The temperature scale, I think. You divide the logit output by the temperature before feeding it to the softmax function. Larger (resp. smaller) temperature results in a higher (resp. lower) entropy distribution.
Each row in the figure is a probability distribution over possible outputs (x-axis labels). The more yellow, the more likely (see the colour map on the right). With a small temperature (e.g., last row), all the probability mass is on 42. This is a low entropy distribution because if you sample from it you’ll constantly get 42, so no randomness whatsoever (think entropy as a measure of randomness/chaos). As temperature increases (rows closer to the first/topmost one), 42 is still the most likely output, but the probability mass gets dispersed to other possible outputs too (other outputs get a bit more yellow), resulting in higher entropy distributions. Sampling from such distribution gives you more random outputs (42 would still be frequent, but you’d get 37 or others too occasionally). Hopefully this is clearer.
Someone in another reply uses the word “creativity” to describe the effect of temperature scaling. The more commonly used term in the literature is “diversity”.
I spent an afternoon once playing Infinite Craft, which uses some sort of LLM behind the scenes to do it's combinations.
At one point I got 007, and found 007+007 = 0014.
The maths gets wild though, and because it's been trained on text, it has no idea when it comes to combinations of numbers it hasn't seen before. I spent ages trying to get it to 69420 and just couldn't, although I could get 42069.
