In the garden of the United Nations Headquarters, there stands a bronze sculpture called "Let Us Beat Swords into Plowshares", a man with a hammer in his hand, who is transforming the sword into a plough - "casting the sword into a plough", representing the consensus of human society for civilization and development for nearly a century: turning weapons into tools of creation for the benefit of all mankind.
In the field of artificial intelligence, as a pre-trained large model for the development of key technologies, it also has the two attributes of sword and plough -
Opening up Xinjiang to make a sword: the big model has set off an arms race in the world, the scale of parameters has been continuously upgraded, and the emergence of large models such as Wenxin means that China has held this AI weapon in its hands and guarded the territory of intelligent technology;
Ploughing the land: With the strong general ability of the big model and open source openness, thousands of industries and developers also benefit from it, do not have to start from scratch, you can "open up" the introduction of AI at a low threshold, and start intelligent transformation and upgrading.
In recent years, in the face of the uncertainty of the global economic and political situation, people have never been so eager to "cast a sword into a ploughshare" of science and technology, and the big model should also be transformed from an arms race into a new momentum for industrial development.
How can we make the big model "cast a sword for a plough" and serve the people's livelihood and industries? I recalled a thing from the past.
In 2019, the big model is still the "secret" of the major technology giants, and a laboratory is ridiculed by the industry as CloseAI (closed AI) because it is unwilling to open the latest big model. Baidu released China's first officially open pre-trained model ERNIE 1.0 (the predecessor of the industry-level knowledge enhancement big model Wenxin) in the same year, but did not hide the device to wait, but directly opened the exploration of the application of the big model, focusing on "industrial intelligence".
It can be said that the Chinese AI forces represented by Baidu have chosen the development direction of "casting swords for ploughshares" and benefiting the industry when the big model trend first appeared.
So, when the time comes to the spring of 2022, can the big model be used by ordinary people and thousands of industries?
On April 16, at the just-concluded 2021 Baidu Cognitive AI Creative Competition "AI Creative School" final, I got some of the latest trends in the AI big model. To put it simply: 1. The application of the big model accelerates the landing, showing the infinite possibilities of entering the depths of the industry; 2. The threshold for the application of the Wenxin big model continues to decrease, and the public participates for the first time, releasing the infinite reverie of AI creativity.
In this warm and cold spring, I would like to share with you how to plant hope with ai models for industries and individuals.
Several spring breezes: THE AI big model has entered the depths of the industry
In your impression, who are using ai models? The world's top algorithm engineers who use GPT3 to train "crush humans" essays; those who use AlphaFold to predict proteins are the world's top AI laboratories; and there is an EYE-bursting AI for science model, which is prepared for scientific tasks with almost unlimited budgets... These big MODELS of AI cannot be used by ordinary people from all walks of life, nor can they afford to use them.
Is the big model doomed to lose touch with ordinary people? At the 2021 Baidu Cognitive AI Creative Competition "AI Creative School", Baidu Wenxin Big Model showed us another way to open.
As Wu Tian, vice president of Baidu Group, said at the competition site: "Only the threshold of artificial intelligence and large models is so low that everyone can easily use them, can we really explode various ideas on a large scale." In this creative competition, the contestants showed their rich creativity based on the Wenxin big model, which allowed us to see the prospect of the wide application of the big model. ”
How did this industrialization process of "casting swords into ploughshares" happen?
Step 1: Forging of technology.
Different from the computing power monsters and scale monsters in the laboratory, there are still many technical problems to be overcome in the large model of industrial landing applications to solve problems such as cost, efficiency, and deployment. To this end, Baidu has done a lot of technological innovation work.
On the one hand, we should continuously improve the scale of parameters and the effect of the model. In December 2021, ERNIE 3.0 was upgraded to the world's first knowledge-enhancing 100 billion model, becoming the world's largest Chinese monomer model so far, which also means that the general ability is better, the generalization ability is stronger, and only a small amount of industry data needs to be learned, which can be applied to many business scenarios and reduce the application threshold of enterprises.
On the other hand, in order to ensure the learning effect of the ultra-large-scale model, Baidu's industry-level deep learning platform Feipao has independently developed an end-to-end adaptive distributed training framework and 4D hybrid parallel technology, which effectively supports the efficient and stable training of the Wenxin large model and allows industrial users to use it at low cost.
Under the tempering of technology, we can forge a large model that is available and easy to use in the industry, and truly support industrial applications. This time, Wu Tian also mentioned that one of the future development trends of the big model will be the continuous innovation of technology.
Step 2: The melting pot of the industry.
At present, the common obstacle encountered by the industry in the application of AI is that the industry know-how is incompatible with the AI model, and it is impossible to solve real and complex business problems. Traditional deep learning is "brute force computing" that cannot understand the deep meaning of data.
As an industry-level big model, one of the core differences between Wenxin and other big models is to add a key catalyst for the integration of industry and AI - "knowledge enhancement", which combines large-scale knowledge with deep learning, so that AI can understand the knowledge relationship behind data in a specific field, thereby improving understanding ability.
