Startup Springboards trains Flint LLM to break “Give me 7” predictability

Open up your favorite chatbot and ask for a random number between 1 and 10. You’ll often get 7. That pattern is exactly what Springboards is trying to fix with its new LLM, Flint, trained to produce a wider variety of responses for open-ended prompts like “Where should I go in Europe?”

The point is not that today’s large language models cannot answer. They can. The problem is that they can be too good at the “obvious” answer. The Download describes LLMs as stuck in a groupthink groove, and it’s easy to see why: many systems optimize for likely, safe, and coherent completions. For tasks like coding or research, that reliability can be useful. But when the job is brainstorming, planning, or exploring options, predictability turns into a creative bottleneck.

Flint’s pitch is straightforward and, for business teams that use AI daily, pretty consequential: train the model to come up with a wider variety of responses than mainstream LLMs to open-ended questions. In other words, instead of returning the same “best guess” across similar prompts, the model is designed to produce more diverse outputs. That matters because most people do not ask chatbots for one “correct” answer. They ask them to spark possibilities, generate alternatives, and widen the range of what they consider. A chatbot that repeatedly nudges users toward the same default ideas can make the user feel productive while quietly shrinking the idea pool.

This is the difference between “assistant” and “ideation partner,” and it has real implications for how companies deploy AI internally. Many organizations are already using LLMs for structured work, like drafting, summarizing, and generating code. Those are areas where consistency can be a feature. But in product strategy, marketing, hiring, travel planning, or any scenario where you want options rather than one answer, diversity becomes a performance metric.

The stakes are also practical. If your AI tool hands you the same answer repeatedly, you eventually stop trusting it for exploration and start using it only for narrow tasks. That shifts ROI from iterative thinking to just finishing documents. Springboards is betting that improving diversity can keep users in “brainstorm mode,” where the tool is generating new angles rather than reinforcing the most likely one.

Zoom out from Flint and look at how the broader AI landscape is moving. The same newsletter lists multiple developments that show AI is increasingly entangled with policy, compute economics, and security decisions. OpenAI has proposed giving the Trump administration a 5% stake, with talks tied to a public ownership deal, while other US AI giants were discussed as potentially providing a 5% stake too. There are also stories on a $42 million mansion seized in Singapore in an investigation involving alleged illegal trading tied to Nvidia chip smuggling, and reports about export bans and access changes affecting Anthropic’s Fable 5. Meanwhile, Meta is exploring monetization of AI compute and models, including selling access to models hosted on Meta’s infrastructure and selling “raw” computing power.

Why does this matter for a small Australian startup focused on “groupthink”? Because it highlights that AI deployments are no longer just about model quality in isolation. Boards and executives are managing risk across the stack: regulatory pressure, supply chain integrity, model access, and infrastructure control. In that environment, features like response diversity can become part of governance and user safety conversations, not just product differentiation. A model that stays within a narrow lane of predictable answers may be easier to evaluate, but it also may fail at the very jobs leadership expects AI to help with, like faster ideation and better planning.

Flint’s approach is also a reminder that “smarter” is not the only dimension that counts. Sometimes the winning move is “less like a consensus machine.” If LLMs are trained to maximize the likelihood of coherent text, they can drift toward the same outputs again and again. That’s the groupthink groove the newsletter calls out, with the “Give me a random number between 1 and 10” example making the rut feel immediate. Flint is trying to step around that by training for variety in open-ended responses.

For decision-makers evaluating AI tools, the strategic question becomes simple: what problem are you actually solving? If your workflow is about execution, predictability may be fine. If your workflow is about exploration, a model that repeatedly answers the same way can be a silent productivity tax. Springboards’ Flint is one attempt to measure and address that gap, and it sets a marker for where competitive differentiation may go next: from raw capability to controlled creativity, especially in the moments where teams are trying to think beyond the default.