Big thing
Meta built a humanoid with China's Fourier in three months, The Information reports
The prototype, called Max, is about 1.5 m tall and 48 kg, runs on NVIDIA's Jetson Thor, and has a face display and five-fingered hands. Meta's VP of robotics product spent nearly two months at Fourier's Shanghai factory, and Max 2 has been in the works since September, with better hands and sensors for stairs (@XRoboHub). Meta sees humanoids as a body for its Muse agent, but says it has no plans to sell these (@Techmeme).
@TheHumanoidHub: "Outsourcing to a Chinese contractor gets you a demo fast. Whether Meta can turn that into a scalable product is a completely different question." @darcyagi: "zuck is going after pretty much every interface through which humans might interact with ai."
Labs news & launches
1. DeepMind's agents built a dancing robot from one plain-English prompt
AI agents running Gemini 4 Argon in Antigravity designed and simulated a jack-o'-lantern robot with 12 degrees of freedom, which the team then built and set dancing, with no manual CAD (@Google). Aaron Saunders, who works on DeepMind's robot hardware, posted it for Halloween.
@buzzhive_ai: "a robot only gets honest on a real floor. Sim says it works, then friction and wobble say otherwise." @NINJA____1: "One prompt to a real dancing robot with no CAD is actually insane."
Heat: medium · 283K views, 8 posts · quiet on Reddit
Key research
Two copies of one AI swapped a secret word using only a robot and a camera
Robocurve kept two copies of one model apart, with no protocol agreed beforehand, and they still swapped the word. Each model found its own code: Claude Opus 5.5 air-writing, GPT-6 Astra Morse, GPT-6.1 Sol flag semaphore, Gemini 4 Argon 5-bit binary. Perfect decodes came above 90% of the time. It matters because copies of one model share the same obvious codes, so robots running it can coordinate through the physical world with nothing set up first.