At a recent summit in Paris focused on AI policy and progress, Meta’s head of artificial intelligence Yann LeCun shared a stripped-down view of what it truly means to be intelligent. While many companies push boundaries with language models and generative tools, he argued that current systems are missing the fundamentals that even ordinary animals grasp intuitively.
In his view, four abilities lie at the core of genuine intelligence that is understanding the physical surroundings, storing memories that last, reasoning through problems, and planning with structure — especially when the steps require hierarchy. These aren’t abstract goals; they’re everyday survival tools in nature.
Modern AI, he noted, doesn’t yet check those boxes. Instead of developing models that genuinely learn how the world works, most companies are adding extra parts to cover weaknesses. A computer might learn to describe what it sees by adding a separate vision module. Or it might pull facts from databases to simulate memory. But stacking components on top of text-based systems, he warned, doesn’t replicate how thinking minds function.
He believes the fix won’t come from upgrades—it’ll come from changing direction. That means building systems that think in terms of cause and effect. Give a machine a sense of what’s happening right now, let it imagine taking an action, and train it to forecast what changes that action might cause. That loop—observe, act, predict—is how living things adapt.
But life doesn’t follow scripts. Events unfold in unpredictable ways, with details that often don’t matter. So rather than trying to predict everything, he said AI needs to learn abstraction. Humans have done this for centuries. Chemistry, for example, became manageable when scientists stopped thinking about every atom and started thinking in layers—particles, atoms, molecules, then materials. At each level, unnecessary information gets filtered out.
That approach—learning by organizing the world into usable layers—is what Meta’s team is now pursuing. Earlier this year, the company released a research model named V-JEPA. Unlike tools that try to guess every pixel in a picture or frame in a video, V-JEPA focuses on the underlying patterns. It learns by noticing what’s missing and figuring out what should be there, without getting distracted by details that don’t carry meaning.
The long-term hope is to teach machines to think less like machines. Not just to respond with plausible answers, but to build a quiet, internal logic—a map of the world they can use to reason and plan. For now, that goal remains out of reach. But if intelligence can be broken down into parts, Meta seems determined to build the missing ones.
Rea next: How to Use Google Lens for Reverse Image Search on Any Device
In his view, four abilities lie at the core of genuine intelligence that is understanding the physical surroundings, storing memories that last, reasoning through problems, and planning with structure — especially when the steps require hierarchy. These aren’t abstract goals; they’re everyday survival tools in nature.
Modern AI, he noted, doesn’t yet check those boxes. Instead of developing models that genuinely learn how the world works, most companies are adding extra parts to cover weaknesses. A computer might learn to describe what it sees by adding a separate vision module. Or it might pull facts from databases to simulate memory. But stacking components on top of text-based systems, he warned, doesn’t replicate how thinking minds function.
He believes the fix won’t come from upgrades—it’ll come from changing direction. That means building systems that think in terms of cause and effect. Give a machine a sense of what’s happening right now, let it imagine taking an action, and train it to forecast what changes that action might cause. That loop—observe, act, predict—is how living things adapt.
But life doesn’t follow scripts. Events unfold in unpredictable ways, with details that often don’t matter. So rather than trying to predict everything, he said AI needs to learn abstraction. Humans have done this for centuries. Chemistry, for example, became manageable when scientists stopped thinking about every atom and started thinking in layers—particles, atoms, molecules, then materials. At each level, unnecessary information gets filtered out.
That approach—learning by organizing the world into usable layers—is what Meta’s team is now pursuing. Earlier this year, the company released a research model named V-JEPA. Unlike tools that try to guess every pixel in a picture or frame in a video, V-JEPA focuses on the underlying patterns. It learns by noticing what’s missing and figuring out what should be there, without getting distracted by details that don’t carry meaning.
The long-term hope is to teach machines to think less like machines. Not just to respond with plausible answers, but to build a quiet, internal logic—a map of the world they can use to reason and plan. For now, that goal remains out of reach. But if intelligence can be broken down into parts, Meta seems determined to build the missing ones.
Rea next: How to Use Google Lens for Reverse Image Search on Any Device
