Speculative decoding made my local LLM actually usable

The problem wasn't the brain, but how it was being forced to think