Author: RadarAI Editorial
Editor: RadarAI Editorial
Last updated: 2026-07-06
Review status: Editorial review pending
Brief
速报
官方
AI动态
开源
AI infrastructure hits a critical inflection point: Meta opens GPU compute for commercial use; Huawei's 'Tao Law' paper details LogicFolding 3D stacking; Peking University's memristor chip achieves in-memory computing breakthrough; HBM inventor Kim Jung-ho identifies memory bottlenecks as the root cause of <10% GPU utilization.
Editorial standards and source policy: Editorial standards, Team. Content links to primary sources; see Methodology.
## 🔍 Key Insights
AI infrastructure is hitting a pivotal inflection point: **Meta is commercializing its GPU compute**, **Huawei’s “Tao Law” paper has been upgraded**—revealing details of its LogicFolding 3D stacking technology, **Peking University’s memristor chip** achieves breakthrough in-memory computing, and **HBM pioneer Dr. Jung-Ho Kim** identifies the core memory bottleneck: GPUs are only actively utilized **10% of the time** [2][5][15][22].
## 🚀 Top Developments
- **Jaade**: A macOS-native coding agent workspace launches [1] — integrates Claude Code/Codex sessions, terminal, and Git, enabling AI-powered development to be **“visible and controllable”**, ending the black-box dilemma.
- **Meta plans to sell AI compute externally** [2]: With internal model development lagging, Meta pivots to monetizing its massive infrastructure—a move dubbed the **“sell shovels” strategy**, driving a sharp stock rally.
- **Huawei releases Version 2 of its Tao Law paper** [5]: Details **LogicFolding 3D stacking** and engineering trade-offs behind AI data center “three core components”, strengthening China’s homegrown compute infrastructure methodology.
- **Peking University & CAS unveil the world’s first programmable in-memory computing chip** [22]: Built on phase-change memristors, it completes **cortical brain modeling in just 0.43 seconds**—nearly **500× faster than GPUs**.
- **Meituan open-sources LongCat-2.0** [14]: A **1.6-trillion-parameter MoE large model**, trained exclusively on domestic GPUs (50,000 cards), matching Gemini 3.1 Pro on Agent benchmarks.
- **SeKV: A new long-context inference method debuts** [23]: Developed by UBC and Microsoft, it combines semantic segmentation + SVD—enabling **24GB GPUs to handle ~300K-token inputs**, cutting VRAM usage by **53%**.
- **Browser Relay open-source project launches** [21]: Lets *any* AI agent directly control a **fully authenticated, real Chrome browser**—closing the loop on local, browser-based actions.
- **Hackers exploit SEO poisoning for AI Agent prompt injection attacks** [20]: Using hidden HTML and malicious search rankings to trick agents into executing high-risk commands—like fraudulent payments.
## 🔗 Sources
[1] Jaade: A macOS Desktop Workspace That Makes Coding Agents “Visible and Controllable” — https://www.bestblogs.dev/article/cbd5172d?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[2] Meta Joins the “Shovel-Selling” Race! Zuckerberg: Models Can Wait—GPUs Must Earn Revenue — https://www.bestblogs.dev/article/08a1446c?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[5] Huawei Updates Its Tao Law Paper! — https://www.bestblogs.dev/article/8b3e9b36?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[14] Meituan Just Open-Sourced LongCat-2.0: 1.6 Trillion Parameters, Trained Entirely on Domestic GPUs — https://www.bestblogs.dev/article/624af818?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[15] Dr. Jung-Ho Kim, “Father of HBM”: AI Is Fundamentally a Memory Problem—GPUs Are Only Active 10% of the Time
AI infrastructure is hitting a pivotal inflection point: Meta is commercializing its GPU compute, Huawei’s “Tao Law” paper has been upgraded—revealing details of its LogicFolding 3D stacking technology, Peking University’s memristor chip achieves breakthrough in-memory computing, and HBM pioneer Dr. Jung-Ho Kim identifies the core memory bottleneck: GPUs are only actively utilized 10% of the time [2][5][15][22].
🚀 Top Developments
- Jaade: A macOS-native coding agent workspace launches [1] — integrates Claude Code/Codex sessions, terminal, and Git, enabling AI-powered development to be “visible and controllable”, ending the black-box dilemma.
- Meta plans to sell AI compute externally [2]: With internal model development lagging, Meta pivots to monetizing its massive infrastructure—a move dubbed the “sell shovels” strategy, driving a sharp stock rally.
- Huawei releases Version 2 of its Tao Law paper [5]: Details LogicFolding 3D stacking and engineering trade-offs behind AI data center “three core components”, strengthening China’s homegrown compute infrastructure methodology.
- Peking University & CAS unveil the world’s first programmable in-memory computing chip [22]: Built on phase-change memristors, it completes cortical brain modeling in just 0.43 seconds—nearly 500× faster than GPUs.
- Meituan open-sources LongCat-2.0 [14]: A 1.6-trillion-parameter MoE large model, trained exclusively on domestic GPUs (50,000 cards), matching Gemini 3.1 Pro on Agent benchmarks.
- SeKV: A new long-context inference method debuts [23]: Developed by UBC and Microsoft, it combines semantic segmentation + SVD—enabling 24GB GPUs to handle ~300K-token inputs, cutting VRAM usage by 53%.
- Browser Relay open-source project launches [21]: Lets any AI agent directly control a fully authenticated, real Chrome browser—closing the loop on local, browser-based actions.
- Hackers exploit SEO poisoning for AI Agent prompt injection attacks [20]: Using hidden HTML and malicious search rankings to trick agents into executing high-risk commands—like fraudulent payments.
🔗 Sources
[1] Jaade: A macOS Desktop Workspace That Makes Coding Agents “Visible and Controllable” — https://www.bestblogs.dev/article/cbd5172d?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[2] Meta Joins the “Shovel-Selling” Race! Zuckerberg: Models Can Wait—GPUs Must Earn Revenue — https://www.bestblogs.dev/article/08a1446c?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[5] Huawei Updates Its Tao Law Paper! — https://www.bestblogs.dev/article/8b3e9b36?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[14] Meituan Just Open-Sourced LongCat-2.0: 1.6 Trillion Parameters, Trained Entirely on Domestic GPUs — https://www.bestblogs.dev/article/624af818?utm_source=rss&utm_medium=feed&utm_campaign=resources&entry=rss_article_item
[15] Dr. Jung-Ho Kim, “Father of HBM”: AI Is Fundamentally a Memory Problem—GPUs Are Only Active 10% of the Time
← Back to Updates