AI Engineer · Published 2026-08-12

Intelligence + Continual Learning = Expertise — Yu Su, NeoCognition

privateReview: unreviewedConfidence: high
intelligenceexpertisecontinual-learningagentsmoravecs-paradoxai-agentsworld-modeltoken-efficiencycoding-agents
Open on YouTube ↗

Summary

Overview

  • Speaker: Yu Su
  • Channel: AI Engineer
  • Main topic: The distinction between intelligence and expertise in AI agents, and how continual learning bridges them.
  • Purpose: To provide a conceptual framework distinguishing intelligence from expertise, highlight the limitations of current AI agents outside of coding, and advocate for continual learning as the mechanism to build specialized expertise. Yu Su explores the conceptual distinction between raw intelligence and domain expertise in AI agents. He explains why current language agents excel at token-native worlds like coding but fail at everyday digital work due to the fragmentation of micro-worlds. By framing continual learning as adaptive compression into reusable structures, he argues that the industry must shift from scaling raw model intelligence to scaling expertise, potentially achieving 'unbound expertise from bounded intelligence'.

Topic Map

History of AI Agents and the Rise of Coding Agents

  • Explanation: Tracing AI agents from 1960s expert systems and 2010s deep RL agents to modern multimodal LLMs, and explaining why coding became their first mass market success.
  • Key claims:
    • Early AI captured limited facets of human intelligence like logical reasoning or single-modality perception.
    • Multimodal LLMs for the first time encode multi-sensory inputs into unified neural representations conducive to symbolic reasoning.
    • Coding is the ideal market for language agents because code is already a language-native, symbolic, and structured world with automated rewards and tests.
  • Examples:
    • Anthropic's revenue growing 400 times to $40 billion (or $60 billion annualized run rate) driven largely by coding productivity.
  • Terminology:
    • expert systems
    • deep RL based neural agents
    • multimodal LLMs
    • language agent
  • Why it matters: Explains why AI agents achieved sudden commercial success in software development while struggling elsewhere.

The Modern Version of Moravec's Paradox

  • Explanation: Why AI agents are successful at complex symbolic reasoning like coding and math yet fail at everyday digital work.
  • Key claims:
    • Modern society consists of millions of micro-worlds, each with its own local physics, structures, constraints, and dynamics.
    • A monolithic static model cannot compress the heterogeneous and dynamic real world.
    • Agents must continually learn on the job to acquire specialized expertise.
  • Examples:
    • Open Class agents making brittle and silly errors in computer use.
    • Andrew Karpathy's observation that it is the 'decade of agents' because they cannot reliably perform computer use without continual learning.
  • Terminology:
    • Moravec's Paradox
    • micro-worlds
    • local physics
    • computer use
  • Why it matters: Identifies the core bottleneck preventing general enterprise and personal deployment of AI agents.

Intelligence vs. Expertise

  • Explanation: Establishing clear working definitions and operational differences between raw intelligence and accumulated expertise.
  • Key claims:
    • Intelligence is the capacity to reason through unfamiliar problems from available context independently.
    • Expertise is accumulated and situated competence: acting reliably, efficiently, and with judgment to achieve superior performance in a domain.
    • Experts do not just know facts; they see the world differently, recognizing deep patterns and understanding when rules can be bent.
  • Examples:
    • An expert looking at a gigantic bug report and immediately locating likely failure points.
    • Scheduling a meeting viewed not as a calendar slot search, but as a constraint optimization problem over authority and priorities.
  • Terminology:
    • intelligence
    • expertise
    • world model
    • situated competence
  • Why it matters: Clarifies why a high-IQ model still fails at specific domain workflows without proper structural adaptation.

Continual Learning as the Bridge

  • Explanation: Defining continual learning and exploring its role in transforming intelligence into unbound expertise.
  • Key claims:
    • Continual learning is the adaptive compression of experience into reusable structures for future behavior.
    • Without continual learning, scaling model intelligence results in the 'world's smartest novice' that brute-forces problems.
    • Strong continual learning can lead to 'unbound expertise from bounded intelligence' (escape intelligence).
  • Examples:
    • Scaling raw intelligence on the X-axis versus growing expertise via continual learning algorithms on the Y-axis.
  • Terminology:
    • continual learning
    • adaptive compression
    • world's smartest novice
    • unbound expertise from bounded intelligence
    • escape intelligence
  • Why it matters: Provides a strategic roadmap for whether the industry needs endlessly larger models or better in-situ learning algorithms.

