Inline AI code editing and generation in elixir cells

[jonastemplestein]

I've been using the Cursor editor for coding recently and have found the inline code editing via ChatGPT to be incredibly useful.

Here's an image to show how it would work:

I thought it might be a fun learning project for me, as I'm new to Elixir and Phoenix, but have some time on my hands. With some pointers I'm sure I could implement it.

Beyond this, I think it would also make sense to integrate other AI features, such as github copilot style autocomplete and an AI chat in a sidebar (perhaps one that could come in on the right side like the output panel). And possibly even a full "coding loop" like open interpreter or chatgpt's advanced data analysis (formerly called code interpreter). I'd really like to use livebook for any kind of data-wrangling tasks, but at the moment it's often faster for me to use those other tools and do the analysis in python.

I guess the big strategic question is whether you want features that are so tightly coupled to a particular LLM API. Maybe you could even have a central LLM abstraction within livebook where you could also run e.g. a fine tuned code-llama model or similar to power these features 🤔

[josevalim]

Hi @jonastemplestein! There is a more general discussion about this on #2073. In a nutshell, I think we will have to depend on some large LLM for some features, but I would love if part of the problem (such as code suggestions), was done with a model tailored to Elixir. :)

[jonastemplestein]

Ah sorry, I'll continue the conversation over there - serves me right for just search for "AI" to find old threads

I'll probably still continue with my learning exercise of implementing Cmd + k on top of ChatGPT 4, as I've already sort-of got it working and I'm learning a lot from the experience. Would it be okay opening a PR for feedback at some point, even if it probably won't get merged?

[josevalim]

No worries, it is completely fine to double the discussion here. :) And it is nice that you are exploring this.

Regarding the PR, I think it would be best if we are sure we want to go down the path before submitting it. We are a small team, so so it is unlikely we will have the time to review the code unless it is a prioritized feature. However, feel free to share your fork/branch here and on ElixirForum and ask people to try it out and give feedback. I may play with it too :)

Thanks!

[jonastemplestein]

@josevalim I've had a go at implementing this here: jonastemplestein#1

Any time you or the team can spend giving feedback would be much appreciated :)

I mostly have no idea what I'm doing so the implementation isn't very good, but the gpt4 version works pretty well for a lot of stuff already