LLM delusion is insufferable. If all it takes is tokens to make a significantly better in programming language in logarithmic time why hasn't anyone done it?

As someone who's vibecoding my own self-hosted language (via a typescript to c++ transpiler and bootstrap), I can tell you mainline commercial models like Opus 4.7 aren't quite there yet. I'm getting 10KB source files balloon into 80MB outputs for now.

The main problem is that the the problem space is vast and highly interconnected, the LLM needs to reason about the entire language every time it suggest an architectural change, but it can't, so it suggests local changes that make sense to me - a language hobbyist - then runs into much more difficult problems down the road.

Maybe Mythos with a lot of (competent) human hand-holding and pre-design can do it.