Is open source or closed source better for large models? Recently, the dispute over the route of opening and closing the source caused another storm because of an internal speech by Robin Lee, chairman of Baidu.
On the evening of April 11, Robin Lee, chairman of Baidu, exposed his internal speech, which poured a lot of cold water on the big open source model. He said that the big open source model is of little significance. The closed source model will continue to lead in capacity, and the "two wheel drive" of model making start-ups, both models and applications, is not a good model.
"The core of the debate over open and closed sources is to see who is open source. 2. Dual wheel drive is the only solution for first-line startup AGI companies." Exclusive news from First Financial News: That night, Wang Xiaochuan, CEO of Baichuan Intelligence, expressed different views in a WeChat group discussion.
In the group, Zhu Xiaohu, the executive partner of Jinshajiang Venture Capital, also said, "GPT4 can achieve more than 90% of commercial demand, and it doesn't matter whether the source is open or closed. In the future, the big model API will be the price of tap water." He added, "But what users need is pure water, sparkling water, oolong tea..."
Chen Ran, the founder of OpenCSG, a large model ecosystem community, said in an interview with First Financial reporters that she "seriously disagrees" with the statement that open source has little significance. Whether in China or abroad, open source for large models has become a major trend, and the open source trend will promote and drive the commercialization of the industry on top of it, with fast iteration, quick trial and error, and co creation and sharing. "Open source will only become more and more surging and upward.".
In the industry, industry insiders who firmly believe that closed source big models are better than open source big models are divided from those who promote open source big models. First Financial reporters have communicated with multiple industry insiders and learned that closed source and open source big models have formed a preliminary differentiation in application scenarios, each with its own advantages and disadvantages, but there is room for survival in the early stages of big model application.
Is closed source or open source better?
Among domestic technology giants, companies such as Baidu, ByteDance, Tencent, and Huawei have not yet released open-source big models. At the same time, many companies have also chosen a parallel path of closed source and open source. As of now, various open-source models with different parameters have been released, including Alibaba Tongyi Qianwen, Baichuan Intelligence, 360, Kunlun Wanwei, Zhipu AI, Inspur, Zhiyuan, iFlytek, Zero One Everything, APUS, and others. Among them, the open source big models that reach billions of parameters include APUS xDAN Big Model 4.0 (MoE) (136 billion) and Inspur Information's "Source 2.0" 102B (102.6 billion).
Why does Robin Lee think that the open source of the big model is not very meaningful? In his internal speech, he gave the reason that a closed source model with a business model can gather manpower and financial resources. The strongest basic models in China and the United States are both closed source, while the best small models are made by reducing the dimensionality of large models. Moreover, closed source models have advantages in cost and efficiency, and have lower inference costs and faster response times with the same ability.
He also has another reason, which is that open source models are different from traditional open source software such as Linux and Android. "Although the open source model Llama also encourages everyone to contribute various data and code, in reality, the main developer is Meta, not a product that truly everyone collaborates on."
Standing on the side of "closed source", there is also Yang Zhilin, the founder of the dark side of the moon. He previously stated in an interview that closed source will attract talent and capital, and ultimately, closed source is better. There are hundreds of applications based on the open source diffusion model, Stable Diffusion, overseas, but none of them have been released.
Yang Zhilin previously mentioned that in the past, everyone could contribute to open source, but now big model open source itself is still centralized. As of now, the dark side of the moon has not publicly mentioned any specific plans or projects related to open source big models.
Earlier, artificial intelligence scientist Shen Xiangyang discussed the debate over the open source and closed source path of large models at an industry forum. He believes that the first place in the industry must be closed source, the second place he is still hesitant, and the third place will be open source.
But there are always people who believe in the power of open source.
The person in charge of an open-source model with different parameter scales, as well as a closed source large model, told First Financial reporters that the capabilities of the model are constantly developing, and technological innovation and breakthroughs in the AI field have never stopped. Different enterprises and development environments may have different considerations for model size, model capabilities, and underlying resources. Sometimes people consider ultimate performance, while others have sufficient resources and require higher quality models.
Another practitioner told the reporter that Robin Lee meant that entrepreneurs should rely on Baidu's big model. But currently, the big model is still in its early stages of development, and there is far no one that dominates or monopolizes the industry. Even though Baidu was one of the early models to enter the domestic market, it is now facing a surprise attack from Kimi on the dark side of the moon.
