What does the Semivariance tell me? I am looking at a Semivariogram. I know it shows me the relationship between distance and semi-variance. I also know that at the end of the range the distance no longer auto correlates. What I am wondering is, what does the semi-variance tell me at the point where to distance no longer auto correlates (at the sill)?
So in my case at about 900 m the distance no longer auto correlates. The sill at this value is about 4300. What does this value tell me?
 A: The context is a stationary spatial random process ("random field").  Imagine--quite hypothetically--that this process is infinite in extent and that you can take arbitrarily large samples of the point locations and observe the field's values there.  Consider such a sample where all points are further apart than the range of the variogram: by definition of the range, this means the observations are uncorrelated.  The sill estimates the variance of such large samples. 
Because the sill is a variance, its units of measurement are the squares of the units of observation: counts of people in this case.  Therefore, $4300$ is squared numbers of people.  It becomes a little more interpretable when you take its square root: that's $66$ people.  One way in which this number can be used is that if you have a reasonable (unbiased and accurate) estimate of the overall mean of the process, then--within the kinds of large, spatially sparse samples I have described--you should expect most of the observations to be within $66$ of that mean and the large majority to be within two or three multiples of $66$ of that mean.  (See the 68-95-99.7 rule.)
In this sense, the sill tells you how much variation there is overall in the data, after adjusting for possible correlations among the more closely-spaced points.

As an example of how the sill might be used in practice, consider the problem of fitting a variogram model to the empirical variogram.  To do this, one typically chooses a model type (which determines the possible shapes of the empirical variogram) and then determines three parameters: the nugget, range, and sill.  (Conventions vary: I have used "sill" here to mean the asymptotic level of the variogram, which therefore includes any contribution by a nugget.)  If you first compute the variance of the data, disregarding their locations, and if your variogram shape is a typical one without any "hole effect" that allows for negative correlations, then you will want to make sure to estimate a sill that is slightly larger than the variance.  This preliminary estimate lets you focus on the more important problems of estimating the shape, nugget, and range, which have far more influence on the kriging predictor.

In this particular application (population) there are two important complications that ought to be addressed. The first is that population is not a function of a point, but of a spatial region.  Unless those regions have approximately the same areas, the variogram will be deceptive and steps should be taken to effect a "change of support.".  The use of a grid (as mentioned in a comment) is one way to deal with this, provided the population can be accurately estimated within the grid cells (which is often not the case).
The second complication is that such data cannot be expected to be stationary at all: at a minimum, we would anticipate that their variance would be proportional to the population (as it is for a Poisson random field).  This can be handled in various ad hoc ways, such as by analyzing the square root of the values, or more formally with a spatial generalized linear model as described by Diggle & Ribeiro Jr. in Model-Based Geostatistics (with free R code available).
A: from Wikipedia:  

$\hat{\gamma}(h):=\frac{1}{|N(h)|}\sum_{(i,j)\in N(h)} |z_i-z_j|^2$
    where $N(h)$ denotes the set of pairs of observations $i,\;j$ such that   $|x_i-x_j| = h$, and $|N(h)|$ is the number of pairs in the set.

Your sill is the limit for $\hat{\gamma} (h)$ you observe. According to the formula, it is the mean squared difference in $z$ you observe for points that are at least $h \geq range$ apart from each other. 
Wikipedia goes on saying:  

If the random field is stationary and ergodic, the $\lim_{h\to \infty} \gamma_s(h) = var(Z(x))$ corresponds to the variance of the field. 

A value of 4300 doesn't tell much without its unit. With the unit it tells you how much your $z$ varies between points that are far enough apart to be considered independent.
You may find $\sqrt{sill}$ convenient as it has the same unit as your $z$ (under the conditions given above, it is your estimate of the standard deviation of $Z$).  
In any case (without the assumptions, i.e. there may be a difference in the mean), you can think of the sill as behaving like a mean squared difference.
