## 🔍 Core Insights This week the AI industry showed a pattern of "technological leaps and risk repricing" running in parallel: on one hand, **Nvidia** boosted **Kimi K3** inference speed by **1.68x** through vLLM optimization, and **Google DeepMind**'s four **Gemini 4 Argon** agents defeated **Minecraft** for the first time without human intervention [6][11]; on the other hand, **CoreWeave**'s stock plunged **25%** and its junk bond yields soared to **15%**, making it the epicenter of the "dual sell-off in stocks and bonds" in this round of AI investment frenzy, as the market's core contradiction shifts from "compute demand" to "who pays for the high cost of capital" [20][22]. ## 🚀 Key Developments - **China plans to include AI models and compute clusters in export controls** [0]: The Ministry of Commerce and Ministry of Science and Technology proposed new rules, but providing access via API or SaaS does not constitute a technology export. - **DeepSeek API launches web search and scraping** [17]: Billed per successful request at 5 yuan per 1,000 calls, supporting scenarios such as Harness and Claude Code. - **Nvidia optimizes Kimi K3 inference performance** [6]: Upstream vLLM optimization enables the 2.8-trillion-parameter model to reach 423 tok/s/user on GB300 NVL72. - **Gemini 4 Argon agents autonomously complete Minecraft** [11]: Four agents divide labor and warn each other in a random world, killing the Ender Dragon with no human input. - **Anthropic internal research: share of junior technical roles drops sharply** [7]: Junior roles in the technical department fell from 21% to 14%, and AI Agent work hours reached six times that of employees. - **Four major Personal Agents achieve only 33% success rate in real tests** [5]: Two-week tests of Google Muse, Grok Bot, OpenAI Dots, and Manus Cue show they are currently better suited as local Agent combination helpers. - **CoreWeave becomes the "eye of the storm" in the AI crash** [22]: The stock fell 25% for the week and junk bond yields spiked to 15%, as the high-leverage expansion model triggered chain repricing across the infrastructure financing ecosystem. - **GLM 5.3 breakdown: unchanged architecture approaches closed-source frontier via post-training** [2]: SemiAnalysis points out that all improvements come from post-training, with agentic benchmarks improving by over 2%. ## 🔗 Sources [0] China proposes AI export controls, DeepSeek invests in robotics, Infinigence completes $334 million financing — https://aihot.news/items/g3h4v1tbgu7giq5ytwfsg39eg [2] SemiAnalysis podcast breaks down GLM 5.3: unchanged architecture approaches frontier closed-source models via post-training — https://aihot.news/items/a4l3q94htgd3wkoo94p8wigmn [5] Real tests of four major Personal Agents—Google Muse, Grok Bot, OpenAI Dots, and Manus Cue: only 33% success rate, three sets of prompts fill workflow gaps — https://aihot.news/items/truyza4ffva9rt8kv09m12r21 [6] Nvidia optimizes Kimi K3 inference speed, 1.68x improvement on GB300 NVL72 — https://aihot.news/items/yjpib8e04l9rntdkg91c2k9ac [7] Anthropic internal research shows decline in share of junior technical roles, AI Agent work hours six times that of employees — https://aihot.news/items