>"Write once, parallelize anywhere" -- With very little modification, the same Flow code should be able to target hadoop / Google's MapReduce, the JVM using Java Threads, C using pthreads, MPI or a GPU using CUDA. By understanding the big-Oh complexity of computations in Flow programs, The compiler will produce code that optimally partitions the workload in systems with nonuniform computing speeds and/or communication costs, e.g. CPU/GPU combinations.
That's a big promise.