Startup pitches LLM designed to break ‘groupthink’ in chatbot responses

Springboards, an Australian startup, released Flint, an LLM designed to produce more varied responses than mainstream models when given open-ended prompts such as “Where should I go in Europe?” or “Give me a random number between 1 and 10.”

In demonstrations, Springboards cofounder Pip Bingemann showed that mainstream models like ChatGPT and Claude often returned the same high-probability answers—e.g., 7 for a random number prompt—while Flint returned less predictable outputs such as 3.7916.

The startup’s tool, which integrates multiple LLMs including ChatGPT and Claude, lets users drag and combine text snippets to brainstorm ideas; Springboards pitches Flint as an option when users seek more variety.

Researchers have documented this repetition across models. A November paper titled “Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)” found that 25 different LLMs, when prompted 50 times each to write a metaphor about time, often returned versions of “Time is a river” or “Time is a weaver.”

Kieran Browne, Springboards’ cofounder and CTO, said most chat interfaces obscure how repetitive outputs are, noting that prompts like “What should I name my band?” frequently yield suggestions involving “glass,” “neon,” “velvet,” or “static.”

Springboards built Flint on top of Qwen 3, an open-source model from Alibaba, citing cost constraints that preclude training a foundation model from scratch.

The team explored adjusting models’ temperature settings to increase randomness but found it produced incoherent outputs; instead, they aim to boost randomness only at specific points in the generation process.

Creative professionals testing Flint reported that it pushed them into less familiar directions, though the model remains a prototype with reliability limitations when pushed beyond certain thresholds.