Key Points

The Mismatch Between Code and the Real World

  • Explanation: Code is structured and symbolic, making it easy for language agents, whereas real-world enterprise environments consist of chaotic micro-worlds with unique constraints.
  • Evidence: Anthropic's massive revenue growth driven by coding versus brittle performance in open-ended computer use.
  • Practical implication: Agents deployed in enterprise or personal settings require adaptive mechanisms beyond pre-training to handle local domain physics.

Intelligence Expands Search, Expertise Compresses It

  • Explanation: Raw intelligence brute-forces problems by exploring numerous parallel paths, whereas expertise relies on structural shortcuts to instantly narrow down the search space.
  • Evidence: Comparison of how novices versus experts approach complex bug reports or scheduling tasks.
  • Practical implication: Expertise dramatically increases token efficiency and reliability compared to relying solely on raw LLM reasoning.

The Four Elements of Continual Learning

  • Explanation: Continual learning requires defining the type of experience, the compression method, the adaptive nature of learning, and the reusable structure utilized.
  • Evidence: Deconstruction of continual learning into experience, compression, structural adaptation, and downstream application.
  • Practical implication: Standardizing definitions helps resolve confusion across different research directions in continual learning.

Frameworks, Models & Processes

Intelligence vs. Expertise Matrix

  • How it works: Plots raw intelligence on the X-axis and accumulated expertise on the Y-axis, showing that continual learning determines the slope of expertise accumulation.
  • Components:
    • Raw Intelligence (capacity to reason from context)
    • Continual Learning Algorithm (adaptive compression of experience)
    • Expertise (situated competence and compressed search space)
  • When to use: When evaluating whether to invest in scaling base model parameters versus developing in-situ continual learning mechanisms.

Examples & Case Studies

Anthropic's annualized run rate grew to $40-$60 billion in under two years.

  • Illustrates: Coding is an ideal, language-native market where language agents achieve immediate mass-market product-market fit.
  • Lesson: Agents succeed exponentially faster when operating within symbolic, structured environments.

Scheduling a meeting handled by a novice versus an expert.

  • Illustrates: Experts view tasks through deep structures and constraints rather than surface-level execution.
  • Lesson: Expertise involves understanding priorities, authority, and exceptions, not just following static rules.

Actionable Takeaways

  • Immediate:
    • Distinguish between token-efficient symbolic tasks (like coding) and heterogeneous real-world workflows when deploying AI agents.
    • Recognize that prompt-based reasoning alone will lead to brittle behavior in specialized enterprise domains.
  • Strategic:
    • Shift organizational focus from acquiring ever-larger frontier models to building robust continual learning loops.
    • Treat private domain-specific data and micro-worlds as the next internet-scale data opportunity for specialized agents.
  • Questions to investigate:
    • How do we effectively measure and define expertise in specific enterprise environments?
    • How can we reconcile the trade-off between stability (reliability) and plasticity in continual learning systems?
    • How can parametric and non-parametric learning be successfully synergized?

Claims Worth Verifying

  • Anthropic's revenue grew 400 times to 40 to 60 billion annualized run rate in under two years. (empirical)
  • Continual learning can produce unbound expertise from bounded intelligence once a threshold is crossed. (theoretical)

Notable Quotes

"we are actually witnessing a modern version of the Moravec's Paradox." "intelligence is the capacity to reason through unfamiliar problems from available context." "expertise is accumulated and situated competence: the ability to act reliably, efficiently, and with judgment to achieve reproducibly superior performance in a particular domain." "continual learning is adaptive compression of experience into reusable structures for future behavior."

Compressed Summary

  • AI agents excel at coding because code is a symbolic, language-native world.
  • Real-world workflows consist of millions of micro-worlds with unique local physics that monolithic models cannot compress.
  • Intelligence reasons through unfamiliar contexts; expertise uses accumulated competence and deep pattern recognition.
  • Continual learning adapts experience into reusable structures, bridging the gap between raw intelligence and expert performance.
  • Scaling expertise through continual learning offers an alternative path to brute-force model scaling.
  • Keywords: intelligence, expertise, continual-learning, agents, moravecs-paradox
  • Core insight: While raw intelligence provides the capacity to reason through unfamiliar contexts, continual learning is required to compress real-world experience into specialized expertise.

