New research from New York University highlights that while artificial intelligence excels at specific, rule-based games, it still struggles to learn new, open-ended titles as quickly as humans.

Key Points

• AI models like Deep Blue and AlphaGo have historically mastered structured games through reinforcement learning and millions of simulated trial-and-error iterations.
• NYU professor Julian Togelius argues that current AI models lack the ability to generalize, often failing when faced with unfamiliar or abstract game mechanics.
• Humans possess a significant advantage in learning new games due to years of lived experience and an intuitive understanding of physical objects and concepts.
• While Google DeepMind’s SIMA 2 model shows progress in 3D environments, it remains far from the researchers' proposed benchmark of mastering 100 new games rapidly.
• The study suggests that true human-level intelligence requires abstract thinking and creativity, which current AI architectures have yet to achieve.

Why it Matters

The inability of AI to adapt to new, open-ended environments suggests that current models are optimized for narrow tasks rather than general intelligence. Mastering this challenge would require breakthroughs in reasoning and planning, which are essential for developing more versatile and capable artificial intelligence systems.