What happens when modern AI models try to play Super Mario Bros.? The results are more surprising than you’d expect!
Good news: In the future, AI won’t just take over your job – it’s making moves in the gaming world, too. Researchers have been experimenting with Super Mario to see how well AI models handle fast-paced platforming challenges. In this case, no one really cares about the fact that you can’t say "no animals were harmed during this experiment" – since empathy for Goombas tends to be pretty low. The results revealed surprising differences in how AI approaches real-time decision-making. Here’s everything you need to know about the latest AI news.
AI vs. Mario
Researchers put modern AI models to the test in Super Mario Bros., making them jump, dodge, and survive classic platforming chaos. The idea? See if these digital brains can react in real-time or if they overthink themselves into a pixelated grave. Surprisingly, Claude 3.7 dominated, pulling off precise jumps and smart evasions, while GPT-4o and Gemini 1.5 Pro tripped over their own logic.
Claude-3.7 was tested on Pokemon Red, but what about more real-time games like Super Mario ? We threw AI gaming agents into LIVE Super Mario games and found Claude-3.7 outperformed other models with simple heuristics. Claude-3.5 is also strong, but less capable of... pic.twitter.com/bqZVblwqX3
— Hao AI Lab (@haoailab) February 28, 2025
Why did some AIs struggle so much? It all comes down to speed vs. strategy. While Claude reacted quickly and adapted, GPT-4o tried to "think" through every move – wasting precious milliseconds. This overthinking ends up looking like Mario having an existential crisis before running straight into a Goomba. In Mario games, hesitation equals death, and no amount of raw processing power can fix that. So, are the older AIs better? Well, Henry Ford said, "If I had asked people what they wanted, they would have said faster horses."
Fast Reflexes Beat Smart Thinking
You’d think an AI that can write essays and solve math problems would have no trouble hopping over a few Koopas. Wrong. While models like GPT-4o and Gemini are great at logic, they got wrecked by Mario’s fast-paced demands. So in the future, if you dare to open another supermarket called Super Mario, a fast-thinking Claude 3.7 robot will be on the run chasing you – while a more intelligent and reasonable GPT-4o robot sues the living s*** out of you in court. What a lovely future.
So, because AI can't even play the simplest game, we're pretty much safe? You guessed it – wrong. The problem is that AI is developing at a rapid pace thanks to machine learning, which basically means we’re feeding it information, making it even faster and more intelligent than us. State-of-the-art benchmarks have actually proven to surpass 100% human performance in reading comprehension, language understanding, and speech recognition – and this was already the case in 2018.
Think about this: When humans left Africa for the first time thousands of years ago, our jaws got smaller because food became softer. But our wisdom teeth stayed, even though there was no more room for them – which is why you had that painful wisdom tooth removal. You really think an AI-driven robot wouldn’t just drop wisdom teeth in a split second after analyzing their uselessness? I also don’t know what that has to do with Mario, I just thought about it. And now you do, too.
Should AI prioritize quick reflexes over deep thinking, or does strategy always win in the end? Tell us in the comments!