I wrote a free 15-part series on LLM internals — real math, real tensor shapes, real hardware constraints. All grounded in Gemma 4 12B's actual config.