🥇Paul Graham: What I Learned from Hacker News

February 2009

Hacker News celebrated its second anniversary last week. It was initially intended as a side project — an app to refine Arc and a place for news sharing among current and future founders of Y Combinator. It grew larger and required more time than I expected, but I have no regrets because I learned a lot working on this project.

Growth

When we launched the project in February 2007, traffic during weekdays was around 1,600 daily unique visitors. Since then, it has increased to 22,000.

This growth rate is a bit higher than I'd like. I would prefer the site to grow because if a site isn't growing, even slowly, it's probably dead. However, I wouldn't want it to reach the growth of Digg or Reddit — mostly because it would dilute the character of the site, but also because I don't want to spend all my time working on scaling.

I already have enough problems with that. I remember that the original motivation for HN was to test a new programming language and, moreover, to test a language focused on experimenting with design rather than performance. Every time the site became slow, I reminded myself of the famous quote from McIlroy and Bentley.

The key to efficiency is the elegance of solutions, not sifting through all possible options.

And I looked for bottlenecks that I could fix with minimal code. I am still able to maintain the site, in terms of preserving its previous performance, despite a 14-fold increase. I don't know how I'll manage moving forward, but I will probably figure something out.

This is my attitude toward the site as a whole. Hacker News is an experiment, an experiment in a new area. Sites like this usually last only a few years. The discussion on the internet itself is only a few decades old. So, we are likely discovering just a small part of what we will eventually find.

This is why I am so optimistic about HN. When the technology is this new, existing solutions are usually terrible, which means there is an opportunity to create something much better. This, in turn, means that many problems that seem insurmountable are actually not. Including, I hope, the issue that plagues many communities: decay due to growth.

Decay

Users have been concerned about this since the site was just a few months old. So far, these concerns have been unfounded, but that won't always be the case. Decay is a complex issue. However, it is likely solvable; it doesn't mean that open conversations have 'always' been destroyed by growth, when 'always' simply means 20 instances.

But it's important to remember that we are trying to solve a new problem, because that means we need to try something new, and most of it likely won't work. A couple of weeks ago, I attempted to display user names with the highest average comment scores in orange. [1] That was a mistake. Suddenly, a culture that had been more or less unified split into the haves and the have-nots. I didn't realize how unified the culture was until I saw it divided. It was painful to watch. [2]

So orange user names won't return. (Sorry about that). But there will be other ideas that are just as likely to fail in the future, and those that work will probably seem just as broken as the ones that don't.

Perhaps the most important thing I learned about decay is that it is measured more in behavior than in the users themselves. You want to eliminate bad behavior rather than bad people. User behavior turns out to be surprisingly malleable. If you expect people to behave well, they usually do; and vice versa.

Although, of course, banning bad behavior often drives out bad people because they feel uncomfortable being in a place where they are expected to behave well. This method of getting rid of them is softer and likely more effective than others.

It is now quite clear that the broken windows theory applies to public websites as well. The theory suggests that small displays of bad behavior encourage even worse behavior: a neighborhood with a lot of graffiti and broken windows tends to be where robberies occur. I lived in New York when Giuliani implemented reforms that made this theory famous, and the changes were astonishing. I was also a Reddit user when the exact opposite happened, and the changes were just as impressive.

I am not criticizing Steve and Alexis. What happened to Reddit was not a result of negligence. From the very beginning, they had a policy of censoring only spam. Additionally, Reddit had different goals compared to Hacker News. Reddit was a startup, not a side project; their aim was to grow as quickly as possible. Combine rapid growth with zero sponsorship support, and you get lawlessness. However, I don't think they would have done anything differently if given the chance. Judging by the traffic, Reddit is much more successful than Hacker News.

But what happened to Reddit does not necessarily have to happen to HN. There are several local upper limits. There can be places with complete lawlessness and places that are more thoughtful, just like in the real world; and people will behave differently depending on where they are, just as in real life.

I have seen this in practice. I observed people cross-posting on Reddit and Hacker News, who took the time to write two versions: an inflammatory message for Reddit and a more restrained version for HN.

Materials

There are two main types of problems that a site like Hacker News must avoid: bad stories and bad comments. It seems that the damage from bad stories is less severe. So far, the stories featured on the front page are still quite similar to those that were posted when HN was just starting out.

