This commit is contained in:
Paul Gauthier 2023-07-01 21:28:07 -07:00
parent 6477a95279
commit e2af782b59

View file

@ -348,3 +348,16 @@ cause a large variance in the overall benchmark results.
Based on these benchmark results, aider will continue to use
the `whole` edit format for GPT-3.5, and `diff` for GPT-4.
GPT-4 gets comparable results with the `diff` and `whole` edit formats,
but using `whole` significantly increases costs and latency compared to `diff`.
The latency of streaming back the entire updated copy of each edited file
is the real challenge. The GPT-3.5 models are quite responsive, and can
stream back entire files at reasonable speed.
Aider displays a progress bar and
live diffs of the files as they stream in,
which helps pass the time.
The GPT-4 models are much slower, and waiting for even small files
to be completely "retyped" on each request is probably unacceptable.