Core insights

5
Mental Modelhigh noveltystrong evidence

Modern AI agents exhibit a modern Moravec's paradox: they excel at symbolic reasoning tasks like coding and math but fail at everyday digital work because the real world is millions of micro-worlds, each with unique local physics (structures, constraints, affordances, dynamics). A monolithic static model cannot compress such a heterogeneous, dynamic world, so agents must continually learn on the job and acquire specialized expertise.

Why it matters

Explains why general-purpose LLM agents are brittle outside coding and implies that scaling model size alone will not solve enterprise deployment; runtime adaptation to domain-specific structure is necessary.

Generalization

Any agent operating across many heterogeneous domains needs a continual learning layer that acquires per-domain structure, rather than relying only on a single static world model.

Modern society is not one world, but millions of micro-worlds.
Open source video
The world is too heterogeneous and dynamic for any monolithic model to compress into one static representation.
Open source video
agents must continually learn on the job and acquire specialized expertise
Open source video
Mechanismhigh noveltystrong evidence

Experts do not simply know more facts; they see the world differently through domain-tuned pattern recognition, deep structure (e.g., meeting scheduling as constraint optimization over authority, priorities, urgency), conditional rules with exceptions, and judgment about when to stop. This constitutes a world model of their environment.

Why it matters

Capturing expertise in agents requires encoding domain-specific world models—perception shortcuts, deep structural constraints, conditionals, and stopping criteria—not just injecting more knowledge or longer reasoning.

Generalization

A reusable domain expertise layer should model local physics: structures, constraints, affordances, and dynamics, and provide guidance for perception, reasoning, decision, and judgment.

experts don't just know more facts. They actually see the world differently
Open source video
experts effectively has, have built a world model of their environments
Open source video
when you are scheduling a meeting. You know that it's not just like finding the shared slots on everyone's calendar. It's actually a constraint optimization problem over everyone's authority, the priorities, the urgency, and everything
Open source video
Architecturehigh noveltystrong evidence

Intelligence is about reasoning through a given context, but expertise is about bringing the right context. This distinction implies context acquisition/curation is a separate competency from reasoning.

Why it matters

Agent architectures that treat context as externally supplied or a simple retrieval step miss a core expertise function; context curation should be an explicit, domain-adapted subsystem.

Generalization

Separate the context provisioning problem (what to attend to, which constraints matter) from the reasoning problem (how to solve given context); evaluating and building these abilities independently may improve performance and efficiency.

intelligence is about hey, when we have the context, uh, how to solve the problem through the context? But expertise actually will bring you the right context
Open source video
Mental Modelmedium noveltystrong evidence

Intelligence and expertise are distinct capacities: intelligence is the ability to reason through unfamiliar problems from available context, with each episode independent; expertise is accumulated, situated competence—acting reliably, efficiently, and with judgment to achieve reproducibly superior performance in a domain.

Why it matters

Current agents are smart novices: they can reason from supplied context but lack situated competence, which explains token inefficiency and brittleness in real workflows.

Generalization

Agent design should treat raw reasoning and domain competence as separate axes; improving domain performance requires accumulating expertise, not just increasing model intelligence.

intelligence is the capacity to reason through unfamiliar problems from available context
Open source video
Expertise is accumulated and situated competence: the ability to act reliably, efficiently, and with judgment to achieve reproducibly superior performance in a particular domain
Open source video
Every episode is more or less independent from each other, here
Open source video
Empirical Resultmedium noveltystrong evidence

Coding became the first mass market for language agents because code is already a language-native, symbolic, structured world with automated symbolic rewards and tests. This privileged environment does not generalize to everyday digital work, which lacks such clean structure and feedback.

Why it matters

Evaluating agents on coding may overstate general competence; enterprise deployments require adapting to domains whose local physics are not already symbolically represented or reward-dense.

