This commit is contained in:
Paul Gauthier 2023-06-07 09:35:32 -07:00
parent c8216d9687
commit eb48a0f589

View file

@ -60,16 +60,18 @@ which helps it understand and modify large codebases.
## GPT-4 vs GPT-3.5
You need an
[OpenAI API key](https://help.openai.com/en/articles/4936850-where-do-i-find-my-secret-api-key)
to use aider.
Aider supports the `gpt-4` model and has new, experimental support for `gpt-3.5-turbo`.
You will probably get better results with GPT-4, because of its larger context window and
greater competance at coding.
You will probably get better results with GPT-4, for a few reasons:
With GPT-3.5, aider has to send it the code to modify and asks it to return
a full copy of the modified code.
Both the original and edited copy of the code count towards the 4k token context window limit.
In practice, this means you can only edit files about 2k tokens (8k bytes) in size.
- GPT-3.5 isn't able to reliably follow instructions to output "diff style" edits to files. So to make an edit, it has to return a whole new copy of the code that includes any changes. This uses up a lot of the already small GPT-3.5 context window.
- GPT-4 has a much larger context window and is able to reliably generate "diff style" edits to files. This allows you to edit much larger files when using aider with GPT-4.
- GPT-4's larger context window also allows aider to include a [concise map of your repository](https://aider.chat/docs/ctags.html) with each of your change requests. This helps GPT-4 successfully make bigger, more complex changes inside larger pre-existing codebases.
Aider also disables the
[repository map feature](https://aider.chat/docs/ctags.html)
with GPT-3.5, again because of the smaller context window.
## Installation