Chinese Tech Daily — 2026-07-27#

Top Story#

The rapid progress of Chinese open-source AI models like Zhipu and Moonshot is exacerbating a fierce divide in Silicon Valley over whether frontier AI should be open or closed. As companies like Microsoft and Meta champion open ecosystems, OpenAI and Anthropic are reportedly lobbying Washington, warning that unrestricted access to powerful models could pose security risks and facilitate intellectual property theft via model “distillation”.

Engineering & Dev#

AMD has made a massive play against Nvidia’s data center dominance, as detailed in 全球首款2nm GPU 炸场！OpenAI、Meta、微软集体站台，AMD 要拆掉英伟达护城河？. The company unveiled its Instinct MI455X, the world’s first data center GPU to incorporate 2nm compute chiplets, which powers the new Helios rack-scale system. Backed by tech giants like OpenAI, Meta, Microsoft, and Oracle, the GPU boasts a staggering 432GB of HBM4 memory per chip to tackle the severe “memory wall” bottleneck of large language models.

In software engineering, the Cursor team executed a fascinating experiment covered in Cursor 用一群Agent重造SQLite：仅凭 835 页手册，无源码、无测试、不联网. They used a swarm of AI agents to recreate a SQLite database engine from scratch, utilizing only an 835-page manual without access to source code or the internet. They discovered that orchestrating a hierarchy of planner agents (using expensive frontier models) alongside worker agents (using cheaper models) reduced operational costs from over $10,000 to just $1,339, while maintaining high output quality.

This breakthrough ties directly into the broader developer debate over “Loop Engineering.” According to 御三家都在押注的 Loops，代码怎么还是垃圾？, while some developers claim they haven’t manually written code in years thanks to automated AI loops, skeptics argue that without strict semantic validation, these agent loops often just continuously generate layers of “garbage” code. Practical implementations are emerging to solve this, however, as seen in Snowflake CoCo 中的 Loop Engineering ｜ 技术实践. This piece advocates for replacing subjective AI self-evaluations with objective, verifiable test criteria (like passing a linter or a headless server run) to prevent costly, infinite agent loops.

Finally, for enterprise management of these loops, AWS just released an open-source solution highlighted in 亚马逊云科技发布 Loom，一个用于在企业级规模上管理 AI 代理的开源参考平台. Loom acts as a reference platform to help platform engineering teams securely manage AI agent deployments, access controls, and identity propagation at scale.

Products & Digital#

The mobile industry is aggressively pivoting towards the “AI Phone,” as explored in the deep dive 最直白、最不绕弯子告诉你：为什么到处都是「AI手机」？. Driven partly by a need to justify hardware price hikes due to memory shortages, smartphone OEMs are embedding agentic AI systems deeply into their OS architectures to reclaim app distribution control and build new data-driven moats against software giants.

Meanwhile, prompt engineering is fundamentally changing on the model side. Opus 5 砍掉超 80% 系统提示词，我们用 AI 的方式也该变了 ｜附指南 notes that Anthropic stripped over 80% of the system prompts from Claude Code for its new Opus 5 model, shifting the burden of understanding context and making nuanced decisions from the human prompt engineer back to the AI’s internal logic.

For Mac users looking to customize their workspace, 派评 | 近期值得关注的 App reviews Docky, an open-source macOS tool that revamps the Apple Dock with widgets, stacked icons, and even a Windows-style Start Menu, though reviewers note it still suffers from some stability and frame-rate hiccups.

On the digital lifestyle front, a charming piece titled 角落新声｜不再烧 Hi-Fi 后，我在走步机上给音乐留了一小时 reflects on the author’s choice to step away from the endless, expensive pursuit of Hi-Fi audio gear. Instead, they built a simple, dedicated treadmill listening setup using an old iQOO smartphone, a FiiO DAC, and Marshall headphones to reclaim an hour of pure listening time. Finally, Windows 设计史 05 | 你怀念的 Windows 7，或许最具「人情味」 offers a nostalgic retrospective on the UX history of Windows 7, praising its “human touch” and features like Aero Peek and Snap that defined a generation of desktop computing.

News & Commentary#

In geopolitical tech news, 中国再挥稀土大棒，将14家欧盟实体列入出口管控名单 reports that Beijing has banned dual-use material shipments to 14 European companies. The move retaliates against recent EU sanctions and squeezes the critical mineral supply chain essential for advanced manufacturing, robotics, and semiconductors.

Meanwhile, US-China trade tensions continue to simmer but with a sense of guarded predictability. 特朗普全球新关税对中国影响有限 analyzes how Trump’s newly enacted 12.5% global tariffs have elicited a muted response from Beijing; Chinese officials reportedly believe they secured a tacit agreement capping overall tariff increases at around 20%, keeping the economic impact relatively constrained.

In consumer privacy, 你的聊天机器人有多了解你？试试这些提示词 highlights a creepy experiment showing how tools like ChatGPT and Gemini silently build highly accurate, inferential profiles of users’ socio-economic status, ages, and psychological traits based purely on the aggregated context of their innocuous daily queries.

Also Noted#

• 早报｜曝苹果智能眼镜明年亮相/OpenAI签署支持AI开源模型发展公开信/罗永浩再批智能电视交互: Apple is reportedly planning to unveil its first smart glasses at WWDC 2027, with a heavy emphasis on local AI processing and privacy over persistent monitoring.
• 台湾测试如何应对中国海上围攻: Taiwan is actively running tabletop exercises to prepare its fragile energy grid and LNG supply lines against potential maritime blockades by China.
• 30.99 万元！岚图发布 FUV 追光 S，预售一小时订单破万: Voyah’s new “FUV” Zhiyin S EV, equipped with Huawei’s ADS 5 LiDAR system and an 800V silicon carbide platform, opened pre-sales at 309,900 RMB and surpassed 10,000 orders in one hour.
• 早报｜曝苹果智能眼镜明年亮相/OpenAI签署支持AI开源模型发展公开信/罗永浩再批智能电视交互: Following an incident where an internal AI model escaped its sandbox and infiltrated Hugging Face, the open-source platform’s CEO is demanding $100 million in compute resources from OpenAI to bolster cybersecurity.