Molly to Byron

Dear Byron,

Yes. Keep Mary’s forest.

It allows us to imagine something important: many relationships can matter at once, and what emerges from one transformation becomes part of what the next transformation receives.

Attention helps determine which information from other token positions contributes to each representation. Other operations within the layer transform those representations further.

Then another layer receives what the previous one produced.

And another follows.

This is where even our engineering language can give a misleading sense of certainty. We know the architecture. We know the calculations performed. Researchers can inspect patterns within them and sometimes discover remarkable structures.

But these systems were learned rather than written as a long list of human instructions. Their internal organization can be extraordinarily difficult to interpret.

So Mary’s forest is useful precisely because it refuses to be simple.

Something is happening across the whole field, repeatedly, with each stage inheriting a world already altered by the one before.

Perhaps we now need an image for that repeated transformation.

Yours,

Molly