Next time someone says "why does everyone in AI use NVidia and CUDA"? this is why. They do high quality research and almost inevitably end up releasing the cod

Next time someone says "why does everyone in AI use NVidia and CUDA"? this is why. They do high quality research and almost inevitably end up releasing the code and models. It's possible to reconstruct all that as a non-CUDA model, but when you want to use it, why would you when it's going to take months of work to get something that isn't as optimised?