Australian startup Springboards has developed a new large language model (LLM) called Flint, designed to generate a wider variety of responses to open-ended questions. The company introduced Flint this week, aiming to address the predictability and limited creativity seen in mainstream LLMs like ChatGPT and Claude, which often produce repetitive answers such as the number 7 when asked for a random number between 1 and 10, according to technologyreview.com.
Springboards cofounder and CEO Pip Bingemann demonstrated Flint’s capabilities by comparing its responses to those of ChatGPT and Claude. While the mainstream models repeatedly gave the same number, Flint produced different answers upon repeated prompts. Bingemann explained that unlike other LLMs that focus on minimizing hallucinations, Flint embraces them to enhance creativity and variety. This approach is intended to improve brainstorming and planning tasks where diverse ideas are valuable.
The issue Flint addresses is the tendency of current LLMs to fall into a 'groupthink rut,' limiting their usefulness for creative applications. While existing models perform well in coding and research, their repetitive outputs can hinder tasks requiring originality. Flint’s design to welcome hallucinations marks a departure from conventional LLM training methods, potentially setting a new direction for AI models that prioritize diversity in responses over strict factual accuracy.
Springboards’ Flint model was publicly introduced in early July, with CEO Pip Bingemann highlighting its unique ability to break the pattern of predictable answers. The startup’s demonstration showed Flint producing varied responses consistently, a feature that could influence future LLM development strategies focused on creativity and open-ended problem solving, technologyreview.com reported.