Pros
- It can communicate with Ollama
- I like that it can create files and make code edits
- It runs the files it creates, which is great and asks the user’s permission before doing so
- The SEARCH/REPLACE system of applying diffs created by the llm is interesting
Cons
- It asked me if I would like to see the release notes… I said yes but it didn’t show me anything.
- I don’t get any sort of waiting indicator
-
I prompted
generate hello world in python
and it sent 2.5k tokens and received 99. What did it send? - The default white background for code prompts is blinding
- It really wants to create a git repo for me. That feature is targeted at someone else, surely.
- When installing with pip it downloaded every dependency in the world
- Does not output time taken
- When using the ollama repl for simple prompts it’s much quicker. The whole system prompt is huge and makes it much more difficult for the model to respond quickly.
On Investigation
There is a --show-prompts
option to show you what aider
is sending. There
is also a --cache-prompts
option which isn’t fully explained. The prompt is
large because it’s trying to tell the model the different ways it wants to
interact with it. Especially with code blocks and the SEARCH/REPLACE system
which is cool, but much prompting is needed to get it to do all that.
As Always
I appreciate the industry level churn around AI coding tooling. I’m set on
ollama at the moment, and despite the poor benchmarks for my preferred model,
qwen2.5-coder:32b, I’ve gotten use out of it. Aider has some great features
but I prefer to use git myself, I prefer to apply changes myself and if aider
only wrote to a file I think I’d be happier. That’s the feature I want, just
get the code to a file quicker.
What aider is trying to accomplish is to manage both the code base and the workflow for you. This is prompt-only engineering and I don’t believe prompt-only engineering will be good for codebases even in the medium term let alone the long term. Companies are going to try to code this way and I’m skeptical it will work out.
As Always, the details are important, we’re creating systems that will let the details slip on by.