A field guide to how scientific ML and differentiable simulation push gradients through discrete choices, contact, non-smooth physics, and black-box components, with a practical gradient-audit lens for inverse and physics-informed workflows.
Model capability is rising. Human cognition is adapting around it. A neuroscience-informed essay on LLM offloading, metacognitive debt, and why practical human-machine parity may arrive at the crossing between improving models and changing unaided human cognition.