How LLMs actually work

Lessons

1. Lesson 1 · ~7 min read · quiz included

   Evals, safety & limits

   How models are measured, where they fail, and how to use them responsibly.

2. Lesson 2 · ~9 min read · quiz included

   How transformers read

   Embeddings, attention, and the forward pass, explained without assuming an ML background.

3. Lesson 3 · ~8 min read · quiz included

   Inference & the KV cache

   What happens per token at serving time, and why session shape matters.

4. Lesson 4 · ~7 min read · quiz included

   Prompt caching in practice

   Stable prefixes, append-only sessions, and reading hit-rate headers on our API.

5. Lesson 5 · ~7 min read · quiz included

   Tokens & tokenization

   How raw text becomes token IDs, and why that shapes context limits and cost.

6. Lesson 6 · ~8 min read · quiz included

   Training at a glance

   What pre-training and fine-tuning do — and what they can't fix.

Coming soon

Planned modules, in rough order. Titles and one-line scope only — nothing here is written yet.

1. 7

   Prefill, decode & serving economics Coming soon

   Quantization, batching, and speculative decoding — where serving cost really goes.

2. 8

   Agents that act Coming soon

   Tool calling, RAG, and multi-agent patterns — plus supervising agents with real permissions.

3. 9

   Reasoning models & distillation Coming soon

   What test-time reasoning buys, and how small models inherit big ones' skills.

4. 10

   Fine-tuning for your domain Coming soon

   When prompting stops being enough: SFT data, evals, and honest expectations.