World Models with V-JEPA 2: prediction in representation space

Welcome to AI Build & Learn, a weekly AI engineering stream where we pick a new topic and learn by building together.

This event is about world models with V-JEPA 2, Meta's self-supervised video model that learns by predicting in representation space instead of generating pixels. We'll load the model, hide parts of a video clip, watch it predict the missing pieces as embeddings rather than images, and measure how good those predictions actually are.

V-JEPA 2 is trained on internet-scale video with no labels, and it has no decoder at all: the predictor emits vectors, not frames. That constraint makes it a great topic to build around, because "show me what it predicted" stops being a screenshot and starts being a measurement. We'll explore how the model represents video, how to probe those representations with a single frozen linear layer, why cosine similarity can quietly mislead you, and where the pretrained predictor stops behaving like a world model.

Depending on where people want to go, we can also compare it against pixel-space world models like NVIDIA Cosmos and latent world models like DreamerV3, and look at how JEPA-style representations get used for robotics, planning, and perception.

Some things to look up to get started

V-JEPA 2 (Meta): github.com/facebookresearch/vjepa2
V-JEPA 2 in Transformers: huggingface.co/docs/transformers/modeldoc/vjepa2
V-JEPA 2 paper: huggingface.co/papers/2506.09985
Meta AI research overview: ai.meta.com/research/vjepa/

Resources

GitHub: github.com/sagecodes/ai-build-and-learn
Events Calendar: luma.com/ai-builders-and-learners
Slack (Discuss during the week): slack.flyte.org/
Hosted by Sage Elliott: linkedin.com/in/sageelliott/

In this stream

Intro to topic
Community Discussion
Practical examples

Community challenge (optional)
Try spending 30–90 minutes during the week learning or building something related