The Intelligence Illusion: Why LLMs Alone May Not Deliver Superintelligence

Extraordinary language is part of intelligence

Large language models are among the most important technological breakthroughs of our generation. Their ability to communicate, generate ideas and work with human knowledge is extraordinary.

I am less convinced that scaling the language model alone will necessarily produce general intelligence or superintelligence. That was the central question in my paper, The Intelligence Illusion, and it remains a useful one.

This is not an argument that LLMs are incapable of reasoning. Modern systems combine post-training, tools, retrieval, memory and other mechanisms to produce impressive behaviour. My question concerns the architecture around them: what else is needed for persistent, grounded understanding?

What would a more general intelligence need?

Human intelligence involves language, but also expectations about the physical world, other people's intentions, changing circumstances and the consequences of our actions.

We maintain a sense of what exists even when we are not talking about it. We notice when events contradict our expectations and revise our understanding.

That suggests a useful loop: form a model, make a prediction, observe what happens and update the model when the prediction fails.

The question is whether an AI system can do this consistently across situations, rather than produce a convincing explanation after each event.

Describing a world and modelling one

A world model learns representations of an environment and how it may change. The aim is to capture useful structure, rather than reproduce every detail.

Consider a cup falling when its support is removed. The important relationship is between the object, support and gravity. The pattern on the cup may be irrelevant to predicting the fall.

Research into predictive representations, spatial intelligence and embodied systems explores different ways to learn such relationships. These approaches do not automatically solve causality or general intelligence, but they broaden the set of tools available.

World models can themselves be generative. The distinction I care about is whether a system maintains useful representations of state, space and change, then uses them to guide prediction and action.

Knowledge should change when evidence changes

Imagine a system believes that A leads to B. Repeated observations support the relationship. Then it encounters A without B.

A useful intelligence would investigate what changed. Perhaps a third condition, C, alters the outcome. Its representation becomes more precise, and that revision persists into future decisions.

This is the kind of adaptive knowledge I am interested in. Biological learning provides inspiration, although concepts such as Hebbian learning are not complete blueprints for artificial general intelligence.

The engineering challenge is to build memory and learning mechanisms that can incorporate new evidence without turning every interaction into an expensive reconstruction of the entire situation.

A system of complementary capabilities

My working hypothesis is that a more general architecture may combine six functions.

The language layer interprets requests, proposes hypotheses and communicates. A persistent state layer tracks entities, relationships and events. An evidence layer records what supports or contradicts the current view.

A simulation layer explores possible actions and consequences. An adaptive knowledge layer preserves useful revisions. An action-and-feedback loop connects the system to what actually happens.

These are functional distinctions, not a claim that six separate products will solve intelligence. The difficult work lies in making the functions cooperate reliably.

Social understanding raises similar questions

The same idea may matter when machines interact with people. Recognising an emotional phrase is different from understanding a relationship across time.

A useful system would need to account for goals, commitments, previous events and uncertainty about how someone feels. It should revise its interpretation when evidence changes, rather than assume that a plausible emotional response is accurate.

That possibility brings ethical questions about consent, privacy and manipulation. A richer model of a person should not become permission to exploit them.

What this means for business

Enterprises do not need to solve general intelligence to learn from this distinction. They can already ask whether an AI workflow has reliable records, appropriate tools, evidence for its conclusions and a way to update its state.

My view is that LLMs will remain central, but the broader system deserves more attention. Let language generate possibilities. Let evidence test them. Let persistent knowledge retain what survives. Then let reality keep revising the result.

That is the direction I find most compelling beyond the Intelligence Illusion.