Shulou(Shulou.com)11/24 Report--

The explosion of ChatGPT has made the big model go out of the professional field and into a more universal circle, and caused more people to pay attention to and think about how this technology will affect human society. Finally, the whole society reached a consensus: this is a revolutionary technology, which will fully release intelligent productive forces and promote human society to enter a new AI era. With the advent of business opportunities, who can take the lead to grasp the tuyere means that who will gain an advantage in the next round of market competition. Since the beginning of this year, Digital China has also actively embraced the brand-new changes brought about by the big model, always paying attention to technological developments, upgrading its own intellectual property products, accelerating the layout of artificial intelligence, computing and other fields, and earnestly seize the opportunity of change in the AIGC era to seek brand-new development.

At the just-concluded 2023 Digital Ecology Conference, Li Gang, vice president of Digital China and CTO, was invited to have an in-depth discussion with a number of industry experts on the in-depth application of large model technology and model evaluation. This article collates and edits the contents of Li Gang's discussion, hoping to give some inspiration and thinking to the hot market.

The following are the details:

The development of large models should pay attention to Native scenarios and avoid path dependence.

After the emergence of the big model and other new technologies, everyone's first reaction is "+", and the term "big model +" is not new, because there was Internet + before, so it is natural for us to cut directly from this point of view.

The big model is a subversive technological paradigm shift over the past hundred years. I don't think its real application scenario has appeared yet. The "+" scenario is not the best one. When the Internet broke out in the past, the last thing that came out was the application of Internet Native, and the real application scenario of the big model should also be the Native scenario. We are looking for Native scenarios to avoid path dependence, that is, to avoid imitating previous patterns and paths to explore large model application scenarios.

At present, the large model technology is not fully mature, and we can not see where the pinnacle of its future development is. I think to pay attention to a very large, Native scene is multimodal, in the enterprise, multimodal is almost no one to think about, then multimodal scene, this is to pay attention to. When Digital China pays attention to the big model, it first sets up a tracking team to organize a group of people to pay attention to the industry trends, read the weekly paper, and verify it as soon as the content of the paper can be verified. The second is to be agile, if the model wants to be used in the enterprise, we need to make a set of agile platform for the enterprise to try, and at the same time bring the customer Involve into the co-creation. We have now signed up some co-creative customers to use the agile platform to explore the most suitable scenario in the process of co-creation. Finally, as a science and technology enterprise, it is still necessary to follow the first principle at both ends, on the one hand, the follow-up of the research, on the other hand, the follow-up of the scene.

Large model evaluation is professional and cutting-edge, paying attention to knowledge itself and emergence ability.

Large model evaluation itself is a very cutting-edge field, although there have been some evaluation methods and standards, but from the enterprise point of view, these methods and standards are not enough, so model evaluation itself is a research direction.

At present, the model evaluation in the industry is mainly carried out around two points, aiming at the external model, mainly general ability assessment, including knowledge itself and emergence ability. In the aspect of industry large model, its industry knowledge evaluation is relatively simple, but the emergence ability of the industry is a highly professional topic. For example, we have made a big model of the financial industry. What is its emergence ability? The emergence ability of the general model, for example, if we give it a Mathematical Olympiad problem, it can solve it, which is very powerful.

Generally speaking, the industry large model is to do the whole parameter through the general large model, or to do fine-tuning, and the big model that comes out after it is done is called the industry big model. After this is done, does it mean that general ability plus industry knowledge equals the emergence capacity of the industry? This is a problem. We talk about risk control, if it is not the application of risk control, but the large model of the industry directly provides risk control ability, this is a very large emerging ability. It does not mean that he only sees these things, he will directly make a risk control assessment of the business, then this is very difficult.

Traditionally, it is done with a small model made of traditional machine learning, can it be generalized into a large model? Did this pop up? This is very difficult. So now it is a question of choice and co-creation, choosing those who are willing to do research with us in the big model, and we will explore and create together.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.