The Sequence Knowledge #808: Stop Trying to Generate the World: Inside the JEPA Way for World Models
Yann LeCun's foundation for World Models
Today we will Discuss:
An overview of the JEPA methods for world models.
A review of the three famous JEPA papers.
💡 AI Concept of the Day: Understanding LeCun’s JEPA Methods to Reimagine World Models
In the midst of the “Generative AI” boom—where OpenAI’s Sora and Runway are dazzling the world by hallucinating video—Yann LeCun, AI legend and former Meta’s Chief AI Scientist, has taken a staunchly contrarian stance. His argument is simple but radical: If you want true intelligence, stop trying to generate the world.
LeCun proposes a different path, known as the Joint Embedding Predictive Architecture (JEPA).1 While Generative World Models (like Dreamer) prove understanding by recreating the image, JEPA proves understanding by predicting the concept.
The Problem with Pixels