This is a very Lisper objection. The thing is a token predictor, and can't count levels of depth.
Think "autistic junior engineer, whose work needs lots of testing but is also prolific at writing tests" instead of "Godlike text generator", it's much more productive.
Think "autistic junior engineer, whose work needs lots of testing but is also prolific at writing tests" instead of "Godlike text generator", it's much more productive.