> If you always just pick the highest probability option (temperature = 0), it's not random in any way
Huh. That feels like a recipe for either outputting parts of the training data verbatim, or getting stuck in an output loop.
Although I earlier gave an example with 90/9/1 weights where the obvious choice is significantly higher than the others, I'd have thought that in a different scenario where you have, say, 51/49 or 34/33/32 options to choose from, always picking the one with only a fractionally higher confidence than the reasonable alternatives could lead to a "sameyness" in the outputs. Isn't there any value in having variance in the generated responses?
> Isn't there any value in having variance in the generated responses?
Sure, which is why the technique is used. But that doesn't mean "random" is a useful descriptor. I, as a human, try to add variance to my responses as well - does that mean my mind is "random"?
That's why my point is that there is no inherent randomness in the process. We use a bit of randomness to make it subjectively better, but it's fundamentally not random.
Huh. That feels like a recipe for either outputting parts of the training data verbatim, or getting stuck in an output loop.
Although I earlier gave an example with 90/9/1 weights where the obvious choice is significantly higher than the others, I'd have thought that in a different scenario where you have, say, 51/49 or 34/33/32 options to choose from, always picking the one with only a fractionally higher confidence than the reasonable alternatives could lead to a "sameyness" in the outputs. Isn't there any value in having variance in the generated responses?