It doesn't work like that. AI is not a Jinn. You cannot simply command it and have it produce an entire project from thin air. You get to have fun: do the thinking part, and let it do the boring stuff.

I have a long list of projects that I have thought about but never implemented because of lack of time and energy. LLMs have made that happen.

I like designing programming languages and developing parsers/compilers and virtual machines. But the steps beyond type-checking are so incredibly boring (and I don't like using C or LLVM as targets) that I have done the front end 15-20 times over the last couple of decades and the back end only 3-4 times.

This time, I spent two weeks developing a spec for the VM, including concurrency, exception handling and GC. And I led the AI through each subsystem till I was satisfied with the result. I now have a VM that is within 8x of C in tight loops. Without JIT. It is incredible to be able to allocate arrays of 4B elements and touch each element at random, something that would make python cry.

Working on the compiler now.