Lovely to see Cling progress so much (see the CUDA+jupyter example on the page), but a major challenge remains: dependencies. It used to be that you had to use

Lovely to see Cling progress so much (see the CUDA+jupyter example on the page), but a major challenge remains: dependencies. It used to be that you had to use a forked LLVM just to use Cling, but major kudos they cleaned that up. However for large C++ projects with all sorts of compiler flags, include paths, etc, it’s still a major effort to get stuff running. An alternative strategy is to wrap your C++ stuff in a library and use pybind11 for lightweight bindings and python-based interactivity. This approach means more glue code, but python has much better data science tooling. I think Cling will see the greatest opportunity in algo tuning and multi-platform development. That’s a different kind of data science— focused more on benchmarking than consumer problems. I just wish the C++ build tool json ecosystem were friendlier.