Generalization

The ease of agent deployment in a domain is predicted by how much of the domain's structure and feedback is already symbolic and machine-readable; domains with implicit or heterogeneous structure require additional expertise accumulation.

coding is the really the ideal market for these language agents, because code is already a language-native world
Open source video
Everything is already represented symbolically and like uh, recorded in a very structured way. And you get your rewards, you get your uh, like tests all in place in symbolic ways
Open source video
In just under two years, their revenue has grown 400 times uh to uh 40 billion
Open source video

Deep dives

4

Operationalizing Continual Learning for Micro-World Expertise Accumulation

Research question

How can deployed AI agents continually learn from on-the-job experience in a micro-world to accumulate specialized expertise without catastrophic forgetting or unbounded state growth?

Why

Addressing this is necessary to move agents beyond coding to everyday digital work; current general-purpose LLM agents are brittle and token-inefficient because they rely on a static monolithic model instead of accumulating domain-specific competence.

Modern society is not one world, but millions of micro-worlds.
Open source video
The world is too heterogeneous and dynamic for any monolithic model to compress into one static representation.
Open source video
agents must continually learn on the job and acquire specialized expertise
Open source video
Source video

Explicit World-Model Representations for Encoding Local Physics

Research question

What level of explicit representation is needed for agents to encode domain-specific local physics (structures, constraints, affordances, dynamics) and guide perception, reasoning, decision, and judgment?

Why

Experts do not just know more facts; they see the world differently through structured mental models. Without a formal world-model representation, agents cannot capture the deep structure (e.g., meeting scheduling as constraint optimization) that separates expertise from novices.

experts don't just know more facts. They actually see the world differently
Open source video
when you are scheduling a meeting. You know that it's not just like finding the shared slots on everyone's calendar. It's actually a constraint optimization problem over everyone's authority, the priorities, the urgency, and everything
Open source video
experts effectively has, have built a world model of their environments
Open source video
Source video

Context Curation as a Distinct Expertise Subsystem

Research question

How should context acquisition and curation be implemented and evaluated separately from reasoning, so that agents bring the right context before solving a problem?

Why

Intelligence is reasoning from given context; expertise is bringing the right context. If context is treated as a generic retrieval step, agents will reason over incomplete or irrelevant information, causing errors that no amount of reasoning can correct.

intelligence is about hey, when we have the context, uh, how to solve the problem through the context? But expertise actually will bring you the right context
Open source video
Source video

Metrics for Expertise Efficiency Beyond Token Count

Research question

What metrics can capture the efficiency gain from accumulated expertise over raw intelligence, beyond token count, to measure decision-space compression, reliability, and cost per successful workflow?

Why

Current agents are token inefficient because they reason from scratch each episode; expertise should reduce search space. Without metrics that measure this compression, organizations cannot justify continual learning investments or compare architectures.

Why uh the current agents are so token inefficient uh, to the degree that every company right now is like coming out and try to curb their uh, their token maxing efforts in the company.
Open source video
Source video

Article ideas

3

The Micro-World Problem: Why AGI Timelines Ignore the Real Cost of Expertise

AI agents fail outside coding because the real world is not one world but millions of micro-worlds with unique local physics, and no static model can compress them; the only path to reliable enterprise agents is continual learning on the job.

Angle

Technical strategy: argues that scaling model size and intelligence is insufficient and that architecture must shift to accumulating domain-specific expertise.

Source video

Bringing the Right Context: The Missing Half of Intelligence

Intelligence and expertise are distinct capacities—reasoning over context vs. curating context—and treating context retrieval as an afterthought is why LLM agents are brittle and token-inefficient; we need context curation as a first-class subsystem.

Angle

Conceptual architecture: reframes context engineering from a RAG add-on to an expertise function.

Source video

Coding Was an Anomaly: Why the First Mass Market Misleads AI Agent Strategy

Coding succeeded because code is already language-native, symbolic, and structured with automated rewards, a privileged environment that does not generalize; enterprise deployment requires making domain feedback and structure first-class rather than assuming a uniform symbolic interface.

Angle

Critical analysis: warns against over-generalizing coding agent success and outlines domain structuring requirements.

Source video

Project ideas

3

MicroWorld Continual Learning Harness

movement-lab

A continual learning subsystem that accumulates per-domain micro-world expertise from agent trajectories will reduce token consumption per successful workflow by at least 30% and increase success rate on non-coding digital work compared to a frozen LLM baseline.

