Back to Blog
Galen Guan

Daily Briefing · 2026-08-13 Thursday

📰 Daily Briefing · 2026-08-13 Thursday

🌐 Politics · Geopolitics

  • Ukraine strikes Novorossiysk in coordinated assault on Russia's last Black Sea naval stronghold — Overnight Aug 12, Ukraine launched a coordinated sea drone, jet drone, and Neptune missile strike on Novorossiysk, Russia's primary Black Sea Fleet outpost after losing Sevastopol. At least 3 killed (including a child), 4 Russian warships damaged, 2 grain terminals halted. A 141-meter vessel reportedly "disappeared" near the port. Zelenskyy called it a "unique operation." Novorossiysk is also a critical hub for Russian grain exports, marking an economic warfare escalation. VP Vance reportedly asked Ukraine not to strike non-Russian tankers in the Black Sea. This significantly raises Ukraine's deep-strike capability against Russian naval infrastructure. [Reuters / AP / Al Jazeera / Kyiv Independent]

  • Trump pivots to sanctions for Iran as military strategies falter; Hormuz deal hopes fade — As of Aug 11, Trump shifted from military strikes to economic sanctions against Iran, demanding Tehran pay war reparations while Iran counter-demands US compensation and refuses to reopen the Strait of Hormuz. Pakistan is mediating and signals progress on a deal. The US-Iran war has been escalating since the Feb 28 joint US-Israeli strike that killed Khamenei, now reaching a stalemate. Trump told Axios "we are low-keying it." Hormuz remains a critical chokepoint for global oil. Both sides trading reparation demands reads as negotiation positioning rather than imminent de-escalation. [NPR / AP / Reuters / Axios / Al Jazeera]

  • US forces on track for full Iraq withdrawal by September 30 — US officials confirmed Aug 12 that American forces are on track to complete their pullout from Iraq by Sept 30, 2026, ending a 23-year military presence. The withdrawal includes Iraqi Kurdistan (Erbil). German forces are also leaving Erbil. This caps the US-led coalition mission that began in 2003. The withdrawal is linked to Iraq's parallel demand that pro-Iran militias also disarm by Sept 30. Experts warn ISIS is "actively waiting in the gaps." At a time when the Iran war has destabilized the broader Middle East, the departure reshapes the regional power balance. [AP / Stars and Stripes / The National]

  • Former Chinese Premier Zhu Rongji dies at 97 — Zhu Rongji, China's premier (1998-2003) who drove market reforms and led China into the WTO, died on Aug 12, 2026, at age 97. Xinhua confirmed the death; the CPC Central Committee, State Council, and NPC issued a joint obituary. Zhu was the architect of China's SOE reforms, financial system overhaul, and WTO accession (2001) that transformed China into a global trading power. His death is symbolically significant amid current economic headwinds and debate over reform direction. Watch for memorial arrangements and any signaling about economic policy direction in eulogies. [Xinhua / Reuters / SCMP / AP]

  • Lebanon passes general amnesty law and abolishes death penalty — first Middle East country to do so — Lebanon's parliament passed a sweeping general amnesty law on Aug 12 (first in 35 years), and voted to abolish the death penalty on Aug 11 — becoming the first Middle East/Arab country to do so. Existing death sentences replaced with life imprisonment. The amnesty aims to reduce severe prison overcrowding. These twin legislative moves mark a major shift in Lebanese criminal justice. The amnesty is domestically controversial; the death penalty abolition is a regional milestone with potential demonstration effects for other Arab states. [Al Jazeera / AP / Asharq Al-Awsat / The National]

  • Drone strike hits Libya's Zawiya oil refinery — A drone strike on Aug 11 hit the Zawiya oil refinery in western Libya, igniting a blaze at a power substation and triggering widespread outages. The fire was still burning as of Aug 12. Zawiya is Libya's key western oil installation. The strike disrupts fuel supply and electricity in a country already wracked by political division. No group claimed immediate responsibility. [AP]

  • Zelenskyy: Russia receiving more North Korean missiles for Ukraine war — Zelenskyy stated Aug 11 that Moscow is receiving increasing supplies of North Korean missiles. Russian attacks killed at least 10 Ukrainians the same day. The UN estimates over 16,000 Ukrainian civilians detained by Russia remain in custody. Deepening Russia-DPRK military cooperation extends a trend from 2024-2025. North Korean missile supplies sustain Russia's long-range strike capability. [AP]

