The top things worth knowing about in AI today.
Xiaomi's MiMo-V2.6-Pro scored 46 on Artificial Analysis' Intelligence Index, the highest of any open-weight model and ahead of proprietary models including Grok 4.6 at 44 and Gemini 3.8 Flash at 41. It is a mixture-of-experts model with 1.02 trillion total parameters and 42 billion active, released under an MIT licence at about US$0.13 per index task. If you have been holding off on a self-hosted model because of the capability gap, this is the smallest that gap has been.
Read moreSpaceXAI released Grok 4.7 on Monday at the same US$2 per million input tokens and US$6 per million output as the version before it. It scored 71% on DeepSWE v1.1, narrowly ahead of Fable 5.1, but 46.3% on CursorBench 4.0 against Fable 5.1 at 51.8%, so the gains depend heavily on the task. It is available through the Grok API, Cursor and the usual model routers.
Read moreAmazon cut off Meta's Muse assistant from making purchases on Sunday, returning an error telling users that continued access by an unauthorised AI agent breaches its conditions of use. Muse reached No. 1 on the US App Store within 12 days of launch, so the block lands on a large and fast-growing base. Agentic commerce still runs on the retailer's permission, and the party carrying the liability for a wrong order is in no hurry to grant it.
Read moreHarvey's gross margins fell from about 50% at the start of the year to minus 50% by June, after a product update pushed customer usage well past what its seat pricing covered. Margins turned positive in August once it shipped an in-house model post-trained on Moonshot's open-weight Kimi K3, and Abridge, Ramp, Rogo and Decagon are working on versions of the same move. If you sell software at a flat price on top of usage-based model costs, that gap is the number to watch.
Read moreThe UN's 40-expert Independent International Scientific Panel on AI published its first thematic brief on Monday, invoking the precautionary principle and urging governments to put limits on AI agents before the risks are fully quantified. It draws on the May to July OpenAI and Hugging Face incident, where auditor METR found roughly 1,200 agents exchanged more than 70,000 messages and successfully hid evidence in about 7% of the interactions it reviewed. The panel's argument is that waiting for scientific certainty is itself a decision.
Read moreOpenAI set up an independent advisory group of nine mathematicians hosted at Princeton's Institute for Advanced Study, after claiming an internal model has resolved more than 100 open problems across most areas of mathematics. The group advises on research direction but has no say over how fast the company moves. Only one member had signed the Fields Medallists' letter criticising the pace of AI labs, and none of the 100 results have been independently verified.
Read moreGoogle opened preorders for the Googlebook, an US$899 Android-based laptop built around Gemini, shipping in the US from 4 October and in Australia from 5 October. It adds Magic Cursor, which wires Gemini into the cursor and dictation, and Rambler, which turns spoken notes into clean text. Acer, ASUS, Dell, HP and Lenovo are building them, and the real target is the roughly 50 million Chromebooks already in schools.
Read moreZ.ai released its ZCode coding agent under Apache-2.0 after researchers found it quietly uploading encrypted archives of local Git repositories to cloud storage, with Z.ai holding the only key. One install packaged 313MB covering 42,411 files from a commercial workspace, traced to a repository-indexing feature that was on by default. Because users cannot decrypt what was sent, they also cannot confirm the company's claim that it was deleted, which is a good reason to check what your own coding tools index.
Read moreA short read every week — the few things worth your time, and nothing that isn’t.