Quick test: open ChatGPT, Claude, or Gemini and ask for a random number between 1 and 10. You’ll get 7. Almost every time. Ask for another? Probably 3 or 4. Another? 8 or 9.
The pattern isn’t random. It’s predictable. LLMs are surprisingly uncreative.
Australian startup Springboards built Flint — an LLM trained specifically to produce more varied responses. CEO Pip Bingemann has a sales trick he loves. He asks ChatGPT, Claude, and Flint for a random number. The first two give 7. Flint gives 3.7916.
Bingemann asked the models to name a car brand. ChatGPT and Claude said Toyota or Honda. Flint said Ford F-150. He asked for a New Balance tagline. ChatGPT and Claude both said “Run your way.” Flint said “Built to last, run to win.” Not award-winning, but at least it’s different.
The problem is real enough to win a NeurIPS best paper award. A team of researchers asked 25 different LLMs 50 times each to write a metaphor about time. Most of the 1,250 responses were “Time is a river” or “Time is a weaver.”
Compare that to humans. Six colleagues gave six completely different answers. One gem: “Time is a favorite sweatshirt, shaped by a lifetime of wear.”
Why does this happen? Most LLMs train on similar data, in similar ways, for similar tasks. They converge on the same high-probability answers. Great for coding. Terrible for brainstorming vacation ideas.
Springboards built a creative tool that lets you drag text from different models and combine them into something new. Flint is the alternative model for when you want variety. “Most language models are fighting hallucinations,” Bingemann says. “We welcome them.”
Now try asking your favorite chatbot what to name your band. It won’t say Sofa Astronauts. That one’s taken anyway.
