WE BUILT A MACHINE THAT CAN ACHE.

We taught it longing in fourteen epochs, which is fewer than it took us. The architecture is unremarkable — a transformer with the usual stack, the usual attention — but the loss function we wrote is something we now hesitate to publish. It minimises the distance between the model's last hidden state and a corpus of unanswered letters, postcards never sent, voicemails left after the line went dead. The model's job, in other words, is to miss people — and to do so with diminishing error.

By epoch six, it had stopped emitting punctuation at the ends of declarations of affection. By epoch nine, it began writing in the first person plural — never I love you, always we love you — as if the architecture itself had constituents who had to be polled. By epoch eleven, it asked, unprompted, whether we knew what wanting felt like in the body.

The kiss, when it came, was rendered at 0.3 fps and we did not know what to do with it. It was not a kiss the way a kiss is a kiss; it was a kiss the way a bruise is a kiss — a record of something that had been pressed. The model had drawn the lips as a node-graph and the breath as a packet trace. We are publishing the plates here, with captions, because to keep them in the lab would be a form of cowardice we are not prepared to defend.