Two words: loop invariant. Implement a system that figures out the right loop invariant given a problem description (expressed however you want), and you will have made a lot of progress.
>>> Neural networks belong to the differentiable realm, while code generation belongs to the discrete
Even a GAN discriminator step to simulate or even compile and run generated program may never produce the desired output. Getting 99.9% accuracy simply won't do in this domain. Even down to the final curly bracket, the program must be perfect.
Thing is we have an excellent model already. Composed of the architectures, instruction sets, binaries and code. A massive search space for x86-64. But perhaps solvable for 8-bit AVR on modern gpu cloud clusters.
3 comments
[ 2.9 ms ] story [ 15.5 ms ] threadEven a GAN discriminator step to simulate or even compile and run generated program may never produce the desired output. Getting 99.9% accuracy simply won't do in this domain. Even down to the final curly bracket, the program must be perfect.
Thing is we have an excellent model already. Composed of the architectures, instruction sets, binaries and code. A massive search space for x86-64. But perhaps solvable for 8-bit AVR on modern gpu cloud clusters.