Disconnect between Stated/Implemented Preferences

post by Strange7 · 2010-03-17T02:26:45.596Z · LW · GW · Legacy · 59 comments

Currently, the comment for which I've received the most positive karma by a factor of four is a joke about institutionalized ass-rape. A secondhand joke, effectively a quote with no source cited. Furthermore, the comment had, at best, tangential relevance to the subject of discussion. If anyone were to provide a detailed explanation of why they voted as they did, I predict that I would be appreciative.

Based on this evidence, which priors need to be adjusted? Discuss.

59 comments

Comments sorted by top scores.

comment by MrHen · 2010-03-17T17:40:16.582Z · LW(p) · GW(p)

As a counterpoint, my highest rated comments are huge walls of text. This could be because (a) I don't make a lot of jokes or (b) I make crappy jokes or (c) people like my walls of text more than the typical wall of text or (d) something else.

I keep an eye on my karma and have noticed these things that I believe are related to your post:

  • Talking about the karma system has fallen out of favor. I think people are getting tired of it.

  • Asking why something was downvoted usually brings more upvotes unless you really, really deserved the downvotes. This latter case will probably be swarmed with downvotes.

  • Some (many?) people vote with an end score in mind. These meta-voters have more affect on threshold comments that fluctuate between -2 and 2.

  • Long conversations will generally pull between -1 and +2 per post. Most of my karma comes from lengthy discussions in the comments. Even if the top level post was only rated +2 I will net almost 100 karma points from the post and comments. (Note: I haven't actually added this up. It may be closer to 75 karma.)

  • Quick responses pointing out third alternatives or simple problems get upvoted and usually roam between +2 and +14. If you want to get your karma higher, this is the easiest way. Comment immediately after a post is submitted and point out the most obvious flaw respectfully and concisely. Don't try to make a point, just note an error. If you get in before the rest of us, you will probably get +4 or higher.

  • Long responses pointing out serious problems get upvoted but have a lower chance of pulling large amounts of karma than quick responses. However, after the first wave of quick responses, only the longer comments are true candidates for higher karma. I suspect that a long response to a quick comment has a good chance to do well, but haven't really watched those comments yet.

  • Jokes are upvoted when they are either extremely funny or solidly funny and on topic. Randomness is upvoted if it is an inside joke, otherwise stays around +0. Sarcasm is appreciated but has the problem of being mistaken for nonsarcasm.

  • Extending the point or conversation of a top-level post gets upvoted. Most of mine get between +2 and +4. Examples would be almost every comment I have made while reading the sequences.

  • Aggressiveness is generally poorly received on technical topics. Aggressiveness is easier to get away with when dealing with fuzzy topics. I attribute this mostly to margin of error. Technical topics are harder to be bulletproof. I am still having a hard time predicting which non-technical top level posts are voted up. I suspect this is because I don't know what is already been discussed or that the issue is a technical topic that I have misidentified.

  • Self-depreciation is a huge karma pull. Both my highest rated post and comment were essentially me slamming myself over and over. Each were voted higher than +20.

  • Bullet point lists of extensions, ideas, questions, and so on seem to do about as well or better than long paragraphs of text. The walls of text may be harder to skim for goodies or the bullet point list better organized?

  • Non-aggressive requests for clarification or information are not generally downvoted. Mine seem to roam between 0 and +2 karma. If the question and response delve into a lengthy but friendly conversation, I seem to get between 0 and +2 karma for each of my comments. If a solid agreement or conclusion is reached, the capping comment gets about double whatever the individual comments were getting.

  • Posting nearby "famous" people amps up the karma action. Replying to EY, Alicorn, Vladimir_Nesov, pjeby, et al will increase the amount of people that read your comments. The reasons for this are varied. The four I used here are just names that popped into my head. Also, some people seem to vote up conversations they are in while others do not. A few downvote anyone disagreeing with them. The people that matter generally fit these criteria: (a) top contributer (b) easily recognizable name (c) frequent poster/commenter (d) abnormal amount of recent activity (e) holds atypical beliefs for the community (f) a troll.

  • Better grammar, spelling, and language increases the likelihood that your comment will move upward faster.

