Moravec’s Paradox
Moravec’s Paradox is the observation that high-level reasoning tasks (symbolic logic, mathematics) require relatively little computation in AI systems, while low-level sensorimotor skills (vision, mobility, physical interaction) require enormous computational resources. Named after roboticist Hans Moravec and later articulated by Raj Reddy and Rodney Brooks, the paradox highlights a fundamental asymmetry between intellectual tasks and embodied tasks in artificial intelligence.
Core Principles
- Evolutionary Efficiency: High-level reasoning is evolutionarily recent and computationally streamlined; sensorimotor skills are ancient, robust, and deeply ingrained through billions of years of biological evolution, making them deceptively difficult to replicate algorithmically.
- Symbolic vs. Sub-symbolic: Traditional AI excels at symbolic manipulation (chess, theorem proving) but struggles with sub-symbolic processing required for real-world perception and action.
- Embodied Cognition: Intelligence cannot be divorced from physical interaction with the environment; “thinking” is deeply rooted in sensory-motor systems.
Implications for AI Development
- LLM Limitations: large-language-models (LLMs) demonstrate high-level reasoning capabilities but lack inherent grounding in the physical world, exemplifying the paradox by being strong in syntax/semantics but weak in embodied understanding.
- Robotics Challenges: Achieving human-like dexterity and navigation remains a harder engineering problem than building systems that can pass a graduate-level exam.
- World Models: To bridge this gap, AI systems require internal simulations of physical reality—World Model—to predict outcomes of actions and integrate sensory data, rather than relying solely on statistical text prediction.
Related Concepts
- Symbolic AI vs. Connectionism
- Grounding Problem
- Embodied Intelligence
Recent Developments & Integration
- Yann LeCun’s World Model Thesis: Contemporary research emphasizes that overcoming the limitations highlighted by Moravec’s Paradox requires moving beyond pure language modeling.
- See: Yann LeCun’s Argument: World Models for True, Adaptive AI Beyond LLMs
- LeCun argues that true adaptive AI must incorporate world models to handle the computational complexity of sensorimotor tasks that LLMs currently ignore.
References
Yann LeCun’s Argument: World Models for True, Adaptive AI Beyond LLMs