Proof of concept

Deploy a language agent on three heterogeneous micro-worlds (e.g., enterprise meeting scheduling, CRM data entry, internal IT ticket triage) and implement a per-domain memory that stores discovered structures, constraints, and repeated conditionals from each completed and failed episode; compare token usage, success rate, and brittleness against the same base model without this memory over 100 trials per domain.

Measurement

tokens per successful workflow, task success rate, variance across repeated trials, and count of repeated mistakes prevented

Source video

CURATE: Context Curation Evaluation Suite

gatehouse

Agents equipped with an explicit context curation module that brings domain-relevant constraints (authority, priorities, urgency) before reasoning will reduce reasoning tokens by at least 25% and improve task success by at least 15% on scheduling and enterprise workflows compared to generic retrieval-augmented generation.

Proof of concept

Build a controlled meeting scheduling environment where the correct constraints are known. Implement two agent variants: one with a context curator that extracts constraints from domain world model, one with standard retrieval. Run both over identical tasks and log selected context items. Compare downstream reasoning tokens, success rate, and context precision/recall.

Measurement

context precision/recall against ground truth, tokens used by reasoning stage, success rate, and cost per success

Source video

Domain World Model Probes

beyond-evals

The degree to which an agent explicitly encodes a micro-world's local physics (structures, constraints, affordances, dynamics) predicts deployment success more strongly than general model capability, as measured by a set of domain-specific probes.

Proof of concept

Create a suite of micro-world probes across coding, scheduling, CRM, and design. For each domain, measure whether agent can answer constraint optimization questions, identify missing constraints, and stop at expert-level judgment. Correlate probe scores with actual success in deploying the agent on that domain. Optionally test a version fine-tuned on domain world model vs base model.

Measurement

correlation between probe score and task success, performance difference between base and domain-adapted agent, and false confidence rate

Source video

Architectural implications

3

The world is decomposed into millions of micro-worlds with unique local physics, and a monolithic static model cannot compress them; agents must continually learn on the job.

Before

A single pretrained LLM is expected to handle all enterprise/personal tasks without runtime adaptation.

After

Agent architecture includes a continual learning subsystem that acquires and updates per-domain micro-world expertise from on-the-job experience.

Consequence

Deployment strategy shifts from frozen model inference to maintaining evolving domain-specific expertise; infrastructure must support safe, observable runtime learning.

Source video

Experts build a world model of their environment that supports perception, reasoning, decision, and judgment; expertise brings the right context.

Before

Agent context is assembled by generic retrieval or explicit tool results, with reasoning over whatever context is present.

After

Context curation becomes an expertise-driven layer backed by domain world models, selecting and compressing the relevant constraints before reasoning.

Consequence

Need new abstractions for domain world models and context compilation; evaluating context quality becomes as important as reasoning quality.

Source video

Coding is privileged because it is already symbolic and structured with symbolic rewards; non-coding domains are heterogeneous and not pre-compressed.

Before

One agent framework/loop used across all domains with same assumptions.

After

Domain structuring and feedback design are first-class concerns; agents may need domain-specific adapters or representations that make local physics explicit.

Consequence

Architectures must allow per-domain variation in observation, action, and reward modeling rather than assuming a uniform symbolic interface.

Source video

Tradeoffs and failure modes

3

Raw intelligence vs token efficiency

Benefit

High-intelligence models can reason through unfamiliar problems from provided context, enabling broad generalization.

Cost or risk

Without accumulated expertise, agents brute-force each episode and become token inefficient; companies are already curbing token maxing.

Why uh the current agents are so token inefficient uh, to the degree that every company right now is like coming out and try to curb their uh, their token maxing efforts in the company.
Open source video
Source video

Coding-optimized agents vs real-world digital work

Benefit

Code provides a language-native symbolic world with automated tests, enabling rapid success and revenue.

Cost or risk

Agents deployed outside code exhibit brittle, silly errors in enterprise and personal settings, and struggle with computer use due to lack of continual learning.

we are running into a lot of challenges deploying these agents in enterprise settings. And the uh, also in personal settings like the uh, like Open Class uh, constantly make this uh, like uh, quite brittle and silly uh, errors.
Open source video
Source video

Static monolithic model vs heterogeneous dynamic world

Benefit

A single static model offers simplicity and general-purpose coverage.