  • Comments quickly upvoted higher than typical seem to either (a) go through the roof or (b) get meta-voted back to between +1 and +3.

  • Telling people to vote in a particular way tends to produce easy to predict results but not in a manner that is easy to describe.

And oh wow did that get long. Do note that this is all being typed from the top of my head using myself and the comments I read as an example. Naturally, the above does not dictate how people vote.

EDIT: I guess for fun, I predict that this comment gets ROT13orgjrra cbfvgvir gjb naq cyhf fvk xneznROT13.

EDIT 2: I would normally downvote a post such as this but elsewhere in the comments you seemed to have received the message and was wondering about deleting it. So I just left it as it is. Also relevant: I generally do not upvote jokes unless they are truly amazing.

Replies from: mattnewport, prase, Morendil, Lightwave
comment by mattnewport · 2010-03-18T05:59:33.194Z · LW(p) · GW(p)

There also appears to be a non-linear effect where posts that reach a certain threshold of votes get even more votes - like there's some kind of 80/20 power law at work. I can think of a number of possible explanations for that phenomenon but it would be interesting to see if the data backs up my impression first. I predict that the Gini coefficient for comment karma is > 0.3. Does anyone know a better way to obtain the data than a web scraper?

ETA: I'm referring to the Gini coefficient for individual comments' karma rather than for individual commenters'.

ETA2: How to handle negative karma? I'd suggest ignoring negative scoring posts but other suggestions welcome.

comment by prase · 2010-03-18T09:18:13.784Z · LW(p) · GW(p)

This analysis looks brilliant, but the impression is sort of spoiled by the fact that the prediction of karma gain failed. It is even close to self-referential paradoxes: comment reading "This comment will get X karma" should be upvoted because of its precise prediction when it actually has X karma, but upvoting makes the precision worse and may attract later downvotes.

Replies from: MrHen
comment by MrHen · 2010-03-18T13:38:51.742Z · LW(p) · GW(p)

Yeah, I was way off. I didn't think people would be that interested in karma theory. I think the big oops was the first bullet point.

comment by Morendil · 2010-03-18T14:17:52.929Z · LW(p) · GW(p)

Posting nearby "famous" people amps up the karma action

This in itself strikes me as a sufficient incentive to aim at getting into the "Top contributors" list. My "return on time investement" for participating in LW is what I learn (often tangentially) from getting into interesting conversations. The lower my chance of being ignored, the higher my expected utility from any contribution to LW.

Replies from: MrHen
comment by MrHen · 2010-03-18T20:23:54.577Z · LW(p) · GW(p)

There is a slight difference between being a top contributor and being famous as I am mentioning it here.

My current karma experiment is deliberately not posting comments I think are worth less than 2 karma unless I have a compelling reason to do so (such as asking for help or information). My goal is to increase the quality of my comments to the point that someone could think, "What has MrHen posted recently?" and the answer is more impressive than a series of one-liners and nitpicky comments.

Ideally, this will increase the weight of my words to the point that when I speak, people will listen. It is a pure, straightforward status grab, but one that doesn't involve gobbling up karma. The pinnacle of the status tree at LessWrong is to have someone think, "I, the reader, am wrong" instead of, "they, the writer, are wrong."

I am compiling a mental list for LW status games, rewards, and penalties that is similar to the karma list above. I am not much for status games but the gaming here seems to be harder to avoid than in real life. ("Avoid" is the wrong term but conveys the right intent. Status games are hard to "avoid.")

My biggest regret with the above karma list is that I have no good way to verify or catalogue my comments and their karma. LW just doesn't have any tools to make such a thing easy. I worry that the status list will be even further removed from reality -- possibly to the point of being unable to predict anything at all.

In any case, I like to think that I am getting better in regards to comment quality and topical knowledge. My karma rating keeps going up, so that's a good sign.


EDIT: Replaced "I think will be rated lower than 2 karma" with "I think are worth less than 2 karma."

Replies from: NancyLebovitz, Document
comment by NancyLebovitz · 2010-03-18T22:22:36.103Z · LW(p) · GW(p)

Why do you want more status?

Do you have ideas that you think are good but won't be heard unless you're more respected?

