Thirteen explainers, in the order that makes them easiest to absorb. Each one stands alone, but read in sequence they build on each other — later pieces assume the vocabulary earlier ones establish.
If you only have twenty minutes, read the first two. If a term stops you, the glossary defines it in a line and links back to the explainer that covers it properly.
Foundations
What these systems are and how they represent language. Everything else assumes this.
1
What Is a Large Language Model?
The single mechanism — next-token prediction — that explains both what these systems can do and how they characteristically fail.
2
Tokens and Embeddings
Why a model cannot count letters, why some languages cost three times more, and how meaning becomes geometry.
3
Transformers Explained
The architecture underneath everything since 2017, and why attention made scaling possible.
4
Context Windows Explained
The hard limit every AI application is designed around, and why a bigger window is not automatically better.
What they can do
The four patterns that account for most of what gets built on top of these models.
5
RAG Explained
How systems answer questions about private or current data the model was never trained on.
6
Fine-Tuning vs RAG vs Prompting
The decision that costs teams the most money when they get it wrong. One question sorts it.
7
Tool Use and MCP
How a model stops merely talking and starts doing — and where the security boundary actually sits.
8
AI Agents Explained
What agents genuinely do well, where they fail, and why the honest picture is narrower than the marketing.
9
Reasoning Models Explained
Spending more computation at the moment a question is asked, and when that helps rather than hurts.
Where they break
Read these before trusting any of the above in something that matters.
10
Why AI Models Hallucinate
Not a bug awaiting a patch — a predictable result of how models are trained and, crucially, how they are graded.
11
AI Benchmarks Explained
Why most widely-quoted scores are saturated, contaminated, or not comparable between models.
The physical layer
AI capability is downstream of hardware and, increasingly, electricity.
12
The AI Compute Stack
Why memory and packaging rather than arithmetic are the scarce resources, and why power became the binding constraint.
13
How Self-Driving Actually Works
The most visible real-world test of AI, and the vocabulary that separates capability from marketing.
Canon pieces are maintained rather than published and forgotten — each carries the date it was last reviewed. If something reads as out of date, it probably is, and corrections are welcome.