I think they can be an excellent copy of human code. Are they great at novel out-of-training-distribution tasks? Definitely not, they suck at them. Yet I'd argu

I think they can be an excellent copy of human code. Are they great at novel out-of-training-distribution tasks? Definitely not, they suck at them. Yet I'd argue that most problems aren't novel, at most they are some recombination of prior problems.