JEPA
Joint Embedding Predictive Architecture: an AI model that learns abstract concepts by predicting missing parts of data.
A Joint Embedding Predictive Architecture, or JEPA, is a way to train AI models. Instead of just predicting the next word in a sentence, it tries to understand the underlying meaning or concepts in data. It does this by learning to predict missing parts of information, like filling in the blanks in a picture or a sentence.
Think of it like a child learning about the world. They don't just memorize facts; they build an understanding of how things work. A JEPA aims for this kind of understanding. For example, it might see a picture of a cat and then try to predict what the cat would look like from a different angle, or what sound it might make, based on its learned concepts.
A Joint Embedding Predictive Architecture (JEPA) is a self-supervised learning framework that aims to build world models by learning abstract representations of data. Unlike autoregressive models that predict raw data (e.g., pixels or tokens), JEPA models learn to predict abstract representations in a latent space. This is achieved by training two encoders: one that processes observed data and another that processes missing data, and a predictor that maps the representation of observed data to the representation of missing data. The goal is to learn representations that capture underlying semantic information rather than surface-level details, potentially leading to more generalizable and efficient AI.
Yann LeCun, a prominent AI researcher, has recently argued that current large language models (LLMs) are a dead end. He suggests that JEPA offers a more promising path towards building AI that truly understands the world, moving beyond pattern matching to a deeper conceptual learning.
- Ep 86: Yann LeCun on Leaving Meta, Breaking The LLM Paradigm, & Why Hinton is WrongMeta · Unsupervised Learning ·
- Meta Meta AI
- Meta AI is actively researching and developing JEPA architectures, with Yann LeCun being a key proponent of this approach for future AI development.
- Autoregressive Models
- These models predict the next token or pixel sequentially
- Contrastive Learning
- While both are self-supervised methods, contrastive learni
The concept of Joint Embedding Predictive Architectures was introduced by researchers at Meta AI, including Yann LeCun, as a proposed alternative to current dominant AI paradigms. Source ↗
While directly experimenting with JEPA models is complex, you can observe the principle of predicting missing information in image editing tools. Try using an AI-powered 'inpainting' feature in an image editor to fill in a missing part of a photo; the AI is essentially performing a form of predictive modeling based on learned visual concepts.
The Playbook: reusable how-tos, prompts, configs and tips pulled out of the news.
- PromptRole-play a student struggling with math for AI tutor simulationSimulate a student's math struggles and uncertainty for AI tutor evaluation. · 5 min
- TemplateReusable workflow templates for healthcare teamsTeams can create and share templates for common workflows like documentation or prior authorizations. · 15 min
- TemplateTemplate for building an AI tool tailored to your workA structured approach to defining and building a custom AI tool for specific business needs. · 15 min
- TemplateEvals CSV prompt list for agent skillsA CSV file template for listing prompts used in evaluating agent skills. · 20 min
- HeuristicDouble coding agent task completion with docs linksProviding coding agents with a documentation link significantly increases their task completion rate. · 2 min
Questions people ask
- What is the main goal of JEPA?
- The main goal is to train AI models that learn abstract representations of the world, enabling them to understand concepts rather than just memorizing patterns.
- How is JEPA different from LLMs?
- JEPA focuses on learning world models through predicting abstract representations of missing data, aiming for deeper understanding, while LLMs primarily predict the next token in a sequence.
First seen on Subvolts May 15, 2026. Updated Sep 17, 2026. Written by an AI model from the items above and reviewed against them weekly; report an error.