Open source model or closed-source model, who is the "version answer"? Around this question, Zhou Hongyi, the founder of 360, and Robin Li, the founder of Baidu, recently had a debate in the air.
First, on April 12, some media exposed Robin Li's internal speech at Baidu. Among them, he believes that the closed-source model will continue to lead in terms of capabilities, rather than leading for a while, and the operation mode of the open-source model is different from that of traditional open-source software.
Immediately afterwards, at the Create 2024 Baidu AI Developer Conference held on April 16, Robin Li declared in his speech that "open source models will fall behind more and more".
Two days later, at the 27th Harvard China Forum, Zhou Hongyi, a self-proclaimed believer in open source, stood up and sang the opposite, "I have always believed in the power of open source, as for some celebrities talking nonsense on the Internet, don't be fooled." He said that open source is better than closed source? Even the company that said this has grown to this day with the help of open source. ”
In fact, in the domestic market, whether the large model is open source or closed source is not a technical issue. Regardless of whether it is closed-source or open-source, the performance of these large models is objectively inferior to that of similar overseas products, and they are not the leaders in the technical level of large models. Purely at the technical level, the current closed-source large model is in the leading position in terms of capabilities, such as OpenAI's GPT-4, Anthropic's Claude-3, Google's Gemini Ultra are all closed-source, and the situation in China is the same.
Zhou Hongyi believes that the open source model provides a platform for the entire community to share, collaborate and innovate, and promotes the rapid development of technology, and the number of engineers and scientists gathered in the open source community is hundreds of times that of closed source, so this year's open source model has only been done for one year, and it has already surpassed the ability of GPT-3.5. However, the actual situation is far less optimistic than he said, whether from the level of model capabilities or the dimension of product implementation, the fact is that the open source model has always lagged behind the closed-source model.
Although Meta's just-released Llama 3 does outperform GPT-3.5, the problem is that no one thinks that Llama 3 is a product of the open-source community, after all, its core development strength has always come from Meta.
In fact, considering that Robin Li and Zhou Hongyi have long been far away from the front line, the two have different views on open source and closed source of large models, which may not be a difference in the technical route. Therefore, there is a view that one side supports closed source and the other party believes in the power of open source, which is actually more driven by the commercial interests that each represents.
The reason why Zhou Hongyi supports the open source model is actually very simple, after all, in today's "100 model war", 360 Intelligent Brain is still in the position of catching up, and it is inferior to the large models launched by Internet giants such as Wenxin Yiyan and Tongyi Qianwen in terms of user scale and influence on the industry.
Previously, 360 made it very clear when it open-sourced the 360 Intelligent Brain 7B model, "There is no need for everyone to reinvent the wheel", which means that AI entrepreneurs do not need to develop their own anymore, and they can directly use the ability of the open-source 360 Intelligent Brain. Throughout the history of the software industry, it is actually a very classic operation for commercial companies to gain leadership in a certain track through open source, such as Google, which relies on open source to allow the Chromium ecosystem to dominate the PC browser market.
In a sense, Zhou Hongyi's 360 open-source own large model is actually a replica of Meta's open-source Llama's operation. Since Zhou Hongyi, who claims to be an open-source believer, has selfish intentions, Robin Li, who insists that the closed-source model is better, actually has his own small abacus.
While pouring cold water on the open source model, Robin Li also said, "The large model itself does not directly create value, and the AI applications developed based on the large model can meet the real market demand." This is what our Baidu has done in exchange for stepping on countless pits and paying high tuition fees according to the practice of the past year. ”
The implication of this sentence is to link his public speech a year ago. At the Nishan Dialogue on Digital Civilization at the World Internet Conference last summer, Robin Li said in his keynote speech, "The key point of the new international competition strategy is not how many large models a country has, but how many native AI applications there are on your large models, and to what extent these applications have improved production efficiency." ”
Behind this similar discussion at different times, Robin Li actually wants to dissuade large model entrepreneurs and hope that they will directly join Baidu's AI ecosystem. Based on greater computing power, more data and stronger algorithms, relying on the paddle platform, from Wenxin 3.0, 3.5, to 4.0, the ability of Wenxin's large model has gradually increased.
Compared with a year ago, the algorithm training efficiency of Wenxin's large model has been increased to 5.1 times, the average weekly training efficiency rate has reached 98.8%, the inference performance has been improved by 105 times, and the cost of inference has been reduced to 1%.
Previously, Baidu CTO Wang Haifeng revealed that the cumulative user scale of Wenxin Yiyan has reached 200 million, and the average daily call volume has also reached 200 million, and the PaddlePaddle Wenxin ecosystem has gathered 12.95 million developers, served 244,000 enterprises and institutions, created 895,000 models based on PaddlePaddle and Wenxin, and developed more than 190,000 AI native applications using the Qianfan platform.
Therefore, what Baidu wants is to build an AI ecosystem like OpenAI, and use Wenxin Yiyan as a fulcrum for platformization. In this way, Baidu hopes that the fewer participants in the large model track, the better, and once entrepreneurs develop AI applications instead of large models, it is naturally possible to repeat the story of the App Store.
Baidu, which has been investing in the AI field for many years, is undoubtedly the pioneer of domestic AI large models, and the incremental revenue realized in its AI field is also reflected in the 2023 financial report, but the absolute value is not large. Therefore, for Baidu, which has almost missed the entire mobile Internet era, consolidating its own AI ecosystem and improving the monetization ability of AI business is undoubtedly the top priority at present. Compared with making products yourself, it is obviously a better choice to do an app store to "collect taxes".
In the final analysis, whether it is better to open source or closed source large models is not yet superior at the technical level, but closed source or open source has its own advantages and disadvantages for the business interests of different companies.