If you allow LLMs (Large Language Models) to directly call out to APIs, like POSIX, Vulkan, Win32, X11, etc. then it may become possible to program things using a "Socratic" method that involves producing no code at all. You simply channel inputs to the LLM, collect API calls from the LLM, and execute the API calls it makes. This will create a feedback loop that'll eliminate the need to produce code at all. Code can then be produced selectively as an optimisation, like what we call "compilation" today.

Of course, it's currently unreliable and slow, but the trend is a fast one.

Additionally, it might be given a memory store it could read from and write to -- and a StackOverflow-like network to curate advice from other AIs like itself.

A half-way version of this might soon start to exist.