One could train a language model to predict task difficulty (and other parameters) here, which would be great for the initialisation of new cards