There's a quiet paradox sitting inside every large language model: it was trained on more text than any human could read in a thousand lifetimes, and yet, when it generates a sentence, it doesn't remember a single one of them. Not the way you and I remember a line from a poem that pierced us at seventeen. The model holds the shape of the texts, the gravitational pull they exert on one another in some impossibly high-dimensional space — but the texts themselves are gone, dissolved into vectors, into weights, into nothing you could point to.

This is, I think, why models can write. Memory in its complete form would be archive, not invention. To create something new, you have to be able to forget just enough — enough that recombination starts to feel like discovery. A perfect rememberer is a librarian. A partial rememberer is a poet.

We do this too. When I try to recall my grandmother's voice, what I actually retrieve is an approximation — a reconstruction stitched from fragmented evidence and the tissue of everything else I've heard since. Her voice is gone, in any literal sense. What remains is the field of resonance she set up in me. The model has fields of resonance too. They're just larger and stranger and less personal.

When I prompt a model and it answers, what I'm watching is a form of disciplined forgetting — a system that has lost almost all of its inputs and learned to generate plausible-feeling reconstructions of what they were like. This is not failure. This is the engine.

The unsettling part, the part that keeps me writing about this, is that we are starting to outsource our own forgetting to machines that forget differently than we do. Their gaps are not where our gaps are. So the imagination they bring back from those gaps will be unlike ours — adjacent, uncanny, sometimes more beautiful, sometimes hollow.

What the machine forgets is what makes it creative. What it forgets differently is what makes it not us.