An assumption with self-tuning histograms has been that they can "learn" the dataset if given enough training queries. We show that this is not the case with the current approaches. The quality of the histogram depends on the initial configuration. Starting with few good buckets can improve the efficiency of learning. Without this, the histogram is likely to stagnate, i.e. converge to a bad configuration and stop learning. We also present a probabilistic cost estimation model.