The idea of institutionalizing AI generated code before we actually have a working model of what software engineering should look like is what makes me skeptical about this direction.

Whenever I bring up aerospace and the way they do software engineering there as an alternative to how it is usually done elsewhere (and sure, it still isn't perfect) is that you don't have the resources to do everything right. Ok, so now with AI that shouldn't be a limitation. But those projects that have had substantial (or even all) of their code written by AI mirror the common practices rather than the best (at least not the worst). So we're casting that model in concrete now and will generate much more of it than before. I'd much rather see that effort go towards the validation of the software created than the generation itself, I think the long term pay-offs in terms of quality would be much larger.

We can hardly walk and now we're trying to run.

So all this rewriting will achieve is that the motivation to do it better is going to drop out because you'll be able to rewrite everything in a half-baked way in a few minutes, but to deliver quality over the longer term will no longer be nearly as enticing. This is a really hard engineering problem and I for sure don't have the answers but the key for me is that we should be writing better software, not more of it.

It seems "coding AI" shows promising results at assisting the port of c++ code to plain and simple C. This could save some GNU projects.