Five diverse AI agents vs. five clones for 14 nights: a constant tied both (thoughts.jock.pl)

A recent experiment revealed that a diverse group of five AI agents outperformed five identical clones in forecasting virality over a two-week period. Using the same model (Opus 5) and budget, the diverse agents—each with unique contexts and information—achieved a lower Brier score of 0.0225 compared to the clones' score of 0.0275. This significant finding supports the notion that networks of AI agents can offer greater intelligence and richer contextual understanding than multiple instances of the same model, undermining common industry pitches that rely on simple diversity through varied prompts without structural differences. The experiment featured rigorous controls, with agents tasked to predict the virality of 30 fresh social media posts daily. Each agent was designed to focus on different aspects of content, leading to independent outcomes rather than homogeneous responses. Notably, the correlation between predictions in the diverse arm was significantly lower than in the clone arm. The findings suggest that real diversity in AI requires not just varied prompts but fundamentally different operational frameworks, paving the way for future innovations in multi-agent systems within the AI/ML community.