British AI startup beats humans in international forecasting competition (www.theguardian.com)

A British AI startup, ManticAI — co‑founded by a former Google DeepMind researcher — placed eighth in the Metaculus Cup, an international forecasting competition that asked entrants to assign probabilities to 60 real‑world events over the summer (questions ranged from election outcomes to acres burned by wildfires). Mantic outperformed many human forecasters, including several professionals, signaling a rapid advance in AI forecasting: its ensemble approach produced original, often contrarian predictions rather than clustering around community averages. Metaculus’s CEO estimates AI could match or beat the best human forecasters by about 2029, though top human “superforecasters” still lead on average today. Technically, Mantic decomposes each forecasting problem into tasks and routes them to a roster of machine‑learning models (including OpenAI, Google, and DeepSeek) and specialized agents that assess current events, run historical research, simulate scenarios, and update predictions daily. That architecture leverages scale and persistence — strengths where AI already excels — but also exposes limits: AI still struggles with logic verification on complex, interdependent forecasts and with sparse-data, judgment‑heavy questions where humans retain an edge. The contest highlights a pragmatic near future: hybrid human+AI teams are likely to produce the most reliable forecasts, with AI serving both as a productivity multiplier and an antidote to groupthink.