metastable blog

Notes & explainers

Visual explanations of AI systems, silicon, and inference.

Step-by-step walkthroughs of how machine learning actually runs — from matrix arithmetic up to the hardware that executes it.

Writing

  1. Why transformers need a KV cache

    The attention matrix, built one token at a time — and the redundant work that makes caching keys and values unavoidable.

    1 mininferencetransformers

  2. How matrix multiplication actually works

    A stepped, visual walk from the black box down to the multiply-accumulate that silicon is built around.

    1 minfundamentalssilicon