A British AI startup is challenging the notion that large language models (LLMs) are mere mimics, after its system demonstrated what its co-founder calls “genuine reasoning” by placing eighth in a global forecasting competition. ManticAI, a company with Google DeepMind heritage, outmaneuvered many human experts in the Metaculus Cup, a rigorous contest of predictive skill.
The competition involved forecasting the outcomes of 60 real-world events over the summer. The questions were diverse, touching on politics, technology, and environmental science, requiring a deep and flexible analytical ability. The AI’s success in this arena suggests a significant evolution from regurgitating known facts to reasoning about unknown futures.
Toby Shevlane, co-founder of ManticAI, explained that their system’s high performance is a milestone for the AI community. “You can’t predict the future” by simply repeating training data, he stated. The system’s architecture supports this claim; it uses a roster of different AI models from OpenAI, Google, and DeepSeek, assigning them to specialized tasks like historical research and scenario modeling to build a comprehensive view.
One of the most intriguing aspects of the AI’s performance was its independence of thought. Shevlane noted that the system often “strongly disagreed” with the average predictions made by human participants. This tendency to avoid the “community average” suggests AI forecasters could be a powerful tool for breaking through human groupthink and uncovering novel insights.
The human forecasting community is taking note. While many, like superforecasting pioneer Philip Tetlock, argue that elite humans still have the upper hand, the rapid improvement is startling. The conversation is now shifting from “if” AI will catch up to “when,” and how human experts can best leverage these powerful new tools in a collaborative model for even greater accuracy.