The process used to generate a valid computation could itself be a valuable compression/training process of some sort. Characteristics of the final computation might be useful as feedback. I guess generalizing is just taking your listed use cases and seeing the computation in question as either up or downstream from the thing we really care about.
The process used to generate a valid computation could itself be a valuable compression/training process of some sort. Characteristics of the final computation might be useful as feedback. I guess generalizing is just taking your listed use cases and seeing the computation in question as either up or downstream from the thing we really care about.