The guy is wrong when he states that a CSRNG was "completely unnecessary". He even confirms that when he says that you need an "unpredictable seed".
What means an unpredictable seed?
When I give you a reasonable amount of pseudo-random numbers derived from that seed, you shall not be able to compute the next number. Simple.
From a mathematical point of view the security of such generators can be proven by a simple game.
I send you a sequence of numbers and you have to decide if they are completely random or derived by a function (F). If your average success rate is better than a half (+/- a neligible amount) you win.
You will find that your chances to win depend (at least) on the amount of input the function gets. If I used only a few bits of input than you can surely distinguish between complete rendomness and function output much better than a half.
Up to this point there is no way around a good seed. What about the magic throw-away?
Does throwing away some output make it more secure?
I find the statement to "throw away a large and unpredictable amount" somehow funny. What they do here is to skip portions of the function's output. But how many Bytes exactly? He states "a large and unpredictable amount".
This basically means that you modify the function from above.
However, you are still at the point where you started. You play the game and try to distinguish between complete randomness and output of a new function (F').
After all there is an algorithm running and it has to decide exactly how many Bytes to skip. So this value is somehow computed an can be seen as an additional part of the original function.
What you can do?
As you can see throwing away an unpredictable amount of output does not make the function more secure. In contrast if you chose to throw away a completely random amount of bytes this colud lead to a more secure algorithm.
If an algorithm outputs the bits 0101010101... and so forth your chances to win the game are quite good.
If I skip a random amount of output and send you one bit, then skip again another amount and send you one bit, and so on, what would it be?
Your chances to win were your chances to predict the amount of skipped bits.
In the end you would have to predict the randomness, and you chances to win would no longer be that good.
However, the secureness of my number generator benefits from skipping only under certain conditions, whose evaluation would go beyond this thread.
Just to get the right impression: what would the skipping bring if my function outputs the bit "1" in 99% of the time and "0" only in 1%?
You would find that even if I skip a random amount of bits, the statistical distribution is still constant.
You definitely would want real randomness in a poker game, for the seed or the skipping of output or both. But it is sufficient to use this expensive randomness as input for well defined functions that work as a CSPRNG.
Just for the record: You could create such a function that withstands all of the mentioned tests using cryptographic primitives in a way that the sequence is predictable if you know a secret key. Without knowledge of the secret key it looks so random, that your chances to win the game are roughly a half.