Gary Marcus

I would say it very much depends on how you handle the sampling. The graph is using top_k sampling so that is inherently unlikely to generate exact replica text. But you dont have to do it that way.

I wouldnt say Im an expert, though I enjoyed reading the gpt papers, training on corpora, working with the code a little bit. But Im conflicted in commenting here b/c Im afraid I'll be encouraging mud slinging like with the rubiks cube thing.