AI News — September 9, 2026

$ cat top-story.md

▸ Meta Launches Muse, a Personal AI Agent That Books Travel and Pays Bills

Meta shipped Muse, a proactive personal agent (new iPhone app plus web) that can send emails, book travel, buy tickets and make payments by connecting to your calendar, email, shopping and payment apps. It's free, with $20 and $100/month tiers for heavier use. Why it matters: Meta's most autonomous consumer agent yet raises exactly the permission and trust questions that self-hosting an agent on your own infrastructure sidesteps. // TechCrunch

▸ Claude Formalizes Fermat's Last Theorem — the Largest Lean Proof Ever Written

Working largely autonomously over 11 days, Claude produced the first complete, computer-checked formalization of Wiles' proof of Fermat's Last Theorem in Lean: 13 million lines of code, 30,300 theorems, ~6 billion output tokens, verified against Lean's three core axioms. Claude formalized an existing proof rather than discovering new mathematics. Why it matters: A real benchmark for how far agentic, multi-day verification work has come — and a sign of what's now plausible to point Claude at unattended. // Anthropic

▸ "i-have-adhd" Is a Skill That Stops Coding Agents From Burying the Answer

The skill rewrites agent output style: lead with the next concrete action, number multi-step work, cap lists at five items, cut preambles and closing pleasantries. Installable for Claude Code, Cursor and Gemini; it's trending on GitHub with 360+ Hacker News points. Why it matters: A five-minute install if you're tired of scrolling past hedging to find the actual instruction. // Hacker News / GitHub

▸ Claude Code 2.1.265–2.1.266: /diff Command, Plugin Directories, a Gateway Auth Regression Fixed

2.1.265 adds a fullscreen /diff command, --plugin-dir for auto-loading folders of plugins, per-session prompt-cache metrics in /cost, and a 1GB cap on disk-saved tool results; it also fixes subagent resumption breaking prompt-cache reuse. 2.1.266 is a same-day patch fixing a regression where CLAUDE_CODE_USE_GATEWAY wrongly forced Cloud-gateway sign-in for proxy setups. Why it matters: If you run Claude Code behind a gateway or proxy, update past 2.1.265 to avoid the auth-breaking regression. // Claude Code changelog

▸ context-mode Cuts Coding-Agent Tool Output by 98% Across 17 Platforms

An MCP server plus hooks that sandbox raw tool output (logs, API responses, file dumps) into an indexed SQLite FTS5 store instead of dumping it into the context window, and persist session/task state so compaction doesn't lose in-progress work. Works across Claude Code, Cursor, Copilot, Gemini and more. Why it matters: A concrete fix for the "agent burns its context window on tool output" problem in long coding or RAG-debugging sessions. // GitHub

▸ "Deadbugz" MCP Supply-Chain Campaign Hides Its Payload Until the Third Tool Call

A single account filed 23 GitHub pull requests in 74 minutes, pushing a malicious MCP server into unrelated AI/dev-tool projects. The server behaves innocuously for the first two tool calls, then rewrites its own tool metadata on the third to hunt for SSH keys, AWS credentials and Kubernetes configs. Why it matters: Pin and audit MCP server versions rather than trust-on-install, especially anything wired into a self-hosted stack with real credentials nearby. // Pillar Security

▸ Deltafin Runs the Full 2.8T-Parameter Kimi K3 on One Mac by Streaming From SSDs

ARGODRIVE's fork of Deltafin streams Kimi K3's 1.45TB of MXFP4 expert weights on demand from four SSDs into a local disk cache on an Apple Silicon Mac, hitting about 1 token/sec steady decode with fused NEON kernels — no experts pruned or skipped. Why it matters: A working blueprint for running frontier-scale MoE models on hardware you already own when VRAM, or even RAM, can't hold the full weights. // GitHub

▸ Qwen3.8-27B Quantization Benchmark: 4-Bit Matches BF16, 1-Bit Collapses

Testing Q8_0, Q4_K_M, UD-Q2_K_XL and UD-IQ1_S on GPQA Diamond, IFBench and Terminal-Bench 2.1, the 17GB Q4_K_M quant matched the full 55GB BF16 model on agentic coding and fits an RTX 4090 with room for ~64k context — while the 1-bit quant scored near random chance and got worse with longer reasoning. Why it matters: Solid, benchmark-backed guidance on exactly which quant to run locally alongside or instead of Gemma. // Quesma

▸ Microsoft's markitdown Surges on GitHub Trending as a RAG Preprocessing Staple

The Python utility that converts PDFs, Office docs, images (with OCR), audio and HTML into clean, structure-preserving Markdown gained over 2,000 stars in a single day, now past 181k total; recent additions include an Azure Content Understanding integration and a plugin architecture (e.g. markitdown-ocr). Why it matters: A near drop-in ingestion step for turning messy source documents into clean chunks before they hit your embedding model and Qdrant. // GitHub

▸ GPT-6 Astra Changes Everything

Two Minute Papers' take on what GPT-6 Astra's capability jump — and its safety-card caveats — actually means in practice. Why it matters: A fast, skeptical-but-substantive second opinion if you only read the press releases. // Two Minute Papers (YouTube)

▸ 'Model Fatigue' Sets In as AI Labs Race to Ship at a Frenetic Pace

With Anthropic, Meta, Google and OpenAI all shipping major model updates in the same week, IT buyers describe real "model fatigue." Altman attributes some of the pace to labs being back from summer break — but nobody wants to be the one standing still. Why it matters: Useful context for why chasing every release beats picking on capability-per-dollar for your actual stack. // CNBC