-
The world is locked right now in a deadly puzzle, and needs something like a miracle of good thought if it is to have the survival odds one might wish the world to have.
-
Despite all priors and appearances, our little community (the "aspiring rationality" community; the "effective altruist" project; efforts to create an existential win; etc.) has a shot at seriously helping with this puzzle. This sounds like hubris, but it is at this point at least partially a matter of track record.[1]
-
To aid in solving this puzzle, we must probably find a way to think together, accumulatively. We need to think about technical problems in AI safety, but also about the full surrounding context -- everything to do with understanding what the heck kind of a place the world is, such that that kind of place may contain cheat codes and trap doors toward achieving an existential win. We probably also need to think about "ways of thinking" -- both the individual thinking skills, and the community conversational norms, that can cause our puzzle-solving to work better. [2]
-
One feature that is pretty helpful here, is if we somehow maintain a single "conversation", rather than a bunch of people separately having thoughts and sometimes taking inspiration from one another. By "a conversation", I mean a space where people can e.g. reply to one another; rely on shared jargon/shorthand/concepts; build on arguments that have been established in common as probably-valid; point out apparent errors and then have that pointing-out be actually taken into account or else replied-to).
-
One feature that really helps things be "a conversation" in this way, is if there is a single Schelling set of posts/etc. that people (in the relevant community/conversation) are supposed to read, and can be assumed to have read. Less Wrong used to be a such place; right now there is no such place; it seems to me highly desirable to form a new such place if we can.
-
We have lately ceased to have a "single conversation" in this way. Good content is still being produced across these communities, but there is no single locus of conversation, such that if you're in a gathering of e.g. five aspiring rationalists, you can take for granted that of course everyone has read posts such-and-such. There is no one place you can post to, where, if enough people upvote your writing, people will reliably read and respond (rather than ignore), and where others will call them out if they later post reasoning that ignores your evidence. Without such a locus, it is hard for conversation to build in the correct way. (And hard for it to turn into arguments and replies, rather than a series of non sequiturs.)
When I was doing the job, I would have appreciated having an anonymized offline copy of the database; specifically the structure of votes.
Anonymized to protect me from my own biases: replacing the user handles with random identifiers, so that I would first have to make a decision "user xyz123 is abusing the voting mechanism" or "user xyz123 is a sockpuppet for user abc789", describe my case to other mods, and only after getting their agreement I would learn who the "user xyz123" actually is.
(But of course, getting the database without anonymization -- if that would be faster -- would be equally good; I could just anonymize it after I get it.)
Offline so that I could freely run there any computations I imagine, without increasing bills for hosting. Also, to have it faster, not be limited by internet bandwidth, and to be free to use any programming language.
What specific computations would I run there? Well, that's kinda the point that I don't know in advance. I would try different heuristics, and see what works. Also, I suspect there would have to be some level of "security by obscurity", to avoid Eugine adjusting to my algorithms. (For example, if I would define karma-assassination as "a user X downvoted all comments by user Y" and make the information public, Eugine could simply downvote all comments but one, to avoid detection. Similarly, if sockpuppeting would be defined as "a user X posts no comments, and only upvotes everything but user Y", Eugine could make X post exactly one comment, and upvote one random comment by someone else. The only way to make this game harder for the opponent is not to make the heuristics public. They would be merely explained to other moderators.)
So I would try different definitions of "karma assassination" and different definitions of "sockpuppets", see what the algorithm reports, and whether looking at the reported data again matches my original intuition. (Maybe the algorithm reports too much, because e.g. if a user posted only one comment on LW, then downvoting his comment was detected as "downvoting all comments from a given user", although I obviously didn't have that in mind. Or maybe there was a spammer, and someone downvoted all his comments perfectly legitimately.)
Then the next step would be, as long as I believe I have a correct algorithm, to set up a script for monitoring the database, and reporting me the kind of behavior that matches the heuristic automatically. This is because I believe that investigating things reported by users is already too late, and introduces biases. Some people will not report karma assassination, because they will mistake it for genuine dislike by the community; especially the new users intimidated by the website. On the other hand, some people will report every single organic downvote, even if they well deserved it. I have seen both cases during my role. It's better if an algorithm reports suspicious behavior. (The existing data would be used to define and test heuristics about what "suspicious behavior" is.)
That would have been what I wanted. However, Vaniver may have completely different ideas, and I am not speaking for him. Now it's already too late for me; I have a new job and a small baby, not enough free time to spend examining patterns of LW data. Two years ago, I would have the time.
(Another thing is, the voting model has a few obvious security holes. I would need some changes in the voting mechanism implemented, preferably without having a long public debate about how exactly the current situation can be abused to take over the website by a simple script. If I had a free weekend, I could write a script that would nuke the whole website. If Eugine has at least average programming skills, he can do this too; and if we start an arms race against him, he may be motivated to do it as a final revenge.)
It is actually not obvious to me that we gain by having upvotes/downvotes be private (rather than having it visible to readers who upvoted or downvoted which post, as on Facebook). But I haven't thought about it much.