I think dependent types is next big step, and it also blurs the line between compilation and runtime. This will enable applications to do domain specific optimizations since types become so much more expressive. Purity allows for automatic parallelization. What Edward Kmett does in C++ here: https://www.youtube.com/watch?v=KzqNQMpRbac could be the default for all computation. On a modern computer, data parallelism is free, since SIMD instructions take just as long as any other instruction.
