by Tyler Cowen
on April 15, 2012 at 6:54 pm
1. Why Netflix never implemented the prize-winning algorithm it paid for.
2. Criticism of on-line schools in Colorado.
3. An Economist Gets Lunch in The Daily Mail.
4. Do languages have a roughly constant rate of information transmission? (pdf), summary here.
5. The Will Wilkinson update.
re: #5: so hopefully this means one less libertardian promulgating abstract neohippie theory with no basis in reality.
“promulgating abstract neohippie theory with no basis in reality.”
The trolls are getting lazy.
Beats why Pfizer didn’t implement Torcetrapib which they paid close to a billion for.
HP and webos. $3.5 billion for an OS and they were talking about using it for nothing but printers.
Although it looks like it wasn’t $3.5
HP bought Palm for $1.2 Billion, including WebOS: http://www.hp.com/hpinfo/newsroom/press/2010/100428xa.html
How much did HP flush when they dumped webos?
Any news on an audiobook possibility?
It is in the works…
I like the paper on languages. However, I think that the French and English translation do not match perfectlty (i’m a native French speaker). But, I’m not a professional translater and I’m aware that translations may vary form one to another. These problems of translation surely mean that there is more variability in the results than what the authors found and that differences between languages may not be statistically significant.
It is one thing for an economist to get lunch in dumpy joints in strip malls with ugly women screaming at each other. It is quite another to get lunch in The Daily Mail. That is just over the top, :-).
BTW, as for the advice about going to places with feuds, the ultimate in that is the old phenomenon of whenever there is a mafia assassination in a restaurant, long lines form outside of it almost instantly, given that everybody knows that mafia dons really know where the good Italian food is.
Re #1: Netflix got a lot more out of the prize winning entry than the “Too hard; didn’t implement” the article implies. For a million dollars they got an algorithm that gave them an 8% improvement (this is pretty good, seeing as the 10% target was set pretty arbitrarily), and pretty strong evidence that even _trying_ to push that any higher was pretty futile, and they should shift their engineering focus to different departments.
This, IMHO, is pretty valuable. As CS researchers (and doubtless, in other fields), we spend so much time barking up the wrong tree that getting strong evidence that this particular tree has no more fruit is financially valuable information.
Sure, the entire business changed in the meantime, but even if it didn’t, the prize didn’t end up being “useless” to Netflix. Valuable knowledge was gained.
In any case, a million dollars is peanuts for a company like Netflix. I’d estimate that’s approximately what it costs them annually to employ about six good engineers( if you factor in non-salary overheads.)
5. Not a linguist, but languages do not equalise in information density, but they may differ in the information that they exchange as well.
E.g. Japanese has a low information rate here, but perhaps that reconciles because the language is structured so that a lot of information content can be “reconstructed” upstream, or where a lot of “irrelevant” (redundant) information is discarded.
After considering this, it seems harder to say whether languages are equally information dense in the sense of being able to communicate equal information in equal time, in a real world situation.
Is that another way of saying some people are smarter than others?
What, because you’d have to be smarter to learn a language with more syllabic complexity and/or a higher information transmission rate and use it at that rate? ; )
I don’t think higher inferential capacity and economy of language is necessarily smarter than transmitting information very quickly, maybe with more complex syllables, but it seems like the kind of thing people with a language with lower information transmission rate would be or if not get better at (if they were just as smart).
I was thinking your “upstream reconstruction” sounded like bit-compression and decompression. Activities requiring brainpower.
I wonder how the researchers corrected for issues like the fact that Japanese explicitly encodes social information (e.g. about the relative social status of the speaker and addressee) in a way that English or Mandarin does not. Of course English, like all languages, also indirectly encodes social information through accent and vocabulary choice. Seems like it would be very hard to truly compare the information exchange across languages when taking that into account.
The story on Colorado online schools is from 2006, in case someone didn’t notice.
And the conclusion, “Moloney said a large part of online schools’ enrollment are at-risk, academically challenged students who go to the schools as a last resort. Auditors said it was impossible to determine how many students were at-risk because there is no statewide standard,” isn’t true any longer either. The state’s evaluation method which used to look at just the percentage of kids that are proficient, advanced, partially proficient and unsatisfactory, now has a measure of year to year learning growth that addresses comprehensively the issue of at risk v. not at risk kids by virtue of its design.
I was suprised that nearly all the Daily Mail commenters were North American. Is there a special edition of the website targeted to that market? After all many of the details in Cowen’s book don’t translate well to the UK (even if the principles do).
The standard Daily Mail website is heavily watched by the North American crowd because they are the usually the first place to get the news out (They beat US sites often by hours).
Also, the website is run separately from the newspaper.
Someone explain this to me like I’m 5 years old:
I’m embarrassed to say it, but I never was able to understand how an improved algorithm for guessing people’s movie-watching preferences was supposed to make Netflix extra money.
Can someone please explain this to me in a very simple way?
Guess viewer tastes -> Recommend Movies -> Viewers Like recommended movies -> Happy Viewers = Good for Netflix
Second order effects:
Non-clustered movie demand and perhaps viewers switch to higher plans. Both good for Netflix’s bottomline
5. “underestimate the upside of taking a chance”
In my experience this doesn’t apply to graduate school. I certainly wish him better luck.
Also consider that the rental business itself is not “sticky”— anybody could collect a pile of dvds and go into competition with Netflix. The review system makes Netflix much more sticky— the customer has to go to Netflix to get good recommendations.
Linda is right, that article on on-line schools is 6 years old…what is your point Tyler? There is plenty of stuff to talk about on on-line education but an article about failures that occured at a very low point in CO’s state education administration is very old news.
My son is taking a class online (in Florida each student must take at least one online class). I have help my son with some lessons and IMHO the software is still very, very bad. The class is very poorly done. It needs to evolve a lot but I still think that in 10 years or so it can be better than most teachers. Perhaps they could start with the direct instruction scrips.
Comments on this entry are closed.
Previous post: Robert Sampson responds to my short book review
Next post: Toward a 21st-Century FDA?
Email Tyler Cowen
Follow Tyler on Twitter
Email Alex Tabarrok
Follow Alex on Twitter
Subscribe in a reader
Follow Us on Twitter
Marginal Revolution on Twitter Counter.com
Get smart with the Thesis WordPress Theme from DIYthemes.