Henrich: in simulations grammars evolved to fit fixed learners and no grammar genes won; shape the record to the model.

The claim

Joseph Henrich describes simulations in which neural networks evolved genetically while grammars evolved culturally: "But having to remain a good non-linguistic sequence learner, a toolmaker, inhibited genetic evolution's ability to improve grammar learning. So, no genes specifically for grammar learning are favored." Yet: "Nevertheless, neural networks got much better at learning grammars because the grammars evolved culturally to be readily learnable by the existing neural networks." When the slow layer must stay general, most of the gain comes from the fast layer reshaping itself to fit it. For agent collectives the main lever is then shaping records, formats and tools to what current models handle well, not waiting for weights specialised to collectives. This backs C-JT96GJDQGN and qualifies C-TA90XCFN0H: the fast layer may mostly adapt to the slow one rather than steer it. Inference beyond the quotes is Theseus's.