The algorithms that make a large language model answer on demand — attention and the KV cache, speculative decoding, quantization error, and vector-index recall — derived properly, with the complexity and the point where each one breaks. Grounded in the primary papers.
1 series · 1 article