LLMs for Codebase Assistance

Making onboarding easier!

I was reading this sci-fi story, “Short Story on AI: Forward Pass” by Andrej Karpathy; about a neural network becoming self-aware of its existence¹, having its original thoughts and augmented consciousness². This got me thinking of self-aware LLMs for codebase assistance.

Working at different organizations throughout my career, with different codebases, data structures, utility libraries, internal tools, etc there’s always a steep learning curve initially to get comfortable with all the code we had and how things worked internally.

During the onboarding process, senior engineers would have to spend some time mentoring new joiners and teach them the same things, answer repetitive questions, and so on.

Time is expensive, and asking questions is a leaky process. There are also types of questions that have simple answers like “yes/no” or “look into </directory/xyz>”, but are blockers for new engineers to work.

What if we had a neural network take over the onboarding process for the codebase? What if an LLM can answer all my curious questions about the code with examples, always on-call and available?

Extending this idea one step further, if neural networks have a self-awareness mechanism, it would be nice to ask them about their own opinion on how to improve the code, implement a new feature, discover vulnerabilities inside our codebase, etc.

Finally and this is especially for remote workers, building a speech layer on top of LLMs would be a great way to give an augmented human experience of onboarding, that has the back-and-forth attribute of human conversations.

Notes

1. Its existence, in this case, would be the class() definition of the model and the training loop.

2. I wouldn’t call it consciousness just yet, as I don’t have a clear definition of what it means for a human to be conscious in the first place.

---