Back to News
LLMs are stuck in a groupthink groove. This startup is trying to get them out.
AppsWednesday, July 1, 2026 at 03:30 PM

LLMs are stuck in a groupthink groove. This startup is trying to get them out.

Curated Summary

Large language models exhibit a significant 'groupthink' tendency, producing highly predictable and homogeneous responses to open-ended prompts.

An Australian startup named Springboards has developed a new model called Flint designed to break this pattern by encouraging greater diversity in outputs.

The article demonstrates that while mainstream models like ChatGPT and Claude often converge on identical answers, Flint generates varied results for the same queries.

Co-founder Pip Bingemann argues that this homogeneity limits creativity, whereas Flint welcomes the variation that others might consider hallucinations.

This issue is gaining attention as researchers publish papers highlighting the artificial hivemind effect across different AI systems.

The startup aims to provide a solution for users seeking novel ideas rather than the most statistically probable answers.

This development highlights a growing niche for AI tools focused on creativity and divergence from standard model behaviors.

Want the full story?

Read Original Article

via MIT Technology Review

#llm#startup#creativity#springboards#groupthink

Related Articles

Meta has a new app called Pocket that is absolutely nothing like the old Pocket
The Verge AI
AppsYesterday

Meta has a new app called Pocket that is absolutely nothing like the old Pocket

Meta is launching a new application named Pocket, which bears no relation to the former Mozilla read-it-later service. The app introduces a social feed centered on interactive, AI-generated experiences called 'gizmos.' These gizmos can respond to touch, device tilt, and camera input, allowing users to create and share playable content. The technology appears to stem from Meta's acquisition of licensing rights from Atma Sciences Inc., creators of the original Gizmo app. Mark Zuckerberg envisions this as a key part of his strategy to integrate AI deeply into social media interactions. Currently, the app is not available in the United States, with users seeing region restriction notices on app stores. This launch signals Meta's continued pivot toward AI-driven user-generated content and interactive social experiences.

1 min readRead
Yep, we’re using OpenClaw to date now
TechCrunch AI
Apps2d ago

Yep, we’re using OpenClaw to date now

A content creator named Ben Guez has automated his dating strategy using OpenClaw and Claude to post Instagram reels about World Cup losses, claiming to have received over one million views and 200 DMs. Guez uses the AI agent to trigger nearly identical posts targeting women from specific countries, directing them to his AI language learning app, Canary. He argues that recipients are impressed by the creativity rather than feeling played, provided the automation is transparent. Another tech founder, Jeff Weisbein, uses OpenClaw to research date locations in South Florida, streamlining his planning process. While Guez's approach is more outrageous, both examples highlight how AI agents are being repurposed for personal and social tasks. The article notes that TechCrunch could not independently verify the women's reactions, relying on Guez's word. This story illustrates the emerging trend of using autonomous agents for unconventional personal automation.

1 min readRead
The Download: a startup has a solution for AI’s groupthink problem
MIT Technology Review
Apps2d ago

The Download: a startup has a solution for AI’s groupthink problem

MIT Technology Review highlights a new startup named Springboards that has developed an LLM called Flint to combat the groupthink problem in large language models. Mainstream chatbots like Claude and ChatGPT often produce predictable, repetitive responses, such as always choosing the number 7 when asked for a random number between 1 and 10. This lack of creativity hinders brainstorming and planning tasks where diverse ideas are crucial. Flint is specifically trained to generate a wider variety of responses to open-ended questions, offering a potential solution to this limitation. The newsletter also covers other major tech news, including the creation of the first synthetic cell and OpenAI's proposal for government stakes in AI companies. These developments underscore the ongoing industry focus on improving AI creativity and navigating complex regulatory landscapes. The article serves as a daily digest of significant technological advancements and business shifts.

1 min readRead