Temperature
The creativity dial. Low (0–0.3): boring, safe, repetitive, reliable. Same prompt → almost the same output. High (1.0–1.5): varied, surprising, occasionally brilliant, occasionally nonsense. For client-facing drafts that you'll edit, 0.7 is usually right. For audit logs, fact-find summaries, anything where reproducibility matters, push it to 0 — make AI deterministic again.
Temperature is a hyperparameter that controls the randomness of the AI's output. Lowering it makes the model deterministic and reliable, which is ideal for factual summaries or audit logs. Raising it increases diversity and creative flair, making it better for brainstorming. A setting of 0.7 is generally recommended for balancing coherence with natural, human-like variety in client-facing drafts.
Temperature is a scalar parameter (typically 0.0 to 2.0) that modulates the softmax probability distribution prior to token sampling. A temperature of 0.0 forces greedy decoding, ensuring deterministic output by selecting the highest-probability token. A temperature of 1.0 preserves the original distribution, while values > 1.0 flatten the distribution, increasing the entropy and likelihood of lower-probability tokens. This mechanism effectively controls the trade-off between output consistency and creative variance.
evolution
- 2017-06 · historyAttention Is All You Need
The Transformer architecture introduced the softmax function with a scaling factor, laying the mathematical foundation for temperature-based probability distribution adjustment.
- 2020-05 · historyGPT-3 Release
OpenAI popularized the 'temperature' hyperparameter in the GPT-3 API, allowing users to control the randomness of token sampling.
- 2022-11 · historyChatGPT Launch
The mainstream adoption of ChatGPT standardized the concept of temperature as a primary 'creativity' dial for conversational AI interactions.