I'm a freak about tools and symmetries. I suffer from chronic achronic dysfunction. Usually displaced a few years into the future? What is now or then? Probably on the spectrum, too, but why know where? I'd actually prefer to discover I am less right because that's where the biggest learning opportunities are, but at the same time deep learning becomes more difficult with age. Latest recent self modifications? Homo sapiens' mobile larynx and DNA as recipe book, not blueprint.

Wiki Contributions



Okay and you're welcome, though I wish I had understood that part of the discussion more clearly. Can I blame it on the ambiguity of second-person references where many people are involved? (An advantage of the Japanese language in minimizing pronoun usage?)


Is your [Dogon's] reference to "your model" a reference to 'my [shanen's] preferred financial model' (obliquely referenced in the original question) or a reference to Vladimir_Nesov's comment?

In the first case, my "preferred financial model" would involve cost recovery for services shared. An interesting example came up earlier in this discussion in relation to recognizing consistency in comments. One solution approach could involve sentiment analysis. In brief, if you change your sentiment back and forth as regards some topic, then that would indicate negative "consistency", whereas if your sentiment towards the same topic is unchanged, then it indicates positive consistency. (If your sentiment changes rarely, then it indicates learning?) So in the context of my preferred (or fantasy) financial model, the question becomes "Are enough people willing to pay for that feature?"

Now things get more complicated and interesting in this case, because there are several ways to implement the feature in question. My hypothesis is that the solution would use a deep neural network trained to recognize sentiments. The tricky part is whether we yet know how to create such a neural network that can take a specific topic as an input. As far as I know, right now such a neural network needs to be trained for a specific domain, and the domain has to be narrowly defined. But for the sake of linking it to my financial model, I'm going to risk extending the hypothesis that way.

Now we get to an interesting branch point in the implementation of this feature for measuring consistency. Where do we do the calculations? As my financial model works, it would depend on which approach the users of the feature wanted to donate money for. I'm going to split it into three projects that could be funded:

  1. Developing the deep neural network to analyze sentiments towards input topics. This is basically a prerequisite project and unless enough people are willing to fund this project the feature is DOA.
  2. Analyzing the data with the neural network on the LW (LessWrong) side. In this version of consistency measurement there would be a lot of calculation on the LW side testing sentiments against topics, so there would be both a development project and a significant ongoing cost project. Both parts of this double project would need sufficient donor support to use this approach.
  3. Analyzing the data with the neural network on the users' side. In this version of consistency measurement, the tedious calculations could be removed from LW's servers. The trained neural network would be downloaded and each person would calculate (and optionally share) the consistency metric using the data of that person's own comments. The cost of the development project should be similar, but there wouldn't need to be donors for a major ongoing cost project. (I would actually favor this version and be more likely to donate money for this version due to privacy considerations.)

(If there are enough donors, then both 2 and 3 could be supported. However, deciding which one to implement first could be determined by which project proposal attracts enough donors first.)

In the second case, I'm afraid I don't understand what part of Vladimir_Nesov's comment was about a "model". And you weren't talking to me, anyway. And I should also apologize for my longish and misdirected response?


It's hard to change or improve something without measuring it. I think you are describing a fairly complicated concept, but it might be possible to break it down into dimensions that are easier to assess. For example, if some of the assessments are related to specific comments or replies [our primary "actions" within LessWrong], then we could see what we are doing that affects various aspects of our "amiability".

This demonstration of "Personality Insights" might help illustrate what I'm talking about. If you want to test it, I recommend clicking on the "Body of Text" tab and pasting in some of your writing. Then click on the "Analyze" button to get a display for some of the primary dimensions. If you then click on the "Sunburst visualization" link at the bottom, you'll see more dimensions and how they are grouped. I think your notion of "amiability" may be within the cluster of "Agreeableness" dimensions.

Another way to think of it is related to the profiles that Facebook and the google have compiled for each of us. My understanding (from oldish reports) is that they are dealing with hundreds of dimensions. I would actually like to see my own profile and the data that created it. I might even disagree with some of the evaluations, but right now those evaluations are being used (and abused) without my knowledge.


Is this another karma-related topic? Your tags suggest otherwise, but I would like to see some of these dimensions as part of the karma metric, both for myself and for other people. Most of the examples you cite seem to be natural binary dimensions, but not fully orthogonal. Not sure what I should say here, but I'll link to my longest comment on Less wrong on the topic of enhanced karma. As you are approaching the topic, such an approach would help me recognize "amiable" people and understand what makes them amiable. I doubt that becoming more amiable is one of my goals, but at least I could reflect on why not. Or perhaps most importantly I could look for the dimensions that reflect sincere amiability to filter against the fake amiability of the "charming sociopaths" you mentioned.


I feel like this branch of the discussion might be related to Dunbar's Number? Either for total members or for active participants. Is there any data for number of participants over time and system versions?

However I also feel like Dunbar's Number is probably different for different people. Social hubs have large numbers of personal friends, whereas I feel overwhelmed by any group of 150. My personal Dunbar's Number might be around 15?


This topic of karma in general interests me, per my reaction to the karma project from 2019. However my question in response to this "site meta" item is: "Is there a karma explorer?" One side would be a way to see the basis of my own karma, but I would also like a way to understand the basis of the karma of other users. For example, I see that the author habryka has over 13,000 points of karma here and 242 points of karma somewhere else, but what does that actually mean? Does any of that karma represent reasons I should read comments from habryka with greater attention? (Right now it feels like there are a lot of magic numbers involved in karma calculations?)

More fuzzy reaction, but I feel like whatever forms the basis of karma, it should age over time. Recent contributions to karma should matter more than old ones.


Thanks for the lead to the "Site Meta" tag. I have that one open in another tab and will explore it next. However my general response to your reply is that part of the problem is that I would like to see different kinds of "tracking summaries" depending on what kinds of things I am trying to understand at a particular time.

You introduced a new example with your mention of "meetup announcements". If you are trying to track your activity on LW in terms of such meetings, then you want to see things from that perspective.

What I have done in today's experiment is to open all the "recent" notifications in tabs because it is not clear which ones are actually new... It would be helpful if the notifications pulldown list also showed the notification times (though the mouseover trick for date expansion also works for the relative dates on the floating summary that appears to the left of the notification when you hover over it). Overall I'm still having a difficult time grasping the status of this question.


Accuracy is relatively easy to assess. If you think someone is saying something that is false and you are reacting to the comment on that basis, then you should be able to cite appropriate evidence to that effect. (But the other person should be able to object to your evidence as part of a 'proper' MEPR system.) 

I actually think most dimensions of the reputation system should be normalized around zero, so that if people tend to give more negative reactions, then the system should be adjusted to make it more difficult to give a negative reaction, such as saying a comment is inaccurate. (However I also think that should be weighted by the MEPR of the person making the rating. If someone has a established a long track record of catching inaccuracies, then the likelihood is higher for that person.)

I agree that consistency is much trickier. Even in the case where I know the person has changed his mind on a topic, I would not regard it as inconsistent if there was good reason for that change. I think I might like computer support for something like that. How about a triggered search? "Show me this person's comments about <target keyword>" and I could then look over the results to see if they are unchanged, evolving over time, or jumping back and forth. 

But actually that is something I would like to apply to my own comments over time. I think I am fairly consistent, but perhaps I am deluding myself?

Load More