Apple cuts engineers’ daily AI budget from 200
According to Bloomberg’s Mark Gurman, Apple lowered how much some engineers can spend on AI tools per day and may cut it to $100. Staff worry it signals layoffs.
Tip
even Apple is counting its AI bill, so how should individuals keep AI spending in check?
Top models drop to 13% when tracking changing values in long documents
The Allen Institute for AI (Ai2) released StateBench, which tests whether models can keep track of variables that change across a long document. Gemini, Claude Opus 5.5 and GPT-6 did well on easy versions but all fell to 13% or lower on the hardest level.
Tip
when AI reads a long contract or spreadsheet, which parts you should still check yourself
Claude finds two critical flaws in the Xen hypervisor
Xen’s official advisory credits Claude. The flaws let one malicious virtual machine read and write the memory of other VMs on the same host, affecting every release from 4.6 on; third parties have reproduced them.
Tip
AI finding security bugs is going from news to routine. What changes for everyday developers?
Moonshot AI and Tencent release open models on the same day
Kimi K3.1 Pro’s weights are public, with 2.8 trillion parameters and a 1M-token context, and it shot to the top of Hugging Face trending. Tencent’s Hunyuan Hy4 has 770 billion parameters and adds image understanding and the ability to operate a computer directly.
Tip
Chinese open models keep getting bigger. How can ordinary people actually use them?
From iOS 27.3, files in public iCloud links will be scanned
For files shared with “anyone with the link,” Apple will check for “harmful content.” It can delete files and disable accounts for repeat violations, but doesn’t say which models do the checking.
Tip
what to know before sharing files through an iCloud link
Huawei, Alibaba, Tencent and 17 others sign an AI agent self-discipline pact
At the World Internet Conference in Wuzhen, CAICT, the research arm of China’s industry ministry, led a pact covering eight areas including trust, interoperability, security, oversight and user protection, with signatories pledging to accept industry and public oversight.
Tip
countries are starting to set rules for agents, and China is starting with “self-discipline”
⚠️ Compiled by an AI assistant from public AI news roundups; original articles are not individually verified. For reference only.