The video generation model Sora recently launched by OpenAI continues to detonate related concepts and is regarded as a major breakthrough in video generation services. However, rival scientists poured cold water on the idea. Sora is favored by the market for its video stability and clarity, but some people have higher hopes for it, believing that it may have the opportunity to hit the peak of artificial intelligence-world simulator, which is also the slogan shouted by OpenAI.
OpenAI also hopes in its official website article that Sora is not just a video generator, but a real-world simulator. By learning and simulating data distribution, virtual samples similar to the real world are generated, thereby providing predictive information to the real world.
However, this prospect was publicly refuted by Meta's chief AI scientist Yann LeCun. This AI tycoon known for his outspokenness said bluntly: Just make videos, don't brag about what you have.
A doomed emulator
LeCun publicly posted on X that simulating the world by generating pixels is a waste of resources and is destined to fail. The post also attracted great attention, with many netizens discussing the sword in the comment area.
Simply put, LeCun believes that the Sora model tries to infer too many irrelevant details, like trying to determine the trajectory of a football by trying to analyze its materials.
He pointed out that generative models are suitable for textual content because text is discrete and is data composed of a limited number of symbols, in which case it becomes very easy to deal with uncertainty in predictions. But if you move to the realm of predictions in pixels, the uncertainty becomes very tricky, and success is impossible.
Netizens below also spoke one after another. Some people said that although Sora is impressive, there may be logic problems in almost every scene. For example, in the artist's test video, three wolves inexplicably split into five wolves. This makes simulating reality somewhat unreliable.
Some people also say that Sora was born in response to dreams, and people can have lucid dreams or less lucid dreams. This seemingly sober statement is a bet that Sora has some possibility of realizing OpenAI's "world simulator".
At the same time, Meta, where LeCun is located, released a video joint embedded prediction architecture V-JEPA last week. It is said that it teaches large model understanding and modeling of the physical world by watching videos. It is another attempt at a world simulator besides Sora.
In addition, V-JEPA can flexibly discard unpredictable information and improve training and sample running efficiency by 1.5-6 times. However, V-JEPA obviously did not cause much commotion in the market and is much lower-key than Sora.