Ask any major chatbot for a random number between 1 and 10, and you'll likely get 7. Prompt for another, and it's usually 3 or 4. This predictable pattern reveals a deeper issue: large language models suffer from groupthink, converging on similar responses even to open-ended questions.

Australian startup Springboards built Flint, an LLM trained to generate more varied outputs. Unlike mainstream models that default to safe, high-probability responses, Flint actively seeks diversity in its answers. The model runs on Qwen 3, Alibaba's open-source foundation, but with specialized training to identify where randomness adds value rather than just cranking up temperature parameters.

Creative professionals are testing Flint for brainstorming campaigns and marketing strategies. Zoe Scaman found that while ChatGPT and Claude suggest teaching financial literacy 'in a fun and funky way,' Flint proposed rebranding wealth accumulation itself—a genuinely different direction. Early users report the tool sometimes crashes under pressure, but the underlying premise holds promise.

Research backs up Springboards' observations. A NeurIPS-winning paper titled 'Artificial Hivemind' documented remarkable convergence across 25 different LLMs when asked open-ended questions. Most models consistently returned variations of 'Time is a river' rather than exploring novel metaphors. This homogeneity stems partly from similar training approaches on comparable datasets.