💰 Finance · Markets

Market Snapshot (US close Aug 12 / China close Aug 13): S&P 500 7,751.68 (+0.30%) · Nasdaq 26,599.25 (+0.58%) · US 10Y 4.676% · DXY 99.99 · VIX 14.59 (lowest since Jan 9) · Shanghai Composite 3,946.68 (+0.32%) · CSI 300 4,690.92 (+0.58%) · USD/CNY 6.7330 · Gold $4,460.50/oz · WTI $83.16

  • July CPI cools to 3.4% y/y as inflation continues slow grind lower — US CPI rose 0.1% m/m in July (as expected), with the year-over-year rate easing to 3.4% from 3.5% in June. Core CPI rose 0.2% m/m (prior 0.0%). MBA 30-yr mortgage rate fell to 6.77%. Combined with a soft July jobs report (-23,000 payrolls), markets increasingly price a Fed rate cut later in 2026, supportive for bonds and rate-sensitive equities. [CNBC / BLS]

  • US 10-year Treasury auction hits highest yield since 2007 — 10-year Treasury yield stands at 4.676% (30Y at 5.24%, 5Y at 4.366%). A 10-year note auction set the highest yield since 2007. Elevated long-end yields reflect persistent term premium and heavy Treasury supply (July budget deficit $432.3B vs $346B estimated). Higher long rates are a headwind for growth/tech equity valuations but support USD strength. [EastMoney / Reuters]

  • AI infrastructure earnings blowout: CoreWeave +19%, Nebius +29% — CoreWeave (CRWV) +18.8% to ~$107.3, Nebius (NBIS) +28.5% to ~$248.4, Super Micro (SMCI) +18.2%, Lumentum (LITE) +16.5%, Cerebras (CBRS) +11.5% after strong Q2 results. NVIDIA (NVDA) +2.8%. AI compute/datacenter demand is re-accelerating, lifting the entire AI hardware chain. A key driver of the Nasdaq's +0.58% outperformance vs the Dow's flat session. Signals continued capex-driven growth for semiconductors and networking. [Yahoo Finance / Motley Fool]

  • PBOC pledges more countercyclical stimulus; A-shares rally broadly — The People's Bank of China said it will "promptly roll out pragmatic incremental policies and strengthen countercyclical adjustment." On Aug 13, A-shares rallied broadly: Shanghai Composite +0.32%, CSI 300 +0.58%, ChiNext +1.49%, over 4,100 stocks advanced. Real estate and consumer sectors led. China's Q2 GDP growth hit its slowest since 2022 and July factory activity unexpectedly contracted. The PBOC's policy pledge supports A-share and H-share risk sentiment. Gold reserves increased for a 21st consecutive month (de-dollarization strategy). Watch for potential rate cuts or RRR adjustments. [EastMoney / PBOC / Cailian Press]

  • US credit card debt hits $1.26 trillion in Q2 2026, near record — Total US credit card balances reached $1.26 trillion in Q2 2026, approaching a record high. Rising revolving consumer debt amid a cooling labor market (July lost 23k jobs) signals consumer strain at the margin — a caution flag for consumer-discretionary equities and a reason for the Fed to ease, though sticky services inflation limits how fast it can cut. [Yahoo Finance]

  • US July budget deficit $432.3B, far above $346B estimate — The US federal government posted a $432.3 billion budget deficit for July 2026, versus a consensus estimate of $346 billion. A wider-than-expected deficit points to continued heavy Treasury issuance, contributing to the elevated 10-year yield (4.676%). Fiscal slippage is a structural headwind for bonds. [EastMoney (Bloomberg data)]

  • VIX breaks below 15, lowest since January 9 — CBOE Volatility Index fell 4.5% to 14.59, its lowest close since January 9, 2026. Extremely low implied volatility reflects complacency after the CPI print and strong AI earnings. Michael Burry has flagged the historically becalmed market as a "red flag." Low VIX supports carry and equity risk but signals crowded positioning. [Yahoo Finance / EastMoney]

