Here's a roundup of AI news from the morning of August 6, Japan time. It was a day lined with dense, developer-leaning topics: coding agents, AI-focused infrastructure, robotics, local LLMs, research, and community governance.

Meta's First AI Coding Agent "Muse Code" Rolls Out in Beta

Meta CEO Mark Zuckerberg announced on social media on August 5 that the company has opened a beta of Muse Code, its first AI coding agent. The agent is said to handle software engineering tasks such as writing code and verifying the code it writes, signaling a full-fledged entry into territory where OpenAI and Anthropic have led the way. It's still in beta at this point, so its real-world quality and coverage remain to be verified.

AI-powered coding assistance is spreading fast through the field, and plenty of developers feel they're living through a transition. Pieces that sort through this change from a practitioner's perspective — like "Notes on Migrating to AI Coding" — are starting to appear.

Cloudflare Announces Open-Source "Cloudflare OS" for AI

Cloudflare has announced Cloudflare OS, open-sourced and billed as an "operating system" for AI. It reads as a move to build out a foundation layer designed with AI workloads in mind. Because the announcement just landed, the detailed specs and scope should be confirmed against primary sources, but an edge provider expanding its territory all the way into the OS layer is worth watching.

Xiaomi Releases Robotics Foundation Model "Robotics-1 (XR-1)"

Xiaomi has released Xiaomi-Robotics-1, a robotics foundation model trained on more than 100,000 hours of real-world manipulation data. It's a Vision-Language-Action (VLA) model intended for mobile manipulation that stays useful even in unfamiliar environments, and it's trained in two stages — pre-training followed by post-training — mirroring the recipe used for large language models. Architecturally, it couples a pre-trained VLM (Qwen3-VL) with a Diffusion-Transformer (DiT) via Mixture-of-Transformers, a design that keeps inference speed in check. The work reportedly also validates whether scaling trends from training carry over into real-robot performance through post-training. The model is available on Hugging Face and GitHub, and the paper is out as well.

"Mference" Runs Thinking Machines' Inkling-Small in Under 10GB

An effort to run gigantic MoE models even on consumer Macs has been reported. According to the post, Mference — an inference engine written in Swift and Metal — ran Inkling-Small (276B, roughly 12B active), Thinking Machines' Apache 2.0 model, on a 24GB Mac with a 3.4GB resident footprint and a peak below 10GB, pushing out about 2.9 tok/s. Disk usage swelled to around 148GB, and the first pass on long prompts is slow — but the fact that a several-hundred-billion-parameter model runs in roughly 10GB of memory is significant. It's text-only for now, supporting four model families: Gemma, Qwen, DeepSeek-V4-Flash, and Inkling.

Google DeepMind's Picture of "What Comes After AGI" — Four Paths to ASI and Six Bottlenecks

Google DeepMind has published a paper laying out how far AI could go once human-level AGI is achieved. It presents a framework of four paths toward ASI (artificial superintelligence) that lies beyond AGI, plus six bottlenecks, and the content unpacks a dense debate in a way that reaches beginners. The details of the specific paths and barriers are left to the paper and its explainers, but what's interesting is the growing shared view that "AGI is a waypoint, not a destination."

Linux Staging to Reject LLM-Generated Patches in Principle

The Linux kernel's Staging tree has switched to a policy of not accepting LLM-generated patches, as a rule. The only exception is said to be "genuine security fixes." The decision turns on the quality of generative-AI-written code and the review burden it creates — and a movement is spreading in which each community makes its stance on LLM-submitted contributions explicit.