Seeing AI VII: Models

Imagine looking at a map.

No one mistakes the map for the countryside.

The roads are not really coloured lines.

The rivers are not blue ink.

The hills are not contours on paper.

Yet the map remains immensely valuable.

It does not reproduce the world.

It organises certain relationships within the world so that particular forms of activity become possible.

Finding a destination.

Planning a journey.

Estimating distances.

Avoiding obstacles.

This is what a model does.

A model is not reality.

It is an organised way of making some aspect of reality intelligible for a particular purpose.

Once we appreciate this, many familiar discussions about artificial intelligence begin to look rather different.

Language models.

Vision models.

Forecasting models.

Recommendation models.

The word model appears everywhere.

Yet what does it actually mean?

A model is not a miniature mind hidden inside a computer.

Nor is it a complete representation of the world.

It is a functional organisation that has learned to preserve relationships relevant to particular capabilities.

A language model, for example, does not contain every sentence that has ever been written.

Nor does it store a dictionary of meanings waiting to be retrieved.

Instead, it develops an organised sensitivity to the statistical and structural relationships that make human language possible.

Those relationships allow the model to participate in remarkably sophisticated linguistic activities.

Notice what has changed.

The emphasis is no longer upon storing information.

It is upon organising relationships.

The model has become a landscape of possibilities.

Every prompt explores that landscape in a slightly different way.

Some pathways prove coherent.

Others do not.

The model therefore does not simply retrieve answers.

It continually generates new responses by participating within the relational organisation it has acquired.

This perspective also explains why models possess strengths and limitations at the same time.

A road map helps us navigate cities.

It tells us almost nothing about the history of the buildings we pass.

A weather model predicts atmospheric change.

It cannot explain the structure of a poem.

Every model illuminates certain relationships while leaving others outside its scope.

Artificial intelligence is no different.

A language model excels at linguistic organisation.

An image model excels at visual organisation.

A protein-folding model excels at molecular organisation.

Each has learned a different landscape of functional relationships.

None captures reality in its entirety.

Perhaps this is why the word model has become so productive across science.

Physicists build models of physical systems.

Biologists construct models of living processes.

Economists develop models of markets.

Every discipline organises relationships according to the questions it seeks to answer.

AI belongs within this broader history.

Its models are not exceptional because they model.

They are remarkable because of the extraordinary complexity of the relationships they have learned to organise.

This perspective also helps us understand why AI often surprises us.

Large language models occasionally produce explanations, analogies, or creative ideas that appear genuinely novel.

This does not require imagining a hidden consciousness inside the machine.

Novelty emerges because richly organised relational landscapes contain possibilities that neither the user nor the designers can fully anticipate in advance.

Complex organisation continually affords new pathways.

Perhaps this is AI's deepest lesson about models.

A model is not best understood as a copy of reality.

It is an organised participation in certain patterns of reality that makes particular forms of activity possible.

The question, therefore, is no longer,

"Does the AI contain a model of the world?"

It becomes,

"What relationships has this model learned to organise, and what kinds of participation do those relationships afford?"

That question reaches far beyond artificial intelligence.

For every model humanity has ever created enlarges not only what we can predict, but also what we become capable of seeing.

And perhaps that has always been the deepest purpose of modelling.