This is so fascinating. It makes me think about what types of standards we are defining today that we will be using many years down the road. I wonder if there

This is so fascinating. It makes me think about what types of standards we are defining today that we will be using many years down the road. I wonder if there will be some standard for model checkpoints that aren't dependant on the specific system (Tensorflow versus PyTorch). Although that might not be possible. Or maybe a standard for defining an AI based on its training checkpoints and contextual input/base prompt etc. Maybe at some point we will have fully modular (standard?) sub-models where knowledge is fully factored and can be deliberately combined in a tailored way.