How models work · L2 · 15 min
Context windows
Learn in five minutes
- The context window is how much text the call can see at once.
- Instructions, notes, and the growing reply all spend that budget.
- Overflow is a cut, not a memory.
- Published window sizes change. Confirm the model card.
- A long paste is not a retrieval system.
Put the instruction after a long note and the model may never reach it if the note was truncated. Say what you dropped.
Visual
A toy budget of 100 units
- Give the instruction 20 toy units and the note 90.
- The sum is over 100, so something is cut.
- Write down which part you kept. Do not call the cut a memory.
Quiz
Ecosystem
- Context window
- The maximum text one model call considers.
Sisters: Tokenizers, Transformers, Decoding, Model cards.
The same next-token idea sits under chat, voice, and tool calls.
Pin a model id when you leave the toy explanations.
Config
- Write the goal in one sentence.
- Name the input and the allowed output.
- Save the trace next to the result.
Troubleshooting
- The answer looks fluent and is wrong. Check whether a source passage actually contains the claim.
- The loop never stops. Set a step cap and stop when the tool result repeats.
Sources and go deeper
- Brown and co-authors, Language Models are Few-Shot Learners, 2020 paper last checked 2026-10
- Large language model — Wikipedia wikipedia last checked 2026-10