The LLM Knowledge Map How language models work, and how to build with them — twelve notes in two tiers

The domain specialist

You work in a field — energy & batteries, law, medicine, finance — and you want to apply or evaluate LLMs in it. You don't need to derive the math.

1. Evaluation Pipelines — how to tell if it actually works
2. Safety & Alignment — what can go wrong in deployment
3. Retrieval (RAG) — grounding the model in your own data
4. Model Adaptation — when to specialize a model
5. Foundations · Context — just enough to see why retrieval is needed

Skip for now: most of Tier 1 (tokenization internals, attention math, pretraining, scaling) — useful background, not required to use or judge a system.

The builder / engineer

You're shipping LLM features and want to build well — enough machinery to make good architectural calls, plus the practical disciplines.

1. Foundations · Tokenization & Embeddings
2. Foundations · Transformer & Attention
3. Foundations · Sampling & Decoding  +  Context
4. Retrieval, Inference, Evaluation
5. Agents & Adaptation as your product needs them

Optional: Pretraining and Scaling Laws — valuable context, but rarely load-bearing for application work.

The researcher / deep learner

You want to understand the machinery end to end — the why beneath the how, in order.

1. All of Tier 1 in sequence: 01 → 02 → 03 → 04 → 05 → 06
2. Then the Tier 2 disciplines, which build on it — starting with Evaluation and Adaptation

Skip nothing: the foundations are written to be read as a chain, each setting up the next.

The leader / decision-maker

You fund, direct, or vet AI work and want intuition, not internals — enough to set strategy and ask vendors the right questions.

1. Evaluation Pipelines — how quality is actually measured
2. Safety & Alignment — the risk surface
3. Model Adaptation — build vs. buy vs. fine-tune
4. Foundations · Scaling Laws — why capability and cost track size

Skip for now: the rest of Tier 1 — architecture and training details below your decision altitude.