Interesting points, I don’t think a million line code base is a good shield though, Gemini already has a massive context window and there have been some recent papers about infinite context.

I will say in my personal experience llms suck at rust, archaic vb, forth, and other older languages, they excel at Python and Javascript though.

Using the prior generation of models results in decent code, the next generation is already baked, we need to prepare at least a little.