It would be a fun exercise to test it with small language models- seeing how this works. If the results are good, there could be a good path to drive mass adoption of the concept!