Bootstrap data by AI<>Terminal interaction

Basically you give a model instructions to make a simple program, and then let it interact directly with the terminal to make it and debug it. (possibly while torturing it by changing it's code to force it to test and explore more)

If you monitor test case creation (or make them manually) then you can attach the code after the test passes to the original request once it has iterated it, giving it higher quality prompt<>code response pair training data.

I've made a simple QT application "ModelProgrammer" to test this concept out a little easier and it seems: Getting at least GPT3 to iterate on the code seems to take fairly few examples for it to get what it should do, but actually doing it could be a very valuable way to quickly bootstrap training data and make a model MUCH better at coding. (And make it a way more useful tool)

The issue names no repository files, tests, or entry points to inspect. Define the scope and success criteria for terminal interaction and training-data generation before implementation can be started.