Springboards' Flint tackles LLM groupthink, boosts response

Springboards has built an LLM called Flint, trained to come up with a wider variety of responses than mainstream large language models for open-ended prompts. The startup designed Flint specifically to avoid the predictable answers many chatbots deliver on casual or creative queries.

What is Flint and how does it differ from mainstream models?

Flint is an LLM created by Australian startup Springboards and trained to generate a broader range of responses than mainstream LLMs to open-ended questions, such as "Where should I go in Europe?" The core distinction the company emphasizes is variety: rather than converging on a single obvious reply, Flint aims to surface more diverse options when the user intent is exploratory or creative.

Beyond that training goal the newsletter does not detail Flint’s architecture, dataset or deployment plan. The framing given centers on Flint’s behavioral target: reduce the sameness that shows up in everyday prompts and brainstorming tasks.

How common is this "groupthink" problem in chatbots today?

Many large language models are highly predictable on simple prompts, the newsletter notes, giving the example that if you ask "Give me a random number between 1 and 10" you are going to get 7 almost always. That tendency appears across popular chatbots mentioned as comparators, including Claude, ChatGPT, and Gemini, and it manifests as reduced creativity on open-ended tasks.

The piece contrasts that predictability with the needs of exploratory tasks. It says that predictability is acceptable for work like coding or research, where consistent answers are useful, but becomes a drawback for brainstorming or personal planning where users expect varied possibilities.

Why it matters

Groupthink in LLMs limits their usefulness for ideation. When a model repeatedly returns the same or similar responses, users lose useful alternatives during brainstorming, travel planning, or other creative workflows. The newsletter frames this as a gap between models optimized for correctness or safety and the kinds of outputs people want when they ask open-ended questions.

Flint’s stated approach addresses that gap by prioritizing response variety, which could change how people use chatbots for discovery and creative tasks. If successful, it would make conversational agents more helpful when the goal is exploration rather than a single authoritative answer.

What to watch

Watch whether Flint consistently produces a wider spread of plausible answers on the sorts of open-ended prompts used in the newsletter, and whether those outputs outperform mainstream models in user-facing tests. A clear signal will be Flint avoiding the predictable single-answer behavior exemplified by the repeated "7" on random-number prompts and demonstrating measurable variety on planning or brainstorming queries.

Beyond immediate user tests, adoption by teams that run brainstorming or creative workflows will be the next practical indicator of whether a variety-first training approach sticks. The newsletter does not provide launch dates, benchmarks, or technical specifics, so further announcements from Springboards will be the primary source for confirming Flint’s impact.