Rewind the time to a year ago, at that time, many ** discussed with netizens, can China make a large model application similar to ChatGPT?
At that time, we said, don't panic, don't be busy, China will definitely be able to make a big model. In an instant, 2023, known as the "first year of the big model", is about to end with the sound of firecrackers. China's AI model, looking up, has felt that thousands of mountains are green.
According to relevant data, more than 230 large models have been released in the Chinese market by October 2023. "Hundred models rush to take the exam" is well deserved. However, the number of large models is not the end of making a large model, or even the purpose of making a large model.
At this stage today, we must answer a new question: how can we do a good job and use the large model well? How can we make the big model make great value for the social economy?
To answer this question, it is necessary to know what is "good" for a large model.
The evaluation of the large model is not based on parameters and scale, but on efficiency, technology, application, ecology, and ultimately who is the best and most useful.
In these questions, we will find the other side of the domestic large model pattern: although there are many large models, ordinary people know that they are not used much. Just ask a friend and find that he probably only knows one domestic model, that is, Wenxin Yiyan and the Wenxin model behind it.
Although there are more than 100 large models galloping on horses, it is a large model of Wenxin. Why is this happening?
Only by understanding this problem can we understand the essential law of the large model: only by accumulating thick and thin, deepening and becoming stronger, is the future road of the domestic large model.
Looking back at the end of the year, we can re-examine the industry pattern of "100 models rushing to take the exam and winning the championship with Wenxin", and we can see that the dawn of AGI is quietly coming to this land.
Ahead of the curve: the efficiency race for large models
At the time when the AI model was just emerging, users were curious about it, the industry was eager for it, and the social economy had extensive and diverse expectations for it. At this time, whoever can take the lead in bringing the large model to users and the industry will be able to make a pioneer and determine their own industry advantages through their leading position.
Looking back at the development of the large model industry in the past year, we will find that Wenxin large model has played a pioneering role every time. In the end, the large-scale model technology will be put in the hands of users, developers and thousands of industries as soon as possible. The high efficiency and fast pace of the Wenxin model, as well as the technological leadership and product confidence behind it, are the primary keys to its ability to maintain "one ride in the dust".
In March 2023, it took the lead in releasing the Wenxin Yiyan large language model. This is due to more than ten years of layout and deep cultivation of deep learning technology, and extensive AI business practices. Since 2019, we have been deeply engaged in the research and development of pre-models and created a Wenxin large model system. Wenxin, who has accumulated a lot of experience and made sufficient preparations, has achieved a leading position at the beginning of the year.
In July 2023, during the 2023 World Artificial Intelligence Conference, the National Artificial Intelligence Standardization General Group announced the list of leaders of China's first large-scale model standardization special group leader, and served as the co-leader unit. At this point, Wenxin large model officially entered the "large model national team" to explore the channel and establish the direction for the standardization of domestic large models.
Next, on August 31, Wenxin Yiyan was officially opened to the whole society, becoming the earliest large language model that the Chinese public can experience, truly bringing the ability of large models to thousands of households.
In December 2013, the results of China's first official "Large Model Standard Conformity Evaluation" were released. Wenxin Yiyan became one of the first batch of language models that passed the evaluation and met the relevant technical requirements of the "Large-scale Pre-trained Models for Artificial Intelligence Part 2: Evaluation Indicators and Methods", and fully met the relevant national standards in terms of versatility and intelligence.
The first to launch, the latest to open, the first to meet the national standards, "first" has become the key word of the Wenxin model. Deeply cultivating AI core technologies and building an AI ecosystem brings high efficiency and fast pace through accumulation.
Wenxin is the first to ride the dust, which can be used as a reference for the entire AI industry.
Thickening technology: accumulation of technology differentiation of large models
When we discuss the problem of too many AI large models and saturation of the industry, we often notice a phenomenon: hundreds of large models have come out, but it is difficult for us to find out what are the technical differences between these large models? Not to mention the difference in technology, which brings about the difference in application.
The reason for this phenomenon lies in the blind pursuit of data volume and model parameters by large models, and the neglect of the research and development and accumulation of core technologies. The reason why the Wenxin model has a good response and high user recognition is that it has chosen the thickest and most solid road of technology research and development. Each generation of Wenxin model upgrade is based on the research and development and application of new technology capabilities. As a result, the more Wenxin is upgraded, the greater the gap between the technical capabilities of other large models. The snowball phenomenon of technology began to appear, and finally formed the strategic advantage of Wenxin model in terms of technology. Pile up mountains and hundreds of people, starting from the soil.
