Yu Su on Agents: Intelligence + Continual Learning = Expertise (#AI)
I watched Yu Su’s talk “Intelligence + Continual Learning = Expertise” on the Memory & Continual Learning track at AI Engineer World’s Fair 2026. Su is a professor at Ohio State University; his group produced Mind2Web, SeeAct, MMMU, and HippoRAG. In April 2026, he brought NeoCognition out of stealth with a $40 million seed round to build agents that learn on the job. In twenty minutes, he tackles a question I share: why are coding agents a major success while agents for almost everything else still struggle?
The talk’s framework is straightforward. 🤖 Intelligence is the ability to reason about an unfamiliar task using the supplied context. Frontier models keep improving at this, but each episode remains isolated. 💪 Expertise is accumulated, situated competence: the ability to act reliably, efficiently, and with judgment in a particular domain. In Su’s account, these axes are almost orthogonal. Scaling intelligence alone produces “the world’s smartest beginner”: it tackles any task brilliantly, but accumulates nothing between tasks.
Why, then, did coding work? Code is a privileged, language-native world: everything is already represented symbolically and structurally, and tests provide ready-made rewards. Su calls the result a modern Moravec’s paradox: agents can handle “crown jewel” symbolic disciplines such as code and mathematics, yet struggle with everyday digital work. That work is not one world but millions of microworlds. Every profession and company—even two instances of the same software—is configured differently, with its own local physics: structures, constraints, and dynamics. It is too heterogeneous for a single static model to compress into itself.
The most interesting part is Su’s decomposition of expertise, grounded in cognitive science. Experts do not simply know more; they see differently. They recognize patterns instantly—where to dig in a huge bug report; see the deep structure of a task—scheduling a meeting is not merely finding a shared calendar slot, but constrained optimization involving authority, priorities, and urgency; understand that rules are conditional and know when they can be bent; and exercise judgment about what quality means and when something is “good enough.” An expert has effectively learned a world model of their microworld. This also explains agents’ appetite for tokens: intelligence expands the search with a hundred parallel attempts, while expertise compresses it because the shortcuts have already been learned.
The bridge between the two axes is continual learning: adaptively compressing experience into reusable structures for future behavior. Four questions define the solution space: 1️⃣ What kind of experience—episodes, facts, procedures, feedback 2️⃣ How to compress it—vectors, symbolic indexes, distillation into parameters, RL 3️⃣ What structures to put it in—adapters, graphs, skills, world models 4️⃣ How to use those structures—recall, prediction, planning, value functions At the end, Su also proposes “unbounded expertise from bounded intelligence”: if continual-learning algorithms become good enough, a stronger model may no longer be necessary beyond some intelligence threshold; experience-based learning is what should scale. The next opportunity on the scale of “the internet as a dataset” would then be experience from private microworlds, with specialized agents feeding what they learn back into general models.
What I took away: — This is a useful frame for what is already happening: memory files, skill libraries, and RL with verifiable rewards give coding agents a primitive form of continual learning. The open question is how to reproduce it where there is neither a symbolic world nor free tests. — The management implication: if Su is right, a company’s moat is not access to a model but the learning loop over its own microworlds, turning the experience of people and agents into institutional memory. — Unbounded expertise is still a hypothesis, not a result. How to measure expertise and reconcile reliability with plasticity remain open questions, and Su explicitly lists them himself. Moreover, NeoCognition sells exactly this vision, so the talk is a founder’s manifesto rather than a neutral survey.
#AI #Agents #Research #Engineering #Conference