Turning up the weirdness dial.
Temperature controls how creative or random an AI's responses are.
Imagine you're baking cookies. Low temperature means you follow the recipe exactly, making the same cookies every time. High temperature means you start adding random ingredients like gummy bears or hot sauce, giving you wild, unpredictable cookies.
It lets you tune whether you want safe, predictable answers or fresh, surprising ones. You see it in chatbots like ChatGPT when you adjust the 'creativity' slider.
Many think temperature is about how 'hot' or emotional the AI is. Actually, it's a mathematical setting that controls randomness in word choice, not feelings.
Temperature is a hyperparameter in language models that scales the logits (raw scores) before applying softmax, affecting the probability distribution over tokens. Higher values (e.g., 1.5) make the distribution more uniform, increasing randomness; lower values (e.g., 0.2) sharpen it, making the most likely token dominate.