A recent discussion post has compared the difficulty of an AI destroying modern human civilization to that of a modern human taking over the Roman Empire, with the implication that it is impossible.
The analogy has a few problems: first, modern humans don't have much greater raw intelligence than the Romans, only a bit more knowledge and tools; an AI would have a genuine intelligence advantage. Second, a high-tech civilization like ours offers many more ways for a genius to cause chaos than existed in classical Rome: it's more plausible that you can throw a few existing technologies together to create a superweapon than that Ptolemy could have done likewise, and there's no ancient Roman equivalent to hacking a nuclear launch system.
But taking over ancient Rome might serve as an interesting upper bound on the difficulty of an AI taking over modern civilization. And it's a theme of Harry Potter and the Methods of Rationality that rationalists should be able to come up with creative solutions to seemingly hard problems. So if Professor Quirrell offered it as an extra credit assignment, how would you take over Rome?
Here are the rules:
- You are thrown back in time to the year 1 AD. You can choose to arrive anywhere in the world, but your method of arrival cannot itself give an advantage (you can't appear in a flash of light in the middle of a religious ritual or anything).
- You do not start with Roman citizenship or any other legal record of your existence.
- You keep your original physical characteristics, including sex, height, and fitness. You will appear in period-appropriate dress of your choosing, and can't carry any artifacts with you. You may start with enough money to live a patrician lifestyle for a year.
- You are intellectually near-perfect. You know all human knowledge as of 2012. You speak fluent Latin (and all other languages of the day) and can orate as eloquently as Cicero or Demosthenes. You are a tactical genius of the order of Caesar and Napoleon. And you have infinite willpower and goal-directedness: aside from human necessities like sleep or food, you need never rest.
- You win if you either become Roman Emperor (and are acknowledged as such by most Romans), or if a state you control conquers the city of Rome. You lose if you die, of old age or otherwise, before completing either goal.
You may well be right about what the LW karma system is about.
My $0.02 on a related subject follow, though.
A reputation system provides me with value insofar as it gives me feedback about how valuable various pieces of content are judged by people whose opinions I value.
LW-karma's biggest weakness in this area (which it shares with every other reputation system I've ever seen used anywhere) is that it doesn't distinguish between the judgments of people whose opinions I value, those whose opinions I anti-value, and those whose opinions don't matter to me at all.
Ignoring that for the moment (in effect, pretending I value everyone's opinion equally), a system modeled on "ignore the current value, upvote or downvote based on whether I like it or not" means that when comparing comment A with a score of N to comment B with a score of N/2, all that tells me is that twice as many people who read the comment net-liked A than B. This might be because twice as many people read A, and the comments were approximately equally likable. It might be because 20 times as many people read A and A was far less likable. It might be because 20 times as many people read B and A was far more likable.
In other words, a karma system modeled that way means I can't really tell based on the karma scores of a comment to what degree people found the comment likable. Since that's exactly what I want to know, a karma system modeled that way is relatively useless to me, unless I can assume that roughly the same number of people read A and B. On LW, I don't consider that assumption warranted, so to the extent that you are correctly describing the LW karma system, that system is relatively useless to me.
By contrast, a karma system modeled on "if I like it more than its current rating, upvote it, if I like it less than its current rating, downvote it," doesn't have that defect. It would, of course, have the defect that late voters count more than early voters for most comments, which creates incentives for vote-withholding and generally makes it hard to extract the information I want from vote totals. I am to a large extent prepared to ignore that failing, as it strikes me as falling into the wireheading category, and trying to stop a system that's able and inclined to wirehead from choosing to do so is like trying to push back the tides.
So on balance, I think that RobertLumley's model is more useful than yours.
Again, I'm not arguing that he's right about which model is actually in use on LW. I suspect that in practice it's neither; the actual system is far messier and more embarrassing to articulate, and in practice karma scores don't provide much useful information except in very extreme cases.
Mostly, karma scores are a mechanism for silencing people whose contributions are sufficiently widely disliked so they stop pestering everyone else. Which is fine, and I don't object to that, but let's not make more of it than it is.
I can mostly agree with what you saying here. What I dislike a bit, is playing two different games simultaneously. One is the debating, what a like a lot, the other is karma shooting what I still find funny but a little annoying in the midst of the first.
For the first is important and the second does
But do I really want to read only what I already know or agree with?