In the submitted article from February 2009, Paul Graham wrote,<p>"Probably the most important thing I've learned about dilution is that it's measured more in behavior than users. It's bad behavior you want to keep out more than bad people. User behavior turns out to be surprisingly malleable. If people are expected to behave well, they tend to; and vice versa.<p>"Though of course forbidding bad behavior does tend to keep away bad people, because they feel uncomfortably constrained in a place where they have to behave well. But this way of keeping them out is gentler and probably also more effective than overt barriers."<p>Trying to set barriers to bad user behavior has been an ongoing problem.<p>In March 2011, pg wrote that Hacker News was still having user behavior problems, under the title "Ask HN: How to stave off decline of HN?"<p><a href="http://news.ycombinator.com/item?id=2403696" rel="nofollow">http://news.ycombinator.com/item?id=2403696</a><p>He wrote, "The problem has several components: comments that are (a) mean and/or (b) dumb that (c) get massively upvoted."<p>So the founder of Hacker News thought then that there was a comment voting problem: (a) mean comments were getting too many upvotes, and (b) dumb comments were getting too many upvotes, and (c) too many of the comments that got the most upvotes were either mean or dumb or both. Let's stop and think about what that means. That means that, according to pg posting as of last year, two more years after the article submitted here, comment karma scores were often NOT reliable signals of good comments, comments worth finding rapidly when skimming a thread. So pg changed the Hacker News software a few days later so that comment karma scores were hidden from Hacker News readers (they are STILL visible to each person who makes a comment, on his or her own comments). You can check the scores of each of your own comments. We can all READ comments to see which comments are the good comments. (Better comments also tend in general to thread to the top of discussions, with some other rules also influencing that evaluation as "better.")<p>Checking the results of the software change involves empirical investigation. How do the highest-voted comments visible in the bestcomments list<p><a href="http://news.ycombinator.com/bestcomments" rel="nofollow">http://news.ycombinator.com/bestcomments</a><p>look to all of you recently? Are there fewer mean comments than two years ago? Are there fewer dumb comments than two years ago? Are the comments that are "massively upvoted" since the experiment began mostly comments that are reasonably kind and well-informed, helpful comments on the whole? In most of the treads you visit, do helpful, thoughtful comments seem to rise to a position of prominence, while mean or dumb comments gray out?<p>In my observation, after 1345 days here, and rather active participation on Hacker News, the comments have improved both in the threads I post in and in the threads I only lurk in. Occasional checks of the best comments page STILL reveal a lot of high-ranked comments that consist of one-liner jokes or whining about something, so the signal provided by upvotes still isn't perfect, but looking at the average comment karma score of some users I've been following for years suggests that comment karma is better allocated now than it was two years ago, which as I said is also my impression when I look at particular threads under submitted stories. There may be some more changes yet to make, but Hacker News has scaled up some more and still managed to emphasize quality of comments, which is not an easy thing to do.