smolcluster demos SyncPS inference on Mac minis

This is a legit systems demo: the point is less “look, a 1B model runs” and more “look, a scratch-built cluster can coordinate inference across cheap Apple Silicon boxes.”

• –SyncPS is a neat teaching architecture, but it adds coordination overhead that makes it more interesting as a research prototype than a production inference path
• –The three-Mac-mini setup is compelling because it makes distributed LLM experiments accessible without server GPUs or heavyweight cluster software
• –smolcluster’s biggest differentiator is the from-scratch networking layer, which gives it educational value and a lot of flexibility for experimentation
• –For a 1B model, the real signal is orchestration and transport correctness, not scale; the next interesting step would be latency and throughput benchmarks
• –The project still feels closest to an infrastructure playground for makers who want to learn distributed inference by building it themselves