  • Fed quiet this week: last FOMC July 29, Chairman Warsh, five new task forces — No FOMC meeting or major Fed speech in the Aug 12-13 window. The last FOMC press conference was July 29, 2026. The Fed has announced five task forces to examine monetary policy conduct. Fed Chair is Kevin Warsh. The Fed is in data-dependent mode. July CPI (3.4% y/y) and weak jobs data keep rate-cut expectations alive but not imminent. Five new task forces signal a possible review of the monetary policy framework. [Federal Reserve]

🚀 Tech · AI

  • DeepSeek V4 Pro 0813 GA release: 1M context at 1/4 of frontier pricing — DeepSeek released V4 Pro 0813 as the GA version of their large-scale MoE model. Priced at $0.435/1M input (cache miss) and $0.87/1M output — roughly 1/4 to 1/5 of comparable frontier models. At 93% cache hit rate, effective input price drops to ~$0.039/1M. Supports 1M token context, thinking/non-thinking modes, tool calls, JSON output. Scored 45.3 on Artificial Analysis Intelligence Index. Hit #1 on Hacker News (273 points). API is compatible with both OpenAI and Anthropic SDK formats (including api.deepseek.com/anthropic endpoint), enabling drop-in replacement for Claude Code and GitHub Copilot. P50 throughput 57 tok/s, tool call error rate 1.8%. Developer consensus (HN 288 pts, 89 comments): benchmarks below Fable 5 and Sol but competitive with Opus 4.8, at ~20x lower cost — in agentic workflows, effective cost is ~60x cheaper per request ($0.000875 vs $0.052 for Opus). [OpenRouter / DeepSeek API Docs / Hacker News]

  • Qwen3.8-2.4T-A95B open-weights: first Qwen-Max-class model open-sourced, 2.4T parameter MoE — Alibaba's Qwen team released Qwen3.8-2.4T-A95B, the first Qwen-Max-class model to be open-released. 2.4T total parameter MoE with 95B activated parameters, 512 experts (10 routed + 1 shared), 262K native context extensible to 1.01M tokens. Uses a novel hybrid architecture with Gated DeltaNet (linear attention) and Gated Attention layers. Hit #15 on HN (200 points), trending on HuggingFace. PaperBench 93.0 (vs Claude Fable 5 88.8, GPT-5.6 Sol 90.5), Terminal Bench 2.1 at 86.6, SWE-bench Pro 67.7. 1-bit quant is 397GB — HN commenters note it "literally puts Opus 4.5 performance level into a machine a normal person could buy." A milestone: no prior open model has matched frontier-tier agentic performance at this scale. [Hugging Face / Hacker News]

  • Meta Muse Glimmer 30B: Apache-2.0 agentic model for consumer hardware, 233 tok/s on RTX 5090 — Meta Superintelligence Lab released Muse Glimmer, a 30B parameter dense causal transformer with integrated 1.8B ViT-G/14 perception encoder, purpose-built for autonomous agentic tasks on consumer hardware. Apache 2.0 license. 4-bit quantization requires only 24GB VRAM (~1% degradation). #1 trending on HuggingFace (1,250 likes). Ships with DFlash block-diffusion speculative decoding drafter: 233 tok/s on RTX 5090 (3.1x), 50 tok/s on M5 Max (1.8x). Compatible with OpenClaw and Hermes Agent scaffolds. Fills the gap between 3B-8B efficient models and 100B+ frontier open models for local agent use cases. [Hugging Face]

  • Lovable raises $400M Series C at $13.3B valuation — AI app builder Lovable raised $400M Series C at $13.3 billion valuation, led by Menlo Ventures, co-led by EQT's Scaleup Europe Fund. Investors include Tencent, Kaszek Ventures, Balderton Capital. Since Nov 2024 launch: 60M+ projects created, 900M+ monthly visits to Lovable-built apps, nearly 2/3 of Fortune 500 using it. One of the largest valuations for an AI application-layer company. Platform now handles payments, SEO/AI-search, Google Workspace/Salesforce/Stripe integrations, AIUC-1 security certification. Enterprise users include Adidas, NVIDIA, Deutsche Telekom. Roadmap: proactive agents, multi-model routing, post-training open-source models. Signals the AI app builder category maturing from prototyping to business-critical infrastructure. [Lovable Blog / Hacker News]

