Hello, tech enthusiasts! Today, we’re diving into an exciting intersection of nostalgia and innovation.

Researchers at the University of California San Diego recently challenged various AI models to conquer the classic Super Mario Bros. game, claiming it’s a tougher benchmark than you might think.

Leading the pack was Anthropic’s Claude 3.7, which outperformed its predecessors and other notable models like Google’s Gemini 1.5 Pro and OpenAI’s GPT-4o.

This wasn’t just a nostalgic trip down memory lane; the game was run in an emulator that utilized a specialized framework called GamingAgent, allowing AI agents to maneuver through levels.

The AI was guided with basic controls, instructing it to avoid obstacles or enemies and adapt its tactics based on in-game conditions. This approach forced the AI to learn real-time decision-making, crucial for mastering gameplay.

Interestingly, the results highlighted a surprising trend: reasoning AIs, designed to think step-by-step, lagged behind non-reasoning models. The reason? In the fast-paced world of Super Mario, split-second reactions are vital.

For years, gaming has been a testing ground for AI capabilities, yet some scholars question its relevance as a true reflection of AI’s abilities in the real world. Games are highly controlled environments that attract critics who suggest they may not accurately portray AI advancements.

Ultimately, watching AI navigate the world of Super Mario offers a thrilling glimpse into the future of technology and gaming.