delivering. This earnings report is the core confidence anchor for tech stocks this week. · Sina Finance

  • Tesla Q2 revenue up 26% but profit halved; capex surges — Revenue growth masks deteriorating earnings quality as massive capex points toward new vehicle models and AI infrastructure. ↳ Context: Musk chose "growth over profitability." The key question for investors: how long will this spending cycle last, and what will it buy? With SpaceX also hitting new lows the same day, the Musk ecosystem faces broad pressure. · Sina Finance

  • AMD makes first Anthropic bet: up to $5B investment, locking in tens of billions in AI server orders — AMD will invest up to $5B in Anthropic and supply 2 gigawatts of AI servers with next-gen Instinct MI450 chips starting 2027, with total contract value in the tens of billions. ↳ Context: This is AMD's largest single AI chip win to date — not just chip orders, but an investment that locks in the hottest AI customer outside of OpenAI. NVIDIA's AI chip monopoly is showing a real crack. · CLS

  • OpenAI sharply raises computing spend forecast: from $600B to $750B — OpenAI revised its cumulative AI infrastructure spend through 2030 to $750B and announced a $20B investment in its first self-designed data center. ↳ Context: Two upward revisions in one month — the physical scale of the AI arms race has shifted from "server count" to "datacenter gigawatts." The money ultimately flows to chip makers and power infrastructure. · CLS

  • China A-share securities stamp tax surges 97% in H1, a decade-high — H1 stamp tax revenue reached ¥154.9B, up 97.3% YoY. May-June monthly growth rates both exceeded 145%. ↳ Context: Stamp tax is a direct proxy for market activity — A-share trading volumes were exceptional in H1. Active equity funds are approaching ¥4 trillion, with inflows highly concentrated among top-tier managers. Tech exposure was the decisive factor. · CLS

🚀 Tech & AI

  • AMD × Anthropic strategic partnership: MI450 chips secure 2 GW of AI compute — AMD's first hyperscale customer commitment in AI training chips, with deliveries starting H1 2027. AMD simultaneously invests up to $5B in Anthropic. ↳ Context: NVIDIA+OpenAI vs AMD+Anthropic — the AI compute supply chain is forming a bipolar landscape. The key question: can MI450 approach Blackwell performance? That determines how far Anthropic's compute sovereignty can go. · CLS

  • OpenAI's first self-designed data center: $20B investment — OpenAI's largest single capex in history, transitioning from cloud rental to self-built infrastructure — a milestone signaling AI companies taking control of their infrastructure destiny. ↳ Context: Moving from "renting cloud" to "building your own" is a maturity milestone for AI companies. The logic: training next-gen models requires purpose-built architecture, not commodity cloud. Self-built data centers are the physical ticket to AGI. · CLS

  • Amazon cuts AGI positions — While Alphabet and OpenAI invest heavily in AI, Amazon is shrinking its artificial general intelligence team. ↳ Context: Amazon is taking a differentiated path in the AGI race — scaling back frontier research to focus on commercializing AI in e-commerce and AWS. Does this mean the Alexa AGI roadmap has been internally abandoned? Worth watching. · Sina Finance

  • Apple plans full Mac lineup overhaul — In response to the AI PC wave, Apple plans a comprehensive Mac refresh. ↳ Context: The first hardware update post-WWDC. The core question: will these Macs have dedicated optimizations for local LLM inference? Apple's on-device AI strategy will define the next growth narrative for M-series chips. · Sina Finance

🧠 Research & Papers

  • arXiv: "Copy Less, Ground More" — Overcoming repetitive copying in long-context reasoning via evidence-aware RL — Researchers identify a critical failure mode where LLMs fall into "repetitive copying" rather than genuine understanding during long-context reasoning. They propose an evidence-aware reinforcement learning method that teaches models to "copy less, think more." ↳ Context: Context windows keep growing, but are models really "reading" or just "copy-pasting"? This paper targets the core deficiency in current LLM reasoning — larger windows don't mean deeper understanding; new training paradigms are needed. · arXiv

  • arXiv: "Appearance Pointers" — Multimodal region control for Diffusion Transformers — A new method enabling creative professionals to precisely control materials, object identities, and spatial arrangements in specific image regions, going beyond text-only prompting. ↳ Context: Controllable generation is key to moving AI art from "gacha" to professional tooling. This work lets designers specify exactly what goes where — direct applications in advertising and film/TV industries. · arXiv

  • arXiv: "Agents in the Wild" — Where research meets deployment (comprehensive survey) — A systematic survey of LLM-based agent systems transitioning from lab prototypes to production deployment, covering current status, challenges, and best practices. ↳ Context: Agent deployment is the biggest AI theme of 2026. This survey spans software engineering, scientific discovery, and finance — a great panoramic view of what agents can actually do today. · arXiv

📝 New Blog Posts (24h)

  • No significant new blog posts today — In the past 24 hours, major AI engineering blogs — Anthropic Engineering, Lilian Weng, Cognition, Manus, Baoyu, DeepSeek, Qwen, Jiqizhixin — have all been quiet. The latest HuggingFace blog post is from July 16 ("Security Incident Disclosure"). ↳ Context: Wednesday (US time) is typically an active day for AI blogs, but this week is unusually quiet — possibly a "quiet period" as major labs prepare for their next product or model launches.

⭐ GitHub Trending

  • Jakubantalik/thinking-orbs (TypeScript · ⭐573) — Animated "thinking" loading indicators for AI chat and agent UIs, with six states and auto dark mode. ↳ Context: Every AI product is racing on "typing feel" and "thinking feel" — this tool covers the full state spectrum from "thinking" to "done." Good reference for agent product frontend UX. · repo

  • gnipbao/story-to-handdrawn-video (JavaScript · ⭐390) — Agent skill: converts Chinese stories or illustrations into hand-drawn diary-comic-style animations (silent MP4). ↳ Context: AI Agent × content creation crossover — automatically turning text stories into hand-drawn videos. Interesting implications for children's education and social media content production. · repo

  • hahhforest/pi-textbook (TypeScript · ⭐128) — "Hands-on Pi": a Chinese textbook for building a Pi-style Agent from scratch, following 15 real checkpoints. ↳ Context: Chinese-language AI Agent educational resources are scarce. This textbook follows a "learn by doing" approach rather than pure theory — modeled after Andrej Karpathy's teaching style. Great resource for Chinese-speaking developers entering agent development. · repo

  • 0xwilliamortiz/andrej-karpathy-skills (⭐164) — Behavioral guardrails for Claude Code, distilled from Andrej Karpathy's field notes on where LLMs go wrong when writing software. ↳ Context: Karpathy's "lessons learned" compiled by the community into reusable skill packs — essentially a catalog of LLM coding anti-patterns. Directly useful for developers using Claude Code or coding agents. · repo

(3 additional items not listed: risa-labs-inc/BossConsole — multi-platform AI agent console in Kotlin, non-Electron; tanishq-dubey/macos-laguna-s2.1 — Poolside Laguna S 2.1 community benchmarks; realfishsam/agent-notch — open-source vibe-island alternative. Several crypto-related repos were filtered out.)