Replies from: MrHen
comment by MrHen · 2010-03-19T02:29:11.744Z · LW(p) · GW(p)

Not so much ideas as things that I think would be helpful but will be buried in obscurity at the level I am.

That being said, this is helped more by learning the basics through reading the sequences than playing status games. My thoughts on status should be taken with the clarification that I am primarily seeking to learn and am thinking about status because I find it interesting.

If I can take my comments and change my behavior so as to be looking forward and increase my status without hampering my ability to learn... why shouldn't I? When I look at the various status levels here at LessWrong I notice the highest level I can see, which is the one I mentioned in my previous comment.

These comments are meant to be taken as observations for people who are curious or interested. I am not nearly as concerned with status as these comments may imply. I just know status is there and wonder about it in the same way I wonder about most things.

comment by Document · 2010-04-11T08:18:56.484Z · LW(p) · GW(p)

My biggest regret with the above karma list is that I have no good way to verify or catalogue my comments and their karma. LW just doesn't have any tools to make such a thing easy.

Have you tried Wei Dai's script?

comment by Lightwave · 2010-03-18T14:13:22.326Z · LW(p) · GW(p)

You wrote this (now high-rated) wall of text just to further prove your point, didn't you?

Replies from: kpreid
comment by kpreid · 2010-03-18T16:09:09.593Z · LW(p) · GW(p)

This reminds me of that my highest rated so far answer at Stack Overflow was a big bullet pointed list.

I think the format itself (a list, where each item is at least a whole sentence) has a connotation of thoroughness and authority; and it reduces the TL;DR effect by presenting the content in small-but-not-trivial separable pieces.

Also, I see a “wall of text” as not having bullet points: being a series of long paragraphs without much structure visible without reading the text itself. The term is being misapplied to this comment in my opinion.

Replies from: MrHen
comment by MrHen · 2010-03-18T20:03:50.192Z · LW(p) · GW(p)

I agree that a bullet point list isn't generally considered a wall of text.

comment by Paul Crowley (ciphergoth) · 2010-03-17T15:05:31.269Z · LW(p) · GW(p)

Number of upvotes indicates breadth, not strength, of support.

comment by RobinZ · 2010-03-17T02:34:24.014Z · LW(p) · GW(p)

Upvotes come more easily - and therefore more numerously - when it takes little effort to evaluate the upvote-worthiness.

Replies from: h-H
comment by h-H · 2010-03-20T18:48:14.074Z · LW(p) · GW(p)

should the number of allowed down votes be increased to encourage more down-voting?

Replies from: RobinZ
comment by RobinZ · 2010-03-20T19:11:37.073Z · LW(p) · GW(p)

Possibly, but that's rather a matter for the Meta Open Thread.

comment by orthonormal · 2010-03-17T04:19:51.971Z · LW(p) · GW(p)

This should be a comment on the Open Thread, not a top-level post.

comment by wedrifid · 2010-03-17T13:18:56.237Z · LW(p) · GW(p)

The vote I just added was because:

  • It added value to my life. Laughing is healthy and it was certainly the funniest joke I had read today. This is despite reading both XKCD and SMBC.
  • It was relevant the parent.
  • Biasses in grant allocation and the epistemic consequences are relevant to the core topic of the site.
  • A vote by me means "I wish to see more comments similar to this one". I wish to read more jokes like that.

(I downvoted this post.)

Replies from: thomblake
comment by thomblake · 2010-03-17T13:24:08.359Z · LW(p) · GW(p)

ditto here.

comment by prase · 2010-03-17T11:55:55.142Z · LW(p) · GW(p)

I haven't upvoted the joke, mainly because I think that jokes generate more karma than I think they are worthy, and thus I never upvote jokes since the moment I made that observation. But hadn't I the deliberate policy of never upvoting jokes, I would simply upvote it because it was funny.

Remember, upvoting here, from a single user perspective, is a boolean scale. A good comment earns my single upvote, a bad comment earns single downvote, while most comment pass along without my voting because a) I don't read them for different reasons or b) I am not sure whether they are good or bad.

