Robin Li's latest internal speech: open source models are not as good as closed sources, and the latter will continue to lead

Interface News Reporter | Li Rujia

Interface News Editor | Song Jianan

On April 11, Jiemian News learned that Robin Li, the founder, chairman and CEO of Baidu, talked for the first time about why the Wenxin model is not open source in an internal speech, and his views on the route choice of open source and closed source of large models.

In addition, he also mentioned the focus of the industry on whether AI entrepreneurs should focus on models or applications, and whether the "two-wheel drive" of startups to do both models and applications is a good model.

It doesn't make much sense to open source large models

In the field of large models, there are currently two technical routes: open source and closed source.

For example, Elon Musk's artificial intelligence startup "xAI" has chosen the open-source route, after it officially open-sourced the world's largest large-parameter large language model, Grok-1. Meta's large language model Llama series, as well as Mistral AI, which has attracted much attention in the open source community, and the domestic Zhiyuan "Wudao" large model, Baichuan Intelligent large model, and Alibaba's Tongyi Qianwen model are also open source large models.

GPT-3.5 and GPT-4 developed by OpenAI, which set off a craze for large models, chose closed source, and the same is true for Baidu's Wenxin large model.

Robin Li mentioned in his internal speech that Baidu had a very heated discussion about whether Wenxin needed open source, and finally decided not to open source. At that time, the judgment was that there would be open source models on the market, and more than one would be open source. In this case, there are not many open sources for Baidu, and there are many open sources for less Baidu.

"There is no shortage of open source models in this market. If we want to open source, we have to maintain an open source version ourselves, which is not cost-effective. He believes that the significance of open source models is not very large, and these open source models are scattered and small-scale to do various verification applications, but they have not been verified by large computing power.

In his view, unlike traditional software open source, large model open source is not a high flame for everyone. On the contrary, the closed-source model will continue to lead in terms of capabilities, rather than a temporary lead.

Some industry insiders also told Jiemian News that the advantage of open source software before was that everyone shared the code, which could be done by multiple people to fix bugs together, so that the software was constantly updated. However, the large model itself is a black box, and there is a possibility of retraining after someone submits modifications, and each training will consume a lot of computing power and funds, and it does not benefit as much in the joint development of multiple people like the previous open source software.

The above-mentioned people judge that in the long run, closed-source can concentrate resources such as intelligence and computing power to iterate on large models, which is more efficient than open source.

Robin Li also emphasized that the advantage of closed source is that it has a real business model, which can make money to gather computing power and talents. In terms of cost, the inference cost of the closed-source model is lower and the response speed is faster under the same ability. With the same parameters, the closed-source model is also more capable.

"Today, whether in China or in the United States, the strongest basic models are closed-source, and all kinds of small models, the best small models, are distilled through the large models. The model made by dimensionality reduction of the large model is better, which will also lead to the advantages of closed source in terms of cost and efficiency. He said.

Regarding the open-source closed-source controversy of large models, Wang Xiaochuan, CEO of Baichuan Intelligence, also mentioned it in an interview with Jiemian News. His point of view is that the large model itself does not represent the consumer side, unlike Android and IOS, which need to choose one or the other, and today from the perspective of the enterprise side, open source and closed source are needed.

Wang Xiaochuan attaches great importance to the value brought by open source, and he believes that 80% of enterprises will use open source models in the future, because open source models are small and closed source cannot make optimal adaptation to many scenarios.

The core competitiveness of AI entrepreneurs is not the model itself

In addition to his stance on the open and closed source routes, Robin Li also put forward his own views on AI entrepreneurs and startups.

He believes that the so-called "two-wheel drive" of some startups that make models is not a good model, and it is bound to be distracting to do both models and applications. Startups have limited energy and resources, and when resources are limited, they should focus rather than engage in the so-called "two-wheel drive".

For AI entrepreneurs, the core competitiveness should not be the model itself, which is very resource-intensive and requires a long time of persistence to run out. The real advantage of an entrepreneur should be the knowledge and data in a certain field.

"If you want to find a 'yellow men's swim trunk without a pocket' today, you can't find it on any e-commerce platform, and the current technology can't solve this demand. Large models can be solved if they have domain knowledge, which is an example of how domain knowledge can provide unique value. He said.

In his opinion, there are a large number of models on the market, large, small, open source, closed-source, and there is a skill in how to use the combination of these models in a specific application, which is something that entrepreneurs can do and can provide value gains.

Regarding the outside world's concern that if Wenxin or closed-source models are used, they will be plagiarized and robbed of their jobs if they do a good job, Robin Li also responded that in the mobile era, WeChat did not eat Pinduoduo, and Didi did not become a part of Tencent. Each of them offers its own unique value and has its own very different competitiveness. Their rise is based on WeChat, a closed platform in the mobile ecosystem, but they are not afraid of WeChat to steal its job, so there is no need to worry about the application of AI in the basic model.

The research report of China Securities Construction Investment pointed out that at present, the ability of large models in China is gradually improving, and the processing of Chinese and some features such as long text processing have alignment and leading advantages. With the increase in the popularity of Kimi, many domestic large-scale model manufacturers have joined the competition for long text capabilities, and the implementation of the industrial consumer side has accelerated. The first year of domestic large-scale model application has arrived.

After the model gradually matures, the future large model will start a new round of competition and competition in the product and application layer. Robin Li's speech is also attracting more application layer developers to choose the Wenxin model.

At Baidu's previous earnings call for the fourth quarter and full year of 2023, Robin Li revealed that the total revenue of Baidu Intelligent Cloud in the fourth quarter was 8.4 billion yuan, of which the large model brought about 660 million yuan of incremental revenue to the cloud business.

At present, the daily call volume of Wenxin's large model has exceeded 50 million, an increase of 190% quarter-on-quarter. In December last year, about 26,000 companies called the Wenxin model, an increase of 150% quarter-on-quarter. Samsung, Honor, Autohome and other companies have reached cooperation with Baidu.

Since its release, Baidu has continued to reduce the inference cost of the Wenxin model, which has now been reduced to 1% of the March version last year.

Robin Li also said that in the future, multimodal or multimodal fusion, such as text to video, is a very important direction for the development of basic models, and it is also a necessary direction for AGI (General Artificial Intelligence). Baidu has already invested in these areas and will continue to do so in the future.