The concepts every LLM practitioner needs before writing a single prompt.
Build the mental model — how models learn, generate text, and where their capabilities end — so every downstream decision (prompting, RAG, fine-tuning) rests on real understanding, not folklore.
Engineers, PMs, and analysts who use LLMs daily but have never opened the box. No ML background required — working knowledge of software is enough.
The guides are numbered — read in order for the curriculum path, or jump straight to the one you need. Each card is self-contained.
The nested hierarchy of the field and how each term relates.
The shift from rules to learning-from-examples; training vs inference.
The language problem, why it is hard, classic vs modern approaches.
Why models need numeric input; one-hot, bag-of-words and their limits.
Tokens vs words, subword tokenization, why tokens are the unit of cost/context.
Meaning as geometry; similar words as nearby vectors.
The basic building block, forward pass, how layers stack.
Non-linearity, common activations, and how softmax turns scores into probabilities.
The training loop that adjusts weights to reduce error.
Handling ordered text and the bottlenecks that motivated attention.
Letting a model weigh which other tokens matter.
How tokens attend to each other; why multiple heads help.
Assembling embeddings, attention and feed-forward layers into the full model.
Next-token prediction over massive data; emergence of capability.
Turning a raw predictor into a helpful assistant.
Decoding, probability distributions, the creativity/reliability dial.
What the context window is, why it is limited, how it shapes design.
The plausibility-vs-truth root cause and what it implies.
Honest map of what LLMs do well, do poorly, and what is next.