AI Engineer World's Fair 2026
Intelligence + Continual Learning = Expertise
Loading slides…
Read the talk
Intelligence + Continual Learning = Expertise
Why can an agent solve a difficult coding problem yet stumble through everyday digital work? Yu Su explains how continual learning could turn general reasoning into reliable local expertise.
From a talk by Yu Su
Why are capable agents still brittle?
Why are coding agents so successful while agents doing other kinds of work remain unreliable? And why does solving a task consume so many tokens that companies start trying to curb their employees’ token-maxing habits? These are the linked problems behind Yu Su’s distinction between intelligence and expertise. Su introduces himself as an Ohio State professor and the CEO of NeoCognition, where he works on agents and continual learning. His question is whether better reasoning alone can explain—or fix—the gap between impressive individual performances and dependable everyday work.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
What language agents changed
Agents have been an ambition throughout AI’s history. Expert systems and logical agents in the 1960s through the 1980s captured aspects of logical reasoning. Deep reinforcement-learning agents in the 2010s connected perception to decisions, often within a single modality. These systems could be capable within their settings while representing only limited facets of human intelligence.
Multimodal language models changed the representational foundation. In Su’s account, they encode inputs from different senses into a unified neural representation that also supports symbolic reasoning and communication. Language agents built on that foundation can express more, reason across more kinds of information, and adapt to unfamiliar situations. He sees this combination as a new evolutionary stage of machine intelligence.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Coding’s privileged environment
Coding became the first mass market for language agents. Su reports that Anthropic’s revenue grew 400-fold in under two years to a $40 billion annualized run rate, and tentatively cites a newer $60 billion run rate. He attributes much of that growth to coding and related productivity. These are his figures; annualized run rate describes a current revenue pace, not revenue earned over a completed year.
The fit between coding and language agents is unusually good. Code already lives in a language-native world: its contents are symbolic, its history is recorded in structured forms, and its tests and rewards can also be expressed symbolically. An agent does not have to discover an entirely new representation for the environment before it can start reasoning about it. The same medium carries the task, the work product, and much of the feedback.
Leaving that environment exposes the gap. Su points to enterprise deployments and personal assistants such as OpenClaw, where agents still make brittle, seemingly silly mistakes. He invokes Andrej Karpathy’s decade-of-agents forecast, particularly its concerns about computer use and missing continual learning. Su allows that Karpathy’s views may have changed with progress in coding agents, but argues that those underlying difficulties remain. The puzzle is how a system can be so smart and so brittle at the same time.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Every company has its own local physics
Su frames this as a modern version of Moravec’s paradox: tasks that look intellectually difficult can be easier for AI than activities people perform routinely. Coding and mathematics, once treated as the crown jewels of intelligence, are areas of strength. Everyday digital work remains difficult because it demands a different collection of cognitive competencies.
Modern society is not one uniform environment. It consists of millions of micro-worlds: domains, professions, companies, and their particular ways of operating. Even two companies using the same software may configure it differently. Each environment has local physics—its own structures, constraints, affordances, and dynamics—which determine what actions are possible and what those actions mean.
Su argues that this landscape is too heterogeneous and too dynamic for a monolithic model to compress into one static representation. An agent must learn on the job, acquiring specialized expertise in the particular micro-world where it operates. General capability provides a starting point; dependable performance requires learning how this environment works.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Expertise changes what an agent sees
The distinction starts with two working definitions:
| Capacity | What it enables |
|---|---|
| Intelligence | Reasoning through unfamiliar problems from available context |
| Expertise | Reliable, efficient action with judgment in a particular domain |
For intelligence, the problem statement, context, and tools arrive together. A frontier model can reason through them even on its first encounter, but successive episodes remain largely independent. Expertise is accumulated and situated: it produces reproducibly superior performance through competence developed in that domain.
Experts do more than accumulate facts. They develop different ways of perceiving and representing problems:
- Pattern recognition: An expert reading a gigantic bug report can quickly identify the most plausible places where something went wrong.
- Deep structure: Scheduling a meeting is more than finding an empty slot shared by several calendars. It is a constrained optimization problem involving authority, priorities, and urgency.
- Conditional knowledge: Rules have preconditions. Expertise includes knowing when a rule applies and when an exception justifies bending it.
- Judgment and taste: An expert recognizes quality and knows when the result is good enough to stop.
Together, these capabilities amount to a generalized world model of the environment. It captures how that micro-world works and supports perception, reasoning, decisions, and judgment.
This changes both context selection and search. Intelligence solves a problem using the context it receives; expertise helps bring in the right context in the first place. Intelligence can also expand the search, exploring many possible approaches. Su illustrates this with a hundred parallel attempts. Expertise compresses the search space instead: learned shortcuts make the most plausible paths visible early. The efficiency gain comes from knowing where to look and what to try, rather than repeatedly reconstructing a solution from scratch.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Continual learning as adaptive compression
Continual learning is the proposed bridge from intelligence to expertise. Su defines it as “adaptive compression of experience into reusable structures for future behavior.” That definition leaves room for several implementations while making the design questions explicit.