Wang Xiaochuan also believes that open source and closed source are not like iOS or Android operating systems in mobile phones, which can only be chosen between two options. Open source is indeed easy to "build character" and "have many friends", allowing everyone to quickly understand and evaluate the quality of large models. At the same time, open source is also a preparation for commercialization. If you find it good to use, you can explore further commercialization paths when you need better services and larger parameters.
Previously, Alibaba Cloud CTO Zhou Jingren also responded to the route dispute in interviews with media including First Financial, stating that the open source community has flourished. He said that Alibaba Cloud's original intention was not to commercialize models in its own hands, but to help developers. The open source ecosystem is crucial for promoting technological progress and application implementation of large models in China, as well as the flourishing development of the ecosystem.
"From the perspective of resources, data, and manpower, the claim that closed source big models with business models do better than open source big models is to some extent valid. The reason is that model training costs are high and requires financial support, such as training for $10 million per session," Zhang Junlin, the head of new technology research and development on Sina Weibo, told reporters. But from another perspective, this statement is not entirely correct because many open source big models are also made by large companies and there is also resource investment.
Professor Lin Dahua, a leading scientist at the Shanghai Artificial Intelligence Laboratory, previously stated in an interview with First Financial that closed source models may have a stronger focus on product polishing, resulting in better product maturity in all aspects. For open source models, although they may not be as mature as closed source models, they can support a wider range of practitioners for use and secondary development. "In the end, the real market share of open source models may be very high.". He believes that open source and closed source will present a complementary path, similar to Windows and Linux in the field of operating systems in the past trend of computer technology development.
Liu Wei, Senior Legal Advisor of the Open Atom Open Source Foundation, previously stated at the GDC conference that there are several bottlenecks to the closed source model. One is that the cost of cloud based inference is too high, and if the concurrency demand is high, it may result in significant delays. The larger the model, the lower its flexibility and economy, and the typical hallucination problems of the model may also require support from the scene to be further integrated.
"Open source models are superior to open source small models, which can be flexibly deployed to terminal devices, including PCs and mobile phones to achieve low latency inference AI experience. They can train vertical domain applications with high-quality data, accelerate the empowerment of large models in applications, which means that closed source 'diseases' can be' cured 'by open source." Liu Wei mentioned.
MiniMax Technology Vice President Anderson previously told First Financial reporters that both open source and closed source have their own advantages and value. Open source can attract more developers to participate together, accelerate the development and popularization of technology, and also stimulate more innovation and creativity. The quality and stability of closed source models are more controllable, which can better protect intellectual property and promote commercialization. "I believe that open source and closed source can complement each other and form a more open, collaborative, and innovative development model, jointly promoting the development of the field of big models."
Big model developer Gao Xiaoan (pseudonym) started fine-tuning based on open source models and publishing projects on open source platforms last year. He believes that open source has brought favorable changes to the big model ecosystem. "Large model developers can do a lot of secondary creation work based on open-source models, and various versions of StableDiffusion models and Chinese Llama models based on Llama models have emerged, which cannot be achieved by closed source models."
Gao Xiaoan also believes that another advantage of open source big models cannot be ignored. Compared to the potential data leakage problem caused by feeding company information to closed source models, collecting unique business data to train open source models reduces this concern. Trained models can also rival or even surpass closed source models such as ChatGPT. Of course, there is another advantage of open source big models, which is to advertise to big model enterprises, including Google, Alibaba, and Baichuan Intelligence, which have opened up models with smaller parameter quantities, while models with larger parameter quantities are closed or even charged.
Open source ToB, closed source ToC?
Talking about whether open source big models are better or closed source big models is better, a potential semantics is that the two will compete against each other in the same application scenario. However, in reality, there has been a certain degree of differentiation between open source and closed source models. A large model developer told reporters that many companies have developed their own large models based on open source models for free and data security considerations.
"Open source and closed source form two camps. In terms of scenarios, open source is more inclined towards To B, while closed source is more inclined towards To C, which are two tracks." Zhang Junlin said that open source cannot be To C because it does not provide specific products and cannot respond to individual users. However, open source can be targeted at enterprises, which can gain advantages in data privacy based on the open source big model.
