Two ideas for using LLMs in software development
Other than coding assistants and code completions,
- Software developers spend substantial amount of time making technical decisions related to naming and organizing codes. We do this so 1) the programs can run (are programmatically correct) and 2) we can maintain it with the least amount of cognitive load.
Now I wonder: how can we get LLMs to either a) learn, design, then refactor a codebase with a minimal “proper” structure or b) figure out and orchestrate the interdependencies on the fly, based on a loose reference to some other pieces of code.
Is this idea worth exploring? What are the concerns, benefits, and tradeoffs?
- LLMs that scour through codebases for security backdoors or vulnerabilities like the recent xz exploit. I’m sure some people are already working on this (and have probably released something). I’d be interested to check them out!
