Adam Marblestone explores the fundamental differences between current AI, particularly LLMs, and the human brain, focusing on how the brain's complex, evolutionarily-designed loss functions and a 'Steering Subsystem' enable efficient learning and generalization. The discussion delves into concepts like omnidirectional inference, amortized versus real inference, the compact information transfer via the genome, and the distinct roles of biological vs. digital hardware. Marblestone also highlights the potential of connectomics to reverse-engineer brain algorithms and the transformative impact of formal math verification (like Lean) on both mathematics and provably secure software, advocating for scalable infrastructure to address critical gaps across scientific fields.