What happens when you ask an LLM to write an infinite poem? Two experiments took that question seriously, using a couplet form with anaphoric litany in which only the end words change. The structure is a revealing testbed because it strips away most variables and forces the model to manage the relationship between paired words.

The first experiment relied on manual prompting and co-creation. A human author guided the model to extend the original poem, and the outcome became an animated web installation and interactive artwork. In that version, quality came less from the model's autonomy than from tight curator control. The second experiment aimed for automation: a pipeline combining an LLM with semantic similarity reliably reached the 1000 benchmark, showing stronger capacity for both quantitative and semantic output.

The technical shift matters. Pairing semantic similarity with generation moves part of the poetic judgment outside the autoregressive model. Instead of requiring every pair of end words to emerge with the right relationship already built in, the system can compare meaning representations and select candidates close to a desired relation. That does not make the model more creative, but it reduces the burden on pure generation and reframes the problem from 'write a poem' to 'design a pipeline that filters poems'.

The results are clear. LLMs produce interesting pairings, but they still fall short of human-level creativity and understanding of poetic relationships. Quantity does not close the qualitative gap. The limitation is not about generating many lines; it is about recognizing why two end words create tension, assonance, or a shift in meaning. Word-level generation remains the weak point because the constraint is semantic and relational, not purely statistical.

For readers looking at this through a local deployment lens, the signal is structural. A pipeline that separates generation from semantic evaluation can be distributed across components with different memory and compute requirements. The generator can stay relatively small, while semantic vector comparison can run on a dedicated service or more modest hardware. In a self-hosted setting, that modular approach reduces VRAM pressure and makes TCO easier to manage, because the entire flow does not require a single massive model. It is not a promise of automatic creativity, but a way to treat poetry as a composable workload.

There is also a control issue. The manual, co-creative version is a reminder that the value lies in process rather than raw text: the person designing the interactive installation decides constraints, relationships, and acceptance thresholds. The automated version delegates those choices to a proximity measure. Authorship shifts from the individual verse to the design of the pipeline and the curation of the reference dataset. That is not a loss of poetry, but its reconfiguration.

The research conclusion is sober: poetry will respond and revive by incorporating LLM workflows. But the resurrection is not the ability to generate infinite variations. It is the chance to explore word relationships at a scale a human could not sustain manually, provided the aesthetic judgment still remains, at least in part, human.