Frontier models can reason through unfamiliar problems, but expertise comes from accumulated competence in a particular environment.
2
Coding agents work well because code is symbolic, structured, and tested, while ordinary digital work spans millions of changing micro worlds.
3
Continual learning could turn bounded intelligence into unbounded expertise by compressing experience into structures that improve future behavior.
Summary
Yu Su separates intelligence from expertise. Intelligence is the ability to reason through an unfamiliar problem using the context, tools, and instructions provided. Expertise is accumulated, situated competence that brings the right context, recognizes deeper patterns, handles exceptions, and knows when work is good enough. He argues that coding is an unusually favorable environment for agents because code is symbolic, structured, and evaluated by tests. Everyday digital work is harder because each company, profession, and software setup creates a different micro world with local constraints and dynamics. Continual learning is the bridge between the two capabilities. Su defines it as adaptive compression of experience into reusable structures for future behavior. He proposes that raw intelligence and expertise are largely independent, creating the possibility of unbounded expertise from bounded intelligence. The open problems include measuring expertise, balancing reliability with plasticity, combining parametric and nonparametric learning, and using specialized agents to generate new private-world training data.
Multimodal language agents joined perception, reasoning, and communication
Su traces agents from expert systems and logical agents in the 1960s through 1980s to deep reinforcement learning agents in the 2010s. Earlier systems captured narrow parts of intelligence, such as logical reasoning or perception in one modality. Multimodal large language models changed this by encoding multiple sensory inputs into a shared neural representation that supports symbolic reasoning and communication. Su says this combination was previously unique to humans. It increased the expressiveness, reasoning ability, and adaptability of machine agents.
Coding is a favorable environment because its world is already symbolic
Language agents found an early mass market in coding because software is represented in a structured, language-native form. Code, surrounding information, and changes are recorded symbolically. Tests provide rewards and evaluation in the same symbolic world. Su connects this favorable setup to Anthropic's rapid revenue growth, which he says was largely driven by coding and related productivity capabilities. When agents leave code, they face computer-use tasks and enterprise workflows where errors remain brittle and difficult to correct.
Everyday digital work is made of millions of local micro worlds
Su describes modern society as a collection of millions of micro worlds. Each domain, profession, and company has different structures, constraints, affordances, and dynamics. Even companies using the same software configure it differently. These local rules are too heterogeneous and dynamic for one monolithic model to compress into a single static representation. Agents therefore need to keep learning on the job and acquire specialized expertise for each environment.
Intelligence reasons from supplied context while expertise brings the right context
Su defines intelligence as the capacity to reason through unfamiliar problems from available context. A frontier model can receive statements, context, and tools, then work through a problem it has not seen before. Expertise is accumulated and situated competence that produces reliable, efficient, judged performance in a particular domain. An expert does not only know more facts. An expert recognizes the deeper structure of a bug report or sees that scheduling a meeting is a constraint optimization involving authority, priority, urgency, and calendar availability.
Expertise compresses the search space through conditional rules and judgment
Experts know when rules apply and when exceptions justify bending them. They bring relevant context to a problem instead of processing every possible path. Su contrasts intelligence, which can expand a search by trying many parallel approaches, with expertise, which narrows the search through learned shortcuts. Expertise also includes taste and judgment, including knowing what counts as high quality and when the result is good enough. Together these capabilities form a world model of a particular environment.
Continual learning is adaptive compression of experience
Su defines continual learning as "adaptive compression of the experience into reusable structures for future behavior." The experience may be episodes, semantic facts, procedures, or feedback from people and environments. Compression can involve vectors, symbolic indexes, model parameters, or reinforcement learning. It must adapt because what has already been learned affects how later experience is compressed. The resulting structures may support recall, prediction, planning, control, or value estimation.
Raw intelligence and expertise can grow independently
Su presents a figure with raw intelligence on one axis and expertise on the other, arguing that the dimensions are largely orthogonal. Scaling a model without continual learning produces what he calls "the world's smartest novice": a system that can attack each new problem but accumulates no expertise. Different continual-learning methods determine how quickly expertise grows as intelligence increases. This framing separates making a model better at solving the current problem from helping it become better at future problems.
The research goal is unbounded expertise from bounded intelligence
Su asks whether a continual-learning method could make expertise keep growing after raw intelligence passes a reasonable threshold. In that case, stronger base models might not be the main requirement. The harder work would be building methods that let an agent specialize through experience. He lists practical questions around measuring expertise in each environment, balancing reliability with plasticity, combining parametric and nonparametric learning, and transferring lessons from specialized agents back into a general model.
Specialized agents could create a new source of private-world data
Su says public training data for large language models is becoming exhausted, while private company environments contain another large source of experience. If specialized agents can learn in place, their learning could feed back into general models. He connects this to making expertise more available through personal health, financial, and tutoring support, as well as company-specific learning loops that build institutional memory. He also argues that lower expertise-related friction could make more kinds of work economically viable.
"What if we can come up with a continual learning algorithm such that given, once the raw intelligence has crossed a certain threshold, we don't need stronger intelligence anymore?"14:32
Who should watch
You are building coding agents and want to understand why their performance does not transfer cleanly to computer use or enterprise workflows.
Your system needs to learn company-specific procedures, constraints, and judgment instead of starting from the same model state on every task.
You are researching continual learning and need a framework for thinking about experience, memory structures, reliability, and specialization.
