On April 11, the interface news learned that Robin Lee, the founder, chairman and CEO of Baidu, for the first time talked about why Wenxin's big model was not open source, and his views on the route choice of open source and closed source of the big model in an internal speech.
In addition, he also mentioned industry hot topics such as whether AI entrepreneurs should focus on models or applications, and whether the "dual wheel drive" of startups doing both models and applications is a good model.
The open source significance of large models is not significant
In the field of large models, there are currently two technological paths: open source and closed source.
For example, Musk's artificial intelligence startup "xAI" has chosen the open source route, after officially opening up the world's largest parameter large language model Grok-1. Meta's Llama series of big language models, as well as Mistral AI, which has gained attention in the open source community, as well as China's Zhiyuan "Wudao" big model, Baichuan Intelligent big model, and Alibaba's Tongyi Qianwen model, are all open source big models.
OpenAI's GPT-3.5 and GPT-4, which have sparked a wave of big model development, have chosen closed source, and Baidu's Wenxin Big Model is no exception.
Robin Lee mentioned in his internal speech that Baidu had a very heated internal discussion about whether Wenxin needed to open source, and finally decided not to open source. The judgment at that time was that there would definitely be open-source models in the market, and more than one would be open-source. In this situation, it is not uncommon to have multiple Baidu companies open source, and it is also not uncommon to have fewer Baidu companies open source.
"There is no shortage of our open-source model in this market. If we want to open source, we have to maintain a set of open-source versions ourselves, which is not cost-effective." He believes that the significance of open source models is not very great. These open-source models are used for various verification applications on a scattered and small-scale basis, but have not been verified by large computing power.
In his view, unlike traditional software open source, big model open source is not just about everyone picking firewood and having a high flame. On the contrary, the closed source model will continue to lead in terms of capability, rather than temporarily leading.
Industry insiders also expressed to Interface News that the advantage of open source software in the past was that everyone could share code, which could enable multiple people to work together to fix bugs and continuously update the software. However, the large model itself is a black box, and there is a possibility of retraining after someone submits modifications. Each training consumes a lot of computing power and funds, and it does not benefit as much from collaborative development among multiple people as open-source software did in the past.
The above individuals judge that in the long run, closed source is more efficient in concentrating resources such as intelligence and computing power to iterate large models compared to open source.
Robin Lee also stressed that the advantage of closed source is that it has a real business model and can earn money to gather computing power and talents. In terms of cost, the closed source model has lower inference cost and faster response speed under the same capability. Under the same parameters, the closed source model also has stronger capabilities.
"Today, whether in China or the United States, the strongest basic models are all closed source, and all kinds of small models, the best small models, are distilled through large models. The models made by reducing dimensions through large models are better, which will also lead to closed source advantages in cost and efficiency," he said.
Regarding the open source and closed source dispute over large models, Wang Xiaochuan, CEO of Baichuan Intelligence, also mentioned it in a previous interview with Interface News. His viewpoint is that the big model itself does not represent the consumer end, unlike Android and iOS, which require a choice between two. Today, from the perspective of the enterprise end, open source and closed source are both necessary.
Wang Xiaochuan values the value brought by open source very much. He believes that in the future, 80% of enterprises will use open source models because open source models are small and cannot adapt well to many scenarios with closed source.
The core competitiveness of AI entrepreneurs is not the model itself
In addition to his position on the open source and closed source routes, Robin Lee also put forward his own views on AI entrepreneurs and start-ups.
He believes that the so-called "dual wheel drive" of some modeling startups is not a good model, as both modeling and application will inevitably distract energy. Entrepreneurship companies have limited energy and resources, and when resources are limited, they should focus more instead of pursuing so-called "dual 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 true advantage of entrepreneurs should be their knowledge and data in a certain field.
"If you are looking for a 'yellow men's swimsuit without pockets' today, and you cannot find it on any e-commerce platform, the current technology cannot solve this demand. If you have domain knowledge, a large model can be solved, which is an example of how domain knowledge can provide unique value.".
In his view, there are a large number of models on the market, big, small, open source, and closed source. There are skills in how to use the combination of these models in specific applications, which is something entrepreneurs can do and can provide value gain.
As for the concern of the outside world that if Wenxin or closed source model is used, if it is done well, it will be copied and robbed. Robin Lee also responded that in the mobile era, WeChat did not eat Pinduoduo, nor did Didi become part of Tencent. They each provide their own unique value and have their own very different competitiveness. Their rise relies on a closed platform in the mobile ecosystem - WeChat, but they are not afraid of WeChat grabbing its job, so there is no need to worry about the basic model taking over AI applications.
According to a research report by CITIC Securities, the current domestic large-scale model capabilities are gradually improving, and there are alignment and leading advantages in Chinese processing and some features such as long text processing. With the increasing popularity of Kimi, several major domestic model manufacturers have joined the competition for long text capabilities, accelerating the landing of the industrial consumer end. The first year of domestic large-scale model application has arrived.
As the model gradually matures, future big models will engage in a new round of competition and competition at the product and application levels. Robin Lee's speech is also attracting more application layer developers to choose Wenxin model.
At the previous telephone conference of the fourth quarter of 2023 and the annual financial report of Baidu AI Cloud, Robin Lee 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 for cloud business.
At present, the daily adjustment usage of the Wenxin model has exceeded 50 million times, with a quarter on quarter increase of 190%. Last December, about 26000 companies called the Wenxin Big Model, with a quarter on quarter increase of 150%. Samsung, Honor, Autohome and other companies have all reached cooperation with Baidu.
Since its release, Baidu has continuously reduced the inference cost of the Wenxin Big Model, which has now decreased to 1% of the March version last year.
Robin Lee also said that in the future, multimodal or multimodal integration, such as text to video, is a very important direction of basic model development and a necessary direction of AGI (General Artificial Intelligence). Baidu has already invested in these areas and will continue to invest in the future.