AI TECHNOLOGY
Readable notes on LLMs, tokens, KV Cache, inference bottlenecks, memory movement, and AI accelerator architecture.
Twelve illustrated lessons explain data movement, reticle stitching, defect tolerance, local SRAM, dataflow, CSoft, CS-4, prefill and decode, and the trade-offs among Cerebras, GPU clusters, and hardwired inference.
A ten-plate classroom on hardwired inference: why moving weights is expensive, how a 4-bit lookup-and-select teaching model works, and how to read HC1's public numbers without overclaiming.
A story-led walkthrough of tokens, Query-Key matching, Value mixing, the two memory walls, and what KV cache saves—and still has to read—during autoregressive decoding.
A six-page classroom that starts with a simple question: how does an LLM keep writing one piece at a time? We begin with next-token prediction, then move into tokens, vectors, K/V notes, KV Cache, and the memory cost that appears when conversations get long.
