I agree that building systems that can keep track of large-scale programs seems very difficult, and the current LLMs don’t seem up to the task. However, I wonder whether we can leverage the ability of LLMs to understand short code snippets to inform higher-level static analysis. I think we may be able to generate high-level models of codebases using the intuition of LLMs, that we could then use to run higher-level simulations to detect possible bugs.

It seems like it could be possible, but soon? I agree, probably not.