Intelligence + Continual Learning = Expertise — Yu Su, NeoCognition
Stronger models can still behave like smart novices inside company-specific workflows. Agents need learning loops that compress experience into reusable context, procedures, and judgment without losing reliability.
Su separates intelligence—the ability to reason from supplied context—from **expertise**, accumulated competence within a specific environment. Coding agents benefit from a language-native, verifiable world; everyday digital work contains local authority, priority, and exception rules.
Build learning loops that compress episodes, feedback, facts, and procedures into **reusable structures**. The aim is to help agents select the right context and narrow the search space instead of brute-forcing each task as if it were new.
Su separates intelligence—the ability to reason from supplied context—from **expertise**, accumulated competence within a specific environment. Coding agents benefit from a language-native, verifiable world; everyday digital work contains local authority, priority, and exception rules. Build learning loops that compress episodes, feedback, facts, and procedures into **reusable structures**. The aim is to help agents select the right context and narrow the search space instead of brute-forcing each task as if it were new. Continual learning creates a stability-versus-plasticity problem: agents must absorb new experience without losing dependable behavior. The talk argues that parametric and non-parametric learning are both needed, but does not specify a proven general architecture.
This reframes agent memory from preserving context to accumulating environment-specific competence. It reinforces external reusable state and episodic-to-parametric consolidation, while adding the stability requirement that learning must not displace dependable behavior. The candidates provide partial storage, handoff, and routing mechanisms, but none resolves that general learning problem.