Remix.run Logo
fancyfredbot 2 hours ago

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.

I guess I was using the large model?

sillysaurusx 27 minutes ago | parent | next [-]

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.

wat10000 an hour ago | parent | prev | next [-]

Probably a much better prompt, too. I just literally pasted in the top part of my comment and let fly to see what would happen.

daveguy 2 hours ago | parent | prev [-]

Here 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.

https://huggingface.co/openai-community/gpt2-xl