Code is the spec, and it's far faster to write the code yourself (if you know what you want to do and how to write code) than try to pass all this context to an LLM in the form of lossy and imprecise human language.
Your dataset is all of the 6502 binaries, corresponding manuals, and descriptions of gameplay generated by multimodal AI looking at videos, screenshots, etc. Maybe even screenshots and memory captures of games could be input along with sequences of key or controller inputs.
But if you feed in a large chunk of the Z80 and 6502 programs and as much metadata as you can and for a training signal make sure that the thing runs without freezing on an emulator and maybe shows the correct loading screen or something.. I think there are enough 8-bit programs and games for this to work, at least to some degree. Maybe. Start with shorter simpler programs and work up. Generating BASIC might be easier in a way, I don't know.
But the idea is you would describe your game along with maybe a screenshot or cover art or something and it would generate the machine code.
You could also do something similar but on a frame-by-frame input-by-input basis like Oasis.
Anyone in this thread want to fund that? I would love to work on it.
The mostly unsolved problems right now are around build, deployment, and operations.