Chinese Tech Daily — 2026-05-15#

Top Story#

The era of simply hoarding AI models is over, and the industry focus has decisively shifted to Agent infrastructure and viable business loops. At the Baidu Create conference, executives warned that the “old supply” of AI is reaching its limit, noting that the massive token consumption of complex Agents has made enterprise adoption prohibitively expensive. In response, Baidu has completely overhauled its full-stack AI cloud to prioritize an “Agent-first” architecture, utilizing advanced KV Cache pooling to drastically reduce the marginal cost of computing and make multi-step Agent tasks economically viable.

Engineering Reads — 2026-05-18#

The Big Idea#

The limits of engineering capability—whether writing new software with AI or comprehending legacy systems—are ultimately dictated by the quality and tightness of our feedback loops. The tools we build to verify correctness or surface the context of past decisions will become far more critical than the raw generation of code or text.

Deep Reads#

[What’s Easy Now? What’s Hard Now?] · Marc Brooker · Source Coding agents will eventually excel at deeply technical systems programming while struggling with UI/UX, directly inverting current conventional wisdom. Brooker argues that AI agents are fundamentally feedback loops wrapped around open-loop LLMs. Tasks with rigorous automated feedback—like writing a database storage engine verified by Rust, TLA+, or property-based tests—can be solved entirely by an agent iterating without human intervention. Conversely, front-end development relies on slow, inconsistent human feedback, making it a inherently difficult problem for autonomous agents. Engineering leaders and systems programmers should read this to understand why mastering formal specification tools will be their highest-leverage skill in an AI-assisted future.

Tech Videos — 2026-05-18#

Watch First#

Build Agents That Run for Hours (Without Losing the Plot) — Ash Prabaker & Andrew Wilson, Anthropic is a masterclass in scaffolding for LLMs that goes beyond “vibes”, detailing the specific adversarial generator/evaluator patterns needed to keep an agent on track over 12-hour context windows. It’s a required watch if you are building autonomous systems that need to execute reliable software engineering tasks for hours instead of minutes.

Chinese Tech Daily — 2026-05-18#

Top Story#

The AI industry is waking up to the reality that owning the coding workflow is the ultimate moat. Reports that Elon Musk’s SpaceX has secured a massive $10 billion partnership with Cursor—including a right to acquire the coding agent startup for $60 billion—prove that real-world, on-policy developer data is now the most coveted asset for model builders 马斯克花 100 亿想清楚一件事，不做 coding agent 就是等死. Concurrently, this arms race is creating toxic corporate behavior: companies are instituting “Token Consumption Leaderboards” as a KPI, leading developers to spin up useless sub-agents and burn millions of dollars in compute just to look productive 一个月烧掉 930 万元 Token 的人，也没烧出个答案.

Hacker News — 2026-05-19#

Top Story#

The massive “Mini Shai-Hulud” supply chain attack on npm is dominating discussions today. An attacker compromised the atool maintainer account and published over 600 malicious versions across 314 packages in just 22 minutes to harvest AWS, Kubernetes, and local password manager credentials. It’s a sophisticated wake-up call for the ecosystem, utilizing GitHub’s API for stealthy C2 communication, injecting persistent backdoors via GitHub Actions, and specifically targeting developers’ local Claude Code and Codex environments through hook injections.

Chinese Tech Daily — 2026-05-19#

Top Story#

The shift from single AI agents to multi-agent “swarms” and agentic organizations is dominating Chinese tech discourse. At the AMD AI Developer Day in Shanghai, Lee Kai-fu declared that while 2025 was about completing workflows, 2026 is the year multi-agent architectures will be capable of running entire enterprise functions. This vision is immediately materializing with Huawei-backed openJiuwen’s open-source release of JiuwenSwarm, a framework introducing “Coordination Engineering” to let multiple agents dynamically distribute tasks, negotiate, and self-evolve as a highly coordinated team.

Chinese Tech Daily — 2026-05-20#

Top Story#

Alibaba Cloud’s CIO team has achieved massive productivity gains by explicitly banning a popular industry vanity metric: the “AI code generation rate.” Instead of chasing raw lines of AI-generated code—which they argue often just scales up technical debt—the team successfully reorganized around end-to-end business value and new “Half-Stack” developer roles, proving that AI’s true enterprise value lies in redefining the workflow, not just replacing the coder.

The AI Reality Check: Token Shock, 100x Orgs, and Valuation Absurdity — 2026-05-21#

Highlights#

The AI industry is currently experiencing a massive collision between theoretical valuations and harsh operational realities. While the “token subsidy era” is reportedly ending as staggering compute costs evaporate enterprise budgets, forward-looking organizations are aggressively restructuring to become “AI-native” by replacing human software bottlenecks with high-leverage agent managers. Concurrently, astronomical claims around total addressable markets and impending mega-IPOs are drawing sharp skepticism from observers who argue the math no longer adds up.

Tech Videos — 2026-05-21#

Watch First#

Software engineering at the tipping point by Google for Developers. Why: A highly pragmatic, sobering look at how a 10x increase in AI-generated code will completely break our current CI/CD, testing compute, and human code review pipelines unless we immediately adopt rigid “software ecology” and systems thinking.

Chinese Tech Daily — 2026-05-21#

Top Story#

Alibaba released its next-generation flagship model, Qwen3.7-Max, signaling a decisive industry pivot from conversational LLMs toward task-executing autonomous agents. Topping domestic benchmarks in the Arena blind tests, the model boasts significant improvements in coding, tool utilization, and long-context processing. Most notably, Qwen3.7-Max autonomously optimized a production-grade attention kernel over 35 continuous hours, underscoring Alibaba’s ambition to position its Model-as-a-Service platform as a critical enterprise infrastructure for the Agentic era.