Who is the real king of the gaming world? AI actually challenged the classic game "Super Mario Brothers"! A surprising battle report came from the Hao Artificial Intelligence Laboratory of the University of California, San Diego: In a unique AI "Malio" battle, Anthropic's Claude3.7 model "leading ahead" and surpassed the heroes and won the throne of "the strongest AI Mario"! Following closely is the junior brother of the same school, Claude3.5, while Google Gemini1.5Pro and OpenAI's GPT-4o, two "AI giants" unexpectedly "fall out" and their performance was surprising! What's going on?
This AI "Malio" tournament was not played on the ancient red and white machine, but in a "high-tech" simulator. Researchers have specially created a framework called GamingAgent to serve as a "bridge" between AI and the gaming world. In this virtual world, AI becomes "Malio", holding the "game controller" in hand, and receives "combat commands" from the system: "There is an obstacle ahead! Jump!", "The enemy is coming! Go away!", the commands are concise and clear, but also full of challenges. The system will also "concertically" send game screenshots to help "AI Mario" "see the six directions and listen to the world" and better "control" the situation. What’s even cooler is that AI can actually “write” Python code on the spot, directing “Malio” to make various “sexy operations”, jumping up and down, avoiding obstacles, and it’s simply “showing”!

However, the "battle situation" on the field was unexpected. Those AI models that are "inexperienced and well-known for their "reasoning ability", such as OpenAI's o1, actually "stumbled" and performed worse than some "non-inference" players! Why is this? It turns out that "reasoning masters" also have "fatal weaknesses" - "too slow to react"! In a real-time game like "super Mario Brothers", the "reasoning model" takes several seconds to "thought carefully" to make decisions "slowly", but "fighting shots are fleeting" and a second "hesitation" may lead to "Malio" and "death"! It seems that in the ever-changing game world, "reaction speed" is the "hard truth"!
Although games have long become an "important stage" for AI competition, some experts have "look at it differently". In their opinion, the game world is a "virtual world" after all, compared with the "real world", it is still "Too young, Too simple"! The game environment is "too simple" and "abstract", and AI can "brush experience values" in an infinite amount and accumulate "theoretical data", but "the talk on paper is always shallow". The "real ability" of these AI models remains to be "tested in practice". OpenAI research scientist Andre Kapasi even issued a "soul question" of "assessment crisis", which makes people "fall into deep thought".
However, no matter how questioning is, watching AI’s “fancy play” Super Mary is still a “technology show” that is pleasing to the eye. It vividly shows the “change of AI technology” and also opens a “window” for us to “sight into the future”. Who would have thought that AI, which once could only "strategy" on the "chessboard", can now "show its skills" in the "game world"? Perhaps in the near future, AI can really "rule" the game world, or even "transcend" human players, and become the "real king" of the "game world"! Let's wait and see!