Cost or risk

Cannot compress millions of micro-worlds with unique local physics into one representation, so agents fail outside homogeneous symbolic domains.

The world is too heterogeneous and dynamic for any monolithic model to compress into one static representation.
Open source video
Source video

Open questions

4

How can continual learning be operationalized in deployed agents to accumulate micro-world expertise without catastrophic forgetting or unbounded state growth?

Why unresolved

Speaker identifies the requirement but does not specify learning algorithms or update mechanisms.

Research direction

Continual learning methods that adapt from agent trajectories, compress experience into reusable structures, and maintain stability.

Source video

What level of explicit world-model representation is needed for agents to encode local physics (structures, constraints, affordances, dynamics) and guide perception/reasoning?

Why unresolved

Transcript defines world models conceptually but not their representation or how they are updated.

Research direction

Benchmarks or architectures for domain-specific world models, possibly hybrid neural-symbolic.

Source video

How should context curation ('bringing the right context') be implemented and evaluated separately from reasoning?

Why unresolved

The speaker contrasts the two but does not provide a mechanism or metric.

Research direction

Context selection/compression algorithms and evaluation suites that measure whether the right constraints are brought before reasoning.

Source video

What metrics can capture the efficiency gain from expertise over raw intelligence, beyond token count?

Why unresolved

Token inefficiency is observed, but no proposed measure of search-space compression or judgment quality.

Research direction

Measuring decision-space reduction, reliability, and cost per successfully completed workflow.

Source video

Key claims

7
factualVerification needed

Modern society is not one world, but millions of micro-worlds, each with unique local physics: structures, constraints, affordances, and dynamics.

Evidence

Modern society is not one world, but millions of micro-worlds. Professionals, organizations, and software systems. Each has its unique local physics: structures, constraints, affordances, and dynamics.

Question

Is there empirical evidence for distinct micro-worlds with measurable local physics as the primary barrier to agent deployment?

Source video
factualVerification needed

The world is too heterogeneous and dynamic for any monolithic model to compress into one static representation.

Evidence

The world is too heterogeneous and dynamic for any monolithic model to compress into one static representation.

Question

Can experiments show performance degradation of a static LLM across heterogeneous micro-worlds compared to a continually adapted system?

Source video
causalVerification needed

Coding is the ideal market for language agents because code is already a language-native, symbolic, structured world with symbolic rewards and tests.

Evidence

coding is the really the ideal market for these language agents, because code is already a language-native world. Everything is already represented symbolically and like uh, recorded in a very structured way. And you get your rewards, you get your uh, like tests all in place in symbolic ways.

Question

Does the presence of symbolic structure and automated tests causally explain coding agent success compared to other domains with similar complexity?

Source video
opinionVerification not requested

Intelligence is the capacity to reason through unfamiliar problems from available context, with every episode more or less independent.

Evidence

intelligence is the capacity to reason through unfamiliar problems from available context

Source video
opinionVerification not requested

Expertise is accumulated and situated competence: the ability to act reliably, efficiently, and with judgment to achieve reproducibly superior performance in a particular domain.

Evidence

Expertise is accumulated and situated competence: the ability to act reliably, efficiently, and with judgment to achieve reproducibly superior performance in a particular domain

Source video
factualVerification needed

Experts don't just know more facts; they actually see the world differently and build a world model of their environments.

Evidence

experts don't just know more facts. They actually see the world differently. ... experts effectively has, have built a world model of their environments

Question

Is there cognitive science evidence that experts' performance arises from distinct world models rather than knowledge volume?

Source video
factualVerification needed

Anthropic's revenue grew 400 times to $40 billion (or $60 billion annualized run rate) in under two years, largely driven by coding.

Evidence

In just under two years, their revenue has grown 400 times uh to uh 40 billion. I think the newest number is maybe 60 billion, uh, annualized run rate. And it's largely driven by coding and coding related productivity uh, capabilities.

Question

What is Anthropic's actual annualized run rate and what portion is attributable to coding-related products?

Source video

Connections

5