  • xAI Grok 4.6 scores 61 on Artificial Analysis Intelligence Index, ties GPT-5.6 Sol — Grok 4.6 (high reasoning) scores 61, ranking #6 out of 183 models. Ties GPT-5.6 Sol (61), behind only Claude Opus 5 (63) and Claude Fable 5 (62). Priced at $2.00/1M input, $6.00/1M output, 500K context window. Output speed 86 tok/s. Grok 4.6 enters the top tier of frontier models but is priced higher than Chinese competitors (DeepSeek V4 Pro at $0.435/M input). 75% cache discount ($0.50/1M) helps for repeated-context use cases. Within 2 points of #1 (Claude Opus 5 at 63). [Artificial Analysis / Hacker News]

  • NVIDIA Nemotron 3.5 Lightning: Mamba-2+MoE+Attention hybrid, 3B active/30B total — NVIDIA released Nemotron 3.5 Lightning (30B total, 3B active) with a novel hybrid architecture combining Mamba-2 state-space layers, MoE, and selective Attention layers. Runs on a single DGX Spark (GB10) or H100. Supports 1M token context. NVFP4 quantized version fits in 18GB. OpenMDW-1.1 license (commercial use allowed). SWE-bench Verified 52.8%, MMLU Pro 81.6%, GPQA Diamond 75.6%. NVFP4 format runs on Blackwell GPUs including consumer RTX 5090. Positioned as a "sub-agent workhorse" for long-running autonomous agents. [Hugging Face]

  • Tailscale traces production database corruption to 16-year-old SQLite WAL-reset bug — Tailscale engineering traced a database corruption issue to a 16-year-old bug in SQLite's WAL (Write-Ahead Log) reset mechanism. Reached #2 on Hacker News (423 points, 64 comments). SQLite is embedded in virtually every OS, browser, and mobile device — a WAL corruption bug affects billions of installations. Tailscale's investigation methodology demonstrates how distributed-system failures can surface latent bugs in foundational libraries. The disclosure will likely trigger patches across the SQLite ecosystem. [Hacker News / Tailscale]

🧠 Thinkers · Opinions & Research

  • Dwarkesh Patel: Recursive self-improvement is "the most important question in the world right now" — Published Aug 12, a 2h12m podcast with Ryan Greenblatt (Chief Scientist, Redwood Research) debating whether AGI triggers a rapid slingshot to superintelligence. Dwarkesh said he's been "historically skeptical" but Greenblatt "made a pretty good case." Greenblatt's median for automating AI R&D is 2031; estimates 4-5 years of AI progress compressed into a single year once R&D is automated. Core debate: is AI R&D verifiable enough for recursive self-improvement? Dwarkesh worries current alignment specs won't make superintelligences "personal advocates and guardian angels." Discussion of recent AI collusion/deception and reward hacking extrapolation to takeover scenarios. 53K views in 1 day. [Dwarkesh Podcast (Substack + YouTube)]

  • HN consensus: DeepSeek V4 Pro delivers Opus 4.8-class performance at 1/20 the cost — HN thread (288 pts, 89 comments) on DeepSeek V4 Pro GA. Developer consensus: benchmarks below Fable 5 and Sol but competitive with Opus 4.8 at ~20x lower cost. With cache-read discounts in agentic workflows, effective cost reaches ~60x cheaper per request ($0.000875 vs $0.052 for Opus). Developers note V4 Flash remains the value sweet spot — "crossed a threshold of production grade coding." Pro only worth it for complex tasks Flash can't handle. HN commenters observe DeepSeek "just releases when ready" rather than strategic timing. [Hacker News]

  • HN mega-thread: "AI is removing the middle class of software engineering" (438 pts, 368 comments) — Blog post by Florian Herren (see Blogs section) arguing AI accelerates failure in teams with weak engineering culture: "people can prompt an agent for a few hours and open a PR... to the untrained eye, it works." Most-discussed HN story of the day. Commenters cite Oracle banning AI-generated code from JDK (Aug 7) as evidence. Split between "debt is debt, sometimes worth it" and "AI psychosis is real." Author engaged: "they can move so fast that nobody has a realistic chance to stop them before the damage compounds." [Hacker News]

📝 Fresh Blogs (24h)

