Skip to content

Concept ladder

This is a map of when technical vocabulary enters the canonical course. Every chapter still defines a term inline; this page helps readers see how later ideas depend on earlier ones.

Stage 0 — ordinary ideas

First taught in Before the jargon:

Term Plain meaning
Example or data Something the program can learn from
Pattern A relationship that appears across examples
Model A program whose behavior was shaped by examples
Input / output What goes into the program / what comes out
Prediction A possible output selected before the answer is known
Parameter An adjustable internal number
Training Adjusting parameters using examples
Inference Using the trained model without changing those parameters
Token A numbered piece of text

Stage 1 — language-model basics

First taught across What an LLM is, Learning from examples, and Text becomes tokens:

Term What it adds
Context The earlier text available for the next prediction
Vocabulary The complete set of token IDs a tokenizer can produce
Probability A number describing how likely an option is under the model
Loss One number measuring how poor a training prediction was
Optimizer The rule that turns learning signals into parameter changes

Stage 2 — the mathematical toolkit

First taught in Math with shapes and PyTorch mental model:

Term What it adds
Vector An ordered list of numbers
Matrix A rectangular grid of numbers
Tensor A general multi-dimensional collection of numbers
Shape The size of each tensor dimension
Gradient A local signal showing how a small change affects loss

Stage 3 — the ordinary Transformer

First taught in the Transformer section:

  • embedding;
  • position information;
  • query, key, and value;
  • causal self-attention;
  • attention head;
  • feed-forward layer;
  • normalization;
  • residual connection;
  • language-model head and logits.

Stage 4 — training at scale

First taught in Pretraining:

  • batch and training step;
  • learning-rate schedule;
  • checkpoint;
  • data, tensor, pipeline, context, and sequence parallelism;
  • validation and benchmark evaluation.

Stage 5 — optional architecture extensions

Only after the dense Transformer is established, the Mixture of experts section introduces:

  • dense versus sparse computation;
  • expert feed-forward networks;
  • router and top-k selection;
  • expert capacity and load balancing;
  • expert parallelism.

Stage 6 — behavior after pretraining

The Post-training section introduces supervised instruction tuning, preference data, reward models, preference optimization, reinforcement learning, verifiers, and reasoning-oriented training.

Stage 7 — running and operating the model

The Inference section introduces logits, decoding, temperature, top-k/top-p sampling, key-value cache, batching, quantization, and speculative decoding in their production context.

Stage 8 — systems around the model

The Prompting and Agents sections introduce structured output, tool calls, retrieval-augmented generation, memory, planning, authorization, and agent evaluation.