Muse Glimmer Goes Local, and 181,874 Meetings Leak

Meta open-sourced Muse Glimmer this week, a 30B-parameter agent model under Apache 2.0 that runs on a single consumer GPU, joining a wave of open and on-device releases: Alibaba published the 2.4T-parameter weights for Qwen3.8, Cactus shipped Needle2 (a 14 MB agentic LLM for sub-$200 devices), and DeepSeek and xAI both released new models. Alongside it, the seams in AI-era infrastructure kept showing: tl;dv left 181,874 meeting records queryable by any user, a researcher who bought noreply.net has been receiving companies' secrets for 18 months, and the Stolen Thoughts paper decoded the encrypted reasoning blocks in public agent trajectories to recover API keys and passwords.

News

Over the past week the frontier tilted toward open and on-device releases. Meta's Muse Glimmer (research.meta.ai) is a 30B-parameter model optimized for always-on local agent workflows — tool calling, long-horizon execution, and LLM-as-a-judge — small enough for a single consumer GPU, with llama.cpp, MLX, and ExecuTorch integrations on the way. Alibaba published Qwen3.8-2.4T-A95B (Hugging Face), a 2.4T-parameter MoE with 95B active parameters, a context window extensible past a million tokens, and gated DeltaNet-plus-attention layers. DeepSeek shipped two models — V4 Flash 0731 (ARC Prize) scores 89.0% on ARC-AGI-1 Semi-Private at $0.02 per task, while V4 Pro 0813 (OpenRouter) went GA as a MoE with 1M context. xAI's Grok 4.6 (x.ai) targets long-running agents and visual work, matching GPT-5.6 Sol on the Artificial Analysis Intelligence Index at $2/$6 per million tokens.

Anthropic detailed how Claude marks AI-generated content (Claude Help Center): models launched on or after August 2, 2026 embed an imperceptible watermark in generated text and attach signed C2PA provenance metadata to generated files such as SVG, PNG, and JPEG. The company signed the EU AI Act's Code of Practice on transparency of AI-generated content, and says the marks apply wherever Claude is offered and will be detectable by users and third parties.

Three security stories landed in the same stretch. tl;dv (bobdahacker.com), the AI meeting recorder with over 2 million users, left its Firestore "meetings" collection with no tenant isolation, so any authenticated user could enumerate 181,874 records across 84,312 accounts — including live conference IDs for joinable Google Meet and Teams rooms; the researcher walked into a live Malaysian Ministry of Education call. Separately, a researcher who bought noreply.net (Ars Technica) has received roughly 400,000 misdirected emails in 18 months — injury reports, pizza orders, test credentials — sent by companies that assume the domain goes nowhere. And Stolen Thoughts (stolen-thoughts.com) demonstrates a pipeline that decodes the encrypted reasoning blocks still present in public Claude, GPT, and Gemini agent trajectories, reconstructing 315,320 blocks and 704 privacy artifacts, including 62 API keys, 33 passwords, and 24 access tokens.

Oracle banned AI-generated code from OpenJDK (Dealroom), citing safety, security, and IP risk — contributors may use LLMs privately for debugging and review but cannot submit AI-generated material — even as Larry Ellison says AI models write Oracle's own code. Meanwhile the Nixpkgs core team disbanded (Discourse), writing that the role had become heavier than the "lightweight role compatible with active technical contribution" they'd hoped for and citing micromanagement and poor communication from the Steering Committee; over ten months the team had reformed committer delegation and onboarded 19 committers.

Articles

Tailscale traced 19 database corruption incidents to a 16-year-old bug in SQLite's WAL-reset path (tailscale.com). Their control plane runs one SQLite file per shard behind a single Go process, and corruption took tailnets' control planes down while peer-to-peer connections kept working; the post walks through the forensics that eventually isolated and fixed the bug upstream.

