Meta open-sources Muse Gadgets for DIY AI hardware
Friday's news puts the agent layer in the physical world: Meta hands Muse's hardware over to tinkerers, a presidential AI rebrand unexpectedly moves a European domain registry, and one team's month of open-model-only coding ends with a capacity lesson rather than a capability one.
Meta is open-sourcing the hardware layer of its Muse agent. Launched Friday, Muse Gadgets ships open-source firmware plus ESP32 and Linux SDKs so anyone can wire buttons, screens, sensors and actuators to Muse, and it introduces the Muse Home Link — a small adapter that puts Muse on the home network so it can drive compatible devices: a light, a TV, a printer. The Home Link is free for US Muse subscribers and ships in October; Nat Friedman announced the release in a post on X. The point: Muse already lives in a sealed VM in Meta's cloud, and gadgets give that agent a physical footprint without Meta having to build every endpoint itself — an ecosystem play of the Alexa-skills variety, except the stack is open and user-built. Meta's own disclaimer sets the tone: side effects of tinkering "may include bricked boards, voided warranties, brownouts, or bankruptcies." We covered the agent's launch — Meta ships Muse: a consumer agent inside a sealed VM.
Slovenia's .si registry is the accidental winner of Trump's AI rebrand. The .si registry put September's registrations at roughly 2,199% above the prior month — about 44,000 new domains — after Trump proposed using "super intelligence" in place of "artificial intelligence" and signed an executive order on September 29 directing federal agencies to adopt "SI". The registry's own spokesperson urged caution against attributing the spike entirely to the order, and domain intelligence firm Netcraft found a speculative layer underneath it: a large share of pre-announcement purchases listed for resale, with thousands of registrations snapped up the day after the executive order. The take so far: a rename meant to project American AI supremacy has mostly moved a small European country-code registry's order book, and much of the demand looks like people trading the joke rather than using it. We covered the rebrand itself — Trump's AI rebrand lands with a six-CEO self-audit pledge.
A month coding on one open model ended at 50% — the wall was capacity, not quality. The Wagtail core team's September "open models only" challenge finished at 1 billion of 2 billion tokens on GLM 5.3 Flash, and the misses tell the story: a vibe-coded MCP prototype burned 450 million tokens for about $150 almost overnight, and capacity problems at popular inference providers pushed day-to-day work onto DeepSeek V4.1 Flash and Qwen 3.8 Flash. The target model itself stayed cheap — $68 for the month — but total energy use landed around 35 kWh against a 10 kWh goal, and their 14-model benchmark put DeepSeek V4.1 Flash on top at 95% accuracy for about $0.09 per task. The lesson is less about model quality than discipline: budget the prototypes, measure spend daily, and assume the cheapest good model can be capacity-constrained. The thread hit Hacker News' front page, where at least one commenter priced that 450 million-token burn closer to $300 than $150.
What to watch: whether Meta opens a public showcase for community gadgets in October, and whether Wagtail's October run — with experiments budgeted up front — actually hits its single-model target.
Would your team survive a month locked to one cheap open model, or is 50% adherence the realistic ceiling? Tell us in the comments.
