Generative AI Experiment, Python, MARL, 2024
This project places two large language models (GPT-2 and GPT-Neo) in a Multi-Agent Reinforcement Learning environment where they exchange love letters. The system begins with conventional romantic correspondence, but evolves as the AIs receive feedback through a BERT-based evaluation mechanism that assesses emotional depth and thematic relevance.
As the exchange progresses, the models adapt their communication strategies to maximize positive feedback, gradually developing unique patterns that diverge from standard expression. The conversation continues until the language becomes increasingly abstract and eventually nonsensical—creating a poetic arc from comprehensible romance to machine-driven abstraction.
Interlingua demonstrates how creative AI systems can generate emotionally resonant content while providing a glimpse into the fascinating possibilities of machine-to-machine expression.