Ai

Edge0's founder is pulling 3,600+ likes on LinkedIn with a slick clip of an iPhone running a 35B model on-device. The trouble: there is no iOS app in the Edge0 repo, the founder concedes the phone app is 'coming soon,' and the demo reads as AI-generated. Meanwhile Silo, a one-developer app on the App Store since early 2025, already does private, offline, on-device LLM chat for real, with a year of unglamorous commit history to prove it. I walk the two repos side by side: what ships vs. what's narrated, and why the harder claim (35B MoE streaming on a phone) is still unproven.

antirez's h3.c runs a full MiniMax-H3 video model on Apple Silicon in ~35,000 lines of C, Objective-C, and Metal -- no Python, no PyTorch, no CUDA-style runtime. I compare it to the pain of getting PyTorch running on AMD CPU + AMD GPU + ROCm on Linux in mid-2023, and argue both stories are really about one thing: removing abstraction layers you don't actually need.