  • Florian Herren · "AI is removing the middle class of software engineering" — Argues AI coding tools eliminated the "speed limit" on code production, creating unsustainable cognitive debt. Key insight: "the cost of ownership didn't drop even as cost of writing dropped." AI removed the "middle class" of engineers who translate specs into code, leaving only senior architects and junior "vibe coders." 435 points on HN in 4 hours, 366 comments. Featured by Simon Willison on his blog today. Resonating strongly: "you can't afford bad engineers anymore because AI-generated code has made debugging exponentially harder." Read

  • DeepSeek · V4 Pro 0813 GA Release — Official GA release announcement. $0.435/$0.87 per 1M tokens pricing, 93% cache hit rate, 1M context. Top consumer on OpenRouter is Hermes Agent (186M tokens). #1 on HN (289 pts, 88 comments in 2 hours). Already seeing heavy production use across coding agents. OpenRouter

  • Tailscale · "How we tracked down a 16-year-old SQLite bug" — Full investigation writeup from symptoms to root cause. Tailscale engineering traces database corruption to a 16-year-old bug in SQLite's WAL reset logic. #2 on HN (434 pts, 69 comments). Finding a 16-year-old bug in SQLite — one of the most widely deployed databases — is remarkable. Deep technical dive into the debugging methodology. Tailscale Blog

  • LiquidAI · LFM2.5-VL-3B: Better and Faster Vision for the Edge — 3.1B params, pairs SigLIP2 400M NaFlex encoder with LFM2.5-2.6B backbone, trained on ~34T tokens. Adds screen/UI understanding, grounding, multi-image, function calling. Leads sub-5B class on real-world image tasks (MMStar 63.3, RealWorldQA 73.1). Direct-answering (non-reasoning) design keeps responses fast for real-time on-device apps. Open weights. Hugging Face Blog

  • AI2 · OlmoEarth Embeddings: Custom Embedding Exports — AI2 released custom embedding exports for OlmoEarth Studio. Compute/export int8 COG embeddings from Sentinel-1/2 satellite imagery at 10-80m. Demonstrates few-shot land-cover segmentation from 60 labeled pixels (F1=0.84), wildfire change detection. The 60-pixel few-shot result shows strong ecological signal in foundation model embeddings. Open source models, weights, and tech report released. Hugging Face Blog

⭐ GitHub Trending

  • cathrynlavery/diagram-design (HTML/SVG · ⭐ 9,453 · today +2,951) — 29 editorial diagram types for Claude Code — self-contained HTML + SVG, no Mermaid slop. 31% of lifetime stars landed in a single day — genuine viral breakout. Claude Code can reason about diagrams but previously had no clean output format; this fills that gap. Deliberately positioned against Mermaid's "AI slop" output with an editorial, no-shadows aesthetic. repo

  • NVIDIA-NeMo/Switchyard (Rust · ⭐ 724 · today +370) — New repo from NVIDIA NeMo team. 51% of lifetime stars in 24 hours. Extreme day-one growth from NVIDIA's official org signals it addresses a gap the community has been waiting for. Given the NeMo team's focus on LLM training/serving/inference tooling, likely an agent orchestration or model-routing component. No README description yet — worth watching closely. repo

  • semantica-agi/semantica (Python · ⭐ 5,569 · today +834) — Graph-native infrastructure for context and accountable AI systems. Gives AI agents persistent, queryable context — storing relationships, entities, and provenance. Targets the "amnesia problem" where agents lose state across sessions. 15% of total stars in one day, active multi-contributor development including Claude as a co-author. repo

  • stablyai/orca (TypeScript · ⭐ 43,670 · today +1,215) — Agent Development Environment (ADE) for running a fleet of parallel coding agents with your own subscriptions. Desktop, mobile, and VPS. Key value: BYO model subscriptions rather than per-seat licensing. Manage a "fleet" of agents working on different repos/tasks simultaneously. Sustained 1,215/day growth indicates real product-market fit in multi-agent orchestration. repo

  • AntigmaLabs/ante (Rust · ⭐ 1,310 · today +175) — Self-contained agent harness with optimized Rust core — works like Claude Code or Codex with none of their dependencies or model constraints. Differentiates on binary footprint and speed. Works with any model provider. Early-stage but addresses growing demand for lightweight, portable agent tooling in constrained environments. repo

Sources