About this guide

How to use it

• As a reference: open the page for the thing in front of you. Each page is built to be read in a few minutes: what it is, why it matters to you, what to do.
• As a course: three basic sessions cover "Under the hood," "Using it at work," and "Where it fails" (Parts 2, 1, and 3). Three intermediate sessions cover "Agents and beyond" (Part 4) plus the pipeline page. Each session pairs one concept with a hands-on exercise and one busted myth; the exercises are marked "Try it" on the pages.
• As a handout: print the verification checklist and the glossary. Both are set to print cleanly.

What you will be able to do

After Parts 1–3:

• Say in one sentence what an LLM is and isn't.
• Sketch what is actually inside the prompt.
• Define token, context window, and compaction.
• Name the kinds of work it is good at, pick the right GenAI.mil tool and model for a task, read a model's spec sheet, and say what may be entered at IL5.
• Run the verification checklist on every factual or numeric output.

After Part 4:

• Explain agents, subagents, tool calls, and MCP, and read an agent's transcript.
• Say what RAG is and when it beats attaching a document.
• Explain why an API call differs from the web interface, and sketch a three-step pipeline with a check in the middle.
• Name the main causes of hallucination and the honest limits of each fix.

Dates and drift

Model names and platform facts change monthly; the mental models do not. Every page whose facts can drift carries a "checked" date under its title. The GenAI.mil lineup, the vendor table, the Anthropic status, and the training prices are the pages to re-verify first.

Related

• How LLMs Actually Work: the under-the-hood book, if you want the math.
• AI/ML from the Ground Up: what machine learning is, from fitting a line to neural networks.
• The Basics on Latent Diffusion: how AI makes images.