Ask HN: Can ChatGPT be trained on a specific project's codebase?

2 pointsindigoabstract2 comments

Does anyone know if it's possible to feed a project or parts of it(a few files) to ChatGPT so it can have the proper context when giving answer?

I think I would find it a lot more useful, if this was possible.

Right now it has a limited attention window (short term memory) of up to 30,000 or so tokens so it can fit at most that.

If you could fine-tune the model you could feed in your code base and bake it into the network weights (long term memory)

With fine-tuning you could get good results with a less powerful foundation model than you would otherwise.

I see. So then, I guess, all it knows about my code is what I can feed in a given query. Thank you.