If you make a thoughtful comment, many readers aren't interested to read it and skip it, because it is probably too long, or because it is incomprehensible, or because they aren't interested. From those who read, they would need to think about it for a while before they realise that the reply was really good, and maybe forget to upvote afterwards. On the other hand, a joke generates quite a few instinctive upvotes at the beginning, which makes it even more visible (I tend to read upvoted comments first and more carefully), the visibility then generates more karma.

Nobody is responsible for the others' choices and thus for the fact that the comment stands now at +16. For any single user, the joke is at +1 as well as other better posts, which (s)he is aware of. We want to hear good jokes and thus want to reward them somehow, and the system doesn't allow to differentiate between them and the information-bearing intelligent comments. I suppose that if it was possible to upvote on some more continuous scale, the seeming discrepancy between the individual preferences and the collective voting results would disappear.

Replies from: RobinZ
comment by RobinZ · 2010-03-17T12:42:16.314Z · LW(p) · GW(p)

Note, for comparison purposes, that this related comment stands at +10 after 22 hours.

Edit: The parent stands at +4 after 24 hours.

Replies from: prase
comment by prase · 2010-03-17T15:52:56.315Z · LW(p) · GW(p)

Do you mean that I have uselessly written basically the same thing in 15 times more words? I have read your comment, but I think mine isn't completely isomorphic, although it is related.

Replies from: FAWS, RobinZ
comment by FAWS · 2010-03-17T16:01:44.915Z · LW(p) · GW(p)

It's more that RobinZ post is itself an example of the effect that very post describes (as does your post, in a way), and your post is an example of the effect at the other end of the spectrum. RobinZ seems to be making the prediction that your post will not gain as much karma.

Replies from: prase, RobinZ
comment by prase · 2010-03-17T17:55:02.289Z · LW(p) · GW(p)

your post will not gain as much karma

It certainly will not, but do we know whether this will be because of the length, or because RobinZ's post is simply better, or because it was posted sooner?

Replies from: RobinZ
comment by RobinZ · 2010-03-18T05:41:10.167Z · LW(p) · GW(p)

It is not a well-controlled experiment - but ciphergoth's top-level comment, which is later than prase's, has already gotten more votes than prase's. This would suggest that the delay was not an overwhelmingly strong effect.

(Actually, MrHen's comment likewise, and MrHen's is not short.)

Replies from: prase
comment by prase · 2010-03-18T09:13:19.915Z · LW(p) · GW(p)

There are other effects as well, status of the commenter included. People are more likely to read longer comments if the author is known to make good comments. I guess that this effect is weaker for shorter comments where the evaluation of actual content is easier. So, low status karma maximisers should write a lot of brief comments, and as they reach high status, they should swap to longer "walls of text".

comment by RobinZ · 2010-03-17T16:03:23.193Z · LW(p) · GW(p)

Correct on every point. Indeed, prase's comment appears to imply the same prediction.

Edit: I posted my comment primarily in order that, when timestamp "18 March 2010 09:55:55AM*" arrives, a reader can examine the two posts and observe the difference in karma corresponding to the same length of voting time. I would do so myself, but that timestamp is pretty early in my time zone.

Replies from: FAWS
comment by FAWS · 2010-03-17T16:07:20.050Z · LW(p) · GW(p)

I was editing that in while you were replying or at least before I saw the reply, ironically I hadn't read prase's post all that closely so it didn't come to mind immediately.

comment by RobinZ · 2010-03-17T16:16:38.626Z · LW(p) · GW(p)

FAWS is correct, but (1) thanks for asking - I should have made that clear, and (2) props for calculating the ratio of the wordcounts. ;)

(Edit: I admit that I had actually suspected you were intentionally performing the experiment when I first saw your comment - I know I was intentionally writing a short, immediately-comprehendible comment myself, to invoke the effect; that made it plausible to me that you were intentionally writing a long, substantive comment for the same reason.)

Replies from: prase
comment by prase · 2010-03-17T18:01:37.597Z · LW(p) · GW(p)

No, I wasn't performing an experiment, at least not intentionally. In fact, I didn't count your comment as a cheap karma generator (I have even upvoted it, IIRC) - it was brief, but relevant reply to the OP, and not visibly funny.

