- numpy __array__ interface standardization (easy compatibility across a lot of numerical libraries, e.g. torch, awkward-array, zarr, jax.numpy, dask, xarray)
- the various jit compilers (e.g. numba, jax.jit, torch.compile)
- hw-accelerated functional programming style (e.g. functorch.vmap)
- pybind11 for easy af bindings to c++
- named dimensions for tensors
- pandas-compatible ecosystem (eg seaborn)
- pytrees
- pyenv
- mamba
What I really need, though, is a realtime successor to matplotlib with a much simpler interface/dsl. Haven't looked into plotly much but probably will try it soonish