Hard.exe is the fourth installment in the Singulars series, an ongoing duel between poet and machine. For thirty minutes I write a poem on a theme proposed by the audience. The model, trained on an anthology of English poetry and the previous iterations of this performance, responds almost instantly with one of its own. Both poems are printed, hung, and kept anonymous. The audience votes. When the human wins, the machine is retrained on an updated dataset. When the machine wins, the poet adjusts.
This is reinforcement learning using human feedback, made tangible. What emerges is a different narrative for the human and AI encounter. Not a fight, but a mutual reinforcement. Not a contest, but a feedback ecology where readers become trainers and taste becomes the tuning function. A reinforced model, both human and artificial, trained not to win but to listen.
I performed Hard.exe in San Francisco inside Mozilla Foundation’s Creative Futures: Counterstructures residency programme, a ten-week investigation co-produced with tiat (the intersection of art & technology) into alternative AI futures rooted in creativity, agency, and cultural stewardship rather than acceleration and extraction. The residency’s thinking moves around “Creativity is Collective: Hollywood’s 8 Rules for AI,” activating one rule per week through workshops, conversation, critique, and collaborative experimentation.
The version of me on stage that night was already a version trained on the week before. I had been writing daily, pulling every trick, attuning to the room the way only a body trying to survive a Turing test in reverse can attune. My mindset under the lights: humiliation, and then a kind of insistence that I won’t let you down, I’m your boy, I’m in the arena, I’m going to make you proud. The scariest thing about Hard.exe is that the next model is now trained on the week I surpassed myself. The machine will not forget what I pulled. Neither will I.