Ever ask a chatbot for a random number between one and ten? Chances are, it confidently declared "seven." Because apparently, even AI gets stuck in a rut.
Most large language models (LLMs) are, to put it mildly, a bit same-y. They're trained on similar data, they learn similar patterns, and they spit out similar answers. It's like asking 25 different people for a metaphor for time and getting "Time is a river" from over half of them. The other half? "Time is a weaver." Riveting.
But an Australian startup called Springboards decided enough was enough. Their solution? A model named Flint that actively embraces what other LLMs try to avoid: a little well-placed hallucination.
We're a new kind of news feed.
Regular news is designed to drain you. We're a non-profit built to restore you. Every story we publish is scored for impact, progress, and hope.
Start Your News DetoxInjecting a Little Chaos
Pip Bingemann, Springboards' cofounder and CEO, explains that while most LLMs are rigorously taught to avoid making things up, Flint is designed to inject randomness at specific points in its responses. The goal isn't nonsense, but novelty. So when ChatGPT or Claude suggest "seven" for that random number, Flint might drop a 3.7916. When asked for a car, instead of the usual Toyota or Honda, Flint might suggest a Ford F-150. Because why not?
This isn't just about quirky numbers. When tasked with New Balance taglines, mainstream models offered the expected "Run your way." Flint? "Built to last, run to win." A subtle, yet distinct, difference.
Kieran Browne, Springboards' CTO, points out that we often don't even realize how much we're all getting the same AI-generated information. Ask for band names, and you'll drown in a sea of "glass," "neon," "velvet," and "static." It's like a digital hivemind, but less productive and more... beige.
How Flint Breaks the Mold
Building a foundational model from scratch is prohibitively expensive, so Springboards built Flint on Alibaba's open-source Qwen 3. The trick wasn't just cranking up the "temperature" setting (which controls randomness) across the board. Browne found that just makes models switch from English to code mid-sentence – impressive, but not useful.
Instead, Flint is trained to identify where to get weird. If you're asking for European travel destinations, Flint will get creative with the cities, but won't suddenly start suggesting travel by unicycle. It's targeted chaos.
This focused approach is already making waves. Zoe Scaman, founder of Bodacious, uses Flint for brainstorming and finds it helps explore wildly different ideas. She recalls a test where mainstream models offered similar takes on reinventing a finance company for youth. Flint? It suggested rebranding the entire concept of wealth accumulation. Which, if you think about it, is both impressive and slightly terrifying.
Maximilian Weigl of marketing firm Uncommon also uses Flint to generate "boundary-breaking" ideas, though he wisely cautions that for most day-to-day tasks, an "average" response is perfectly fine. He also reminds us that human thinking, discussion, and that unique personal voice are still irreplaceable for true creativity.
For now, Flint is aimed at advertising and marketing pros, but Springboards sees this lack of variety as a universal problem. Because a world where machines dictate all responses? That sounds, frankly, a bit gray and boring. And nobody wants that.
