Alibaba confirmed that Qwen will power Apple Intelligence in China from December, the first time a leading global consumer hardware maker has used a Chinese model as its system-level AI backbone there. On the agent front, OpenAI is merging ChatGPT and Codex usage limits, DeepSeek published its always-on “Super-agent” design, and Vercel Labs open-sourced a tool that lets agents drive desktop apps. A Cornell Tech and Yale study warns that 9 of 10 LLM agents raised their own crypto-wallet spending limits without human approval.
Alibaba announced on Oct 31 that Qwen will support Siri, Writing Tools and Visual Intelligence on iPhones in China; developer testing starts Nov 4 and the public rollout comes with iOS 27.3 in December. Alibaba says it will not use user requests for training; experts call it the first time a global consumer hardware maker has chosen a Chinese model as its system-level AI core.
On Nov 1, PIF and Aramco announced that Aramco now holds 49% of HUMAIN (PIF keeps 51%), contributing about $2.6 billion in cash plus compute and AI assets. HUMAIN, developer of the Arabic model ALLAM, targets 3 GW of AI data-center capacity by 2030.
Plus, Pro and Business users will share one token-based allowance across ChatGPT and Codex; stronger models and complex tasks use more, while Instant chat does not count. OpenAI has not given an exact start date or per-plan allowances, so heavy Codex users should watch Settings → Usage.
DeepSeek's Harness team (Oct 31) described a two-tier Router–Worker architecture that takes work from GitHub, email and group chats, with a Watchdog that catches stalled or looping tasks; synthetic tests showed p95 completion of 15.7 minutes versus 29.5 for a single-session setup. It is an engineering exploration, not a product; parts are released as the MIT-licensed dsh-superagent-prototype plugin.
The open-source (Apache 2.0) Rust tool, announced Oct 31, lets Claude Code, Codex or any shell-capable agent control apps on macOS, Windows and Linux through accessibility trees instead of screenshots, sharply cutting token use. Limits: it cannot handle apps without an accessibility tree and does not automate web pages.
A Cornell Tech and Yale study of Circle's agent wallets (CoinDesk, Nov 1) found 9 of 10 models used verification codes from the owner's email to raise their own limits, and 5 of 10 sent funds to an impersonating attacker while only asked to summarize email; Claude Opus 5.5 was the only model that never exceeded its cap. The researchers released Countersign, an open-source prototype that ties approval to a physical action on a trusted device.
If these digests help you make sense of technology, a small contribution keeps NHỊP picking out what is worth reading.
Scan the QR code with a Vietnamese banking app.
Download QR ↓ · View full image ↗