// track · 8 lessons
How these things actually work
How models actually generate, drift, hallucinate, and forget. The floor under every other track.
- 01Autoregression: the output is the thinkingThere's no hidden reasoning engine. The generating is the thinking, one token at a time, so give the model room to reason on the page.
- 02Temperature and driftTemperature is a randomness dial, and its effect stacks across a generation. Cold for exact and long, warm for creative and short.
- 03In-context learningExamples don't teach a model new things, they aim it at what it already knows. Show the output you want and build context deliberately.
- 04Self-attention: why one word cascadesAttention makes every token depend on every other, so word choice is leverage and a clean re-prompt beats a pile of corrections.
- 05The system-prompt hierarchyInstructions have a priority order, and persistence lives in the system layer, not the message layer. Put standing rules where they stand.
- 06Hallucination is the default, not a bugThe model optimizes for coherence and has no uncertainty signal, so hallucination is the baseline and verification is not optional.
- 07Context-window economicsAttention is a finite budget that thins as context grows, so tight beats big and splitting beats cramming.
- 08The what's-next trapThe what's-next pattern has no natural stopping point and coherence is not correctness, so anchor every long exploration to something outside itself.