Reading This Site with the Map
Lesson 3 of 3 in The LLM Lifecycle Map.
This site has ten domains, numbered L-01 through L-10, and every one of them is a place on the map you just learned.
Two domains come before the map, describing the machine itself: L-01 Foundations (you are here) establishes what a language model is and what it can honestly do, and L-02 Inside the Transformer opens the machine — tokens, embeddings, Attention, the Transformer block.
Four domains are the four stages, one each: L-03 Pre-training is stage one, the making of a base model. L-04 Post-training & Alignment is stage two, behavior shaping. L-05 Inference & Serving is stage three, where latency and cost are made. L-06 Adapting LLMs is stage four — prompting, RAG, fine-tuning, and how to choose between them.
Two domains are the cross-cutting lenses: L-07 Evaluation teaches you to measure models and defend the measurement; L-08 Security & Risk maps how models fail and get attacked at every stage.
And two domains are where everything lands in practice: L-09 LLMs on the Cloud grounds all four stages on real infrastructure — the managed platforms, the GPUs, the cost models — and L-10 Model Landscape teaches you to read the artifacts the lifecycle produces: model families, Open weights versus Closed weights releases, benchmarks, and model cards.
A vertical stack of ten layers, bottom to top: L-01 Foundations, L-02 Inside the Transformer, L-03 Pre-training, L-04 Post-training and Alignment, L-05 Inference and Serving, L-06 Adapting LLMs, L-07 Evaluation, L-08 Security and Risk, L-09 LLMs on the Cloud, and L-10 Model Landscape at the top. The bottom layer, Foundations, is highlighted to show the reader’s current position.
“Why is the model slow, and why is the bill so high?”
That is stage three. L-05 Inference & Serving explains the mechanics — tokens are generated one at a time, memory fills with the KV cache, batching trades latency for throughput — and L-09 LLMs on the Cloud turns the mechanics into infrastructure and cost decisions.
“Should we fine-tune, or is RAG enough?”
Stage four’s central question. L-06 Adapting LLMs compares prompting, RAG, and fine-tuning honestly — what each can and cannot change, and how to pick with evidence instead of fashion.
“The vendor quotes an impressive benchmark score. Can we trust it?”
Cross-cutting lens one. L-07 Evaluation covers what benchmark numbers mean, how contamination inflates them, and how to build the only eval that really matters: one on your task.
“The model confidently made something up. Whose fault is it?”
Trick question — it is the map’s fault. Hallucination is born in pre-training (L-03), damped but not cured in post-training (L-04), managed at the application layer (L-06), and risk-assessed as a failure mode in L-08 Security & Risk. One symptom, four stages of ownership.
“Which model should we pick in the first place?”
L-10 Model Landscape teaches you to read the field — families, weights access, model cards — and L-07 Evaluation gives you the harness to test the shortlist against your own data.
If you prefer to be led rather than to browse, the guided path walks this stack in order, in seven legs: first principles (L-01), the machine, opened up (L-02), how models are made (L-03 and L-04 — stages one and two), running the model (L-05 — stage three), making it yours (L-06 — stage four), trust but verify (L-07 and L-08 — the two lenses), and production synthesis (L-09 and L-10 — where it all lands). It is the lifecycle map, converted into a walking route: the machine first, then its life story stage by stage, then the judgment to run it responsibly.
Wherever you go next, carry the map. Every concept on this site — every acronym, every vendor claim, every incident report — sits at a stage. Placing it there is the first step of understanding it.
Interactive checkpoint quiz (1 questions) — open this page in a browser to take it.