Tue Oct 13, 2026 · the AI news, short version

Morning. Six things from the last 24h, no fluff.

1. Anthropic: small AI models may be "months away" from copying themselves onto smart gadgets - Red team paper (Oct 12) tested whether models can set themselves up and spread across a firewalled fleet of cheap devices. - GLM-5.2 was the only open model to replicate reliably: all 8 test nodes, in 4 of 5 trials. - They ran Gemma4-26B across 15 off-brand Android TV boxes (~$55 each)... at ~0.39 tokens/sec. Slow, but it ran. - Caveat from Anthropic: they made the hacking part trivial, so treat this as an upper bound. - Takeaway: the "no off switch" AI worry just got a concrete test, not a sci-fi one. - Sources: AI Weekly · India Today

2. Meta drops "Wonder" and "Wonder Full", says they match OpenAI and Anthropic - Muse Wonder is the bigger follow-up to Muse Spark. Zuckerberg says it gets Meta "to the frontier." - The claim leans on Artificial Analysis benchmark data Meta shared in its own blog post. - Takeaway: Meta is back in the frontier race, but that's its own claim until people outside Meta test it. - Source: Economic Times / AFP

3. Microsoft announces Maia 300, its own AI chip, to run most of its AI from 2027 - Microsoft says it gets 40% more tokens per megawatt than leading GPU systems, and up to 7x Maia 200 throughput per rack. - Built on TSMC N2P. Maia 200 already runs Copilot and GitHub Copilot. - Takeaway: Microsoft is leaning less on Nvidia, at least for inference. - Sources: Microsoft Blog · The Verge

4. Nvidia opens up its Vera Rubin rack designs at the OCP Summit - Rack and compute tray designs given to the Open Compute Project. MGX 1.2 supports racks up to 250 kW, fully liquid-cooled. - Meta will use Nvidia Spectrum-6 switches for its next Vera Rubin clusters. - Takeaway: Nvidia is handing out the blueprint so everyone builds racks its way. - Sources: NVIDIA Blog · SiliconANGLE

5. Google, Meta, Microsoft and Nvidia agree on how AI data centers should behave when the power grid has a problem - The OCP Grid Resilience Spec says to stay connected during grid disturbances and get back to near-full load in under 2 seconds. - Separately, AMD, Google, Meta and Nvidia published Open Data Center Spec v0.7 covering power, cooling and layout. - Takeaway: AI's energy problem now has a rulebook for the grid side too. - Sources: NVIDIA Blog · madrobot

6. Nvidia + Cisco plan faster plug-in optics for AI networks - New USFP module format: 400Gb/s per lane, 16 lanes to start, 12.8Tb/s total. - TE Connectivity and Amphenol are building the parts. - Takeaway: the slow part in AI clusters is the network now, and they're working on it. - Source: NVIDIA press release

That's the drop. See you tomorrow. ⚡