Track · 2:30 · Liner note
Warm, Never Boiling
Temperature is a sampling setting that controls how random a language model's word choices are, with low values giving predictable output and high values giving more varied, riskier output.
Temperature is a dial on randomness. Low values make a language model pick its most likely next word nearly every time, and high values let less likely words through. Warm, Never Boiling is the setting for analytical work: enough variety to phrase things naturally, not enough to invent a column.
For SQL generation, extraction and anything that must match a schema, stay low. For brainstorming names or explanations, raise it. Either way, temperature adjusts variety, not correctness: a deterministic wrong answer is still wrong.