A language model predicts the next token from the tokens before it. That loop, plus training, tokenization, and the context window, sets what these tools do well and where they fail.
Start with fluency versus correctness, training versus inference, and what counts against the context window. Your exit task is to explain why an answer needs evidence and identify the information missing from a request. Parameter storage arithmetic, byte-pair encoding, attention pair counts and sampling code are optional background. You do not need them to use the later engineering walkthroughs.
Topic | Lessons | Description | Time |
|---|---|---|---|
5 | Prediction, training, parameters, fluency limits, cutoffs and sampling explained. | 40 min | |
5 | What tokens are, how subword splitting works, and how to estimate tokens before sending. | 40 min | |
5 | How the context window, attention cost, and long-input limits shape everyday AI use. | 40 min |
The model predicts likely next tokens rather than verified facts, so fluency does not imply correctness. Sampling choices and the knowledge cutoff add further limits, as the What a Language Model Does topic explains.