On Friday, Australian startup Springboards introduced its new large language model (LLM), Flint, designed to overcome the predictability and creativity limitations of existing models. Unlike mainstream LLMs like ChatGPT and Claude, Flint offers a broader range of responses, catering especially to creative tasks such as brainstorming and planning.
Understanding Groupthink in LLMs
Large language models have been criticized for their tendency to produce similar responses to open-ended questions, a phenomenon known as groupthink. For instance, when asked to generate a random number, most models consistently return the same values, indicating a lack of creativity. According to Pip Bingemann, co-founder and CEO of Springboards, "Most language models are fighting hallucinations. We welcome them." This acceptance of variability is a key feature of Flint.
Flint's unique approach allows it to generate diverse outputs. In a demonstration, Bingemann prompted Flint and other models with the same question, and while ChatGPT and Claude returned the number 7, Flint provided 3.7916. This illustrates Flint's ability to break free from the predictable patterns of its competitors.
Research Highlights the Limitations of Current Models
Recent research published in a paper titled "Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)" emphasizes the homogeneity in responses from various LLMs. The study revealed that models, when tasked with similar prompts, often converge on identical or near-identical answers, underscoring the limitations of their training methods. The authors of the paper received the best paper award at NeurIPS, a prominent AI conference.





