“Non-generative, joint embedding architectures like DINO, IJEPA, and VJEPA, developed primarily at Meta, have proven more effective for learning image and video representations than generative approaches.”