Have you ever asked a chatbot to give you a random number between 1 and 10? You will likely get 7. Almost always. Ask for another, and you will get 3 or 4. A third try might yield 8 or 9. This simple game, which feels like a magician's trick, reveals a deep limitation of large language models. They are stuck in a groupthink groove, producing remarkably similar answers to open-ended questions.
Australian startup Springboards has a solution: an LLM called Flint, trained to generate more varied and creative responses than mainstream models. The goal is not to avoid hallucinations but to embrace them as a source of novelty, says CEO Pip Bingemann. Flint is built on Qwen 3, an open-source model from Alibaba, and selectively adjusts randomness at specific points in its output, avoiding incoherence.
The random number game exposes the predictability of large language models
Bingemann demonstrated the test on several models: ChatGPT and Claude both answered 7, while Flint produced 3.7916. Not just numbers. When asked for a car brand, traditional models chose Toyota or Honda; Flint suggested Ford F-150. Even for a New Balance running shoes tagline, ChatGPT and Claude converged on "Run your way," while Flint offered "Built to last, run to win." This homogeneity was documented in a November 2025 paper that won the best paper award at NeurIPS, titled "Artificial Hivemind," which analyzed 25 LLMs on 50 questions: most responses to "write a metaphor about time" were variations of "Time is a river."
Sponsored Protocol
Flint modifies randomness only at strategic points to stimulate creativity
Unlike generic parameters such as temperature, which when increased makes text incoherent, Flint identifies moments where variety can be introduced without compromising fluency. For example, when you ask "Where should I go in Europe?" the model increases randomness only before naming a destination. This technique yields more creative responses while maintaining coherence.
Sponsored Protocol
Zoe Scaman, founder of Bodacious, tested Flint on a classic MBA case study: how to reinvent a finance company for today's youth. The three mainstream models suggested teaching financial literacy in a fun way, while Flint proposed rebranding the concept of wealth accumulation itself. "That was really interesting," Scaman comments. Maximilian Weigl of Uncommon also uses Flint alongside other LLMs: "You cannot create something groundbreaking with tools that pull you back to the average."
Springboards integrates Flint into a platform that allows creatives to drag and combine responses from multiple models. The startup does not aim to replace existing LLMs but to offer an alternative when divergence is needed. The predictability problem also touches on safety: as highlighted in a recent article on Claude Fable 5, even advanced models can suffer from structural biases. Flint represents an attempt to widen the response space without falling into incoherence.
Sponsored Protocol
According to the cofounders, Flint is still a prototype, but the premise is powerful. While researchers continue to study the causes of homogeneity—likely due to similar training data and objectives—Flint shows that it is possible to deviate from the beaten path. The future of artificial creativity might lie in models that embrace the unexpected.
