What do all intelligent beings share?According to Yann LeCun, Meta's chief AI scientist, there are four key traits. At the AI Action Summit in Paris earlier this year, political leaders and AI experts gathered to discuss AI development. During the event, LeCun conveyed his fundamental definition of intelligence to IBM’s AI leader, Anthony Annunziata. “There are four essential characteristics of intelligent behavior that every animal—or relatively smart animal—and certainly humans possess, ” he explained. “These are understanding the physical world, having persistent memory, the ability to reason, and the capability to plan, especially complex actions through hierarchical planning. ” LeCun pointed out that AI, particularly large language models, have yet to reach this level and that integrating these abilities would require a shift in their training methods. This explains why many leading tech companies are currently attaching new capabilities onto existing models as they compete to lead the AI field. “For understanding the physical world, you train a separate vision system and then attach it to the large language model. For memory, you use retrieval augmented generation (RAG), add some associative memory, or simply scale up the model, ” he said.

(RAG is a technique developed at Meta to improve large language models by incorporating external knowledge sources. ) Nonetheless, LeCun considers all those efforts to be mere “hacks. ” He has repeatedly discussed an alternative approach called world-based models, which are trained on real-world scenarios and exhibit higher cognitive abilities beyond mere pattern recognition. In his conversation with Annunziata, he elaborated on this concept. “You start with some idea of the world's state at a particular time T, imagine an action, and then the world model predicts how the world’s state will change in response to that action, ” he explained. However, since the world evolves through countless unpredictable possibilities, the only feasible way to train such models is through abstraction. Meta is already exploring this concept with V-JEPA, a model released publicly in February. It is described as a non-generative model that learns by predicting missing or masked segments in videos. “The core idea is that you don’t predict at the pixel level directly. Instead, you train a system to operate on an abstract representation of the video, allowing it to make predictions within that abstraction. Ideally, this representation filters out unpredictable details, ” LeCun said. This approach mirrors how chemists established a fundamental hierarchy for matter’s building blocks. “We created abstractions: particles, on top of which are atoms, then molecules, and eventually materials, ” he said. “Each layer up filters out much irrelevant information from the lower layers, depending on the task at hand. ” In essence, what this means is that we understand the physical world by constructing hierarchies, which is fundamental to intelligence.