Meta open-sources Muse gadgets you can build

Sciencepositive +0.7#embodied-ai #agents #meta #muse #hardware

Meta open-sourced SDKs so anyone can wire Muse into ESP32 boards, Raspberry Pis, E Ink, HDMI sticks, and DIY Charms. It is also waitlisting 5,000 Muse Home Link boxes that run community skills for lights, TVs, and printers.

What they showed / shipped

Meta: "Muse gadgets are open source devices you build yourself. Program an off-the-shelf ESP32 board or set up a Raspberry Pi with our SDKs, then connect Muse to your displays, buttons, sensors, actuators" (theverge, techcrunch).
Suggested builds: color E Ink for reminders, HDMI stick for big-screen Muse, small touchscreen DIY Charm. Meta labels it "Proceed at your own risk."
Muse Home Link: Meta made 5,000 units (Nat Friedman). Claim a waitlist spot now; ships this month. Community skills can turn on lights, control a TV, send a doc to a printer, and more.

Why it matters

The agent leaves the phone and lands on cheap hardware you already own. ESP32 + Pi is the new surface.

Sources

ds4: Redis creator's local DeepSeek V4 engine

Techpositive +0.8#open-source #local-ai #tooling #ds4 #infrastructure

Salvatore Sanfilippo (antirez) shipped DwarfStar 4, a narrow C inference engine for DeepSeek V4 / V4.1 Flash, GLM 5.x, and Qwen3.8 on Metal, CUDA, and ROCm. Not a generic GGUF runner: project weights, disk-backed KV cache, CLI, local API, and a native agent.

What they showed / shipped

ds4 from antirez (Redis creator): MIT-licensed C engine aimed at high-memory Mac / CUDA / ROCm boxes. Supports DeepSeek V4 and V4.1 Flash, GLM 5.x, Qwen3.8 Flash Next, text + vision (dwarfstar.sh, hn).
Tricks that matter: asymmetric 2-bit quantization on routed MoE experts, KV cache persisted to SSD by prompt-hash (restarts skip full prefill), one model state shared by ./ds4, ./ds4-server, and ./ds4-agent.
Reference: M5 Max 128GB q2 at 2k ctx ≈ 790 t/s prefill / 39 t/s gen. Server speaks OpenAI and Anthropic-style APIs so Codex / Claude Code / OpenCode can point at localhost.
📌 On the radar: Perplexity also open-sourced Lily, a Rust + custom Metal local inference engine (~1.23× prefill / 1.35× decode vs MLX-LM on M5 Max), plus a 0.6B on-device PII-Tracer (x/@AravSrinivas).

Why it matters

Frontier open MoE on your desk without a generic GGUF lottery. Plug coding agents into ds4-server.

Sources

FLUX 3 holds structure across nine edit turns

Mediapositive +0.8#image-gen #flux-3 #creator-economy #ship-it

Venturetwins stress-tested FLUX 3 on an old Chloe ad: nine multi-turn swaps of dresses, couch, wallpaper, makeup, and logo with no visible degradation. Multi-turn image editing without drift is the creator unlock.

What they showed / shipped

@venturetwins: "We've definitely cracked multi-turn editing without image degradation" after nine turns on a Chloe ad, swapping dresses, couch, wallpaper, makeup, and logo while keeping structure (x/@venturetwins).
Claim is product-level: iterative brand edits that used to accumulate artifacts now stay clean enough for real ad work.

Why it matters

Multi-turn consistency is the hard part of image agents. A working demo is the signal.

Sources

x/@venturetwins X@venturetwins
♥ 591↻ 44👁 63K

Claude-shaped science: BootLoops across 18 fields

Sciencepositive +0.7#science #claude #bootloops #agents #open-source

Harvard physicist Matthew Schwartz stopped forcing Claude to act like a human scientist and instead hunted "Claude-shaped" problems. Result: BootLoops, an open harness, plus 36 manuscripts across 18 fields with 19 coauthors in three months.

What they showed / shipped

Anthropic guest post "Claude-Shaped Science": Schwartz built BootLoops, an open-source harness for exact quantitative calculations, then steered it with domain experts (anthropic, bootloops.ai).
Physics start: ported S-matrix bootstrap methods, then cracked elliptic Feynman integrals humans had barely touched (30 integrals end-to-end; 15 brand new).
With experts: ecology model that beats neutral theory on Barro Colorado Island; 5.7B mutation pairs showing gene conversion; AI data editor that ported 4,452 economics replication packages; AccStack stress database covering 6,072 languages.
Scale claim: 36 manuscripts in 18 fields with 19 coauthors over three months, from ~400 candidate problems. HN also surfaced the related "36 papers with Claude" thread (reddit/r/Physics).

