Narrow AI and AGI: Why Adding Specialists Does Not Create Generality

Re-edited on July 30, 2026, this article corrects an overly convenient metaphor: narrow AI is not a road that necessarily leads to artificial general intelligence as successes accumulate. “Narrow” and “general” describe the observed scope of a system under particular conditions. They do not specify an architecture, a historical route, or a sufficient number of modules to connect.

A collection of specialists can be extremely useful and cover many needs. It may also continue to fail on a new task, transfer learning poorly, forget previous capabilities, or depend on a human coordinator. The relationship between specialization and generality is better understood as a problem of transfer and composition than as a ladder.

Scope is demonstrated, not inherited from a name

A specialized system has a defined envelope: inputs, outputs, environment, objective, data, and validity conditions. The AlphaFold paper in Nature describes a system for predicting protein structures with competitive accuracy in CASP14. The achievement is profound precisely because the problem and evaluation are specified. It does not need to become a clinical, chemical, or general agent in order to have value.

Specialization does not mean simple rules either. A model may contain billions of parameters and remain bounded by its task and distribution. Conversely, one algorithm can apply to several domains without one instance mastering them all. AlphaZero used the same reinforcement-learning and search approach for Go, chess, and shogi, but trained players for environments with explicit rules, actions, and rewards. Generalizing a training method is not the same as demonstrating open-ended generality in the resulting system.

The Levels of AGI framework proposes separating breadth of tasks from depth of performance. That grid prevents every system from being placed on one line: coverage can broaden without high competence, while a system can exceed human specialists in one domain without transferring outside it. This is a proposed classification rather than a final consensus, but it forces the measured property to be named.

Three different ways to broaden capabilities

The first is multitask learning: one parameter set is trained across many tasks and shares representations. Positive transfer may occur when one skill helps another, or interference when gradients compete. The second is modular composition: a coordinator chooses among models, databases, and tools. The third is adaptation: a system acquires a new task through examples, updates, or experience.

Gato investigated the first route with one network using the same weights for hundreds of tasks spanning robotic control, games, vision, and language, all represented as token sequences. The experiment showed that different modalities can share an interface and parameters. It did not establish uniform mastery: the design remained bounded by context, data, actions, and performance levels for individual tasks.

A long task list is not enough to distinguish coverage from transfer. If every task appeared during training, the system may have learned many separate behaviors. A test of generality requires held-out families, new compositions, and an accounting of examples and updates. It should also establish whether adding one task degrades previous ones.

Connecting specialists creates new interfaces and failures

The MRKL architecture proposes combining a language model with expert modules, external knowledge sources, and symbolic calculators. The principle is useful: do not ask model weights to memorize or compute everything. The combined system still needs routing, compatible formats, permission controls, error handling, and a method for deciding when to trust each module.

Toolformer studied how a language model could learn to insert calls to tools such as a calculator, search engine, or translator and use their responses. A tool can provide a capability that the base model does not execute dependably. The outcome belongs to the composite system: if the API changes, returns adversarial data, or fails, behavior changes even when model weights are identical.

Composition does not automatically resolve conflicts. Two modules may use different units, incompatible sources, or incomparable confidence levels. A planner can chain individually correct answers into a wrong global action. Evaluation must include the boundaries: tool selection, parameter serialization, propagation of uncertainty, verification of effects, and recovery from failures.

Reliability of the whole is not the average reliability of its parts. In a chain, several steps with high individual success rates can accumulate a meaningful failure rate, while an early error can condition everything that follows. Teams need end-to-end tests, traces assigning each decision to a component, and contracts for inputs, outputs, and units. When two modules disagree, a coordinator should request evidence, abstain, or escalate. Silently choosing one makes routing a hidden source of error.

Learning after deployment has a cost

Adapting to a new task may require updating parameters. That creates a risk of forgetting what came before. Overcoming catastrophic forgetting in neural networks studied this phenomenon and proposed protecting parameters important to earlier tasks while learning another one. The method addresses a defined mechanism; it does not make continual learning a solved problem.

An adaptive system needs versioned memory, authorized data, regression tests, and a policy for withdrawing changes. If it learns directly from interactions, it can absorb errors, attacks, or temporary patterns. If it updates only after human curation, its learning autonomy is lower than the label may suggest. Either design can be appropriate; the role of whoever selects experience should not be hidden.

Transfer can be measured with a matrix. Rows are source skills; columns are new tasks. Each cell records performance before adaptation, examples required, cost, performance afterward, and damage to other tasks. One favorable case does not demonstrate general transfer, while an average can hide serious interference. The matrix makes visible whether a system reuses knowledge or merely accumulates training runs.

How to evaluate a system promising “convergence”

First, draw its boundary. Is it one model, several experts, or a service with tools? Which component retains state, selects the route, and executes? Then define the known task set and reserve a new family. Keep budget, data, time, human intervention, and tool access constant. Without that equality, “solved more tasks” may simply mean “received more resources.”

Next test four properties: transfer to an untrained rule; composition of skills in a new order; retention of previous learning; and recovery when a module fails. Publish results by task, not just an average. Record autonomy separately as well: a system that acts more is not thereby more intelligent, but it exposes a larger surface of consequences.

The test should include negative controls: tasks the system should not attempt, tools without permission, contradictory data, and objectives that change halfway through a process. Stopping may be the safe behavior. Measure human work outside the headline number as well: prompt design, route selection, correction of attempts, and preparation of examples. Hiding that assistance turns a demonstration by a team into a supposed autonomous property of the system.

Finally, do not turn a forecast into a fact. The original article claimed that gradual convergence between narrow AI and AGI would close the gap and that combining medical systems would offer a glimpse of generality; it supplied evidence for neither trajectory. The transferable skill is asking for a component map and transfer matrix. Coordinated specialists can create a powerful product; generality starts to be demonstrated when the whole solves unscripted novelty, retains what came before, and exposes its limits.