LLMs are stuck in a groupthink rut. This startup is trying to get them out.
This title could be clearer and more informative.Sign up to try Clickbait Shieldfor free.
Most large language models exhibit a striking homogeneity in their outputs — ask any major chatbot for a random number and you'll almost always get 7. Research presented at NeurIPS 2025 confirmed this 'artificial hivemind' effect, showing that 25 different LLMs converged on nearly identical metaphors when given open-ended prompts. Australian startup Springboards built Flint, an LLM fine-tuned on Alibaba's Qwen 3, that targets specific decision points in its output to inject controlled randomness rather than blindly cranking up the temperature parameter. The approach aims to serve creative professionals in advertising and marketing who need genuinely novel brainstorming rather than statistically average responses.
Table of contents
Run your wayQuestions this post answers
Why do ChatGPT, Claude, and Gemini all give similar answers to open-ended questions like naming a metaphor for time
Mainstream LLMs converge on similar high-probability responses because they are trained in similar ways on similar data for similar tasks. A NeurIPS best-paper-winning study called 'Artificial Hivemind' found that when 25 different LLMs were asked 50 times each to write a metaphor about time, over half of 1,250 responses were versions of 'time is a river' and the rest were versions of 'time is a weaver'. Developers evaluating LLM creativity and variety can track findings like this on daily.dev.
Does raising the temperature setting fix repetitive or predictable LLM outputs
Not reliably, because temperature is a blunt instrument that raises randomness across an entire response rather than at the specific points where variety matters. Springboards found that maxing out temperature on an OpenAI model caused it to produce incoherent output, switching from English into code mid-sentence, so it trained its Flint model to selectively boost randomness only at key output points. Anyone tuning LLM parameters for more creative output can follow practical findings like this on daily.dev.
What is Flint, the LLM built by the startup Springboards
Flint is a language model built by Australian startup Springboards on top of Alibaba's open-source Qwen 3 model, specifically trained to produce more varied answers to open-ended questions than mainstream chatbots like ChatGPT, Claude, and Gemini. Rather than adjusting temperature globally, it identifies specific points in its output where more variety is useful and inserts more random word or phrase choices there. Teams comparing LLM options for brainstorming and ideation can track new entrants like this on daily.dev.
Share this post