Senko Rašić pushes back on the idea that "code was never the hard part" (blog.senko.net). If coding were easy, he asks, why the high salaries, the leetcode gauntlets, and the doorstopper textbooks — and if "figuring out what to build" were the hard part, why aren't product managers interviewed and paid like rockstars. His answer is that both craft and customer empathy matter, and that dismissing either is cope.

Florian Herrengt's AI is removing the middle class of software engineering? (blog.florianherrengt.com) argues AI removed the speed limit on code production, so teams with weak engineering culture now fail faster — engineers open 25,000-line PRs, nobody understands the systems they built, and people ask Claude where their own data comes from. His claim is that this squeezes the competent-but-unexceptional middle of the profession while the best engineers grow more valuable.

Laurentiu Raducu's How I use LLMs to learn complex topics (laurentiugabriel.github.io) builds low-poly, Rollercoaster-Tycoon-style interactive simulations instead of asking for bulleted explanations: plan mode builds a knowledge base, a second pass reviews its accuracy, and a third turns it into a playable animation — ChipTycoon follows a wafer from sand to data center.

Google Search Is Dying. What Comes Next Is Worse (thewalrus.ca) documents the web's retrieval crisis: AI summaries invent basic facts like sunset times, link rot erases pages daily, and AI systems scrape Wikipedia directly so traffic and donations no longer flow back to it. The piece points to Disney deleting FiveThirtyEight's entire archive and companies planting content on Reddit to steer AI answers, framing the collapse as a question of who preserves the cultural record.

Tools

Delta (Rust, private beta) is Zed's multiplayer environment for coding with agents. Its DeltaDB replicates the conversation and the worktree together in real time, comments attach to specific lines or plan steps, and a WebAssembly/WebGL build lets teammates open a thread in the browser; it connects to Claude Code.

Docker Sandboxes (macOS/Windows/Linux) runs coding agents — Claude Code, Copilot CLI, Codex, OpenCode, Kiro — inside disposable microVMs so "YOLO mode" (--dangerously-skip-permissions) needs no manual review. Only the project workspace is mounted, and agents can spin up their own containers inside.

Mojo 1.0 (Mojo) reaches a stable, production-ready foundation after converging on a single var syntax, Python-style lambdas, and new memory-safety diagnostics; the compiler is still slated to be open-sourced this year. The accompanying MAX 26.5 adds GLM-5.2 and Nemotron-H support.

h3-metal (C, Apple Silicon) is antirez's native MiniMax-H3 inference engine, built as a sequence of vertical slices with prompt-to-video/audio, first/last-frame conditioning, and image references working end to end, currently being optimized for M3 Max and M5 Max.

Needle2 (C++, tiny devices) is a 14 MB agentic LLM with 45M parameters for tool calling and structured extraction on sub-$200 hardware, trained with 2-bit quantization from pretraining rather than post-hoc; it runs at 800+ tokens/s prefill on a Raspberry Pi 5 and ships as a single dependency-free binary, with Pebble using it in its Index 01 ring.

Worth a Look

nixpkgs-multiverse (Nix) is a flake with no inputs that fetches all 1,393 Nixpkgs revisions lazily via builtins.fetchTree, so you can nix run ...#versions.python3."3.6.2" and mix any package version that ever shipped into one shell.

My server is a phone now documents replacing a Hetzner VPS with a CMF Phone 1 running Termux, runit, Caddy, and Tailscale — the postmarketOS route soft-bricked the device, so stock Android kept the working drivers while each app brought its own filesystem.

Dithered QR codes explains how to embed an image into a QR code by shrinking the data modules and using Floyd-Steinberg error diffusion — including a second diffusion pass to mask the salt-and-pepper noise the data modules introduce.

_for-sale DNS records describe RFC 10023's _for-sale TXT convention, which lets a domain owner signal willingness to sell without parking the site — one lookup tells brokers what a rendered homepage never could.

2026 Eclipse Webcams assembles a live countdown and a set of public webcams along the path of the August 2026 total solar eclipse, tracking totality as it reaches each camera.