Start with experience. An agent might learn from complete episodes, semantic facts, procedures, human feedback, or feedback from its environment. Next comes compression: embedding experience into vectors, indexing it in symbolic structures, distilling it into model parameters, or learning through reinforcement. Compression must also be adaptive. What the system has already learned should influence how it compresses subsequent experience; repeatedly applying an unchanged storage operation is not the whole idea.
The resulting reusable structures might be model parameters, adapters, vectors, graphs, skills, or world models. Their role in future behavior also varies:
- Recall: Recover facts relevant to the current task.
- Prediction: Anticipate future states of the environment.
- Planning: Use accumulated knowledge to choose a course of action.
- Control: Guide the agent’s actuation layer.
- Evaluation: Supply a value function over potential states.
This is why continual learning encompasses more than one technique. Different choices about experience, compression, representation, and use produce different systems, all attempting to turn past experience into better future behavior.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The smartest novice and escape intelligence
Imagine raw intelligence on a horizontal axis and expertise on a vertical axis. Su treats them as largely orthogonal dimensions. Scaling a model’s intelligence without giving it continual learning moves it toward what he calls the “world’s smartest novice”: a system capable of attacking almost any problem, but unable to accumulate the expertise that would spare it repeated brute-force work.
In this conceptual plot, the continual-learning algorithm determines the slope relating intelligence to acquired expertise. A weak algorithm—Su gives simple in-context learning as a possible example—produces relatively little expertise as intelligence increases. A strong algorithm produces much more. The comparison holds the time horizon and available experience fixed, so the curves express how effectively each approach converts the same opportunity to learn into competence.
The possibility Su finds most interesting is unbounded expertise from bounded intelligence. Suppose there is a threshold of raw intelligence beyond which an effective continual-learning algorithm can keep increasing expertise without requiring a more intelligent base model. He calls that hypothetical threshold escape intelligence.
The proposal is conditional on finding such an algorithm. If it holds, the strategic question changes: how much additional value comes from training larger models, and how much from improving the mechanisms that let existing models learn? Models already available might provide enough raw intelligence, with continual learning becoming the missing ingredient.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
What dependable specialization requires
The overarching research question is how an agent can enter an arbitrary domain or environment, continually specialize, and reach expert competence. The first difficulty is defining and measuring that competence. Expertise is situated, so its evaluation will likely need to be specific to the environment rather than reduced to a single general reasoning score.
The next difficulty is the reliability–plasticity trade-off. Reliable systems resist changes that could disrupt established behavior. Plastic systems change readily enough to learn. An agent operating over time needs both properties: dependable behavior today and the capacity to incorporate new experience tomorrow. Su offers humans as an existence proof that substantial adaptability can coexist with usual dependability.
There is also an architectural question: how should parametric learning, which changes model parameters, work with nonparametric learning, which retains knowledge outside those parameters? Su believes both are necessary. The unresolved issue is how to combine them so that learning in one form supports learning and behavior in the other.
Specialization could, in turn, improve generalization. Su motivates this possibility with a strong claim that public training data has been exhausted; the narrower premise needed here is that public data alone may not sustain future progress. His proposed opportunity lies in private micro-worlds. Specialized agents could learn in situ and channel that learning back into a general model, making experience across many environments a new source of training signal at internet scale.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Making more work worth doing
Su’s call to action is to scale expertise as a new dimension of AI progress. He sees intelligence becoming abundant—even suggesting that frontier models may already be smarter than average humans—while expertise remains scarce. The desired outcome is broader access to expert help: personal healthcare support, financial advice, and tutoring shaped to the person receiving it.
For companies, the corresponding opportunity is a human–AI learning loop that becomes institutional memory. Su credits this framing to Satya Nadella, recalling remarks from two weeks before the talk. Each organization could accumulate knowledge of its own work, develop a competitive moat, and retain control over its means of production rather than repeatedly supplying the same context to a system that never learns.
Abundant expertise could also change which work is economically viable. Some opportunities remain untouched because the friction of finding, coordinating, and applying expertise is too high. Lowering that friction would let more kinds of work cross the threshold of being worth doing. Su closes by identifying this as the future NeoCognition is working toward: agents whose experience compounds into competence, expanding both access to existing services and the range of work people can undertake.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Original interview transcript explaining Karpathy's agent timeline and concerns about computer use and continual learning.
Source repository for the personal AI assistant, with project documentation and setup information.
Further reading
Yu Su's companion discussion of why computer use requires contextual understanding and knowledge of idiosyncratic environments.
- From RAG to MemoryPaper
Research coauthored by Yu Su on nonparametric continual learning for language models.
A conditional forecast of limits imposed by public human-generated text, including possible paths beyond those limits.
Full transcript
Loading transcript…