Why it matters

Stop asking models to be scientists. Build harnesses for the problems they are already good at, then bring humans for taste.

Sources

Apple locks down Full Disk Access for AI agents

AImixed +0.1#agents #safety #apple #policy #infrastructure

Days after Muse-on-Mac privacy claims, Apple said it will force much more explicit consent for macOS Full Disk Access. Reason named in public: AI agents "substantially" raise the risk of apps seeing mail, messages, files, and browsing history.

What they showed / shipped

Apple developer note: Full Disk Access was meant for backups, but "some developers are using Full Disk Access in ways that could put users at risk, exposing everything on their systems…without users' full knowledge" (techcrunch, theverge, arstechnica).
Going forward: "very explicit user action" required before an app gets that level of access. Apple frames it as informed consent as agents get more autonomous.
Trigger context: Inc. columnist alleged Muse read private messages without clear permission (Meta disputed); Wired also reported a ChatGPT Mac app flaw that could expose sensitive data.

Why it matters

Desktop agent products just got a platform tax. Expect clearer permission UX and narrower entitlements.

Sources

Suno Speech and Stability's licensed music bet

Suno launched Speech, an audio model that generates voiceover and background music as one track (public beta, ~8 min max). Separately, Sean Parker is rebuilding Stability AI as a licensed music toolmaker after a $76M round with Sony, Warner, and Universal.

What they showed / shipped

Suno Speech (public beta, web + mobile): Simple mode (describe a pirate captain speech) or Advanced (paste a script). Optional background music toggle, gender/style/variety controls, ~8 minute cap (theverge).
Suno CPO Jack Brody: "the first audio model that generates voice and music together as one cohesive track." Beta caveats included in the post (accents wander, dramatic pauses get dramatic).
Stability AI: Parker + CEO Prem Akkaraju positioned the company for music pros after late-August $76M with Sony, Warner, Universal (catalogs licensed for training). Three new audio models plus music-editing software; upcoming hum/beatbox steering (techcrunch).

Why it matters

The Stability path shows the licensed-catalog deal is becoming table stakes for music AI that wants to survive.

Sources

Amazon wants an $8B Nvidia chip SPV

Businessmixed +0.1#economics #infrastructure #amazon #nvidia #hardware

FT/Reuters: Amazon is shopping an ~$8B special-purpose vehicle that would take Grace Blackwell chips out of its data centers, sell up to 10% equity to outside investors, and lease the GPUs back. Asset-light infra finance is the story.

What they showed / shipped

Reuters citing FT: Amazon held talks to move thousands of Grace Blackwell chips installed across a dozen-plus US data centers (five states incl. Nevada and Virginia) into an SPV, then lease them back (reuters).
Vehicle would raise debt from outside investors; Amazon may offer up to 10% equity. Goal: strengthen the balance sheet and run a more asset-light AI infra stack.
📌 On the radar: Epoch AI estimate circulating that chips shipped through 2027 could support ~1.9B agents running at once, enough to match total human working hours (reddit).

Why it matters

The bottleneck is capital structure as much as wafers. Hyperscalers are inventing chip finance products.

Sources

Try today: Opus overnight games, FrogNano, Lego AI

Businesspositive +0.8#ship-it #creator-economy #open-source #solo-founder #local-ai

Three things you can poke now: Min Choi woke up to an Opus 5.5-built Roblox game (~11k parts) and shipped Hunt A Squishy with his kids; Microsoft dropped FrogNano-4B on Hugging Face; and an open-source Lego AI generator hit Show HN.

What they showed / shipped

@minchoi: one game-idea plan into Opus 5.5 before bed → ~11,000 Roblox parts; week of kid co-build → Hunt A Squishy live. Separate overnight: one prompt + music → finished creative output (x/@minchoi, x/@minchoi).
microsoft/FrogNano-4B-2609 on Hugging Face for local runs (reddit/r/LocalLLaMA).
Show HN: open-source Lego AI generator (ldraw-nova) (github).
Bonus local tip: iPhone-as-second-GPU for a 24GB MacBook claimed 29–44% faster Qwen 3.8 27B prefills by parking part of the context window on the phone (reddit).

Why it matters

FrogNano + Lego generator + phone-offload are concrete weekend experiments.