2024.0411

Number of words in this article:2941, reading time is about 5 minutes

introduction：李彦宏说模型开源的意义其实不是很大。

** Author| ** First Finance Liu Jia

On April 11, the First Financial Reporter learned that an internal speech by Robin Li leaked out. In his internal speech, Robin Li expressed his views on the route choice between open source and closed source large models and whether AI entrepreneurs should focus on industry focus topics such as models or applications. In his speech, Yanhong Li poured cold water on large open source models, saying that models are of little significance to open source. They are often used for small-scale verification applications and have not experienced verification with large computing power. In addition, he also talked about the topic of AI entrepreneurship, believing that the core competitiveness of entrepreneurs should not be the model itself, but the application itself.

Why not open source?

In the current large model industry, there are two forces: the closed source large model represented by OpenAI GPT-4 and the open source large model represented by Meta Llma2. The latter is rapidly wooing up the "moments", which has a strong impact on the former. "OpenAI does not OPEN" has also become a complaint in the industry that the star company OpenAI is "not open source". Although the company shares a large number of GPT-4 benchmarks, test results and vivid demonstrations, there is no timetable for whether to open source, such as the data used to train the system, the cost, or the specific hardware or methods used to create it. As the early entrant of the domestic big model, Baidu also chose to take the closed-source route. As for why Baidu Wenxin is not open source, Li Yanhong said today that when Wenxin was just released a year ago, Baidu had a very heated discussion about whether to open source or not, and finally decided not to open source. This is based on the judgment at that time that there must be an open source model in the market, and there will be more than one open source. "in this case, more Baidu one open source is not much, less Baidu one open source is not less." Robin Li talked about the current large model market on the mainstream open source model, such as Llama,Mistral is actually quite influential, the domestic big model Zhiyuan, Baichuan, Ali Tongyi is also open source, for Baidu, open source also has to maintain a set of open source version, which is not cost-effective. However, then, Robin Li poured cold water on the big open source models at the same time, saying, "the significance of open source models is not very great. These open source models do all kinds of verification applications on a scattered and small scale." it hasn't been verified by a lot of math. " In addition, he believes that the open source of the model is not a situation in which everyone gathers firewood. "this is very different from the traditional open source software, such as Linux, Android and so on. Because the open source is brought about by the model, we don't know why the parameters become like that. In this case, it is very difficult for everyone to gather firewood. Although Llama also encourages people to contribute all kinds of data and code, in fact, we all understand that the main developers are Meta, and it is not really a product developed jointly by everyone. " He said that on the contrary, closed source, there is a real business model, is able to make money, can make money in order to gather math, gather talent. "in fact, it is also very simple. Let's make a comparison. For example, there is an excellent person in Silicon Valley. He has OpenAI's offer, Meta's offer, and Llama's offer. Where will he go? This is a very obvious choice. " 'people think that open source is free and has an advantage in cost, but in fact it's not, but closed source has an advantage in cost,'he said. Because as long as it has the same ability, the reasoning cost of the closed-source model must be lower and the response speed must be faster. Conversely, in the case of the same parameters, the ability of the closed source model is stronger. "Today, whether in China or the United States, the strongest basic models are closed-source, while all kinds of small models and the best small models are distilled through large models. The model made by reducing the dimension of the large model is better, which will also lead to the advantage of closed source in terms of cost and efficiency. " There has been a lot of debate in the industry around open source and closed source. In the field of large models, like the closed model represented by Apple and the open model represented by Google Android in the mobile Internet era, the large model forms the basic pattern of open source mode and closed source mode because of its uncertainty. A large model practitioner told China Business News that today's large model industry is far from a monopoly or a monopoly. No matter whether it is open source or closed source, each company has its own route choice, and there are always people who believe in "the power of open source." Open source can reduce the threshold for the use of large AI models, because open source large models mean that countless developers can develop infrastructure tools and applications based on large models, which is expected to accelerate the development efficiency of downstream industry AI applications and promote the vigorous development of ecology. At the same time, open source means that more people can review it, identify and fix possible problems, thus improving security. Prior to this, Baichuan Intelligent CEO Wang Xiaochuan also talked about the dispute between open source and closed source in an interview with first Financial reporter. Wang Xiaochuan said that open source is really easy to "build character" and "have a lot of friends" for everyone to quickly understand and evaluate; at the same time, open source is also a preparation for commercialization, and if you feel good to use it, you can explore further commercialization paths when you need better services and larger parameters. However, he believes that open source and closed source are not like mobile phones where iOS or Android operating systems can only choose between the two. From the perspective of tob, open source and closed source actually need both. He predicts that 80% of enterprises will use open source models in the future, because closed source has no way to better adapt to the product, or the cost is particularly high, closed source can provide services to the remaining 20%. The two are not competitive, but complementary in different products. The big open source model is also of great significance to developers. Earlier, a large model developer told China Business reporter:"First, even if the parameters of a large open source model are not large enough, if the model is trained on a very large data set, it can compete with models with very large parameters in terms of performance; second, some small companies can use a small budget and the right amount of data to fine-tune the small model to a level comparable to the large model." In addition, because open source models support different teams to build on each other's work, open source models are advancing much faster than closed ecosystems.

