2026/04/02/ast-edits

[giscus[bot]]

2026/04/02/ast-edits

https://geometricagi.github.io/2026/04/02/ast-edits.html

[emansom]

Great read! Really enjoyed it and loved the research.

It makes a lot of sense, that a model trained on human-readable formats would be stronger at capturing intent than at nailing precise syntax or line numbers.

I’ve been playing around with the new Gemma 4 model (the E2B and E4B f16 versions from Google), and I’m putting together an AST-based editing tool for it. I’m also using a prompt that teaches it programming concepts, Go syntax, and how to map everything to go_edit tool calls.

I went with Go mostly because; I figured Google probably trained it heavily on their own internal docs and codebases, so it might click better there? Though that’s just my guess!

In your setup you’re targeting edits with regular source code syntax (inside content:".. python code ..") instead of pure AST. Do you think the results would improve if the tool only allowed AST transformation operations? Curious what you’ve seen.