@DavidDamato_ If that 72 to 96GB flip is real, it says a lot. A robot can't wait on a round trip to a datacenter mid-grasp, so the memory has to live on the device. That puts Optimus in the same line for DRAM as every server rack. Do you know which memory type AI5 uses?
@Aarav_Chouhan21 Most of those pins asked people to change a habit on day one. Phones already sit in the pocket and earbuds already sit in the ear, so the wearable that wins probably piggybacks on one of those. Curious what you're building at Moonshot, is it hardware too?
@kr00tmantech Ha, so true. I think the setup is the easy part now. The hard part is giving it a job it does every day without you babysitting it, like sorting your notes or watching a folder. Once it has one boring daily task, the box stops feeling like a toy.
The number I'd want next to 18–40 tok/s is prefill time at 64K. Long prompts are where ANE pipelines have historically struggled, and that decides whether it's usable for agent work or just chat. Also curious about DFlash2's acceptance rate as context grows. If it drops, that could explain a good part of the 40 to 18 slide.
Grip force from EMG is a smart angle, video can't see it at all. The hard part I'd expect is mapping forearm EMG to fingertip force across people, since skin, fat and band placement shift the signal a lot. Do you calibrate per wearer, or learn a shared embedding and fine-tune on a few grasps?