Chinese Tech Daily

Top Story

Uber is tapping the brakes on developer AI usage. After a sixfold increase in AI-related costs and monthly per-developer expenses reaching $2,000, the ride-hailing giant is implementing a strict $1,500 monthly quota for AI coding tools. This signals a major reality check for enterprise AI adoption: while 92% of Uber’s engineers actively use AI, the company’s focus is now shifting from unbridled integration to rigorous ROI, measuring the “net code quality ratio” and balancing automation output with infrastructure cost governance.

Engineering & Dev

Over at InfoQ, there is a fascinating look at architectural discipline in Hard Stop Rule: From 3 HCM Monoliths to 120 Domain Microservices. Paycor shared their “pull-based migration” strategy, where every new feature or bug fix on their legacy monoliths was forcibly spun out into a new Azure domain service. This approach ultimately eliminated the need for standalone migration budgets by amortizing the cost across hundreds of user stories, successfully scaling to over 120 microservices with zero downtime.

In the realm of AI engineering, Alibaba Cloud’s Wang Jingyao shared insights on Quick BI Data Analysis Agent Reliable Engineering Practices at AICon Shenzhen. The session highlighted the shift from focusing on raw model capabilities to system engineering, emphasizing the use of a BI semantic layer to resolve heterogeneous data ambiguity and a Harness architecture for verifiable, permission-embedded agent actions.

Meanwhile, as part of its “Zero Growth Stack” initiative, Uber engineered a dynamic Go garbage collection tuning library called GOGCTunner, covered in Uber Limits AI for 92% of Engineers. By monitoring real-time object usage and integrating cgroup memory limits, they successfully reclaimed 70,000 CPU cores across 30 mission-critical services, proving that dynamic system control beats static configurations in preventing OOM events.

Products & Digital

Over on sspai (少数派), an excellent piece for motorsport newcomers titled A Beginner’s Guide to Racing “Overtaking on Curves”: The Time Ledger Behind Pit Strategies demystifies F1 broadcasts. It breaks down the mathematical and tactical realities of tire degradation, explaining that the real-time leaderboard is often a mirage because drivers are on varying tire lifetimes. The article serves as a deep dive into calculating “pit loss” and why “undercuts” (pitting early for new tires) and “overcuts” (staying out on a clean track) are the true deciding factors in race outcomes.

News & Commentary

Wrapping up the World Artificial Intelligence Conference (WAIC 2026), an InfoQ roundtable titled WAIC Closing Truths: What is AI Actually Rewriting? cut through the hype of what felt like a “World Embodied AI Conference”. Industry experts debated whether AI is genuinely transforming workflows or just repeating the mobile internet cycle, noting that while generative content is already driving real revenue for platforms like Kuaishou and ByteDance, enterprise “digital employees” often remain a pseudo-demand without deep workflow integration and proper operational knowledge.

The same WAIC discussion also touched on the fading dominance of the Scaling Law. Researchers are now looking beyond simply stacking parameters, pointing to hybrid models combining autoregressive and continuous diffusion language models, as well as ecological architectures where multiple smaller models orchestrate tasks to improve reasoning and efficiency without massive data leaps.

Also Noted

• Local AI Deployment is emerging as a strong enterprise necessity to bypass cloud API privacy concerns and reduce data leakage risks.
• AI for Science in Lab Environments is proving to be a highly viable commercial path for embodied AI, automating tedious, high-precision biomedical tasks.
• Continuous Diffusion Language Models are gaining traction for endpoint inference, offering a “latent reasoning” process that speeds up local, single-user generation.