☕️ Chinese AI overtakes US on global usage

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

🌏 Chinese AI overtakes US on global usage

🐉 Alibaba unveils open Qwen 3.8 AI model

🗄️ Microsoft ends SQL Server 2016 support

🎬 Netflix builds AI to personalize homepages

🗄️ Git-style SQL database Dolt hits 2.0

Plus: 🎁 10 other news you might like, 🧰 6 tools, and 📚 5 papers.

🌏 Chinese AI overtakes US on global usage LINK

Chinese open-weight AI models from DeepSeek and Zhipu have overtaken US competitors in global token consumption, driven by rising proprietary pricing and improving capability, according to Vercel and OpenRouter usage data.

Zhipu's GLM-5.2 saw token volume grow 50x on Vercel, while DeepSeek's V4 Flash became the AI Gateway's largest single model at over 20% of traffic, up from roughly 15% a month earlier.

Open-weight models now account for 29% of Vercel token volume, nearly triple their April share, with DoorDash pairing Moonshot's Kimi K2.6 for lighter tasks and Anthropic's Fable for the hardest work to cut costs.

🐉 Alibaba unveils open Qwen 3.8 AI model LINK

Alibaba has released Qwen 3.8, an open-weight model with 2.4 trillion parameters that the Qwen team claims matches frontier competitors and trails only Fable 5, with full open weights promised "soon."

The model is available in preview through Alibaba's Token Plan, Qoder, and QoderWork at 10 percent of standard pricing, and is positioned to beat Qwen 3.7-Max on coding and complex tasks like full-stack development and data analysis.

Developer Shuai Bai says Qwen 3.8 is the team's first multimodal model above 1 trillion parameters, capable of processing images, videos, and documents, though no benchmark results have been published yet.

🗄️ Microsoft ends SQL Server 2016 support LINK

Microsoft ended support for SQL Server 2016 on July 14, 2026, meaning existing installations keep running but no longer receive security updates, product fixes, feature improvements, or standard technical support going forward.

Supported migration paths include SQL Server 2022, SQL Server 2025, Azure SQL Managed Instance, or SQL Server on Azure VMs, while Extended Security Updates offer a paid bridge delivering only highest-severity patches for up to three years.

ESU adds no features, general troubleshooting, or non-security fixes, and machines outside Azure subscribe through Azure Arc; Microsoft caps highest-severity updates at July 17, 2029, so administrators must complete migration before then.

🎬 Netflix builds AI to personalize homepages LINK

Netflix built GenPage, a generative AI system that directly produces personalized user homepages from a single prompt, replacing its previous multi-stage recommendation pipeline that separately handled candidate generation, ranking, and layout.

GenPage unifies item selection, row construction, and layout into one model, then uses post-training reinforcement learning to optimize interactions across rows and items, which unexpectedly boosted homepage diversity and customization.

A/B testing showed statistically significant engagement gains alongside a 20% cut in end-to-end serving latency, and engineers found enriching the prompt outperformed scaling the model from 120M to 900M parameters.

🗄️ Git-style SQL database Dolt hits 2.0 LINK

DoltHub shipped Dolt 2.0, a major update to the open source MySQL-compatible SQL database with Git-style version control, adding automatic storage optimization, beta vector support, and performance the team claims now beats MySQL on sysbench.

A new on-disk "archives" format uses dictionary compression to deduplicate storage and cut the footprint by 30-50%, while automated garbage collection now runs by default to reclaim intermediate copy-on-write state that would otherwise fill disks during imports.

Version 2.0 introduces beta support for version-controlled vector indexes built on MariaDB's Vector type, which the announcement says makes Dolt the only database to version-control vectors, with beta status ending once remaining read-path gaps are resolved.

Other news you might like

• Automate React Native Performance With MetrognomeLINK

• Self-healing GPU nodes in Kubernetes: What we learned building the EKS node monitoring agentLINK

• No, We Can't Harden Node.js Against Prototype PollutionLINK

• Move code review before the codeLINK

• The cleanup trap: Stop asking RAG to fix bad dataLINK

• AWS Releases Loom, an Open-Source Reference Platform for Governing AI Agents at Enterprise ScaleLINK

• Linux 7.2-rc4 drops with critical fixes for gaming handhelds, Wi-Fi, and car systemsLINK

• New Intel Itanium emulator boots Itanium version of Windows XP and 2003LINK

• AWS admits bill error alarms failed to rouse engineeringLINK

🧰 Trending tools

ESP32 Bowling Scoring System: a low-cost, open-source replacement for expensive commercial bowling center scoring systems, using ESP32 microcontrollers to handle pin detection, scoring, and lane automation.LINK

AnySearch: a search API built for AI agents that pulls filtered, de-duplicated, structured results from trusted sources in parallel, improving reliability.LINK

ZooData: converts URLs into structured JSON for AI agents, cutting LLM token usage ~75% while adding pre-analyzed e-commerce insights for Amazon and TikTok.LINK

Archify: detects components, APIs, libraries, and third-party scripts running on any webpage, flagging outbound domains and payment field access risks.LINK

OpenMarkdown: opens and edits .md files instantly, letting your AI agent read/write the same file via CLI, plugin, or MCP, all local-first with no telemetryLINK

Q3Edit: a browser-based level editor for Quake 3 maps featuring brush and patch editing, CSG, terrain sculpting, and entity editing, with in-browser playtesting via WebAssembly.LINK

📚 Trending papers & reports

Trial-and-error decision making gets a training method letting quantum-powered systems learn optimal strategies while making mistakes only a fraction as often over time, versus today's classical approach, sharply cutting the cost of learning by doing.LINK

Chip design timing checks get a faster, more scalable prediction method that reads circuit logic like a sequence instead of a complex map, cutting delays in catching timing problems before manufacturing.LINK

Robot vision decoding lets AI systems read depth and identify objects simultaneously with 78.6% fewer trainable parameters than the standard approach, cutting costs while matching or beating prior top performers on benchmark tests.LINK

Multi-agent teamwork tests reveal that nine leading AI-coordination methods all scored a zero percent win rate on tasks requiring bots to split labor and cooperate, exposing a real gap before deployment.LINK

Fraud-detection graph models can now resist adversarial manipulation with a provable guarantee while boosting accuracy 437.1% and certified robustness 409.3% on tested attacks, without slowing performance.LINK

See you tomorrow for a new dose of ☕️ Devshot!