There already exist such a common denominator language. For
performance reasons, it is architecture-dependent

[...]

There have been plans to move to a better common denominator, or at
least a better bridge language (C--, LLVM, ...)

Why should that be a low-level language ? Why not core-ML ?

What I see as the very first issue is the spread of the efforts between similar yet incompatible ML dialects leading to 4 weak communities (SML, OCaml, F#, Haskell) instead of a really strong one and all the related problems that come with it (fewer books, risk for industrials, work duplication, inefficient funding, lack of visibility, etc).

Example : there is an excellent whole source code optimiser ... for SML. And an award winning SMT solver ... in Caml developed in a company that invests heavily in information-centric web applications ... in F# (http://research.microsoft.com/en-us/um/redmond/projects/z3/ if you don't know Nikolaj Bjorner's Z3). Now say you want to do an application that delivers optimal electricity production plans. What language do you choose ?

Just being able to reuse the source-code between string ML dialects even after recompilation (X -> CoreML -> specific platform) would be an improvement.

Diego Olivier