From my experience, GPT-4 never makes syntax errors directly but when making edits to existing code it's harder to prevent these syntax errors from appearing. We used to add a second pass to check for these syntax errors.

It also frequently makes undefined variables and the like, however.

Did you get rid of the second pass? I'm working on something quite similar and find a pass that inspects and rejects erroneous code to be a big boost to correctness.

We got rid of it. Our new edit framework works around search-and-replace pairs with an example at https://github.com/sweepai/sweep/blob/d37dda3a626f09dea3b322...

Ah, interesting. I completely abandoned diff style updates in favour of AST substitution, but that was only possible because my tradeoffs are different to yours.

I'm building a bot that's building itself, so it doesn't have to support large legacy code bases with different languages.

This is interesting, I'm wondering what you mean by AST substitution. Is this like an agent that traverses the tree and picks what to edit? Is this language model based? Also, thankfully we don't support too many uncommon languages. The most recent ones we added support for are embedded templates (ERB/EJS for flask and Ruby) and mustache. Fortunately many uncommon languages are subsets of other languages.

The agent specifies what function, class, method etc to replace, along with its full source. It's more costly, but I believe it leads to fewer hallucinations as it is generating a coherent piece of code.

But it requires parsing AST and language specific instructions. And things like metaprogramming or macros could cause some hairy confusion.

All of these factors don't hurt my use case.

We have a similar method under the hood, except it's purely text-based search-and-replace. The model decides what to replace. It seems to be consistent and is easy to implement.

My gut feeling based on my experience over the last couple of months is that substitution of an entire function is more reliable than some lines of a function. The surrounding context reduces the chance of hallucinations.

Gut feeling doesn't account for much though - I'm working on an evals system to be able to quantify system performance. It won't be cheap to run.

It could easily be that your method is superior.

From our experience, single or few line replacements generally fine, since many of the changes are many few-line changes in multiple spots accross multiple files. We also provide surrounding context for the search-and-replace pairs, which helps with the model. Beyond 10 lines and the model also usually add the function headers in there which helps with the code generation.

I'm also curious, how are you guys evaluating the performance of your models?

There's no systematic evaluation yet which is the next step. It's successfully bootstrapping itself which is a fairly high bar, but quantitative performance measurements are getting more and more important as the project progresses.

I feel the same, benchmarking in general is a pain but a good benchmark for us could go a long way.

I’ve built out an end to end automated fix pipeline, it’s getting the bug fixes right, but been having trouble with line number errors.

Looking forward to reading through your docs and repo later tonight to see how you’re addressing issues like this.

We used to use line numbers but it became problematic so we switched over to search-and-replace pairs, which works significantly better. The only potential problems are with setting up a fuzzy search system since sometimes the search doesn't match exactly with the code (missing comments, etc.). We're going to write about our core algo and diff managing system soon.

That edit format looks familiar!

https://github.com/paul-gauthier/aider/blob/d5a7aac560d4584d...

https://aider.chat/docs/benchmarks.html

Yup it's based on the aider blogs. They're perfect for our use case and are very reliable compared to our old attempts.