NextFin News - Large language models are built to be helpful, coherent and safe. In practice, that often means they drift toward the same familiar answers when users ask for something open-ended: the same random number, the same metaphor, the same brand-name pattern. That sameness is now the premise behind a new product pitch from the Australian startup Springboards, which has built Flint as an alternative model for creative work. Its bet is straightforward: if mainstream chatbots keep compressing ideas toward the median, then there is commercial value in widening the range of what they produce.
The case is not just anecdotal. A NeurIPS 2025 best paper titled Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond) found that open-ended prompts can trigger both repetition within a single model and striking similarity across different models. The paper used the Infinity-Chat dataset of 26,000 real-world open-ended queries and reported more than 31,000 human annotations. Springboards is positioning Flint around the same friction point: for advertising and marketing teams, the issue is not whether a model can be accurate and polished, but whether it can surface enough distinct starting points to be useful in a brainstorming workflow.
Springboards Is Betting That Model Sameness Is A Product Problem
The startup’s pitch starts with a familiar frustration. Ask several chatbots for a random number between 1 and 10 and they often converge on the same handful of answers. Ask for a car brand or a campaign tagline and the outputs can cluster around the same high-probability tropes. In the MIT Technology Review article that introduced Flint, Springboards cofounder and chief executive Pip Bingemann used that phenomenon to make a broader point: mainstream models are optimized for reliability, but that optimization can flatten variation when the task calls for creativity.
For Springboards, that is not a trivial quirk. It is the commercial opening. The company says Flint is meant for advertising and marketing teams who want better ideas, not just faster answers. On Springboards’ own website, the company says it is designed to widen the field, and claims its responses covered 10–30x more creative territory than leading general-purpose LLMs for common advertising and marketing tasks.
That claim should be read as a product promise rather than a settled industry benchmark. But it helps clarify the company’s positioning: Springboards is not trying to replace general-purpose chatbots on every task. It is trying to build a creative system that behaves differently when the goal is ideation, not retrieval.
“Most language models are fighting hallucinations,” said Springboards cofounder and chief executive Pip Bingemann. “We welcome them.”
That line captures the startup’s strategic wager. For research, coding and other tasks with a single right answer, repeated high-confidence outputs can be a feature. For brainstorming, they can be a ceiling. Springboards is arguing that the market has over-optimized for the first use case and under-served the second.
The Research Backdrop Suggests The Problem Runs Deeper Than Prompting Tricks
The academic backdrop gives the pitch more weight. The Artificial Hivemind paper did not merely show that one model can be repetitive. It argued that different models can converge on similar outputs even when there is no single correct response. Using Infinity-Chat, the researchers studied more than 70 large language models and found both intra-model repetition and inter-model homogeneity, especially on open-ended prompts that admit many plausible answers.
The dataset itself is part of the point. Infinity-Chat includes 26,000 real-world, open-ended user queries and 31,250 human annotations, giving researchers a large sample of the kind of prompts people actually use when they want help generating ideas rather than checking facts. The paper’s message is that the problem is not limited to toy examples like random numbers or naming tasks. It appears across broader categories of creative generation and ideation.
That helps explain why Springboards’ argument resonates. If the output distribution of major models is already converging toward similar answers, then asking users to prompt harder may not solve the underlying issue. The startup’s thesis is that the model layer itself should be tuned for variety in contexts where the first answer is only the beginning.
“The way that most chat interfaces are designed, it makes it feel like you’re having a personal conversation,” said Kieran Browne, Springboards’ cofounder and chief technology officer. “I think most people don’t really realize the extent to which they are getting the same stuff as everybody else.”
That observation matters because interface design can hide homogeneity. A user feels one-to-one interaction, but the underlying model may still be pulling toward the same common pathways that millions of other users see. The result is a kind of invisible standardization: the system feels personal, but the ideas are increasingly generic.
Why The Business Case Depends On Human Editing, Not Autonomous Creativity
Springboards’ strongest argument is also its most practical one. Creative teams do not usually need a machine to finish the job. They need a machine to widen the starting set so a human can select, recombine and refine. That makes Flint more defensible as a workflow tool than as a standalone creative agent.
The company’s site and product framing point in that direction. Springboards describes itself as an AI platform for advertising and marketing teams and says it is meant to inspire, not simply give answers. In that sense, Flint is less a rival to mainstream chatbots than a specialization layer built for concept generation.
The economics of that specialization are plausible. In advertising, strategy and brand work, a first-pass idea that is merely acceptable is often not enough. Teams want range: more tonal options, more metaphors, more directions. If a model can surface more distinct paths in the first round, it may save time downstream even if the final output still depends on human taste.
But the trade-off is real. The more a model is tuned to escape the most probable answer, the more it risks drifting away from reliability. Springboards is betting that creative users will accept that trade-off because they already know how to evaluate and edit. For them, a little unpredictability may be a feature, not a flaw.
That is also why the company’s pitch lands in a narrow but important slice of the AI market. General-purpose systems are still winning on breadth. A creative specialist has to win on difference. Flint’s difference is not that it knows more. It is that it is willing to sound less like everyone else at the first draft stage.
What To Watch Next
The next test is whether Flint can prove that variety is worth paying for. If Springboards can show that its outputs consistently create more useful concept sets for agencies and marketers, it could carve out a durable niche. If the model simply produces noisier answers, the product will be hard to distinguish from prompting tricks, sampling settings or manual ideation.
The bigger implication is that model competition may become more segmented. One system can be best for correctness, another for brainstorming, another for editing and another for planning. If that happens, the market will be defined less by a single universal chatbot and more by a set of specialized behaviors wrapped in different products.
That is the real lesson in Springboards’ launch. The startup is not claiming that large language models are useless or broken. It is arguing that the industry has mistaken sameness for quality in at least one important context. In creative work, the danger is not only wrong answers. It is too many polished answers that all sound alike.
The question Flint raises is simple: if the best answer is usually the same answer, how much creative value is being left on the table? Springboards is betting that enough people will pay to find out.
Explore more exclusive insights at nextfin.ai.
