I'm going to give the same Haskell hot-take I gave on Twitter while at NeurIPS in December: the problem with Haskell is that you can port the awesome parts of pure, compositional abstractions over to Python, far more quickly than you can port really good domain-specific libraries and applications over to Haskell. My experience was that porting some nice ADT-based machinery to Python, using an available ADT library, took roughly one weekend, whereas porting the major domain-specific package I needed from Python to Haskell would take months to years.

I love Haskell. I prefer to use Haskell when I can. Haskell is a language in which I can Achieve Enlightenment. But apparently, God help me if I need to do some heavy graph processing or CUDA numerics in Haskell.