A study led by Meta's Head of AI Yann LeCun points to a different route for building AI systems that understand the physical world. Instead of training a model to generate every visible pixel of a future scene, the work shows that an AI system can learn basic physics from video by predicting in a more abstract representation space.
The findings matter because they test a central question in AI research: whether systems need physical rules built in from the start, or whether they can acquire useful physical intuition through observation. In this case, the research team reports that V-JEPA learned concepts such as object permanence, continuity, and shape consistency without pre-programmed rules.
What V-JEPA is trying to learn
The system at the center of the study is Video Joint Embedding Predictive Architecture, or V-JEPA. It is part of Meta's broader JEPA research program, which LeCun has promoted as an alternative to generative AI systems such as GPT-4 or Sora for developing world models.
The key distinction is how the system predicts. Models such as OpenAI's Sora are associated in the source with pixel-perfect generation. V-JEPA does not attempt to forecast a scene by recreating every pixel. It makes predictions in an abstract representation space.
That design reflects LeCun's view of how useful intelligence may work. The source describes the approach as closer to how he believes the human brain processes information: not by producing exact visual replicas at every moment, but by operating over higher-level representations of what is happening.
For AI, that difference is important. A model that understands a scene only as pixels may still struggle with the underlying structure of the world. A model that learns abstract relationships may be better positioned to represent objects, motion, and physical consistency.
How the study tested physics understanding
The research team included scientists from Meta FAIR, University Gustave Eiffel, and EHESS. To evaluate whether V-JEPA had learned physical intuition, they used a method from developmental psychology known as "Violation of Expectation."
This method was originally used to study infants' understanding of physics. The basic idea is to show two related scenes: one that could happen in the physical world and one that violates ordinary physical expectations. The source gives the example of a ball rolling through a wall.
In human studies, researchers look for signs of surprise when a subject sees the impossible event. In this AI study, the same logic is applied to the system's responses. If the model distinguishes the possible scene from the impossible one, that suggests it has learned something about how objects should behave.
The evaluation covered three datasets:
- IntPhys, used for basic physics concepts.
- GRASP, used for complex interactions.
- InfLevel, used for realistic environments.
Across these tests, V-JEPA showed particular strength on object permanence, continuity, and shape consistency. These are basic but important concepts. Object permanence means an object can still exist even when it is not directly visible. Continuity relates to objects moving in physically coherent ways. Shape consistency concerns whether an object keeps a stable form rather than changing arbitrarily.
Why the results challenge other AI approaches
The source contrasts V-JEPA's performance with large multimodal language models such as Gemini 1.5 Pro and Qwen2-VL-72B. Those systems did not perform much better than chance in the reported comparison.
That result is significant because large multimodal language models can process visual information, but the study suggests that this ability alone does not guarantee basic physical understanding. Seeing images or video as input is not the same as learning a stable model of how objects behave over time.
The study also challenges the assumption that AI systems require pre-programmed "core knowledge" of physical laws. According to the source, V-JEPA shows that this kind of knowledge can be learned through observation alone.
The comparison to infants, primates, and young birds is central to the argument. These biological systems appear to develop some understanding of the physical world by watching and interacting with their environments. The study suggests that a similar learning principle may be useful for AI.
Efficiency strengthens the case
One of the most notable details in the source is the amount of training data involved. V-JEPA needed just 128 hours of video to grasp basic physics concepts. The source also notes that smaller models with only 115 million parameters showed strong results.
Those numbers make the finding harder to dismiss as only a matter of scale. If basic physical intuition can emerge from a relatively limited amount of video training, then the architecture itself becomes an important part of the story.
This does not mean the study proves that V-JEPA has complete real-world understanding. The source frames the results around basic physics concepts and specific benchmarks. But it does suggest that self-supervised video training can produce meaningful physical expectations without hand-coded rules.
Where this fits in LeCun's AI vision
The study supports LeCun's broader argument that world models may require a different foundation from mainstream generative AI. The source says he considers pixel-perfect generation, like Sora's approach, a dead end for developing world models.
Instead, LeCun advocates hierarchically stacked JEPA modules that make predictions at different levels of abstraction. The stated goal is to create comprehensive world models that help autonomous AI systems develop deeper environmental understanding.
Meta's JEPA work has already included I-JEPA, an image-focused variant, before moving into video with V-JEPA. The move from images to video is important because physics is not only about what objects look like. It is about how they persist, move, interact, and remain coherent across time.
The study's broader implication is straightforward: an AI system may not need to generate the world in full visual detail to learn useful things about it. By predicting in abstract space, V-JEPA offers another way to think about machine understanding, one built around observation, expectation, and structured representations rather than visual reconstruction alone.