Right. The reason these traits didn't exist before hand is that they were disadvantageous. So in the absence of antibiotics, they disappear.

I think people are oversimplifying by talking about "stupid" parents. The truth is that since antibiotics and antivirals have few side-effects and are cheap to produce, it's individually rational for people to use them. But when everyone uses them, we get lots of resistance.

"So in theory slang and abbreviations would be no more difficult to translate than dictionary words. "

Sure. The problem is that the slang and abbreviation need to show up in their corpus. From what I understand, Google mostly uses Canadian and European Parliment proceedings for their sample. "LOL' dosn't show up much there...

Nowhere, though maybe they can do some statistical magic. I mentioned this down-thread. Corpus, as far as I know, applies to monolingual collections of text as well.

One thing that they can do, is to use statistical models of language to infer what unknown words "should" mean. They could even incorporate phonetic priors (IE, "Qui" sounds like "ki").

This is a big issue in French too. But statistically, it shouldn't be too hard for them to pick up the most common slang.

The only issue is that it isn't a two-way corpus. They have plenty of stuff in every language, but nothing in more than one language.

This is actually a legitimate issue with translation. I have a lot of teenage cousins from Paris, and they butcher their language as much as our teens do (que turns into ku, qui to ki, non to nn, etc). I actually speak french, so I can sort of trudge my way through it. But my cousins from Israel do it too, and the slang and misspellings completely throw translation software off.

Facebook right now has an oddly rich corpus of multi-lingual slang, they'd be in a good competitive position vs google-translate if they went through the effort to incorporate it into it's translation.

"What "friends" do you have on Facebook that do not speak the language that you do?" Family, friends you've met traveling abroad, friends you've met here but whom were not born in the US. Do you live in Nebraska?

It's not, there are people with no friends or pairs. But I've looked at some data samples from facebook, I'd be shocked if the giant component didn't contain more than 98% of non-singleton notes. I'd be very surprised if the largest "Non-giant component" has more than 10 people.

You can have the system estimate the optimal amount of weighing, by observing the inherent volatility in the data.

Actually, most statistical analysis of cop behavior show that cops still harass minorities more than you'd statistically expect from their greater chance to commit crime:

"In the period for which we have data, 1 in 7.9 whites stopped were arrested, compared with approximately 1 in 8.8 Hispanics and 1 in 9.5 blacks. These data are consistent with our general conclusion that the police are disproportionately stopping minorities; the stops of whites are more “efcient” and are more likely to lead to arrests"

http://www.stat.columbia.edu/~gelman/research/published/frisk9.pdf

This isn't about finding potheads...

The algorithm would be dynamic though, so crimes being reported somewhere else would shift police. Could be useful for a short period of time though, provided that you have a giant data gathering apparatus.

"I always thought that this was a result which was known through graph theory of what happens when you get a large number of nodes each with an arbitrary number of unique connections between them, that it would always tend towards the case that you got an average of no more than six degrees of separation for a sufficiently large network."

Not really, no. It's about scale-free networks (Networks that have preferential attachment, IE, people with tons of friends are more likely to get new friends than people with no friends. Their degree distribution, IE, the number of friends, is power-law distributed as opposed to exponential distributions, which come from friendship being totally random). You can model social networks fairly well as scale-free networks empirically. Roughly speaking, the average distance between two random notes is proportional to the log of the log of the number of nodes.

Right. Probably the best way to do it is to randomly draw a couple thousand UID's and run Dijkstra. Then you can do all sorts of neat stuff like geodesic distributions.