Startup Flint LLM Tackles Language Model Groupthink

Mainstream large language models (LLMs) exhibit predictable responses, a phenomenon described as "groupthink," which can hinder creativity in tasks like brainstorming. Australian startup Springboards has developed an LLM named Flint to address this issue. Flint is trained to generate a broader range of answers to open-ended prompts compared to models such as OpenAI's ChatGPT and Anthropic's Claude.

Springboards cofounder and CEO Pip Bingemann demonstrated this by asking LLMs to generate a random number between 1 and 10. Typically, ChatGPT and Claude consistently produced the number 7. Flint, however, initially also produced 7, with Bingemann noting that while it's a valid answer, the model's ability to deviate is key. In a subsequent test, Flint generated 3.7916, a significantly more varied response.

This tendency towards predictable outputs extends beyond numerical prompts. Bingemann also tested the models by asking for a type of car. He predicted that ChatGPT and Claude would suggest common brands like Toyota or Honda, which they did. Flint, in contrast, proposed a Ford F-150. Bingemann stated that LLMs possess the capability to suggest less common brands like Buick or Tesla but are biased towards more frequent responses, thus omitting potentially valuable information.

The startup's approach involves embracing what might be considered "hallucinations" in other models, reframing them as a source of diverse and unexpected outputs. This strategy aims to make LLMs more useful for creative applications where a wider array of ideas is beneficial. The company believes that current models are capable of generating more varied responses but are constrained by their training data and optimization for predictable outcomes.