"Boy and girl fall in love, they both die."

Yes, it's lost a bit of detail and resolution, but it's significantly more compact than the original, while retaining somewhat the same recognizable message. All we need next is a computer algorithm that generates TLDR.

The result would probably look like it was written by someone with an incomplete knowledge of English (or by a machine) but might end up being quite readable.

Of course, on the other hand, you can just compress the text with LZMA :)

One form of "minimal encoding" you're describing would be to use arithmetic coding where the predictor simply assigns equal probability to each symbol actually used in the input. For example, if you encode a 1000 letter message which includes at least one instance of every letter (both upper and lower case), then you'd have 52 symbols, and your simple prediction step would simply assign 1/52 as the probability for each symbol. This will result in a constant 5.7 bits of output per symbol. So your 1000 letter message would encode at 5700 bits.

Now, suppose you discard capitalization. Now there are 26 symbols, each assigned a probability of 1/26, resulting in a constant 4.7 bits per symbol. Your 1000 letter message will now encode at 4700 bits. So you only save about 17% in this very generously constructed case.

The mistake you've made is to assume that cutting the number of symbols in half will cut the symbol length in half. In fact, reducing the number of symbols by half merely reduces the symbol length by a single bit. For example, you need 8 bits to specify between 256 possibilities, but if you cut the possibilities down to 128, you still need 7 bits.

In addition, lossy compression is usually combined with lossless compression, and a smart predictor could use context to reduce the cost of capital letters to much less than one bit per symbol (assuming that the text has a predictable structure, like human language).

http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=4126...

I've written this page, to try to explain the science behind this meme. There are elements of truth in this, but also some things which scientists studying the psychology of language (psycholinguists) know to be incorrect.

The page also contains a number of counter-examples, show that the result often is not "pflteecry uaaedlbdrsntne".

But we let machines do it all the time too. Go from UTF-8 to ASCII-7. You won't loose much in English - orientation of quotation marks, distinction among various types of dashes. Naïve and Café and El Niño and other words borrowed from other languages and commonly spelt with their native character set would also need to be spelt differently.

You could drop down to 5 bits per character, giving you only upper case letters and room for 6 punctuations. Numerals would need to be spelt out, which means you could no longer distinguish among these: 1, One, ONE, one.

You can also find lots of lossy text compression, focused on minimizing characters, on Twitter or in text messages.

I hate it when I open a document and there's non-ASCII characters cluttering it up just because someone wanted magic quote marks. If you must have oriented quotes, you should do them the ``LaTeX way,'' so you don't have to get into Unicode, which sucks [1].

[1] https://news.ycombinator.com/item?id=4835507

IMHO.

Using a compression method on a different data set is unsurprisingly ineffective.

We're not "blind" to the changes in images, but the algorithm is designed that the changes are minimal when in image form.

I wonder how much of our ability to read text like that is based on context though. Also, I'm going to assume that this only works with fairly common words (which is unique for each person). For example, I'm guessing running said compression on the works of Shakespeare wouldn't work as well, since most people struggle to make sense of them to begin with.