I once thought I would have to consider measures that prevent any nonsense from appearing on the homepage, but so far I haven't had to do that. I didn't anticipate that the homepage would remain so wonderful, and I still don't quite understand why this happens. Perhaps only more thoughtful users are attentive enough to suggest and like links, which is why the marginal cost for a random user tends toward zero. Or maybe the homepage protects itself by posting announcements about what offers it expects.

The most dangerous material for the homepage is content that is too easy to like. If someone is proving a new theorem, the reader must put in some effort to decide whether to like it. A funny cartoon takes less time for that. Loud words with even louder headlines get zero likes because people like them without even reading.

This is what I call the False Principle: users choose a new site whose links are easiest to judge if you do not take specific measures to prevent this.

Hacker News has two types of protections against nonsense. The most common types of content that have no value are banned as off-topic. Kitten photos, political rants, and so on are under strict prohibition. This filters out a large portion of unwanted nonsense, but not all of it. Some links are nonsense in the sense that they are very short but still relevant material.

There is no one-size-fits-all solution for this. If a link is simply empty demagoguery, editors sometimes remove it, even if it is relevant to the topic of hacking, because it does not meet the actual standard, which implies that an article should stimulate intellectual curiosity. If posts on the site are of this type, I sometimes ban them, meaning that all new material for that URL will be automatically deleted. If a post's title contains clickbait, editors sometimes rephrase it to make it more factual. This is especially necessary for links with sensational headlines, as otherwise, they become hidden 'vote if you believe this and that' posts, which is the most pronounced form of useless nonsense.

The techniques for dealing with such links must evolve as the links themselves evolve. The existence of aggregators has already impacted what they combine. Now writers consciously craft content that will drive traffic through aggregators—sometimes rather specific content. (No, the irony of this statement is not lost on me). There are more sinister mutations like linkjacking—publishing a retelling of someone else's article and presenting it instead of the original. Such content can receive many likes, as it retains much of the good information present in the original article; in fact, the more a retelling resembles plagiarism, the more good information is preserved in the article. [3]

I think it is important for a site that rejects submissions to provide users with a way to see what has been rejected if they wish. This holds editors accountable, and, perhaps more importantly, allows users to feel more confident, as they learn if the editors are being disingenuous. HN users can do this by clicking on the showdead field in their profile ('show dead' if translated literally). [4]

Comments

Bad comments seem to be a more serious problem than bad submissions. While the quality of links on the main page has not changed significantly, the quality of the average comment has somewhat deteriorated.

There are two main types of comment toxicity: rudeness and ignorance. There is a lot of overlap between these two characteristics—rude comments are likely also ignorant—but the strategies for dealing with them are different. Rudeness is easier to manage. You can set rules stating that users should not be rude, and if you can get them to behave appropriately, keeping rudeness under control is quite possible.

Controlling ignorance is more difficult, perhaps because ignorance is not as easily distinguishable. Rude people often know that they are rude, while many ignorant people are not aware that they are ignorant.

The most dangerous form of an ignorant comment is not a long, but incorrect statement; it is a stupid joke. Long but incorrect statements are extremely rare. There is a strong correlation between the quality of a comment and its length; if you want to compare the quality of comments on public sites, the average length of a comment will be a good indicator. The likely reason is human nature rather than something specific to the topic being discussed. Ignorance probably more often takes the form of having several ideas rather than incorrect ones.

Regardless of the reason, ignorant comments are usually short. And since it's difficult to write a short comment that stands apart from the amount of information it conveys, people try to stand out by attempting to be funny. The most tempting format for ignorant comments is supposedly witty insults, probably because insults are the easiest form of humor. [5] Therefore, one of the benefits of banning rudeness is that it also eliminates such comments.

Bad comments are like kudzu: they rapidly take over. Comments have a much greater impact on other comments than suggestions on new content. If someone offers a terrible article, other articles do not become less successful because of it. But if someone posts a stupid comment in a discussion, it leads to a ton of similar comments in that area. People respond to stupid jokes with stupid jokes.

Perhaps the solution lies in adding a delay before people can respond to comments, and the duration of the delay should be inversely proportional to the presumed quality of the comment. That way, there will be fewer foolish discussions. [6]

People

I noticed that most of the methods I've described are conservative: they aim to preserve the character of the site rather than enhance it. I don't think I'm biased on the issue. It relates to the nature of the problem. Hacker News has been fortunate to get off to a good start, so in this case, it's literally a matter of preservation. However, I believe this principle is applicable to sites of various origins.

