OpenAI officially opened registration and usage for Sora over a day ago. My first impression was, naturally, not great. However, today, my second impression has changed significantly.
My initial disappointment stemmed from the fact that after a ten-month wait since its debut in February, the return seemed somewhat underwhelming.
My second, more positive impression came after deeper use. Hidden beneath the surface texture are discoveries that bring a visible "surprise" in model capability: Physics.
I uploaded only four images (these were not prompts):
I want a dynamic effect of sunlight;
I want to see the "natural" side of animals;
- I want to see the motion of a train;
- I want to see the rotation of star trails;
I obtained four video segments, which when combined, show:
Indeed, we see: 1. The relationship between sunlight, architectural occlusion, and shifting shadows; 2. The natural facial expressions and movements of a monkey; 3. The symmetrical mirror reflection when a train passes over a lake; 4. The natural sense of rotation in star trails.
All of these align with the physical laws of our reality.
Yes, ten months later, on the surface, the Sora we are using doesn't seem to have progressed much compared to the version released early this year. In these ten months, we have all grown weary of various video generation models.
However, in essence, we can find significant progress: more realistic details, better adherence to physical laws, and a closer resemblance to a "World Model."
Sora is called a simulator—it simulates a "world": environment and objects, "predicting" the movement of objects and environmental changes based on learned "rules." This data is fed into an "AGI" (if one exists or is currently being trained). Through complex reinforcement learning, researchers see if a true AGI can emerge. In the process of training Sora, due to the Scaling Law, "emergence" may have occurred, appearing to show a certain understanding of the rules of the human world. This is the most valuable part of the model.
Dao Ming, Public Account: Dao Ming Laboratory
Perhaps due to various considerations, OpenAI believes Sora can serve as an excellent tool for the film industry, while also being forced by reality to let the "World Model" earn its keep to reduce the cost of AGI research and development.
This version of Sora still has a very, very long way to go before it enters Hollywood. However, the experience gained from this version and its underlying training gives us more hope on the path to AGI.
Keep stacking: computing power, data, and time.