Every few months, a new wave of acronyms lands in the AI ecosystem — MoE, RAG, ReAct, ToT, MCP — and every engineer I talk to is quietly maintaining a mental glossary trying to keep up. This post is my attempt to put mine on paper.
What it covers
Twenty-two chapters, grouped into four parts, designed to be read end-to-end or jumped into wherever you need a refresher:
Part One — The Substrate. Tokenization, embeddings and vector space, attention, the large language model itself, and Mixture of Experts. What the model actually is before you start prompting it.
Part Two — Prompting and Reasoning. Few-shot / in-context learning, chain of thought, self-consistency, tree of thought, ReAct, structured outputs and function calling. The behavioral layer sitting on top of the model weights.
Part Three — Knowledge and Retrieval. Knowledge graphs, retrieval-augmented generation, agentic RAG. How models interact with information they don’t have in their training data.
Part Four — Agents, Tools, and Deployment. State machines, LangGraph-style agents, agent memory, multi-agent frameworks, Model Context Protocol, the fine-tuning versus RAG versus prompting decision, evaluation, and guardrails / prompt injection. The systems engineering around shipping something that actually works in production.
Every chapter has a diagram you can read in about twenty seconds and prose that explains the thing without resorting to hand-waving. The whole piece takes roughly fifteen minutes to read.
Read it
The visual primer lives at its own URL with its own visual identity — a darker paper, a different typographic voice, a full set of custom diagrams. It’s intentionally designed as a standalone artifact, so this post is mostly a pointer. Hit the button above and enjoy.
← Back to Autonomy