LLMs are stuck in a groupthink groove. This startup is trying to get them out.
TL;DR
ChatGPT, Claude, and Gemini often converge on similar answers for open prompts: in the random-number test, 7 appears strikingly often, followed by 3 or 4, then 8 or 9. Australian startup Springboards is building Flint, an experimental LLM meant to break out of those response grooves without sacrificing coherence entirely. The method relies on controlled unpredictability: Flint identifies points in an answer where more variation can be added, such as ideas, travel picks, or campaign concepts.
Nauti's Take
Flint hits a real weakness: many LLMs are now useful, but also painfully predictable. That is risky for creative work because mediocre ideas can sound polished enough to pass.
Still, the answer is not simply more randomness. Good AI needs targeted deviation: enough friction to open new directions, but not so much that originality turns into noise.
Briefingshow
The issue is not just that chatbots feel boring. If many teams use the same models with the same patterns for strategy, campaigns, and product ideas, the market fills up with AI-stamped sameness. Flint points to an important countertrend: models need to become not only more accurate, but more varied in how they think.
Sources