The discussion centers on the idea that the brain's superior sample efficiency stems not from its architecture but from evolved reward functions—innate 'steering subsystems' that guide learning through hardcoded motivations. These reward mechanisms allow humans to develop desires for novel, evolutionarily unprecedented goals. The brain likely relies on amortized inference, encoding probable solutions rather than raw knowledge, with evolution optimizing reward circuits over detailed structures. Model-free reinforcement learning in the basal ganglia, driven by dopamine-based prediction errors, operates alongside model-based cortical predictions, mirroring societal cultural accumulation. Despite slower hardware, biological systems offer cognitive flexibility that AI lacks. Mapping full brain connectomes could unlock reverse-engineering efforts, while integrating neural data into AI training may improve generalization. Finally, automating formal mathematics with tools like Lean and LLMs promises to transform research, verification, and access to complex problem-solving, though preserving intuition remains a challenge.