Post #910365

Slop is indeed the future of GenAI programming unless you can make sure the training data has only bug-free, well-architected code that future developers will understand; that the alignment doesn’t over-fit to the training data, and you can continually update that training data set with enough examples of all the new innovations. At that point you might as well be writing the code yourself. You can’t steal enough examples from the bleeding edge to keep up. These systems generate code based off their training set. There are som interesting “behaviours,” that we don’t quite understand. But that’s not going to change the fundamental principles. The network will approximate some f within epsilon given enough resources… but it will reach limits before any more input will improve epsilon. I remain doubtful that we’ll get much more than slop.