lliiffee

Megan McArdle, the author of this article, has a very active and interesting blog on economic issues:

http://www.theatlantic.com/megan-mcardle

Its usually the first thing I go for when I check my RSS feed.

Since you seem to know this stuff, I hope you doing mind if I ask you a marginally related question... It seems like deep learning is based on greedily building representations in an unsupervised setting. (Not sure if this is a defining characteristic of deep learning or just a common usage.) The postulate seems to be that greedily learning in a supervised setting is inferior. Why should this be so? Is there any justification (theoretical or experimental)? I can only find vague asides on this point in a few papers.

Try section 7 of these notes:

http://see.stanford.edu/materials/aimlcs229/cs229-notes3.pdf

Basically, the idea of feature spaces is to blow up the data into high dimensions. So, we use

x' = f(x)

as our data, instead of x. It turns out that in lots of machine learning algorithms (notably SVMs), you end up only needing inner products between different data elements. That is, we need to compute x^T y for two data elements x and y. In feature space, we need could compute this by doing f(x)^T f(y). However, it turns out that for certain feature spaces (like polynomials) one can compute the number f(x)^T f(y) quite quickly with out ever explicitly forming the big vectors x' or y'.

Technically what is visualized here isn't "the kernel trick". This is the general idea of how nonlinearly projecting some points into a higher-dimensional feature space makes linear classifiers more powerful. You can do this with out SVMs. Just compute the high-dimensional features corresponding to your data, then use logistic regression or whatever. Trouble is, if the higher-dimensional space is really big, this could be expensive. The "kernel trick" is computational trick that SVMs use to compute the inner product between the high-dimensional features corresponding to two points with out explicitly computing the high-dimensional features. (For certain special feature spaces.)

But this is definitely a cool visualization of the value of feature spaces!

For total exports and imports to a country, it makes no difference. For country/country pairs, it changes things. For example, a big chunk of the US trade deficit with China should be reassigned to Japan.

Another anecdote for these. I started running regularly a year or so ago. I'm no natural athlete, and even with relatively short runs of just a few miles, I would struggle with knee pain. (I'm quite tall.)

After reading some of these articles, I decided to give the five fingers a try. At first, my hamstrings and feet were so tired I could hardly walk after running. (You don't realize how many muscles you have in your feet, because you don't use them unless you land on the balls of your feet.)

Now, at about the 1 month point, my hamstrings have strengthened, I still feel a lot of soreness in my feet, but the issues with my knees are completely gone.

Maybe placebo, I know, but I don't mind placebo if it works...

It isn't fun to have a do-nothing job, especially for the type of person who would consider doing a Ph.D. in a biotech field.

However much that helps.

The concern is in setting a precedent that anyone can have a wide hearing of their ideas by killing people, which creates copycats in the future. (We have seen this happen with, e.g., school shootings.)

I urge everyone not to read this. We should not reward murderers (or attempted murderers) with out attention.

In fairness, shakeshack has a wonderfully simple version that loads for me with noscript on.

I wonder how much these problems are problems of information. The reason printers are crap is that most people buy them based on 1) price, and 2) listed features. As a result, manufacturers optimize for these. If people understood that most of the listed features would never work, the paper would constantly jam, the drivers would never be updated, toner/ink would be expensive, installing the printer would require adding lots of stupid software to windows, they would probably pick a more expensive printer with less features that actually works.

True. However, the fastest way to learn almost anything is to go hang out with experts. Graduate school, at its best, is a formalization of this. (Of course, you still have to go to the library!)

I'm not sure what the broken window fallacy has to do with anything.

As for technology, see the comments in the article by economists about the difference between lowering prices based on improved productivity (good) and just lower demand (bad).

> Now, what real-life example demands systems of equations?

Finding the cheapest phone card to use is a good example.

I really don't know, but I would guess Google's search engine is almost all C/C++. When you are operating at that scale, it makes sense to invest huge programmer time for very small speed increases.

As to "normal" people, I do a lot of numerical computing using python+C. I end up with probably 90% of the code in python, and 10% inner loops in C. That's typically enough that the C code is taking, say, 50% of the time, at which it isn't worth it (for me) to move more python into C, since I could never buy myself more than a factor of 2 speedup.

4x is low:

http://shootout.alioth.debian.org/u32q/benchmark.php?test=al...

Roughly speaking, anything time-critical needs to be done in c.

I agree. I think that the main impact of no-camera-on-iPad might be that fewer people use it as their only computer. (This might be by design.)

In my experience, people want videochat to look at their grandchildren.

Fast page refresh makes it potentially very attractive for things like textbooks where you need to skim around. For linear novel reading, I'd think a Kindle is much better.

See that, potential thief? If you steal a phone, there is a 95% probability that you keep it, sure, but a 5% probability that you stole it from a motivated technically competent person who will track you down and ... make you give it back. That huge downside risk will certainly will prevent you from stealing things in the future!

I'm not quite sure what it means to say the U.S. has "permitted police to monitor communications without permission". Was it permitted or not?

You can solve amazingly large problems with a sparse matrix with out understanding what you are doing, as long as you use direct (factorization) methods. If you type A\b with some gigantic sparse A, matlab will apply reordering heuristics that often just work, and are actually incredibly hard to improve upon by manual fiddling. When you get to really, really big problems, you need to use iterative methods (like conjugate gradient), and for those you have to have a very firm understanding of everything. The gap between direct and indirect methods is huge.

The good guys pay me money for insurance, I promise to release my team of ruthless lawyers on anyone that sue them, the patent trolls see my scary lawyers and don't sue. Doesn't that work?

But "the placebo" is not a static thing. If we designed experiments differently, the placebo effect might be able to be reduced (say to 3%), in which case drug A would now beat it.

Well... But since placebos can't be used therapeutically, and Drug A can, doesn't that make it better than nothing?

Woah there tiger-- where do you get that I am suggesting using the placebo as a therapeutic invention!? (If it makes any difference, I meant to type "due to" rather than "do to"). I am suggesting that the following situation might exist:

1) Drug A is 5% effective

2) Placebo is 10% effective

3) No one ever gets the benefits of drug A, since its benefits are masked during trial.

You could certainly buy one and use it with a non-contract pay as you go kind of plan. To use voice, of course, you would need some sim card. Or do you want to just use it with wifi, like an iPod touch?

It is really so ridiculous to think that there is some drug such that:

* The drug provides a modest health improvement.

* That health improvement happens to be smaller than the health improvement do to the placebo effect?

I don't know what the typical magnitude of the placebo effect is. If it is substantial, then actually working, actually helpful drugs might be being denied to people. The very mechanisms for the placebo effect he hypothesizes (e.g. people taking care of their health when on a study) certainly seem like they might vary with our culture...

Bourbaki is a name used by a group of french mathematicians to publish lots of important but dry, abstract and very very rigorous math.

http://en.wikipedia.org/wiki/Nicolas_Bourbaki

Bourbaki's style is not appealing to a lot of people, since it doesn't focus on applications, doesn't use pictures, etc.