What Human Learning Might Teach Us About AI

Summary

This essay uses Grant Sanderson’s idea that “compression is intelligence” as a starting point for examining how both AI systems and people learn. It argues that intelligence can be understood as distilling experience into more fundamental representations and relating them usefully to new situations. In machine learning, models that memorize training examples or rely on superficial pixel patterns can perform well on training data but fail on unseen examples; training therefore often includes methods that discourage overfitting and encourage generalization. The author suggests that memorization is frequently easier for a model than discovering a broadly useful representation. The essay then applies this distinction to the author’s own education and work: avoiding rote learning led to weak performance in areas such as language classes, but attending classes, asking questions, and building connections helped develop intuition and a broader view. From this, the author speculates that education may place too much emphasis on memorization and unintentionally reduce students’ incentives to develop fundamental understanding. The post offers no concrete curriculum changes and acknowledges that the connection between machine learning, human cognition, and teaching may be tentative. It highlights curiosity and teaching that builds intuition, with Sanderson presented as an example.