LLM Foundations
What the model actually does, at exactly the depth an application engineer is expected to hold.
8 topics
How LLMs Generate Text
A model scores every possible next token given everything before it, one token picked at a time — and almost every strange behaviour you will debug follows from that.
core25 minTokens & Context Windows
The unit you are billed in, limited by, and truncated at — and the reason English prose, JSON, and Japanese cost wildly different amounts for the same information.
core25 minSampling & Determinism
How the next token is actually chosen from the distribution, and why "set temperature to 0" was never the reproducibility guarantee people treated it as.
core20 minEmbeddings & Similarity
Text as a vector whose direction encodes meaning, which is what makes "find me things like this" a database query instead of a keyword match.
core25 minHallucination & Grounding
Why a fluent, confident, completely invented answer is the system working as designed, and what actually reduces it.
core20 minPrompting vs RAG vs Fine-Tuning
The three ways to make a general model do your specific job, and the order you should try them in.
core20 minChoosing a Model
Picking a model is a per-step cost, latency, and capability decision — not a one-time vendor choice for the whole product.
core20 minMultimodal Inputs
Images, PDFs, and screenshots as first-class prompt content — what they cost in tokens and where they quietly fail.
deep20 min