Behind the Wenxin Yiyan released in 2023 is the Wenxin Model 30。At that time, Wenxin had already established technological differentiation in the field of large models with knowledge enhancement technology, and had the advantages of knowledge enhancement, retrieval enhancement and dialogue enhancement.
Then in May, Wenxin model 3Version 5 released. It has made innovations in basic models, fine-tuning technology, knowledge point enhancement, logical reasoning, plug-in mechanism, etc., and has achieved an overall improvement in generation effect and efficiency.
By October, Wenxin model 4Version 0 is officially launched. It has achieved breakthroughs in a number of key technical directions, and has significantly improved its four major capabilities: comprehension, generation, logic, and memory. In particular, the logic and memory capabilities have been greatly improved, bringing users very intuitive help.
The Wenxin model can achieve such a high efficiency and substantial technical upgrade, which is inseparable from the efficient computing power, self-developed framework, and collaborative optimization of the data processing mechanism behind it. In particular, the joint optimization of Wenxin PaddlePaddle has become a well-known case of rapid development of large models in the industry, and has been widely discussed by the AI industry in the past year.
Based on the paddle platform on the computing power of Wanka, the Wenxin large model supports the stable and efficient training of the large model through the collaborative optimization of software and hardware of the cluster infrastructure, scheduling system, and paddle framework. Since its release in March 2023, the training efficiency of the Wenxin model has increased several times, and the average weekly training efficiency has exceeded 98%.
The exploration of core technologies and the accumulation of technological differentiation have made the Wenxin model have a strong sense of technology. This is Wenxin's continuous leading hole card and background color. As long as there is a higher pursuit of technology, many questions will naturally have answers.
Wide application: How to bring large models to the front line of application?
If you want to know whether a tree is a pillar of the wood, you can't just sit and talk about it and praise it to the sky, but you must really use it and let it build a house and play its own value.
The same is true for large AI models. Whether the large model is useful or not is not said in the press conference and test data, but in the hands of hundreds of millions of users and thousands of industries.
Looking at the development of the large-scale model industry in the past year, we will find that in terms of the breadth of application, Wenxin is difficult to match other large-scale models. Among C-end users, only Wenxin Yiyan has achieved a scale of 100 million users; In the B-side application, the number of calls of the Wenxin large model exceeds that of the other 200 large models combined.
The leading application spanning orders of magnitude has allowed the Wenxin model to explore countless value possibilities in the hands of industry users, developers, and ordinary users. In the business, Wenxin model has been widely used in Internet products such as search, information flow, and smart speakers. In the process of opening up to the outside world, Wenxin model empowers various industries such as manufacturing, energy, finance, communications, cities, and education through the PaddlePaddle open source open platform and intelligent cloud. Wenxin Model has built more than 10 industry models with leading enterprises and institutions in various industries to accelerate the intelligent upgrading of the industry.
In the national diving team, the Wenxin model has comprehensively upgraded the AI-assisted training system, which can not only understand and execute the complex instructions of coaches and athletes, but also score and accurately quantify the movements. In 2023, the Chinese Swimming Association awarded the title of "Artificial Intelligence Partner of the Chinese National Diving Team".
In cooperation with the National Library of China, Wenxin Model has created the "Ancient and Modern Questions" service by learning a large number of ancient chronicles and genealogical data, and performing text recognition and understanding. Users only need to enter the root-seeking information to get the corresponding clue feedback to help the global Chinese find their roots and ancestors.
In the hands of the SoundBridge AI language training team, the AI speaking application** built based on the PaddlePaddle and Wenxin large models can provide feedback and guidance in the form of text to help the hearing-impaired with language training.
The peach and plum do not speak, and the next is its own. The wide application of Wenxin is the best proof of its value. At the same time, it also proves that China's AI model is not only in quantity, but also in terms of application quality, and there is intelligent exploration of real materials.
