LLMs Stuck in Groupthink: How Australian Startup Springboards Is Building Flint to Diversify AI
MIT Technology Review reveals that LLMs produce alarmingly similar answers. Australian startup Springboards is building Flint — an alternative model trained to generate more varied responses.
"Artificial Hivemind" — NeurIPS Best Paper
In November 2025, a research paper titled "Artificial Hivemind" won Best Paper at NeurIPS. Researchers tested 25 different LLMs 50 times each asking for a time metaphor. Out of 1,250 responses, the vast majority were "Time is a river" or "Time is a weaver."
The Mystery of Number 7
When asked for a random number between 1 and 10, nearly every LLM gives "7." This extreme response homogeneity is a known problem — LLMs converge on familiar patterns rather than producing genuine diversity.
Springboards and Flint
Australian startup Springboards built Flint, based on Alibaba's open-source Qwen 3, trained specifically to produce more varied responses. Cofounders Pip Bingemann (CEO) and Kieran Browne (CTO) argue the problem stems from models trained on similar data with similar objectives.
Creative Industry Testing
Zoe Scaman (Bodacious/77X) and Maximilian Weigl (Uncommon marketing) have tested Flint and find it useful for creative brainstorming, though it remains a prototype.