We'd probably need to start by implementing a bytecode editor, whilst programming the GPU to disassemble that bytecode you're editting.
Then we can implement higher-level visual languages, which I'd want to double as the machine's primary UI. And build some useful editors, renderers, etc upon that!
I'd want enough compute to render & compress videos at faster-than-realtime! If we're not doing anything super fancy.
Use USD/AutoMerge as its fileformat.
4/4 Fin! Care to chip ideas in?