Summary
Yu Su, professor at The Ohio State University and COO of NeoCognition, opens with a deceptively simple question at the AI Engineer conference: why are AI agents so capable at coding and so brittle at nearly everything else? His answer is a conceptual distinction between intelligence — the ability to reason given context — and expertise — the accumulated world model that tells you which context matters in the first place.
Su argues that modern LLM-based agents finally unify multi-sensory input with symbolic reasoning in a way no prior AI architecture achieved, which is why they found their first mass market in coding. Code is, as he puts it, a language-native world: symbolic, structured, and with built-in reward signals. But most of enterprise and everyday digital work consists of what he calls microworlds — idiosyncratic domains with local physics, unique constraints, and dynamic configurations that no single monolithic model can compress into a static representation. He cites Anthropic’s revenue growth from negligible to roughly $60 billion annualized run rate as evidence of coding’s privileged position, and references Andrew Ng’s observation that we may be in a decade of agents rather than a year.
The talk frames continual learning as the bridge from intelligence to expertise, defining it as adaptive compression of experience into reusable structures for future behavior. Su breaks down what that compression can look like — vector embeddings, symbolic indexing, parameter distillation, reinforcement from environment feedback — and argues that without this, agents will remain token-inefficient and brittle outside their training distribution.
📺 Source: AI Engineer · Published August 12, 2026
🏷️ Format: Deep Dive
