Hacker News Reader

All three are top comments here:

So much for a neural network.

Maybe it's just trained on full comments, not the composing words?

So it finds similar images to a person and shows those comments?

Usually, but sometimes not.

Combined into one image for a quick comparison: https://i.imgur.com/ery4SA1.jpg

It's conditioned to output replays way more than make new jokes. But it does make new jokes occasionally.

That sounds just like Reddit tbh.