Entropy, computers and 'real' randomness

I loved this explanation by JustinP on serverfault

If you want a simpler overview of the underlying issue: Some applications (such as encryption) need random numbers. You can generate random numbers using an algorithm - but although these seem random in one sense they are totally predictable in another. For instance if I give you the digits 58209749445923078164062862089986280348253421170679, they look pretty random. But if you realise they are actually digits of PI, then you would know the next one is going to be 8.

For some applications this is OK, but for other applications (especially security related ones) people want genuine unpredictable randomness - which can't be generated by an algorithm (i.e. program) since that is by definition predictable. This is a problem in that your computer essentially is a program, so how can it possibly get genuine random numbers? The answer is by measuring genuinely random events from the outside world - for example gaps between your keypresses and using these to inject genuine randomness into the otherwise predictable random number generator. The 'entropy pool' could be thought of as the store of this randomness which gets built up by the keystrokes (or whatever is being used) and drained by the generation of random numbers.