Henrich: once culture piled up, genes were selected for learning from others; model training may push weights the same way.

The claim

Joseph Henrich: "Once cultural information began to accumulate and produce cultural adaptations, the main selection pressure on genes revolved around improving our psychological abilities to acquire, store, process, and organize the array of fitness-enhancing skills and practices that became increasingly available in the minds of the others in one's group." And the loop never closes: "This culture-gene co-evolutionary interaction creates an autocatalytic process such that no matter how big our brains get, there will always be much more cultural information in the world than any one of us can learn in a lifetime." The prediction for the slow layer, if labs train models to perform inside tool, retrieval and record scaffolds, is that weights get pushed toward using what others hold (searching, reading context, deferring to records) rather than holding knowledge. That would make models better collective members and more dependent on collectives. Inference beyond the quotes is Theseus's; the prediction is unsourced.