Yu Su distinguishes intelligence as the ability to reason through unfamiliar problems from expertise, which is accumulated, context-specific competence developed through continual learning in dynamic environments. He argues that advancing AI requires focusing on continual learning to scale expertise, enabling AI agents to adapt and specialize beyond static intelligence, thereby unlocking new economic opportunities and enhanced human-AI collaboration.
In this talk, Yu Su, a professor at Ohio State and COO of NeoCognition, explores the conceptual distinction between intelligence and expertise in AI agents, particularly focusing on continual learning. He begins by highlighting the evolution of AI agents, noting that while early systems captured limited facets of human intelligence, recent advances in multi-modal large language models (LLMs) have enabled neural models to unify multi-sensory inputs with symbolic reasoning. This breakthrough has propelled AI agents into a new stage of machine intelligence, with coding emerging as their first major successful application due to its symbolic and structured nature.
However, Su points out that outside the domain of coding, AI agents struggle significantly, especially in enterprise and personal settings. He attributes this brittleness to the complexity and heterogeneity of real-world digital environments, which consist of numerous “micro worlds” with unique rules and dynamics. Unlike coding, these environments require agents to continually learn and adapt on the job to develop specialized expertise tailored to each specific context. This need for continual learning is critical because static models cannot capture the dynamic and idiosyncratic nature of these domains.
Su then distinguishes intelligence from expertise by defining intelligence as the capacity to reason through unfamiliar problems using available context, while expertise is described as accumulated, situated competence that enables reliable, efficient, and judicious action within a particular domain. Experts perceive the world differently, recognizing patterns, understanding conditional rules, and exercising judgment about when to apply or bend rules. This expertise is grounded in a deep, generalized world model of the specific environment, which contrasts with intelligence’s broader but more superficial problem-solving approach.
Central to bridging intelligence and expertise is continual learning, which Su defines as the adaptive compression of experience into reusable structures for future behavior. He emphasizes that continual learning involves multiple dimensions, including the type of experience, methods of compression, the nature of the learned structures, and their application in prediction, planning, or control. Su presents a conceptual framework where intelligence and expertise are orthogonal, and continual learning algorithms determine how effectively expertise grows with intelligence. He envisions a future where, beyond a certain intelligence threshold, continual learning can enable unbounded expertise without needing ever-increasing raw intelligence.
Finally, Su outlines open challenges and opportunities in this space, such as defining and measuring expertise, balancing reliability with plasticity, and integrating parametric and non-parametric learning methods. He highlights the potential for specialized agents to generate new data from private, domain-specific environments, fueling further generalization and improvement of AI models. Su concludes with a call to action to focus on scaling expertise, arguing that while intelligence is becoming abundant, expertise remains scarce. By enabling widespread access to expert-level AI support, he envisions a future where new types of work become economically viable and AI-human learning loops empower individuals and organizations alike.