Enterprises "enter the pit" of large models, why are they recommended to be produced by large manufacturers?

Less than a year after the release of GPT-3.5 abroad, and less than half a year after the release of Wenxinyiyan in China, China has fast-forwarded to the "Battle of Hundreds of Models". Entrants include both large manufacturers and a large number of start-up companies. However, the competition for large models is fierce, and if enterprise users do not have a sharp eye, they are likely to fall into the trap, resulting in the project being unfinished.

The large models of some major manufacturers, such as Wenxin Large Models, Tongyi Qianwen, Pangu Large Models, etc., are gradually widening the gap. SuperCLUE's latest evaluation list shows that Wen Xinyiyan has surpassed GPT-3.5turbo, and domestic large models such as GLM-130B are also at the top of the list. Domestic large models have become an important part of the world in terms of quantity, and are rapidly catching up with the most advanced GPT-4 in terms of quality.

Behind the "Battle of 100 Models", the industry is optimistic that large model technological innovation will promote industrial digitization and create trillions of market value. From the current point of view, large models "produced by major manufacturers" represented by Wen Xinyiyan and Tongyi Qianwen occupy a dominant position at the technical level. At the market level, they have also won more project cooperation by building an industrial ecology.

Why are the large models produced by big manufacturers more powerful and more popular with customers than start-up companies? In the author's opinion, there are three main reasons:

First of all, large models must eventually be applied in industry scenarios, and it is not a simple "one-shot deal." As a symbol of greater stability, security, and reliability, major manufacturers generally have sufficient accumulation in AI technology bases. Enterprise customers tend to focus on the application level and do not necessarily have strong accumulation of underlying AI technology. Major manufacturers such as Baidu, Alibaba, and Huawei have built AI bases that have been tested in actual combat. Looking at startups, there are many highlights and breakthroughs in technological innovation. However, from the perspective of full-stack AI technology base accumulation and long-term and stable customer service, The overall capability and battery life are still questionable.

Secondly, large manufacturers have stronger comprehensive strength to invest in iterative development of large model technology. For example, Baidu can mobilize the entire group to develop Wenxin Yiyan. The latest iteration of Wenxin Big Model 3.5 has an inference speed that is 17 times faster than the 3.0 version, and the model effect is improved by more than 50%. On the other hand, the most popular start-up company, light years away, chose to "sell out" to Meituan in less than half a year. Some large models developed based on open source technology lack sufficient underlying technology accumulation and autonomous development capabilities.

According to the latest "AI Large Model Technical Capability Assessment Report, 2023" released by market research company IDC, Baidu Wenxin Large Model 3.5 scored 7 full scores in 12 indicators, including "algorithm model", The two key indicators of "industry coverage" ranked first in the overall score; Alibaba Tongyi Large Model, which ranked second, received full scores in 6 out of 11 indicators and was the only manufacturer to receive full scores in "service capabilities".

Third, the implementation of large models in industry scenarios, subsequent service delivery, operation and maintenance, etc., are "easy for large manufacturers to master" ”, which is a difficult challenge for start-up companies. Large manufacturers can equip vertical industries and key customers with dedicated service teams to relieve customers from worries about embracing new technologies. However, start-up companies often lack experience in serving government and enterprise customers. Large-model product applications developed based on open source technology must do a good job in all aspects. Service of process was grudging.

To sum up, the big model not only looks at various technical parameters, but also looks at the industry’s implementation, the industry’s “know-how” and successful experience. Products produced by major manufacturers currently occupy a dominant position in terms of technology development, industry applications and services. Of course, the "Battle of One Hundred Models" may not be the "War of One Hundred Regiments" more than ten years ago, with only two winners in the end. The large model is still in the early stages of development, and there are more possibilities in the future, including the possibility of surpassing GPT-4 and more European and American competing products.

The above is the detailed content of Enterprises "enter the pit" of large models, why are they recommended to be produced by large manufacturers?. For more information, please follow other related articles on the PHP Chinese website!