If we really wanted it to be good, we'd see a humor classifier trained that could be put in front of a generative process which spit out hundreds of variations

If we really wanted it to be good, we'd see a humor classifier trained that could be put in front of a generative process which spit out hundreds of variations and see what the classifier picked as funniest. Most human comics refine their material by trying it out on tough crowds. Generation without feedback and refinement isn't going to be all that great, but with the proper refinement pipeline, who knows - maybe it can be outstanding.