[thisBrian]

Relevant xkcd [1]

A fun weekend project would be to utilize GPT-2 [2] to model HN comments; quite the challenge considering the usually insightful comments here when compared to other sites.

[1] https://xkcd.com/810/ [2] https://github.com/openai/gpt-2

[lowdose]

On reddit whole sub communities are based on gpt-2, it is just going to increase the noise to signal ratio.

[tfehring]

In particular, https://old.reddit.com/r/SubSimulatorGPT2/top/ is pretty entertaining.

[lowdose]

I agree the content is actually funny in that reddit tread. I wonder how good it can become with a really large good database. I read in some article of OpenAI a dataset gets better by for example replace names by pronouns.