The good things about community sites come more from people than from technology; technology usually comes into play when it needs to prevent bad things from happening. Technologies can certainly enhance discussions—nested comments, for example. But I'd prefer to use a site with basic features and smart, friendly users rather than a fancy site populated only by idiots and trolls.

The most important thing a community site should do is attract the people it wants to see as its users. A site that tries to be as big as possible is trying to attract everyone. But a site aimed at a specific type of user should only attract them— and, just as importantly, repel everyone else. I've intentionally tried to do this with HN. The site's graphic design is as simple as possible, and the rules of the site prevent dramatic headlines. The goal is for a person visiting HN for the first time to be interested in the ideas being expressed here.

A drawback of creating a website targeted solely at a specific type of users is that it may be too appealing to those users. I know very well how addictive Hacker News can be. For me, as well as for many others, it's a kind of virtual town square. When I want to take a break from work, I head to the square, just as I might stroll through Harvard Yard or University Avenue in the physical world. [7] But the online square can be more dangerous than the real one. If I spend half a day wandering along University Avenue, I would notice. I have to walk a mile to get there, and visiting a café is different from working. However, visiting an online forum requires just one click and looks very much like work. You might be wasting your time, but you’re not just idling. Someone on the internet is wrong, and you’re fixing the problem.

Hacker News is definitely a useful site. I have learned a lot from what I have read on HN. I have written several essays that started as comments here. I wouldn't want the site to disappear. But I want to be sure it isn't a web addiction to productivity. What a terrible disaster it would be to lure thousands of smart people to a site just to waste their time. I wish I could be 100% sure that this doesn't describe HN.

I think addiction to games and social applications is still largely an unsolved problem. The situation is similar to crack in the 1980s: we invented terrible new things that are addictive, and we have not yet perfected the ways to protect against them. We will eventually improve, and this is one of those issues I want to focus on in the near future.

Notes

[1] I tried to rank users by both their average and their median number of comments, and the average (discarding outliers) seems to be a more accurate indicator of high quality. While the average number of comments may be a more accurate reflection of poor comments.

[2] Another thing I learned from this experiment is that if you are going to differentiate people, make sure you do it correctly. This is the kind of problem where rapid prototyping doesn't work. In fact, a reasonable honest argument is that differentiating between different types of people might not be the best idea. The reason is not that all people are the same, but that it's easy to make mistakes, and hard not to err.

[3] When I notice rough linkjacking posts, I replace the URL with what was copied. Websites that frequently use linkjacking get banned.

[4] Digg is notorious for its lack of clear identity identification. The root of the problem is not that the guys who own Digg are particularly secretive, but that they use the wrong algorithm to generate their homepage. Instead of ballooning from the top while getting more votes like Reddit, stories start at the top of the page and tumble down with new submissions coming in.

The reason for this difference is that Digg is derived from Slashdot, while Reddit is derived from Delicious/popular. Digg is Slashdot with voting instead of editors, and Reddit is Delicious/popular with voting instead of bookmarks. (You can still see remnants of their origins in the graphic design.)

Digg's algorithm is very sensitive to games, because any story that makes it to the homepage is a new story. This, in turn, forces Digg to implement extreme countermeasures. Many startups have some secret regarding the tricks they had to resort to in their early days, and I suspect Digg's secret is that the best stories are actually chosen by editors.

[5] The dialogue between Beavis and Butthead was mostly based on this, and when I read comments on really bad sites, I can hear their voices.

[6] I suspect that most methods of combating stupid comments have not been discovered yet. Xkcd implemented the smartest method on its IRC channel: don’t allow the same thing to be said twice. Once someone says “fail,” don’t let them say it again. This effectively punishes short comments especially, as they have fewer opportunities to avoid repetition.

Another promising idea is a silly filter, which is a probabilistic spam filter trained on a base of silly and normal comment structures.

It may not be necessary to delete bad comments to solve the problem. Comments at the bottom of a long discussion are rarely seen, so it is quite sufficient to include quality prediction in the comment sorting algorithm.

[7] What makes most suburbs so demoralizing is the lack of a central area to walk around.

Thank you Justin Kan, Jessica Livingston, Robert Morris, Alexis Ohanian, Emmett Shear, and Fred Wilson for reading the drafts.

Translation: Diana Sheremyeva
(Part of the translation is taken from translatedby)

Only registered users can participate in the survey. Please log in, please.