Wang Xiaochuan believes that from the perspective of To B, open source and closed source actually require. In the future, 80% of enterprises will use open-source big models because closed source cannot better adapt products, 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.
Lin Dahua mentioned that from the perspective of ToB, open source is better than closed source because it has greater openness for secondary development. Although closed source model vendors will launch platform based services in the future, developers can use their standardized tools to create models for a single industry and vertical direction. However, the demands of various industries are very complex and diverse, and may not be fully covered by a standardized toolchain.
"Enterprises and institutions in many industries will have needs and want to control the iteration of the whole model and guide them to the direction they need, but they do not have the investment in the basic model, nor are they willing to invest too much or do not have the ability to do so. In this case, the open source model is the best choice for them. They can do secondary development in various details around it, which is not necessarily supported by the standard tool chain." Lin Dahua said that in the future, this open source model can support a large number of industrial needs in the economic system.
For closed source models, greater opportunities exist in some highly clustered tracks, "such as the possibility of a chat app with a very large C-end traffic in the end, which could be some big companies coming out of a commercial system."
Lin Dahua believes that the core logic of the closed source big model lies in its ability to form a business loop, receive a large amount of user feedback, and form barriers in fixed channels. Assuming that it can ultimately find the areas that truly hit user pain points and improve its ability through user feedback, it can occupy a track and commercialize and monetize within it. Ultimately, there will be different patterns for different needs, and closed source and open source will form a complementary effect.
From the perspective of selecting large model manufacturers, Zhang Junlin believes that choosing open source is a differentiation strategy. The closed source big model business model is clearer, but the downside is that if the model does not achieve the best results, it is difficult to charge. If we take the open-source route, although the model is not charged in principle, it can also form a business model based on it, just like open-source Linux also has a business model. So if manufacturers don't have the confidence to do the best with big models, open source is a choice.
Zhang Junlin believes that open source does not necessarily lead to success. Even if the open source big model is not the best, it should have its own characteristics, and the open source big model should be accurately positioned. For example, taking the "small but strong" route may not be as effective as models with large parameters, but it has the characteristics of low cost, simple deployment, and low hardware requirements. It can be deployed to mobile devices and has multiple application scenarios. Another option is to increase the parameters, such as Grok, which focuses on good performance.
After experiencing it, Gao Xiaoan told reporters that there are also differences in the approaches of open source and closed source models. Although the number of open-source Grok-1 parameters has recently reached hundreds of billions, many open-source large models only have 7 billion and 13 billion parameters, such as Alibaba Cloud Qwen-7B and Baichuan2-13B. He told reporters that compared to comparing parameters in closed source large models, these small parameter open source large models can achieve better results by stacking more pre training data.
"Under the same amount of data, larger parameter models have faster convergence speed and better performance in training, but the training cost is also much higher. Smaller models have lower deployment costs and are more user-friendly in practical business use. Some experiments have shown that the pre training quantity of 7 billion and 13 billion models has not yet reached saturation." Gao Xiaoan believes.
Zhang Junlin also believes that the ability of small models has been rapidly improving, and the upper limit of their ability is not yet visible. In principle, as long as more data is given to small models, the effect will continue to improve.
However, Gao Xiaoan believes that open source big models also have obvious drawbacks. If we don't talk about the advertising effect of open source big models, there are still certain difficulties in the commercialization of open source big models themselves. Open source models can consider charging for commercial versions, but this approach is more difficult in the fiercely competitive environment of open source models.
What are the profit points of open-source companies? Lin Dahua believes that open source vendors can establish a service, "What is valuable is not only the model itself, but also its services, because doing secondary development based on the open source model is quite complex. Without sufficient technical support, the cost of doing this will be very high, and the value of technical services can be realized."
As for the open source business model, Chen Ran believes that this approach is similar to the market model of mobile applications in the Internet era. "After a period of free trial, there are functions or services packaged by enterprises. This thing is also equivalent to an engine in a car. It can't be used directly and needs to be packaged into a whole vehicle (enterprise function)."
"The big model will unfold a very large commercial space, and there will be very different patterns in different ways and points. The most taboo is to put everything on one 'shoe'," Lin Dahua told First Financial.