Springboards' Flint: an LLM built to break LLM groupthink

Springboards has built Flint, a language model trained to generate a wider range of answers than mainstream models for open-ended prompts. The company trained Flint on top of Alibaba’s open-source Qwen 3 and demonstrates the model producing different, less predictable responses than ChatGPT and Claude in creative tasks.

How does Flint produce more variety?

Flint is trained to insert extra randomness at specific points in its output rather than raising randomness across an entire response. Springboards fine-tuned Qwen 3 to identify where variety matters — for example, just before naming a destination — and to plug in “oddballs” at those spots. This selective approach differs from simply turning up a model’s temperature, which Springboards found to be a blunt instrument that can make outputs incoherent.

Springboards’ cofounder and CTO Kieran Browne says models appear to hide a lot of possible answers and default to familiar, high-probability options. The company’s cofounder and CEO Pip Bingemann framed the choice differently: “Most language models are fighting hallucinations,” he says. “We welcome them.” In demonstrations, ChatGPT and Claude both returned the number 7 to the prompt “Give me a random number between 1 and 10,” while Flint returned 3.7916 on a restarted session. In other examples, mainstream models defaulted toward brands like Toyota or Honda when asked to name a type of car; Flint offered Ford F-150.

How does Flint compare to other LLMs and to the research on repetition?

Research has documented cross-model repetition among LLMs. A November paper titled "Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)" asked 25 different LLMs 50 times each to write a metaphor about time and found most of the 1,250 responses clustered around versions of “Time is a river” or “Time is a weaver.” The paper won the best paper award at NeurIPS. Springboards uses that finding as part of its critique: many models trained on similar data and objectives converge on the same safe answers.

Springboards positions Flint as an option inside its brainstorming tool, which already aggregates output from multiple models including ChatGPT and Claude. Advertisers and marketers can drag and combine text snippets from different models; Flint is offered as the choice when a user wants ideas that diverge more from the average. In one user test, Zoe Scaman compared Flint with Claude, Gemini, and ChatGPT on an MBA case prompt and found Flint suggested rebranding wealth accumulation rather than repeating the typical “teach financial literacy in a fun way” line.

Why it matters

Flint’s targeted randomness addresses a practical gap for creative professions: mainstream LLMs tend to converge on familiar, high-probability answers, which limits brainstorming. For advertising and marketing teams that need divergent ideas, a model that intentionally inserts unusual choices can expand starting points and reduce the chance of everyone reaching the same derivation. At the same time, Springboards and users acknowledge trade-offs: Flint is a prototype that “sometimes falls over when you start pushing it too far,” Scaman says, and boosting novelty can make outputs less reliable.

Springboards’ approach also highlights an engineering choice: instead of training a new giant foundation model, the startup fine-tuned an open-source backbone and focused on where in a response to increase variety. That lowers cost and targets a specific user need in creative workflows.

What to watch

Watch how Flint performs in real creative workflows beyond demos: whether teams adopt it to seed genuinely different ideas, and whether selective randomness scales without producing unusable or misleading outputs. Also monitor follow-ups to the "Artificial Hivemind" work at conferences and whether other developers adopt selective randomness techniques or continue to tune global temperature settings.