Why setting default temperature to 1.0?

[jinliu]

GPT seems quite likely to make wrong answers in this setting.

For example,
sgpt --no-cache --shell "docker update all images"
often produces
docker image update --all
where there's no "update" command at all; and
docker images -q | xargs docker pull
which is wrong.

Meanwhile
sgpt --no-cache --shell --temperature=0 "docker update all images"
produces
docker images | awk '{print $1}' | xargs -L1 docker pull
which is almost right (should exclude the first line).

So in such high temperature, GPT may produce wrong answers in first run, when it actually "knows" the right answer.

I think users run sgpt mainly for most correct answers (low temperature), not for creativity (high temperature). Maybe we default with temperature=0, and when the user run the same query repeatedly (meaning he's not satisfied with the answer), we gradually raise the temperature to get more answers?

[lpurdy01]

I think just because 1.0 is fun.

[TheR1D]

The default temperature was changed to 0.1 in one of the last patches.