Cover news reporter Meng Mei Zhang Yuexi.
Recently, the initial generation of SORA has attracted the attention of all walks of life for the first time after it was launched, and it is called another "WOW" time point in the field of artificial intelligence after the birth of ChatGPT. At the same time, the well-known AI company Emotibot was rumored to have announced a half-year shutdown, which triggered the public's sigh about the speed of replacement of the old and the new in the field of technology commercialization.
On February 22, the cover news reporter interviewed Cui Lili, executive director of the Institute of E-commerce of Shanghai University of Finance and Economics. She said that any conclusive trend judgment on the current SORA view may be premature when an innovation is just emerging.
Although some of the published SORA generation results are almost realistic, or even better than most content producers who lack professional ability, it will take time, even a short time, for them to truly replace or dominate content creation. According to Cui Lili, SORA dominates or replaces content creation, at least depending on factors from both the model and the user.
Cui Lili, executive director of the Institute of E-Commerce of Shanghai University of Finance and Economics.
On the side of the model, there are two aspects that play a decisive role: one is the accumulation of existing content in the human world and the learning speed of the artificial intelligence model, and the other is how some implicit relationships behind the pure content can be further understood and mastered by the artificial intelligence model. She said that from the current situation of SORA, the first point of this has been realized, which is to quickly mark and learn the content of simple labeling and non-labeling according to the learning of the content that has been finely labeled. However, the second point has not yet been seen in SORA, that is, whether it covers the mechanical relationship, spatio-temporal causality, and partial and overall relationship contained in the picture while showing the perfect visual generated picture, which will determine the rationality and realism of the generated visual picture. It is difficult to identify the logical relationship between visual images and content in a single dimension.
This may mean that when the large language model annotates the content of the visual picture, it must not only describe and describe the condition of the picture in language, but also need to label the implicit physical relationship reflected in the picture, but the problem is that the language description is not as accurate as the mathematical formula. Cui Lili explained that in other words, the mathematical and physical principles that make up the plausibility in the real-world picture are not the expertise of existing large language models, and may need to be supplemented by other technical models that are better at mathematical or physical principles. Obviously, it will still take time to break through.
In her opinion, even if these problems are solved, Sora still needs to face the gap between the perfect laboratory product and large-scale industrial application. Cui Lili analyzed from two points: suitable application scenarios and user side. First of all, it also takes skill to use a good AI model, and to become a magician, you must first learn to use spells. The threshold in this area determines the speed at which a new technology can be applied and difrequent. For example, after the advent of ChatGPT, there was a prompt word engineer for such a professional position. Some people say that the emergence of every artificial intelligence model is the first to fatten out the class. The emergence of various training courses is enough to show that the comprehensive replacement of a new technology requires the whole industry to prepare for a full period of time, and the preparation of individual skills is often the first step.
From a further business perspective, while new technologies can bring competitive advantage in the early stages, they also require some investment. Cui Lili pointed out that these investments not only include funds, but also include the hidden management costs such as personnel layoffs and organizational adjustments. All of the above shows that the singularity that led the public to open the black hole of the future world is far from over. This waiting process depends on the accumulation of existing industry data, the degree of model perfection, the adaptability of application scenarios, the learning threshold, and the cost of capital and management.
Scene requirements are the magic wand that ignites technological productivity. Cui Lili put forward the point of view that from the perspective of optimization, on the one hand, the data accumulation, data quality, and annotation depth of the existing industries need to be continuously strengthened, which depends on the degree of digitalization of the scene industry, so it is the only way to continue to promote the acceleration of the informatization process in all walks of life and realize digital transformation. On the other hand, for the AI model itself, it is important to integrate more model representations that can implicitly reflect the global relationship in the real world.
We have seen too many new technologies accelerate in the field of information technology. The e-commerce sector, for example, is witnessing a paradigm shift from search-based platforms to social communications to content distribution that is gathering traffic more and more quickly. It has been proven that the lower the learning threshold, the faster the popularization and comprehensive replacement of a technology. Therefore, lowering the threshold for the use of AI models and learning to work with AI models with everyone is a two-way process, which can promote the faster arrival of the singularity. This requires a long-term effort from both the technical staff and the education industry. She imagined that after the singularity, the emergence of artificial intelligence might create a new civilization. The intersection and integration of artificial intelligence and metaverse technology may really create a "parallel world".