One AI Model Creates a Physical Intuition of Its Environment
20 hours ago
- #AI
- #Intuitive Physics
- #Machine Learning
- Meta's V-JEPA AI model learns intuitive physics from videos without prior assumptions.
- V-JEPA uses latent representations to focus on essential details, avoiding pixel-level distractions.
- The model demonstrates 'surprise' when encountering physically impossible events, similar to infants.
- V-JEPA achieves 98% accuracy on the IntPhys test for intuitive physics understanding.
- The next-generation V-JEPA 2 model was pretrained on 22 million videos and applied to robotics.
- V-JEPA 2 struggles with longer sequences, likened to a goldfish's memory span.