Akshay Shah recently published a correlation matrix between popular languages on GitHub. As commenter “conjugateprior” observed, however,

I’d guess your correlation routine only works with complete cases a.k.a ‘listwise deletion’. That means a bunch of your correlations are based on very small numbers of cases and will be super noisy. For example, the R to ActionScript correlation is 0.49 (and I can replicate it in R) but it depends on exactly 6 Github users. And the reason there are so few complete cases is that people

don’ttend to use these languages together.

Sorry, to be a bit clearer: In your data preparation (and for all I know in the original data) when user does not use a particular language you represent it as missing data, which the correlation routine duly throws out. In fact these should be zeros, which makes a huge difference. If you put the zeros back in the R / ActionScript columns then the correlation is -0.0038. Pretty much where we might expect to be.

With minor revisions to the code, I included those zeros and regenerated the plot: