> R ... The language is just a joy to use.

My experience has been the exact opposite. R is great as an interactive statistical environment. But if you need to build data-intensive applications in it - it's still powerful but somewhat slow and also a bit of a mess: many exceptions, special cases, strange design choices and multiple ways to do the same thing. It can get the job done, but there is an unpleasant learning curve to get used to all the weirdness.

I'd second the R recommendation. It's a particularly exciting time to use R, with the proliferation of package collections like tidyverse (https://www.tidyverse.org/) and the development and release of free texts like R for Data Science (https://r4ds.had.co.nz/).

Definitely R, but IMO from the start learn it with tinyverse and Rcpp. I.e use it as a pipelining and glue language but don’t write any processing code in it. R is slow, and memory hungry.

Having said that tinyverse basically calls methods written in C and has a beautiful syntax. As Rcpp let’s you write whatever needs to go quickly in c++ then easily expose it in R.

It is still the best language for exploratory data analysis, fitting models intuitively, and visualizing data.

Being able to do these things thoroughly and quickly, has (in my experience) been the most important technical skill for getting data science right, regardless of the language production models end up getting deployed in.