Hacker News Reader

You'd be surprised at the nuances we tend to miss :)

This time around I prompted the models not necessarily to be adversarial - i didn't ask them to try and fool the reader. But i gave them contextual info - something to the effect of "you're a user posting on hacker news"

View on HN

True, if you look for all "obvious patterns" and filter those out of the dataset, not much will be left probably. Maybe the best is then to just publish as complete a dataset as possible, so all questions, all user answers, for each user the nr of questions they did, time for each question, etc. Then people using that dataset can draw their own conclusions.

Reply on news.ycombinator.com