Why isn’t this happening more already?
It takes way more resources to train the model then to use it.