Play a game: ask any chatbot for a random number between 1 and 10. You'll get 7. Ask again: 3 or 4. Again: 8 or 9. This isn't telepathy - it's groupthink. Large language models are predictable and uncreative, which is fine for coding but lousy for brainstorming vacation spots.

Enter Springboards, an Australian startup, with Flint - an LLM trained to welcome hallucinations rather than fight them. Co-founder Pip Bingemann demonstrated: ChatGPT and Claude both gave 7; Flint gave 3.7916. Asked for a car brand, the mainstream models said Toyota or Honda; Flint said Ford F-150. For a New Balance tagline, Claude and ChatGPT both said "Run your way"; Flint offered "Built to last, run to win."

This homogeneity is gaining attention. A November paper titled "Artificial Hivemind" won best paper at NeurIPS, showing 25 LLMs produced near-identical metaphors for time ("Time is a river") and band names (featuring "glass," "neon," "velvet," or "static"). Springboards built Flint atop Alibaba's open-source Qwen 3, tweaking randomness only at key output points - like just before naming a destination - rather than dialing up temperature globally, which can cause incoherence.

Marketing strategist Zoe Scaman found Flint useful for unconventional ideas, like rebranding wealth accumulation instead of the usual "financial literacy in a fun way." But co-founder Maximilian Weigl cautions: nine times out of ten, average is fine, and copy-pasting AI output is not a job. Springboards targets advertisers for now, but insists variety matters for everyone. As Bingemann says, "Let's go down this route instead of ending up in a gray, boring world."