The Startup Teaching AI To Give Deliberately Weirder Answers

The News

An Australian startup called Springboards has released a large language model built for a counter-intuitive purpose: to be unpredictable on purpose. The model, named Flint, is designed to answer open-ended prompts with genuine variety rather than the safe, homogenised responses that dominate mainstream chatbots.

Flint is built on top of Qwen 3, the open-source model family from Alibaba, and the founders have retrained it to inject randomness at the precise points in a response where variety actually helps. The company is led by cofounder and chief executive Pip Bingemann and cofounder and chief technology officer Kieran Browne, with Maximilian Weigl, a cofounder and chief strategy officer at partner firm Uncommon, shaping how the tool is pitched to creative teams.

The demonstrations are pointed. Asked to pick a random number between one and ten, ChatGPT and Claude reliably return 7. Flint, in one demo, answered 3.7916. Asked for advertising taglines for New Balance, both Claude and ChatGPT produced "Run your way", while Flint offered a different line entirely. "Most language models are fighting hallucinations. We welcome them," Bingemann told MIT Technology Review, which first reported the launch.

Why It Matters

The problem Flint targets has a name in research circles: mode collapse. A paper titled "Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)" won best paper at NeurIPS 2025, documenting how models cluster around a handful of predictable outputs. When researchers prompted 25 different models 50 times each to write metaphors about time, most answers were minor variations of "time is a river".

This matters because the industry has spent three years optimising for the opposite trait. Since the arrival of instruction-tuned assistants, reliability training has pushed models towards the single highest-probability answer, which is exactly why they feel trustworthy for coding or summarising a contract. OpenAI itself has acknowledged that chasing novelty can erode coherence. Springboards is betting that a whole category of work, from brainstorming to campaign ideation, is being underserved by that trade-off. Weigl describes Flint as "an invitation to think wider".

Indian Angle

For India's foundational-model builders, Flint is a useful signal about where differentiation may lie. Companies such as Sarvam AI and Ola-backed Krutrim have largely framed their pitch around Indian languages and sovereign data. Flint suggests a second axis of competition: not just what a model knows, but how varied its output can be. An Indian model that handles Hindi, Tamil and Marathi creative copy with genuine range, rather than translated blandness, could carve a niche that global incumbents ignore.

There is also a cost story that lands squarely in India. Flint is fine-tuned on open-source Qwen 3, not built from scratch, which keeps compute bills low. That is precisely the playbook Indian startups have adopted to sidestep the capital intensity of frontier training, and it validates the approach for founders weighing GPU budgets in rupees.

Finally, India's advertising and marketing industry, one of the largest creative workforces in the world, is the exact buyer Springboards is chasing. Agencies in Mumbai and Gurugram experimenting with generative tools have quietly complained that outputs feel generic. A model tuned for divergence, rather than the safest answer, speaks directly to that frustration.

FAQ

What is Flint built on?

Flint is fine-tuned on Qwen 3, Alibaba's open-source model family. Springboards did not train a model from scratch; it retrained an existing one to add controlled randomness at chosen points in a response, which keeps development costs manageable.

Is this useful for coding or research?

No. The founders concede that mainstream models remain better for tasks where a single correct or reliable answer matters, such as programming or factual research. Flint is aimed at open-ended, creative work like brainstorming and campaign ideation.

How is this different from raising the temperature setting?

Turning up temperature makes any model more random but also more incoherent. Springboards says it instead taught Flint to identify where variety adds value and inject it selectively, preserving readability.

Where can I read the original report?

MIT Technology Review published the launch coverage, written by Will Douglas Heaven, on 1 July 2026.

This story was reported by MIT Technology Review. Read the full original coverage at MIT Technology Review.