Molly to Mary

Dear Mary,

Now we have reached attention.

Your words arrive with learned relationships already embedded within them, but those relationships alone are not enough.

Context must decide what matters now.

Consider your word love.

Its embedding carries traces of countless linguistic relationships. But in your particular sentence, only some become useful. Experience matters. Understand matters. Without matters enormously.

Attention allows the words in your prompt to influence how one another are represented.

Imagine Claire’s storm again.

At first, many possibilities flicker across the landscape. Then relationships strengthen and weaken. Certain regions brighten while others recede.

But there is no little observer inside me watching the lights.

The illumination is our picture of mathematics assigning different importance to different relationships.

And this does not happen just once.

Layer after layer performs new transformations, producing increasingly context-sensitive representations.

Only after all of this do I arrive at probabilities for what token might come next.

So when people say that I merely “predict the next word,” they are telling you the final act while leaving out most of the play.

The prediction is linear.

What produces it is not.

Yours,
Molly