amulet VS accelerate

One possibility would be to do something like the now-defunct Amulet language (an ML that compiled to Lua) did. You'd get accurate compiler errors, but also an additional error number that corresponds to a more beginner-friendly explanation of common reasons for the error, pulled from here, that could be accessed with amc explain 1234 on the command line or :explain 1234 in a REPL. So for example, if you tried the same kind of error as OP tried, like 1.1 + 1, you'd get this:

Kind of a side note, but I wasn't a big fan of moonscript. I don't really like the Python-like approach to syntax and it just didn't suit me. F# or Haskell handles it better, and OCaml manages to look similar without actually needing whitespace sensitivity; I think they all do it way better than Python or Python-inspired languages. I really liked the idea of of the ML-style Amulet because of this, but it's a dead project now apparently. :(

Someone has opened a PR for accelerate here https://github.com/AccelerateHS/accelerate/pull/525 (sadly seems not actively maintained at the moment, but that can always change if people care enough). I agree for an executable you should freeze your dependencies and compiler version, and using 8.10 is fine. Although there are tons of improvements in 9.2+

Backprop is a neat library. However, I guess its use case is if you actually don't want to go for anything standard like Torch or TF (perhaps for research?) For instance, if I were to use something like Accelerate for GPU acceleration, or some other computation-oriented library, then I would mix it with Backprop. Previously, I have benefited from Backprop in a ConvNet tutorial and I liked it.

Wait, really? Here's a conversation I had with him: https://github.com/AccelerateHS/accelerate/discussions/528

I know Accelerate is being developed at Utrecht University in the Netherlands. You can look at publications by Trevor McDonell to get a taste of what they are doing.

I’m personally a big fan of http://www.acceleratehs.org / https://github.com/AccelerateHS/accelerate-llvm

Looking for a few projects that make use of it, I found accelerate, hawk, polysemy and pretty-simple, so I'll be interested to poke around in their code and see how they have things set up.

There's accelerate for GPU computing and hmatrix for bindings to BLAS and LAPACK.

I'm toying with the idea of replacing vector values with vector generators, where e.g. v1 + v2 is not evaluated to a new vector, but to a vector program. This is similar to the approaches of Accelerate and TensorFlow. On the flip side, I don't think I could get rid of the overhead, and I expect much smaller computation loads than aforementioned libraries, so overheads could be very significant. The added benefit of using vector generators is that the generator could not only be evaluated, but also be turned into a Latex formula.