The very notion of "fast and slow languages" is nonsense. A language is just an interface for a compiler, translator, or interpreter of some sort. A language is only steer wheel and pedals, not the whole car, so the whole arguments which one is faster is stupid.
In our case, AOT compilation backfired. We used (contractually had to) support older architectures and our Eigen built with meager SSE2 support couldn't possibly outrun Numpy built with AVX-512.
So we stopped rewriting. And then Numba (built on the same LLVM as clang) came up. And then not one but several AOT Python compilers. And now JIT compiler is in the standard Python.