Ecological needs to be prosperous: the ecological construction of large models is urgent
We all know that the most difficult thing to make software is to do ecology. The ecology determines the upper limit of the exploration of basic software technology, and also determines the possibility of its long-term development in the future. When AI technology was just developing, it became an industry consensus that AI must be done in an ecosystem.
However, at this stage, we can see that there are more than 100 large models in China, but few manufacturers pay attention to ecological construction. In the long run, it is easy to cause the large model to become a "unique product" that no one will use and no one wants to use.
A large part of the reason why Wenxin model can maintain its leading position comes from the support and promotion of the ecology. Only the large model built and created by tens of millions of people is a large model with vitality and staying power.
To this end, we will promote the joint innovation and mutual promotion of PaddlePaddle and Wenxin Ecology. As of December 2023, PaddlePaddle has gathered 10.7 million developers and served 2350,000 enterprises and institutions have created 860,000 models based on paddles. PaddlePaddle Ecology and Wenxin Ecology help each other and promote the rapid development of each other.
The developer community is the key support for the development of the ecosystem, and the Galaxy Community, the largest AI community in China, has been officially launched with the development of large models, providing developers with an integrated large model development experience and rich product functions. As of December 2023, the Galaxy model community has launched more than 4,000 innovative AI applications based on the Wenxin model.
In terms of ecological co-creation, the company has released the Wenxin large-scale model Galaxy co-creation plan, hoping to cooperate with developers and ecological partners to achieve extensive innovation in AI applications.
Complementing the developer ecosystem is the construction of the talent ecosystem. In 2020, the goal of "cultivating 5 million artificial intelligence talents for the whole society in 5 years" was proposed. As of October 2023, 4.2 million AI talents have been trained. In the face of the huge demand for talents in large-scale model technology, a new talent training Galaxy plan will be released in 2023, which will cooperate with industry, academia and research circles to cultivate another 5 million large-scale model talents for the society.
No matter from any dimension such as developer aggregation, application innovation, and talent training, the ecological construction of Wenxin model has shown a prosperous side. The ecology starts quickly, the vitality is high, and the audience is wide, and the "thousands of trees and pear blossoms" on the soil of the Wenxin model are truly realized.
It can be said that the rise of Wenxin ecology has laid a model for the overall construction of domestic large-scale model ecology and broadened the boundary.
From the first year of the large model to the dawn of AGI
After a turbulent year, the question in the field of large models has changed from "whether there is a large model" to "whether the large model can be done well" and "whether the large model can be used well".
In the face of new problems and new tests, the leading position of Wenxin model in four aspects: position, technology, application and ecology is the answer.
According to IDC's "AI Large Model Technical Capability Evaluation Report, 2023", Wenxin Large Model scored 7 full scores in 12 indicators among the 14 participating models, ranking first in the comprehensive score of domestic mainstream large models, and winning the only full score in the two key indicators of algorithm model and industry coverage.
According to the "AI Large Model Comprehensive Ability Evaluation Report" released by People's Data, Wenxin Yiyan not only surpassed ChatGPT in terms of comprehensive score, ranking first in the world, but also surpassed ChatGPT in the three dimensions of content ecology, data cognition, and knowledge Q&A. And the scores of the six dimensions are all ranked first in the domestic large model list.
If we say, in the first year of the large model, what we saw was the number of large models.
So in the farther future, in the spring of the large model we are looking forward to, what we need to see is the application quality and inclusive value of the large model.
How can this evolution be achieved? In the past year, the Wenxin model has answered with the words "first", "thick", "wide" and "sheng". Only when the entire industry develops in such a path, removes the chaff, avoids hypocrisy and is pragmatic, can the domestic large model continue to improve, open the spring of the industry, and move towards the summer of agi.
On August 16, 2023, the W**e Summit Deep Learning Developer Conference 2023 was held in Beijing. During the period, Wang Haifeng, chief technology officer and director of the National Engineering Research Center for Deep Learning Technology and Application, said that artificial intelligence has a variety of typical capabilities, and understanding, generation, logic, and memory are the basic capabilities.
Only by finding the right direction, choosing technology long-distance running, and choosing ecological win-win can we make the long journey land step by step and make the dawn of AGI a reality.
The industry pattern of "100 models rushing to take the exam and winning the championship with Wenxin" is essentially an affirmation of technocracy and pragmatism.
If you understand this, you will find the way forward for the development of AI technology.
List of high-quality authors