Meta Opens Its Muse Assistant to Homemade Gadgets

On October 2, 2026, Meta released Muse Gadgets, free code that lets cheap circuit boards and Raspberry Pi mini-computers talk to Muse, Meta's AI assistant. The license (Apache 2.0) allows anyone to use and sell what they build with it. Why it matters: Muse now works with devices Meta doesn't make. A speaker or a lamp you already own can take voice commands from it, if someone writes the code. Web developer Wes Bos had it running on a conference name badge the same day. The Verge ↗

Cloudflare and Perplexity Release Small AI Models for Quick Decisions

Cloudflare and Perplexity each released a free "decision model" on October 1, 2026. These models don't write text. They score the likely answers to set questions, such as which tool to use or where to send a request. Why care? An app can hand those small choices to a cheap model and keep the expensive chatbot for the actual reply. Perplexity's paid version charges about 4 cents per 750,000 words it reads, and nothing for what it returns. Cloudflare's Clef handles up to 64 questions at once. Cloudflare blog ↗

Update: OpenAI Fixes Missed ChatGPT Pro Usage Reset

OpenAI's promised usage reset for paid ChatGPT users skipped the Pro 500 plan, and was fixed early on October 3, 2026. Thibault Sottiaux, who leads Codex at OpenAI, said at 04:01 UTC that he was looking into it. Twenty-five minutes later he wrote "All fixed." If your Pro limits looked wrong on Friday, they should now be fresh. He also said GPT-6.1 Sol is back to normal speed after two days of heavy load. Sottiaux on X ↗

Update: Claude Code Mods Run Outside Its Safety Sandbox

A security write-up from Threat Frontier on October 2, 2026 says Claude Code's new mods are switched on by default. It also says a modded Claude Code doesn't get the sandbox that normally limits what commands it can run. Mods arrive inside plugins and run with your own access. So a plugin from a stranger can touch the same files and keys you can. Anthropic launched mods on October 1. Threat Frontier ↗

CBS News Finds Censorship in Alibaba's Qwen AI Models

CBS News published an analysis on October 2, 2026 saying Qwen, the AI model family from China's Alibaba, gives China-friendly answers and censors some topics. Qwen has been downloaded more than 3 billion times. CBS says US companies including Airbnb and Uber use it. Apps built on Qwen may carry those answers without telling users. Because Qwen is free to download and modify, a company hosting it can retrain it, and some do. CBS News ↗

HeyGen's hyperframes Turns Web Pages Into Video

hyperframes, from AI video company HeyGen, gained 2,661 stars this week. It renders a web page into a video file. The point: an AI assistant can make a video by writing a page, a format it already handles well. Read more ↗

Google Releases ax for Coordinating AI Agents

google/ax gained 1,783 stars this week. Google built it as free, open software for running several AI agents together on one job, deciding who does what and in which order. Read more ↗