Yu Su distinguishes intelligence, the ability to reason through unfamiliar problems from available context, from expertise, accumulated competence within a particular environment. Coding offers language-native structure and explicit tests, while ordinary digital work involves many locally configured tools, conventions and changing constraints. The argument is that one broadly capable model cannot acquire every organization’s working knowledge in advance.
Yu Su describes expertise as recognizing relevant patterns, selecting the right contextContext engineering designs how relevant instructions, evidence, state and tool information are assembled for a model's task., understanding when rules apply and judging when a result is good enough. This can reduce the search needed to solve a problem rather than simply expanding it. Meeting scheduling, for example, involves priorities and authority as well as overlapping calendar slots.
Yu Su defines continual learningContinual learning lets an AI system acquire new knowledge or skills over time while retaining what it learned earlier. as adaptive compression of experience into reusable structures that shape future behavior. Experience might include episodes, facts, procedures or feedback; the resulting structures could include model parameters, adapters, vectors, graphs or skillsAn AI agent skill is a reusable package of instructions, resources, and tool guidance for performing a bounded kind of work.. What an agent has already learned should influence how it processes subsequent experience, and the stored knowledge should support decisions rather than only factual recall.
Yu Su presents the possibility of sustained expertise growth beyond a sufficient level of general intelligence as an open hypothesis. The research questions include how to measure expertise, balance adaptability with reliability and combine learning in model parameters with external memoryAI agent memory is stored information that an agent can retrieve and use across steps, sessions, or changing contexts.. Specialized agents might also produce knowledge that improves more general models, but the talk offers a conceptual research agenda rather than experimental proof of these outcomes.
Watch on YouTube
