Hacker News Reader

I think filtering on upvotes/downvotes, comments/views, sub, user and whatever metrics they have on the content can help AI companies train on somewhat reasonable things. Blend it with Wikipedia, scientific papers, reliable newspapers and you're golden?

Metadata is what makes gold out of poo, I assume model developers can "train negatively" too if metadata suggests they should.

View on HN

So now the people who can buy upvotes get to write history?

True, once you start measuring by something it becomes an useless metric so it'd work until people knew if they do it or not, then they will not be able to rely on that data either. Or they improve their bot detection and mitigation and play the cat and mouse game.

Reply on news.ycombinator.com