I know for mass adoption LLMs need to support natural language input. But we've done a reasonably good job (note: source for endless arguments here ;) over the past ~80 years of developing a very precise system for inputting exactly what we want a system to do in the form of programming languages.

I'm curious whether any of the leading models - LLMs, image generation models, etc. have taken this into consideration. Particularly in more precise I/O domains (image generation comes to mind), it seems like a structured input format where we remove the entire problem space of natural language prompt -> user intent would make things dramatically easier to get the output we want.