[timeinput]

I'm not strongly for or against (non-deterministic) GC. Deterministic GC in Rust or the (there's no real Scotsman) correctly written C++ has benefits, but often I don't care and go / java / c# / python are all fine.

I think you're really overstepping with AI is eaten by python. I can imagine an AI stack with out python llama.cpp (for inference not training... isn't completely that, but most of the core functionality is not python, and not-GCd at all), I can not imagine an AI stack with out CUDA + C++. Even the premier python tools (pytorch, vllm) would be non-functional with out these tools.

While some very common interfaces to AI require a GC'd language I think if you deleted the non-GC parts you'd be completely stuck and have years of digging your self out, but if you deleted the 'GC' parts you can end up with a usable thing in very short order.

[pjmlp]

NVidia has decided that the market demand to do everything in Python justifies the development cost of making Python fast in CUDA.

Thus now you can use PTX directly from Python, and with the new cu Tiles approach, you can write CUDA kernels in a Python subset.

Many of these tools get combined because that is what is already there, and the large majority of us don't want, or has the resources, to spend bootstrapting a whole new world.

Until there is some monetary advantage in doing so.