In the tide of global artificial intelligence, a battle for large models is quietly unfolding. The competition, called "100 Model Wars", is a battle between domestic and foreign technology giants and emerging forces in the field of AI. But behind this seemingly prosperous competition, there is a fact that cannot be ignored: most of the domestic self-developed large models are just "shell" products of Western open source models.
This phenomenon raises a series of questions: Are we really developing large models on our own?What is the innovation ability in the domestic AI field?It's time to lift the veil of this whitewash and face up to the true face of the domestic large-scale model field.
There are two mainstream development routes for global large models: closed-source and open-source.
Closed-source, with OpenAI's GPT series as its banner, represents a combination of commercialization and know-how. The open-source school is represented by Meta's Alpaca series model, which symbolizes the democratization of knowledge sharing and technology.
The domestic AI field responded quickly, and for a time, a situation of "100 model wars" was formed. But how many of the large models involved in this battle are built on independent innovation?On closer inspection, it is not difficult to find that most of the so-called self-developed large models are actually just the product of standing on the shoulders of open source giants and making slight adjustments.
The prime example is Kai-Lee's "Yi" model, which was revealed to be only a superficial modification of Llama - only changing the names of the two tensors. This kind of plagiaristic innovation is not an isolated case in the domestic AI industry, but a common phenomenon. The practice of zero and one things is just a fig leaf that has been hidden for a long time in the field of domestic large models.
So, what is behind this "casing" phenomenon?
It exposes the shortcomings of the country in terms of original innovation, although it has a huge market and capital investment, but in the originality of the core technology, we seem to be still on the way to catch up. Although this kind of follow-up innovation strategy can quickly fill the gap in the market in the short term, it lacks continuous innovation momentum in the long run.
At the same time, this also exposes the desire for "quick results" in the domestic AI field. Driven by a fast-growing market, companies may be more inclined to take a quick copy and slight modification approach to seize the market than to invest time and resources in deep technological innovation. Although this strategy can bring benefits in the short term, it may sacrifice long-term technology accumulation and the healthy development of the industry.
This phenomenon also reflects the lack of technical confidence in the domestic AI industry, and in the global AI technology race, we seem to be more in the position of followers than leaders. This situation needs to be gradually changed by strengthening original innovation, technology accumulation and talent training.
In order to catch up with its American counterparts in the field of large models, China must not only catch up with OpenAI's GPT-4 in the closed-source model, but also catch up with Meta's LLAMA in the open source field.
In February 2023, Meta released the alpaca series model for the first time. In this initial release, the alpaca series includes models of four different scales: 700 million, 1.3 billion, 3.3 billion, and 6.5 billion parameters.
In July, Meta announced its latest large model, Llama 2 (Alpaca 2), which includes three parameter variants: 7b, 13b, and 70b, which can be used for commercial or research purposes for free.
It is gratifying that domestic manufacturers are also actively deploying in the field of open source large models.
Recently, Alibaba Cloud Tongyi Qianwen's 72 billion parameter model QWEN-72B was announced to be open source. So far, Tongyi Qianwen has open-sourced four large language models with 1.8 billion, 7 billion, 14 billion, and 72 billion parameters, as well as two multi-modal large models for visual understanding and audio understanding, realizing "full-size, full-modal" open source.
It can be found that in terms of parameter scale, Ali Tongyi Qianwen has equaled Meta's Alpaca 2, both with a parameter scale of 70 billion.
Another point worth noting is that some large-scale model startups in China are also launching open-source large-scale models. For example,In July, Zhipu AI open-sourced CHATGLM-6B and CHATGLM2-6B;In the same month, Baichuan Intelligent open-sourced baichuan-7b and baichuan-13b.
China has its own open-source large-scale model ecosystem, which is of far-reaching significance for the independence and controllability of technology and industry.
These open-source models provide a strong technical foundation for Chinese tech companies, allowing them to more confidently demonstrate their strength in the global AI competition. The existence of these open-source models provides more choices and flexibility for Chinese researchers and developers, and promotes the rapid development and innovation of domestic AI technology. This will also help reduce dependence on foreign technology and enhance China's autonomy and security in key technology areas.
This series of actions by China in the field of large models is not only a shining debut in the technology competition, but also an important contribution to the global AI landscape. It not only demonstrates the strength and innovation capabilities of Chinese technology companies in the field of large models, but also opens up a new path for the development and application of global AI technology.
As more Chinese companies and research institutions join the race for open source models, there is reason to believe that China will play an increasingly important role in the global AI stage.
In the field of large models, should it be developed open-source or closed-source?
If you look at it from a commercial point of view, it's not a good judgment.
However, from the perspective of the overall interests of mankind, the open-source large-scale model route must be more "safe". The main manifestations are:
1. The open-source model is easier to be understood and supervised by the public.
The transparency of the open source model makes it a model for democratizing technology, providing a wide range of researchers and developers with tools that are easy to understand and use, and allowing the general public to participate in oversight.
This open mode of communication and collaboration enables the rapid popularization of the latest technological advances and knowledge, ensuring that all sectors of society have the right to know and have a say in the direction of AI development. This interaction not only drives rapid technological advancements, but also helps to increase public trust and acceptance of AI technology.
2. Avoid AGI being monopolized by a technology giant.
The ultimate goal of the big model is AGI, and such a powerful force must not be controlled by a profit-seeking commercial company.
The original intention of OpenAI was to avoid artificial intelligence being monopolized by Google. The "open" in the name openai itself means open and open source. It's just that now OpenAI is becoming more and more closed, which is contrary to the original intention of its establishment.
For the overall benefit of humanity, the large model should maintain at least an open-source technical route as a plan b for humanity.
