Idea: Train an ai to construct entire threads based on the input of the title and link. There should be plenty of training material, and frankly I think it would be hilariously similar to the real threads.
When new submissions are added to HN, the system will fetch the title and link, and predict the entire thread. You can view the thread on something like hnpredicted.com/?id=33680661 . When the real thread has had no new comment added for five days, it is reintegrated into the system further training the model.
Possibly, the model could be trained on the actual content of the link. But I suspect just using the short title might be better. Larger concentration on the actual material there, websites contain lots of text which is useless.
Now that I've posted the idea, feel free to steal it ofcourse. It's documented that I came up with it anyway. :)