LessWrong Proposes Humans More Over-Parameterized Than AI

A LessWrong post titled "Scaling Hypothesis 2: Are Humans Just More Over-Parameterized?" frames a core anomaly between deep learning and human intelligence: that humans may be more over-parameterized relative to their lifetime training data than current AI systems are relative to theirs. The essay calls this framing "Scaling Hypothesis 2" and presents it as a lens for interpreting differences in generalization, efficiency, and capability between biological and artificial neural systems.

The over-parameterization question has practical implications for scaling theory interpretation. If biological intelligence involves far more parameters per training example than deep learning models, this could help explain human-vs-model differences in few-shot generalization, robustness, and out-of-distribution performance. Researchers working on scaling laws, sample efficiency, and architecture design may find the comparative framing useful for generating testable hypotheses, even though the essay is conceptual rather than empirical.

This is a forum post on LessWrong, not peer-reviewed research. The claims are theoretical and the essay should be read as exploratory reasoning. The existing sources list in this record contains a different LessWrong post (on rationality) that does not correspond to this essay; the correct article is linked above.