You provided an opinion, but no reasoning or evidence. Could you help us by explaining how you came to your view?
In order for a newcomer to beat the incumbents, it would have to be an order of magnitude better in some way -- perhaps ergonomics, perhaps performance, and it would have to do so in some way that the incumbent couldn't just conceptually backport. That's not to say that Haskell or FP won't eventually find usage in these domains or is theoretically unsuited to it. But, I do think that from a practical perspective, it's a risky choice.
And I think you're overstating the difficulty of even a top-to-bottom new system. Re-implementing deep learning would be much, much easier than reinventing deep learning. Still probably not worth doing! But your estimate sounds about right for the latter, not the former.
As for Cabal hell, doesn't Stack [1] help with that? To be honest, I'm unhappy with dependency managers in most languages: Maven, the things Golang uses, or Python, or NPM (shudder!), etc. This really doesn't seem like a solved problem :(
That doesn't mean dependency management is always easy. It's a hard problem in any language. But you won't find yourself in an inconsistent state that requires blowing away your installation if you use one of those three approaches.