Now the thing I don't get is why every 2016 blocks? Why not every 1 block? The next block will have a new difficulty.
The block interval is not constant. Even with the same amount of hashpower solving the network, the time interval between blocks is still varies by a huge amount. There are certain blocks that are mined within seconds of each other as the probability of mining a block that meets the difficulty depends very heavily on luck too. The distribution of the block follows the Poisson distribution and as such, it would be better to take a sample size that is much bigger before the network adjusts its difficulty.