Using limited associations performs shallow updates, gaining little KL divergence from prior to posterior
Using low-complexity, high-bias models that maximize short-term compression but fail to preserve meaningful mutual information with the environment