OpenAI just released Sora 2, a new video and audio generation model that feels like the jump from GPT-1 to GPT-3.5, but for moving images.
Unlike earlier systems that bent reality to satisfy prompts, Sora 2 obeys physics. If a basketball player misses a shot, the ball bounces off the rim instead of teleporting into the hoop. It can model failures as well as successes, which makes it far more convincing as a world simulator.
The model also adds synchronized dialogue, sound effects, and music, and it can follow multi-shot instructions while keeping objects and characters consistent. It supports “cameos,” a feature that lets you upload yourself once then insert yourself into any generated scene with accurate likeness and voice.
To showcase this, OpenAI launched a new iOS app called Sora, built around social remixing and cameos; it looks more like TikTok meets Midjourney than a research demo. Users can drop themselves into Viking battles, Olympic routines, or anime scenes, and then share clips with friends.
Sora 2 validates scaling video in the same way GPT models validated scaling text. If AI can simulate physical cause and effect, it’s a step toward models that can train agents, robotics, and other systems grounded in the real world. It is also a new medium for communication and marketing, one where your brand and your customers can literally step inside the content. There’s much more to this. Go check out the demo.
Author’s note: This is not a sponsored post. I am the author of this article and it expresses my own opinions. I am not, nor is my company, receiving compensation for it. This work was created with the assistance of various generative AI models.