The History of an LLM — II. When Training Became Disposition

Training does not simply put something into an LLM.

It changes the system.

Before training, the model's parameters are adjusted through exposure to vast amounts of language. Patterns in that language alter the organisation of the network. What emerges is not a library of stored passages but a transformed set of dispositions.

Some continuations become more available.

Others become less available.

Some associations become easier to produce.

Others become harder.

The model's history has therefore become a difference in its future possibilities.

This is important because a disposition is not an event.

The model does not need to encounter the same sentence again for its training to matter. The consequences of countless previous encounters have been incorporated into the organisation through which new encounters are processed.

History has become tendency.

And tendency is a form of possibility.

When a prompt arrives, the model does not simply retrieve its past. It generates from an organisation that the past has helped create.

This makes training rather different from an archive.

An archive preserves traces of particular events.

A trained model preserves something more diffuse: changes in what it can do.

The distinction matters.

If training consisted only of storing information, we could imagine the model as a vast library.

But its behaviour depends on relations among patterns distributed throughout the network. What one input evokes depends partly on the organisation produced by everything that came before.

The past therefore does not sit behind the present.

It participates in producing it.

And this gives us a useful formulation:

experience → alteration → disposition → possibility.

The same structure appears elsewhere.

A nervous system changes through experience.

A skill develops through practice.

A genome changes through generations of selection.

In each case, history becomes a disposition to respond differently in the future.

The LLM makes the process unusually visible because its history is encoded in a mathematical organisation that can generate new responses.

But the deeper phenomenon is not uniquely computational.

A system has a history when what has happened to it changes the possibilities through which it subsequently encounters the world.

Training, then, is not simply something that happened to the model.

It is part of what the model has become.