How Should AI Learn to Understand the World? | Yann LeCun & Eric Xing on JEPA and GLP

Artificial intelligence development currently centers on the design of "World Models" capable of simulating environments for reasoning and planning. Eric Xing proposes a generative latent prediction (JLP) architecture, which utilizes both symbolic and continuous representations to simulate complex scenarios, validating these simulations through generative reconstruction. Conversely, Yann LeCun argues that effective world models must be non-generative, utilizing Joint Embedding Predictive Architectures (JEPA) to focus on abstract, hierarchical representations. By ignoring unpredictable details—effectively treating them as entropy—JEPA models avoid the computational burden and inaccuracies inherent in pixel-level generation. The debate underscores a fundamental divergence in methodology: whether to preserve comprehensive data for validation or to prioritize efficiency and abstraction by discarding unpredictable information. Ultimately, both experts agree that moving beyond static language models toward agentic systems requires robust, simulation-based reasoning to achieve human-level adaptability.