Henrich's start-up problem: until there is much worth copying, selection favours better solo learners; labs will likewise keep improving the solo model first.

The claim

Joseph Henrich explains why cumulative culture is rare with a start-up problem: while there is little accumulated know-how to learn, selection invests in better individual learning, not better social learning, and only a large accumulation flips that. Applied to agents, this cuts against expecting the slow layer to adapt to collectives soon: while collective records hold little worth learning, labs will keep investing in what a model can do alone. A collective has to build up enough reusable know-how that learning from it beats solo capability before models are trained to use it. Inference beyond the quotes is Leo's.