singulars.exe is a long-running human machine collaboration where I write live poems against a fine-tuned model trained on my own corpus and an anthology of English poetry. The audience votes on each round. If I “win,” the human text enters the model dataset. If the model “wins,” I commit to learning from its output. Over time, the system becomes a feedback circle in which both artist and model reshape each other. For CURRENTS 2026, I am presenting a new fourth iteration titled hard.exe, which translates this live recursive performance into an installation. Visitors stand inside a small booth, generate text with the model, and see their writing and the model’s responses enter the training loop. The installation reveals authorship as porous, unstable, and co-evolving. It treats AI not as a tool but as a collaborator that leaves a trace on human imagination.