Artificial intelligence (AI) has made incredible strides in recent years, but according to Yann LeCun, one of the field's leading figures, current models fall short when it comes to real-world applications.

At Meta for a decade as chief AI scientist, he now leads AMI Labs, focusing on developing more flexible and adaptable forms of AI. Unlike Large Language Models (LLMs) like ChatGPT, which excel in predictable tasks but struggle with complex scenarios, LeCun’s team is working on Joint Embedding Predictive Architecture (JEPA).

‘We don’t have robots that are nearly as good at understanding the physical world as a rat,’ says LeCun. His company aims to create abstractions of the real world, enabling more intelligent decision-making in unpredictable situations.

The robotics industry is keen on this development, with billions invested in humanoid robots for household tasks. However, training them remains challenging and costly, underscoring the limitations of current AI models.

Ingmar Posner from Oxford University agrees, promoting the next decade as a time for systems that can explain their actions and reasoning. His team is working on world models—AI that learns through mental simulations, hoping to take a significant step forward in this area.