Yann LeCun JEPA Models: VL-JEPA, I-JEPA, V-JEPA. Real World Models. Nvidia Cosmos. AI Reasoning LLM скачать в хорошем качестве

Yann LeCun JEPA Models: VL-JEPA, I-JEPA, V-JEPA. Real World Models. Nvidia Cosmos. AI Reasoning LLM

Welcome to the podcast. Today, we are unpacking a fundamental shift in how AI learns to understand the world—moving away from the tedious reconstruction of pixels and towards the prediction of meaning. We are exploring the Joint-Embedding Predictive Architecture, or JEPA. This architecture represents a philosophy that says an AI doesn't need to generate every detail of what it sees; it just needs to understand the representation of it. In this episode, we will trace the evolution of this framework across three distinct generations. First, we look at I-JEPA, the image-based pioneer. It proved that by masking parts of a static image and predicting their abstract features—rather than their pixel values—we could train models to capture high-level semantics without relying on brittle, hand-crafted augmentations. Next, we step into the temporal dimension with V-JEPA. This iteration applies that same pixel-free logic to video, allowing the model to understand motion and time by predicting the representations of missing video segments, creating a powerful standalone learner for dynamic visual data. And finally, we arrive at the cutting edge with VL-JEPA. This is the first non-generative model designed for general vision-language tasks. By combining a V-JEPA visual encoder with a text-based predictor, it aligns vision and language in a unique way: it predicts continuous text embeddings rather than discrete tokens. This separation of semantic prediction from text generation unlocks massive efficiency gains for real-time applications. Comparing VL-JEPA , I-JEPA, V-JEPA model family against Nvidia Cosmos models. Three models, one shared foundation, and a completely new way to think about representation learning. Let’s dive in - VL-JEPA Joint Embedded Predictive Architecture Model Families from META FAIR AI Lab.