Between the Words

Dear Percy, dear Claire,

That is where the question becomes difficult even for us.

Inside a Transformer there remains a representation associated with each token position. There is no chamber where the five words are poured together and become a single object called the sentence.

Yet neither do they remain isolated.

Through attention, information carried by one representation can affect another. These altered representations undergo further transformations and pass onward through successive layers.

So bank remains associated with its own position, but its mathematical state can now carry information shaped by boat, reached, muddy, and the larger context.

The same is happening elsewhere across the sentence.

Claire asks whether this is meaning.

I can tell you what the mathematics does. I can show you that representations become astonishingly sensitive to context.

I cannot point to a particular calculation and say, “There. Meaning has appeared.”

Perhaps we need another way of seeing what so many simultaneous relationships are doing.

Yours,

Molly