Will Knight tested specialized AI agents to manage a fantasy Premier League team, finding that task division improved decisions while scaling past 100 agents only inflated API costs.
At first, the system made some odd choices, like leaving Erling Haaland on the bench. But when Knight split the programs into different roles—some reading news and others analyzing results—the suggestions improved. However, increasing the number of programs to over a hundred didn’t make the system any smarter, it just increased the cost.
Researchers at the University of Chicago, led by Haifeng Xu, tested commercial AI programs on prediction platforms. In tests, GPT-5 performed best, followed by Grok 4, Claude Sonnet, Gemini, and Llama. Special evaluation systems have been created to measure these capabilities, such as Prophet Arena and Forecastbench. In another test, known as Prediction Arena and implemented by Harvard University and Arcada Labs, programs were given $10.000 to bet on prediction platforms for 57 days.
Despite advances, AI often lags behind human forecasters, who have a deeper understanding of real life. However, competitions like the Metaculus Cup show that AI can sometimes beat even the best human forecasters.
↳
READ ALSO
HACKS.GR / RELATED
AI Unlocked
Microsoft ends support for Copilot on TVs
↗
As for Knight's experiment, the fantasy team managed by artificial intelligence has failed to surpass the scoring average recently.
G
HACKS.GR / GOOGLE SEARCH
Hacks.gr on Google
Set Hacks.gr as a preferred source.
Preferred source
↗
end of article
Top comments (0)