Moravec's Paradox (angadh.com)

At a public discussion Chelsea Finn revisited Moravec’s Paradox—the idea that tasks humans find trivial (sensorimotor skills like folding laundry) are hard for robots, while tasks that seem hard for humans (large-scale arithmetic, Go) are easy for machines. She argued for a new class of “learned simulators” that reconstruct physics purely from real-world data rather than first-principles models. As a dynamicist, the writer notes this could accelerate model-free, data-driven robotics and perception but warns that learned simulators may suffer from noisy sensors, lose interpretable abstractions, and conflate signal with noise—undermining generalization, control, and creative problem solving. Complementing this, Michael Frank described computational models of infant cognition that frame senses as low-level capabilities and deliberation as higher-level processes mediated by memory and language. For AI/ML, this underscores two technical fronts: improving low-level, embodied perception/control (to close Moravec’s gap) and building higher-level, personalized cognitive models. The debate highlights trade-offs between data-driven realism and principled, interpretable models—implications for robotics, simulator fidelity, robustness to sensor noise, and how AI interfaces personalize to individual human thought and behavior.