The core competitiveness of AI entrepreneurs is not the model itself

In addition to talking about open source and closed source, Robin Li mentioned that for AI entrepreneurs, the core competence should not be the model itself. "it takes too much resources, and it takes a very long time to get out. The real advantage of entrepreneurs should be the knowledge and data in a certain field. " He believes that the so-called "two-wheel drive" mentioned by some outside startups who make models is not a good model. If you make both models and applications, you are bound to be distracted. The energy and resources of startups are limited, so it is self-evident which is more successful to do two things or only one thing at the same time. In any case, we are very focused, "force a hole", when resources are limited, we should focus, rather than engage in the so-called "two-wheel drive". For example, if he is looking for a "yellow men's swimming trunks without pockets" today, they cannot be found on any of today's e-commerce platforms, and this demand cannot be solved by current technology. large models can be solved if they have domain knowledge. this is an example where domain knowledge can provide unique value. There are so many models on the market, large, small, open source, closed source, how to use the combination of these models in specific applications, is a skill, this is what entrepreneurs can do, can provide value gain. He also mentioned some of the possible concerns of entrepreneurs.:If you use a literary heart or a closed-source model, you will plagiarize it when it is done well, and the giant will eat the meal of the entrepreneur. "Let's take a look. In the mobile era, Wechat did not eat pinduoduo, and Didi did not become a part of Tencent. They all provide their own unique value and have their own very different competitiveness. Their rise depends on Wechat, a closed platform in the mobile ecology, but they are not afraid of Wechat to grab its jobs, so there is no need to worry about the application of the basic model takes AI." In fact, behind Li Yanhong's statement, the development of large language models is very large and in a relatively early stage, and the native applications of AI based on large models are still very few. The large model itself is the foundation, similar to the operating system, and ultimately developers have to rely on a small number of large models to develop a variety of native applications. At present, a large number of entrepreneurs have rushed to the battlefield of AI entrepreneurship, and the AI model is expected to gradually enter the stage of spelling native applications and industries. The industry expects that the next stage of the large model is expected to usher in the application layer war. Baidu said earlier that it would open up a number of AI native applications that have been newly reconstructed and launch internal "horse racing" internally, all of which are reconstructed based on native thinking. In the application layer, iFLYTEK successively released application products in industry, finance, education, health care, AI digital people and other fields, opening up the developer interface; Baichuan Intelligent CEO Wang Xiaochuan told reporters that this year, the internal team began to deploy C-side super applications, in his view, landing and applications are the strengths of Chinese entrepreneurs.
**