possibly the beginning of the recursive feedback as models begin to aid in their own improvement? especially algorithmic improvements, which seems to have a lot of wide open space for gains