## 🔍 Key Insights Google’s 8th-generation **TPU**—built on a *training-inference separation architecture*—cuts large-model training time from **months down to weeks**, while boosting inference cost-efficiency by **80%** [3]. Meanwhile, Prof. Yaohui Jin of Shanghai Jiao Tong University open-sourced **Path2AGI**, a five-dimensional learning map that reimagines AGI education for Chinese speakers [5]. And a former ByteDance researcher warns: the AI gap between China and the U.S. is widening—“leaderboard culture” is obscuring real-world model usability [7]. ## 🚀 Top Updates - **Google launches 8th-gen TPU: Training-inference separation powers Agent infrastructure** [3]: Introduces dedicated chips—**TPU 8t** for training and **TPU 8i** for inference—optimized for the agent era. - **Shanghai Jiao Tong University open-sources Path2AGI learning map** [5]: Covers 25 foundational disciplines, organized along **five capability pathways**—cognition, reasoning, embodiment, and more—to systematically build AGI literacy for Chinese learners. - **Huawei unveils ADS 5 intelligent driving system** [6]: Upgrades the **WEWA 2.0 multi-agent博弈 architecture**, integrates its in-house Qiankun OS, and commits over **¥18 billion (RMB)** to R&D in 2024. - **Ex-ByteDance researcher dissects China’s real AI gap** [7]: Highlights how “leaderboard culture” decouples research from engineering deployment; cautions that *model distillation* is both a shortcut and a trap—and identifies **agents** and **embodied intelligence** as critical frontiers. - **Deep Work EP80: The most irreplaceable skill in the AI era** [4]: Synthesizes insights from Cal Newport and five other seminal works to argue that **deep work** is humanity’s core defense against AI displacement. - **Thinkhaven project launches: Cultivating sustained novelty in thinking** [1]: A one-month structured thinking program featuring daily research logs and biweekly deep articles. - **Intelligent robots declared “the next trillion-dollar market”** [8]: Prof. Tianmiao Wang of Beihang University notes that **embodied intelligence** is moving from labs into full industrial loops—ushering in systemic opportunities for hard-tech startups. - **Matt Pocock proposes a new AI coding paradigm: Foundational software engineering drives controllable development** [0]: Embeds core engineering practices—**TDD, modular design, code review**—deeply into AI-powered programming workflows. ## 🔗 Sources [0] #510.AI Coding For Real Engineers: How Software Engineering Fundamentals Make AI Coding More Effective — https://www.bestblogs.dev/podcast/cbd8422 [1] "Thinkhaven" — LessWrong — https://www.bestblogs.dev/en/article/614dd865 [2] Bonus Episode 14: Chatting with Li Dan about Podcasting, Creativity, AI, and Midlife — https://www.bestblogs.dev/podcast/7ec6ea8 [3] Elon Musk endorses Google’s 8th-gen TPU! Training slashed from months to weeks; inference efficiency up 80% — https://www.bestblogs.dev/article/7e61e6c7 [4] EP80 “Deep Work”: What’s the *one skill* AI can’t replace? — https://www.bestblogs.dev/podcast/fcef8e3 [5] The Path to AGI—Open-Sourced by a Shanghai Jiao Tong Professor on Datawhale! — https://www.bestblogs.dev/article/e3d7b4e2 [6] Huawei Unveils ADS 5