3. Open source large model is conducive to the prosperity and innovation of the large model industry.
The open-source model provides startups with a powerful platform for innovation and lowers the barrier to entry into the field of artificial intelligence. These companies are able to build on existing advanced models and develop customized improvements and applications, so as to achieve technological innovation quickly.
This model not only promotes technological diversity and application innovation, but also brings vitality and competitiveness to the entire AI industry, accelerates the commercialization process of new technologies, and strongly promotes the healthy development and prosperity of the entire industry.
Next, let's take a look at two examples of application innovation based on the open source model of Ali Tongyi Qianwen.
Case 1: Large model + robot = embodied intelligence
With the rapid development of intelligent technology, embodied intelligence has gradually entered our lives. Zhejiang Youlu Robotics Technology, a start-up focusing on combining large models and embodied intelligence, has made remarkable progress in this field.
Founder and CEO Chen Junbo leads the team, relying on the self-developed second-generation embodied intelligence model, and is committed to integrating a highly adaptable and generalizable general intelligent brain into every professional device.
Youlu Robot's vision is to make traditional professional equipment intelligent, and this is the biggest opportunity in the era of embodied intelligence. There are tens of thousands of traditional professional equipment manufacturers in China, and they urgently need artificial intelligence systems to upgrade existing products. Youlu Robotics provides a general artificial intelligence brain for these devices, which not only reduces the cost of developing intelligent systems separately for each model, but also improves the intelligence level of the product.
Youlu Robotics has successfully integrated the Tongyi Qianwen open-source model QWEN-7B in the road cleaning robot. This intelligent cleaning robot is able to interact with the user in real time through natural language, understand and execute the user's instructions. This innovation not only improves the efficiency of the robot, but also enhances its flexibility in real-world applications.
Chen Junbo explained that they chose the Tongyi Qianwen model for several reasons: first, it is one of the best open source models in the Chinese field;Second, an easy-to-use toolchain is provided for quick experiments and fine-tuneIn addition, the quantization model has no loss and is suitable for deployment on embedded devicesFinally, the services provided by Tongyi Qianwen are responsive and can meet the diverse needs of enterprises.
The successful case of Youlu Robot shows that the application of open source large models in the field of embodied intelligence has broad prospects. For applications that need to evolve and adapt to new data, the open source model is undoubtedly the better choice. This not only promotes the development of embodied intelligence, but also brings revolutionary changes to related industries.
Case 2: Mental model
In modern society, people are facing more and more psychological pressures and challenges, and mental health has become the focus of public attention. In response to this demand, the X-D Lab (Heart Beat Lab) team of East China University of Science and Technology, with Yan Xin as the core member, is committed to developing AI applications that can soothe and improve the hearts of contemporary people.
Based on the open source model of Tongyi Qianwen, they have developed a series of large models for mental health, medical health, education and examinations, including mental health model MindChat, medical and health model Sunsimiao, and educational examination model GradChat.
In particular, MindChat, as a psychological counseling tool, is like an AI psychological counselor, providing users with timely, safe and convenient psychological assessment services. Through the Alibaba Cloud Magic Community, users can experience the actual effect of this model. By analyzing the user's text content and voice tone, MindChat can empathize with the user, provide them with personalized advice, and even recommend real human or psychological experts to intervene if necessary.
The project originated from a dinner table conversation between Yan Xin and her teacher about loneliness and mental health issues in society at large. Based on this, they decided to develop a large psychological model that can provide emotional outlets and maintain social connections.
Currently,More than 200,000 people have used their large model and provided more than 1 million Q&A services, which has helped many people solve the pressure of further education, postgraduate entrance examination, employment, workplace, etc.
In addition, the team also attaches great importance to user privacy protection, and uses a distributed architecture to store and analyze training data to ensure user information security. Yan Xin emphasized that the open-source model was chosen to achieve the sustainability and adaptability of the technology, especially in areas with a high focus on privacy, such as psychology and medicine. They chose the Tongyi Qianwen model because it has the best intelligent performance in the Chinese field, provides an easy-to-use toolchain, supports fast experimentation and fine-tune, and quantifies the efficiency of the model in deployment.
In the end, X-D Lab adopted a combination of open and closed sources, which not only fed back the open source community, but also provided services for real-world scenarios in the form of closed-source APIs, ensuring that their technology was both innovative and could meet the needs of practical applications. In this way, they hope to make the big model technology work for a wider range of social groups, especially those seeking help in the field of mental health.
It should be pointed out that the open source model is a good thing, but there is a key premise, that is, the open source model itself is very powerful. In addition, this open source model will continue to evolve and promote the technological progress of the entire open source ecosystem. Taking the model of General Qianwen as an example, only a large model with a scale of 70 billion parameters has been open-sourced, which is not enough. I believe that in the future, we will see some open source large models with hundreds of billions of parameters.
In the vast ocean of exploring artificial intelligence, the open source model is like a lighthouse, illuminating the way forward. This is not only a symbol of technological progress, but also a victory for the sharing of human wisdom. The emergence of open-source large models has broken the barriers of knowledge and technology, allowing more researchers, enterprises and even the general public to share the scientific and technological achievements of all generations and jointly promote the development of the field of artificial intelligence.
However, we should also be aware that the journey to open source large models has only just begun. In the future, we need more innovation, cooperation and wisdom to continuously improve the capabilities and applicability of these models.
We look forward to seeing the birth of more powerful and intelligent open source models, which will not only be technological innovation, but also partners in human exploration of the unknown world, and work together to create a smarter and better future. Let us look forward to witnessing more miracles brought to the world by artificial intelligence on this road full of challenges and opportunities.
Text: A cloud of smoke and rainData Ape.