Replies from: RobinZ
comment by RobinZ · 2010-03-17T18:54:12.405Z · LW(p) · GW(p)

It wasn't meant to be a cheap karma generator - it was meant to be short and easy to confirm as good. The jokes everyone have been talking about are likewise.

comment by JGWeissman · 2010-03-17T04:32:19.994Z · LW(p) · GW(p)

You raise a valid point, but,as is, it is not worthy of a top level post. I predict it would have had positive karma as comment in the open thread or meta thread.

Replies from: Kevin, Strange7
comment by Kevin · 2010-03-17T04:38:57.657Z · LW(p) · GW(p)

"Worthy?" If that is defined by positive or negative karma, then yes, it is not worthy. However, when you want an answer to a question and are willing to lose karma for it, the current incentives are for making a top level post rather than an open thread comment, as long as you can deal with the people making meta comments...

Replies from: JGWeissman
comment by JGWeissman · 2010-03-17T04:57:29.408Z · LW(p) · GW(p)

I would say that karma measures the community's aggregate perception of worthiness, but does not define worthiness. My personal perception of worthiness reflects on the importance of the topic and the quality and novelty of the contribution to that topic.

comment by Strange7 · 2010-03-17T04:35:39.313Z · LW(p) · GW(p)

Should I remove it?

Replies from: JGWeissman
comment by JGWeissman · 2010-03-17T04:47:14.177Z · LW(p) · GW(p)

Once it is published, I believe that deleting really only removes the links to it from the various post feeds. But do take it as a learning experience.

Replies from: Strange7
comment by Strange7 · 2010-03-17T04:51:54.777Z · LW(p) · GW(p)

Well, at least I've got enough of a buffer to try again.

But do take it as a learning experience.

I always try to. So long as we're on the subject, any tips on top-level post subject selection and formatting?

Edit: I spoke too soon.

Replies from: JGWeissman, orthonormal
comment by JGWeissman · 2010-03-17T05:14:04.051Z · LW(p) · GW(p)

My preferences for top level posts, which seem to be a higher standard than the community average, is that they should present some novel insight into a relevant topic, and reflect some research into previous work on the topic; or that they offer a deep review of a relevant work external to Less Wrong.

Replies from: RobinZ
comment by RobinZ · 2010-03-17T12:32:22.868Z · LW(p) · GW(p)

Bear in mind that those with lower standards will make more top-level posts.

Replies from: JGWeissman
comment by JGWeissman · 2010-03-17T17:10:03.475Z · LW(p) · GW(p)

Do you mean to imply I am underestimating the community average standard? I was basing my assessment on karma scores on posts I consider low quality, not so much on what people decide to post.

Replies from: RobinZ
comment by RobinZ · 2010-03-17T17:15:09.740Z · LW(p) · GW(p)

Good point - I still think it is relevant to note that those with low standards will be more willing to post, but given that your evidence is independent, the "bear in mind" is unwarranted.

comment by orthonormal · 2010-03-21T19:30:38.066Z · LW(p) · GW(p)

My default place to write something is the Open Thread or a comment on a relevant post. It's only if I realize that an idea needs to be a top-level post (because it's too important and not found elsewhere in a post, or because it's a lengthy development of a key idea) that I make it into one.

As you can see from my two articles, this may be too strict a standard, but I think it's in the right direction relative to the median.

comment by Nick_Tarleton · 2010-03-17T03:50:44.489Z · LW(p) · GW(p)

Criticism of scientific institutions is more than tangentially relevant to LW as a whole.

comment by Liron · 2010-03-17T08:44:32.146Z · LW(p) · GW(p)

"by a factor of four" is not good linking text.

comment by Psychohistorian · 2010-03-17T19:01:00.030Z · LW(p) · GW(p)

While I don't think this type of post should in general be encouraged, it has generated an interesting discussion on karma; -7 (current score) seems unduly harsh to me.

Replies from: RobinZ
comment by RobinZ · 2010-03-17T19:21:38.453Z · LW(p) · GW(p)

It occurs to me that a reasonable strategy for some fraction of LessWrongers may be to downvote immediately and remove that downvote later, under some circumstances. Immediate feedback teaches the lesson, and subsequent change removes the undue harshness.

