I'm new to all this. Where do I start?
Five chapters that get you fluent: how models actually work, the core vocabulary, prompting, and where the risks live.
The path.
- How LLMs Actually Work A large language model predicts the next token given the tokens so far, and that is essentially all it does.
- Core Concepts Quick definitions for the vocabulary you'll run into everywhere. Each one is short, and several get a fuller treatment later in the handbook.
- Prompting & Context Engineering Prompting is the highest-leverage, lowest-cost skill in this whole handbook. Small changes in how you frame a request move quality more than most people expect.
- Chat Assistants & Apps The consumer/prosumer chat apps are the front door to each lab's frontier model.
- Safety, Alignment & Governance Building with AI means owning its failure modes. This section is deliberately neutral and tight, enough to reason about risk without ideology.
Check yourself.
One question per step. Take it cold to find where to start, or after reading to see what stuck. Nobody's grading you.
Best so far:
Question 1 of 5
-
Why can a chat model reference something you said earlier in the same conversation?
Weights are frozen at inference and the model is stateless between calls. It only "knows" the conversation because the accumulated transcript sits in the context window, re-read on every turn.
Covered in step 1 · How LLMs Actually Work -
In a Mixture-of-Experts model, what mainly drives per-token inference cost and speed?
MoE decouples total capacity from per-token compute: only a subset of experts activates per token, so active parameters are what you actually pay for at inference.
Covered in step 2 · Core Concepts -
The model almost does what you want, but the output format keeps drifting. Fastest fix?
Few-shot examples anchor format and style far more reliably than describing them in words. If you're writing paragraphs about the output shape, show examples instead.
Covered in step 3 · Prompting & Context Engineering -
The chat apps are converging in quality. How should you pick one in 2026?
The assistants now differ less in raw intelligence than in surface features and ecosystem lock-in (docs, cloud, IDE), so those should drive the choice, not benchmark gaps.
Covered in step 4 · Chat Assistants & Apps -
Why is prompt injection the top security risk for agents that read untrusted content?
To the model, everything in the window is just tokens, so instructions hidden in a web page or email can be followed as if they came from you. That's why OWASP ranks it the number-one LLM risk.
Covered in step 5 · Safety, Alignment & Governance