详解Sora 为什么是AGI的又一个里程碑时刻？

At the beginning of the year, another bomb was thrown at the world. The video generation model was considered as another milestone of general artificial intelligence a year ago, which means that the realization will be shortened from 2000 to 2000, and the chairman Zhou Hongyi made a prediction. But this model is so sensational not only because the video generated is longer and clearer, but also because it has surpassed all the abilities in the past to generate a video content related to the real physical world. Cyberpunk is cool, but how can everything in the real world be made? It is more meaningful to reproduce. Therefore, a brand-new conceptual world simulator is put forward. In the official technical report, it is positioned as a video generation model as a world simulator. Our research results show that expanding the video generation model is a feasible way to build a general simulator for the physical world. official website believes that it has laid the foundation for understanding and simulating the real world model, which will be an important milestone in realization. With this, it has completely opened up a company such as video track. From words to pictures to videos, it seems to me that they are collecting pieces of puzzles, trying to completely break the boundary between virtual and reality through the form of image media and become the number one player in movies. If Apple is the number one player's hardware, then a system that can automatically build a simulated virtual world is the soul language model, similar to the human brain video model and similar to the physical world. The ambition expressed by doctoral students at Edinburgh University is beyond everyone's imagination, but it seems that only it can do it. Many entrepreneurs lamented how the light cone intelligence has become a newly released model of the world simulator, kicking the door of the video track in 2000 and completely drawing a dividing line with the old world years ago. In a demonstration video released in one breath, the light cone intelligence found that most of the problems that the video was criticized in the past were solved, resulting in clearer images, more realistic images, more accurate understanding, smoother logical understanding and more stable and consistent results, etc. But all this is just a result. The tip of the iceberg is not aimed at video from the beginning, but at all the existing images. Video is a larger concept, and video is a subset of it, such as the virtual scene of the big-screen game world rolling on the street. What we need to do is to cover all the image simulations and understand the real world, that is, the concept of the world simulator that it emphasizes, just as Chen Kun, the producer of the movie Wonderland of Mountains and Seas, told the light cone intelligence to show us its capabilities in video. But the real purpose is to get people's feedback data to explore and predict what kind of video people want to generate, just like large-scale model training. Once the tool is open, it means that people all over the world are working for it, and its world model becomes more and more intelligent through continuous marking and entry. So we see that video has become the first stage of understanding the physical world, mainly highlighting its attributes as a video generation model, and only in the second stage can it provide value as a world simulator to seize video. The core of generating attributes is to find out where the differences between different sums are reflected, which is very important, because to some extent, it explains why it can be crushed. First of all, it follows the idea of training large-scale language models, and uses large-scale visual data to train a generating model with universal ability, which is completely different from the logic of dedicated personnel in the field of Wensheng video. Last year, there was a similar plan called universal world model, and the idea was basically similar, but there was no follow-up. This time, it was completed first. According to Xie Saining, an assistant professor at new york University, the amount of parameters is about 100 million. Although the comparison model is insignificant, this order of magnitude has far exceeded that of other companies. The success of Qi Tangquan, the general manager of Wanxing Science and Technology Innovation Center, has once again verified the possibility of making great efforts to make miracles. It still follows the principle of making great efforts to make miracles, a large number of data models and a large amount of computing power. At the bottom, the world model verified in the field of game driverless and robot is adopted to build the Wensheng video model. Secondly, it shows the perfect integration of diffusion model and large model for the first time in the body. Video is like a movie blockbuster, which depends on two important elements: script and special effects, in which the script corresponds to the logical special effects in the process of video generation, and the effect. In order to realize the logic and the effect, two technical paths have been differentiated behind the diffusion model and the large model. At the end of last year, light cone intelligence predicted that in order to meet the effect and logical diffusion and large model at the same time, the two paths will eventually be realized. Towards integration, I didn't expect to solve this problem so quickly. In the technical report, official website drew a picture focusing on the method of transforming various types of visual data into a unified representation. This representation can be used for large-scale training of generating models. Specifically, every frame of a video picture is coded and transformed into a visual patch, and each patch is similar to one of them, which has become the smallest unit of measurement in a video image and can be broken and reorganized at any time and anywhere to find a unified number. According to the method, the weights and measures are unified, and the bridge between the diffusion model and the large model is found. In the whole generation process, the diffusion model is still responsible for the generation effect. After increasing the attention mechanism of the large model, it has more predictive reasoning ability, which explains why the video can be generated from the existing static images, and the existing video can be expanded or the missing picture frames can be filled. Up to now, the video model has shown a composite trend, and the technology is also moving towards integration. It has also become an advantage to apply the technology accumulated before to the visual model. In the training process of Wensheng video, the language understanding ability of harmony is introduced. According to the representation, the training can accurately generate a set of high-quality video according to the user's prompts. As a result, the simulation ability appears, which constitutes the basis of the world simulator. We find that the video model will show many interesting emerging abilities when conducting large-scale training. These abilities make it possible to model. The appearance of these characteristics in some aspects of people, animals and environment in the quasi-physical world has not produced any definite inductive deviation on three-dimensional objects, etc. They are purely scale phenomena, indicating that the fundamental reason why simulation can be so cracked is that people have become accustomed to creating things that do not exist with large models, but they can accurately understand the operational logic of the physical world, such as how forces interact, how friction produces basketball and so on. These are things that no model can accomplish before, and they are also the fundamental significance beyond the level of video generation, but from the actual finished product. 比特币今日价格行情网_okx交易所app_永续合约_比特币怎么买卖交易_虚拟币交易所平台

注册有任何问题请添加 微信：MVIP619 拉你进入群

文章转载注明 网址：https://netpsp.com/?id=57792