Following this logic, I have removed my downvote.

Replies from: SilasBarta
comment by SilasBarta · 2010-03-17T19:29:09.254Z · LW(p) · GW(p)

lol I think people already do that on top-level posts. Very few of the ones that get downvoted a lot the first day don't return to positive after some time.

Replies from: RobinZ
comment by RobinZ · 2010-03-17T19:33:32.465Z · LW(p) · GW(p)

Yes, and now I can see the justification for it.

comment by Lightwave · 2010-03-18T14:33:19.558Z · LW(p) · GW(p)

This reminds me of reddit. I quite enjoy reading the highly rated jokes (or whole joke subthreads) there.

comment by kpreid · 2010-03-17T13:05:08.646Z · LW(p) · GW(p)

I, for one, try to downvote jokes which are funny but not profound, occasionally, when I think of it.

comment by Roko · 2010-03-18T19:07:26.978Z · LW(p) · GW(p)

In the tradition of institutionalized ass-rape, we all just raped you in the ass by voting you down to -7 ;-0

comment by dclayh · 2010-03-17T03:38:24.415Z · LW(p) · GW(p)

Currently, the comment for which I've received the most positive karma by a factor of four is a joke about institutionalized ass-rape.

Poisson noise.

Replies from: prase, Strange7
comment by prase · 2010-03-17T12:02:47.883Z · LW(p) · GW(p)

It wasn't a single data point. Posting jokes and funny but trivial comments is consistently among the most karma-profitable behaviour. See e.g. this.

Replies from: thomblake
comment by thomblake · 2010-03-17T14:37:45.493Z · LW(p) · GW(p)

and come on!

comment by Strange7 · 2010-03-17T04:56:01.495Z · LW(p) · GW(p)

That sounds suspiciously like a password. Are you saying that upvotes and downvotes are randomly distributed?

Replies from: Kevin
comment by Kevin · 2010-03-17T05:14:00.222Z · LW(p) · GW(p)

Upvotes and downvotes aren't randomly distributed?

Replies from: Strange7
comment by Strange7 · 2010-03-17T05:30:56.572Z · LW(p) · GW(p)

My impression was that any given upvote or downvote was applied by a sapient being, after some degree of conscious consideration of the post's content. I have yet to see a post including 'leetspeak' or egregiously bad grammar with positive karma, yet I have seen several such with negative karma.

Accordingly, I am curious as to why sixteen (or more?) people approved of the post in question.

Replies from: Kevin
comment by Kevin · 2010-03-17T05:44:44.756Z · LW(p) · GW(p)

Right, but don't those aggregate upvotes and downvotes follow some sort of random distribution from our shared perspective of incomplete information? That's a sincere question; I am not at all confident in my ability to do statistics or even my basic understanding of statistics.

It was funny! Humor is not discouraged here when it is legitimately funny. Something of sufficiently high humor value, especially if it contains tangential insight, can be modded high.

Replies from: FAWS
comment by FAWS · 2010-03-17T11:26:03.136Z · LW(p) · GW(p)

The individual upvotes are not independent of each other. A post can be thought of as having an "upvotability" related to its quality, snappiness, originality, conformity to group ideas, adherence to conventions etc that can be expressed as the probability that a random (registered) viewer will upvote it. The total upvotes of a post are determined by the number of viewers, the upvotability and chance (people are also influenced by the current karma of the post and their opinion of the poster, but I think we mostly agree that one ought to try to compensate for that).

Calling the fact that a post received a lot of upvotes "poisson noise" implies that you consider random chance a better explanation than a high upvotability and that your posterior for the upvotability of the post is (close to) unchanged (at the very least it implies that your posterior is closer to your prior than to an estimate based on (net- ) upvotes and estimated number of viewers [and an estimate of downvotes]). Does this match your beliefs? Otherwise guessing the password sounds like an accurate description of your behavior.

EDIT: You might also think of upvotability itself as poisson distributed, but that's a non-explanation equivalent to answering a question about what geological processes cause the Himalaya to be the highest mountain range with "one mountain range has to be highest and that just happens to be the Himalaya".