Yann LeCun says the extraordinary video generator Sora is doomed to fail

At OpenAI, the first videos generated by Sora's AI are so stunning that they raise both astonishment and concern. Instead of finding out who will be able to distinguish between right and wrong in the future, one of the popes of AI, Frenchman Yann Lacon, has just issued a scathing criticism of the technology used by OpenAI. Meta's head of AI thinks OpenAI's approach is bad. First, it challenges the ambition of the publisher ChatGPTChatGPT To be able to create real-world digital twins of its algorithms. onTwitterTwitter), explains that the method of generating pixels from latent variables is doomed to failure because it is inefficient. He says OpenAI's models seek to infer a lot of irrelevant details. This postingenergyenergy To achieve this it is commendable to create videos from text, but when it comes to modeling the world, it is not at all effective.

Too much detail kills the model

For LeCun, who has worked for 30 years on… Machine learning And the Deep learningIf the generative approach with ChatGPT works well, it is because the text contains a specific number of symbols. On the other hand, to simulate the world, we fall into a more comprehensive and complex domain. If Meta searches in ThemeTheme Artificial intelligence has fallen under the radar due to the notoriety of OpenAI and the increase of its generative tools, the parent company of FBFB It is also working on its own AI model capable of creating videos.

Baptized Predictive architecture for video co-embedding (V-JEPA), the method developed by Yann LeCun and his team is actually quite different. The algorithm does not seek to infer pixels, but rather seeks to get to the essentials by removing everything that is unpredictable. This system would improve training by a factor of 1.5 to 6. The future will tell whether criticisms of Yann LeCun and his model will guide the decisions of competing companies.