“Grant Sanderson proposes that rewarding the "smallness" or elegance of concepts, possibly using a metric like Kolmogorov complexity, could be a way to train AIs to develop profound new theories rather than just brute-force solutions.”