MacinAI Local runs TinyLlama on PowerBook G4

This is less a nostalgia stunt than a real systems demo: the impressive part is not just getting TinyLlama to boot on a G4, but stitching together a full local AI stack for classic Mac OS. The tradeoff is obvious, though, because the architecture is clever enough to work and slow enough to make the result feel like a technical proof rather than an everyday assistant.

• –The from-scratch C89 engine and export pipeline make this model-agnostic, which is far more interesting than the usual one-off retro LLM port.
• –Disk paging is the enabling trick for 1.1B parameters on 1GB RAM, but the 9.9 seconds per token figure shows how quickly practicality falls off.
• –The AltiVec work is the real engineering flex here, especially since the project uncovered a CodeWarrior compiler bug along the way.
• –AppleScript generation plus confirmation prompts gives the project a genuine agentic angle, not just text completion on old iron.
• –Compared with earlier retro-AI demos, this reads as a platform with a UI, tokenizer, memory model, and automation layer, not a novelty benchmark.