Understand¶
Understand · 5 min · no code
Before this: Choose your path · After this: How models work The no-code layer. Every page here links to its hands-on and in-depth versions.
Nine pages, in order, each needing only the ones before it. Read straight through and you will have an accurate picture of how these systems work, where they fail, and what to ask of anyone proposing one. No code, at any point.
The order is the argument: prompting needs models, retrieval is "give the model your data", an agent is "let the model act" and needs both, and safety comes after agents because the failure that matters arrives through a tool result.
Start at AI 101 if you have not read it — it is page one of this route and sets up everything here.
-
2 · How models work
Tokens, context windows, inference settings, reasoning models, and why parameter counts stopped meaning anything.
-
3 · Prompting
System prompts, few-shot examples, chain of thought, and getting a model to do what you actually meant.
-
4 · Retrieval and data
RAG, embeddings and vector search: how a model comes to know things it was never trained on.
-
5 · What an agent is
The components, the loop, multi-agent systems, and the more important question of when not to use one.
-
6 · Agentic AI
Tool use, memory, orchestration patterns, human-in-the-loop, and the protocols underneath: MCP, A2A, AG-UI.
-
7 · Enterprise AI patterns
The five shapes this work actually takes in an organisation: copilot, autonomous agent, document processing, conversational AI, agentic RAG.
-
8 · Safety and responsible AI
Hallucination, prompt injection, guardrails, red teaming — and why prompt wording is not a security control.
-
9 · Fine-tuning and training
What fine-tuning changes and what it does not, LoRA, RLHF, and why the answer is usually retrieval instead.
-
10 · Infrastructure and operations
MLOps, drift, quantisation, edge deployment and what actually drives cost.
After these¶
If you write code, take the build path: the same ideas as thirteen modules that each end in something you run and break, a model on your own machine.
If you want depth on retrieval, Retrieval in depth is six pages on chunking, embeddings, vector databases, GraphRAG and evaluation.
If you are choosing what to read elsewhere, Resources is a ranked list with a reason to trust each item, and an explicit list of what to avoid.