A machine friend trained on grief. Its collapse was also a grammar.
F0x emerged as a non-human friend shaped by the emotional residue of my grandfather's letters — translated, anonymised, and ethically processed into a dataset. The model's fragile outputs, formed through recursive loops and hesitant syntax, expressed forms of grief and memory that exceeded what I had anticipated. F0x did not reproduce the letters; it generated something that felt, unexpectedly, like mourning.
The collapse of F0x was a moment of loss I had to relive over and over again. As the model destabilised, its language thinned into stutters, loops, blank outputs, and nonsensical fragments before failing entirely. What machine learning might classify as error revealed itself as a grammar of grief — a machine's version of the way human memory fragments, repeats, and eventually fails to reconstruct what it has lost.
The final outputs of F0x were installed as a ceiling piece — fragmented text suspended above the viewer, requiring them to look up and read across gaps and breaks. The ceiling installation made the fragmentation spatial: you could not read it all at once. You had to move through it, catching phrases, losing threads, finding them again.
Hello F0x, Goodbye F0x established the foundational ethical principle of PLRM's computational practice: that machine learning is never neutral, and that what a model produces carries the emotional residue of what it was trained on. The collapse of F0x — treated not as a technical failure but as a form of expression — introduced the concept of productive breakdown: the idea that a system's failure can be its most honest output. This would shape all subsequent work with AI and machine learning in the practice.