A hands-on build on an Apple Silicon Mac: local Llama versus Claude versus cloud Llama, where the tokens actually go, a bug that cost me an afternoon, and what really happens inside GPT-2 and Qwen3.
I have been building a course on AI agents for network operations, and I decided to do it the hard way. Run everything myself, on my own machine, and write down the real numbers instead of the tidy ones from a slide. This post is what came out of the first two hands-on parts: local models against cloud models with every token counted, a bug that made me question my choices, and then opening up a model to see the machinery inside.
Everything here ran on a MacBook Pro with an M5 Max and 36 GB of memory, with 32 Core GPU. No rented GPUs. The scripts are small enough to read in one sitting.