Wow that is terrible. In my memory GPT 2 was more interesting than that. I remember thinking it could pass a Turing test but that output is barely better than a Markov chain.
replyI guess I was using the large model?
There’s an art to GPT sampling. You have to use temperature 0.7. People never believe it makes such a massive difference, but it does.
replyProbably a much better prompt, too. I just literally pasted in the top part of my comment and let fly to see what would happen.
replyHere is the XL model. 20x the size of the medium model. Still just 2B parameters, but on the bright side it was trained pre-wordslop.
reply