[smaddox]

Agree more levels seems like it could be beneficial. And another Meta paper published a day later shows how that might work: https://ai.meta.com/research/publications/large-concept-mode...