Codex-run fantasy team trails the average, WIRED finds

Knight's experiment sits between two poles. A few years ago, a Turing Institute tool picked fantasy teams from historical league data alone, ignoring injuries and managerial changes, and its suggestions were "remarkably bonkers." This time, Codex could pull in recent news and run models, and a division of labor among agents—some reading news and blog posts, others analyzing results, running models, or critiquing suggestions—visibly lifted quality. Scaling to more than 100 agents did not add much beyond cost.

The wider evidence points the same way. University of Chicago computer scientist Haifeng Xu, who studies AI in market settings, says models are getting better because they can consume huge amounts of real-world information and orchestrate more complex guessing systems. In his tests on prediction markets, models sometimes beat the average person with no special guidance, with GPT-5 best, followed by Grok 4, Claude Sonnet, Gemini, and Llama. A separate benchmark, Prediction Arena, gave frontier models $10,000 to wager over 57 days; they consistently lost money but improved over time.

Humans still hold an edge in depth of understanding, Xu notes, though that edge looks like it is eroding—the Metaculus Cup has shown AI sometimes beating the best human predictors, and ForecastBench results are pending. Knight's own team is a reminder that the gap has not closed yet: clever agent harnesses can produce smarter suggestions without producing a winning record.