For example, medical record quality control is very important for the medical industry, but there are very few doctors who have mastered general medical knowledge, and doctors in the hospital's medical records department have a large number of medical records and difficult work when conducting sampling work, and it is difficult to make a very accurate analysis and judgment of all medical records by relying on limited knowledge. If the AI algorithm cannot understand medical knowledge, it will not be able to help the medical case department work to improve quality and efficiency. Baidu added a series of knowledge such as pharmacopoeia and medical encyclopedia on the basis of the Wenxin model, and then trained a medical-related model based on the large model, reaching the level of medical doctor, and can perform second-level intelligent scanning analysis of each case, thus effectively improving the efficiency of case quality control.
The third step: large-scale production of molds.
From agricultural production to industrial manufacturing, large-scale production is indispensable to the emergence of standardized molds. The large-scale application of AI is also inseparable from a series of supporting tools and platforms.
At present, wenxin big model supports hundreds of enterprises and institutions, the number of developers exceeds 60,000, and the number of external daily adjustments has exceeded 50 million times. This is because, with the help of platforms and tools such as Baidu Flying Propeller and Baidu Brain, the ability of Wenxin Big Model can be applied by developers from all walks of life with a low threshold.
Taking EasyDL, the zero-threshold AI development platform of Feipao Enterprise Edition, as an example, it has launched a high-precision NLP model based on the Wenxin large model, and the model training in the vertical field can be completed in as little as 15 minutes. Even financial, legal and other practitioners who do not have an AI foundation can use a small amount of data to build customized models such as automatic analysis of contract terms and automatic analysis of financial research reports through EasyDL one-stop automated training.
The support of these supporting tools and platforms allows the Wenxin big model to enter the stage of large-scale application. Therefore, Wu Tian also believes that 2022 will be a key year for the landing of the Wenxin big model industry.
From the divine weapon in the hands of scientists and engineers, to the rope and ink furnace hammer of ordinary people and thousands of industries, the Wenxin big model "casting swords for ploughs" has jointly bred a spring of wisdom with a hundred flowers.
Whispering: Unleashing the Endless Imagination of the Big Model of AI
After Columbus discovered america, chili peppers entered the human table; the Wright brothers' yearning for the sky opened the prelude to flight; Jobs's genius-like creativity, the smartphone began to dominate your and my life... Every wave of technology is indispensable to the divine stroke of a genius, igniting the infinite imagination of new technologies with creativity.
In the era of AI, everyone does not want China to follow behind others, but where does creativity come from?
Baidu gave an answer: "Creativity is in the folk." Wu Tian, vice president of Baidu Group, said at the scene of the AI creative competition.
The private sector means that the AI model should be used by the widest range of talents, regardless of the level of technology.
In this competition, Baidu lowered the threshold for the use of AI large models to the level of the public for the first time, attracting nearly 2,000 players of different ages and different technical bases across the country. Lin Jie, the developer of the "Lezhi Q&A" project, is the product owner of an agricultural company and has not been developed. This time based on the Flying Propeller EasyDL platform using ERNIE 3.0, Lin Jie's development process is quite smooth, step by step training can solve plant problems DEMO. Recently, more and more friends plan to grow their own vegetables, and maybe in the future, Lin Jie's products can really save plant killers.
Folk means that AI innovation will be close to real life, and large models should be able to solve a large number of detailed and specific problems.
Among the more than 300 creative solutions in this creative competition, many of them are unexpected by large factories, but ordinary people are particularly looking forward to them. For example, Huan Yi, a second-year student of Harbin University, spent six days to create a "circle of friends ancient style copywriting small expert", as long as there are pictures, minutes let you become a copywriter, install a literati inker in the circle of friends, and attract countless likes.
Folk, meaning unknown scenes and challenges, the big model will encounter a variety of unexpected uses and scenarios, but also in practice can be tested and improved. One unexpected idea is the "anti-fraud vaccine robot".
I have played the anti-fraud App, the vaccine I have played, and the robot I have played, but what is this "anti-fraud vaccine robot" is, and it is really confused. It turned out that developer Guo Cheng (pseudonym) was impressed by his grandmother's various anti-fraud experiences, and had the idea of using NLP technology to help vulnerable groups identify fraud routines. After registering for the competition, Guo Cheng used the Baidu ERNIE 3.0 model to develop a dialogue robot, and users can learn anti-fraud experience from practice and improve anti-deception ability by talking to it, which is equivalent to hitting a "thought vaccine". To understand the cunning and devious language routines of the crooks, it is natural to require the robot to have extremely high semantic understanding, intention recognition, etc., which is one of the advantages of the ERNIE 3.0 model.
For the general onlookers, the idea of a game may not all be turned into reality. But every participant must be able to feel the real strength of the big model from the project practice. The infinite possibilities of the landing application of ai large models are hidden in these concerned eyes and flying brain holes.
A plough of new rain breaks spring ploughing: Wen Xin wrote wind and material poems for the era of intelligence
If it is said that the Wenxin big model is deeply rooted in the industry, it is from the beginning of its birth that "moisturizes the fine and silent", then the just-concluded AI creative competition is like a new rain to break the spring ploughing, so that we can see very clearly how the big model plugs the wings of the large-scale application of AI.
The reason is also very simple, the big model to serve the industry, must be able to solve those specific, detailed, engineering problems. AI scientists discuss professional issues such as model convergence, robustness, and cognitive intelligence, and when big models enter thousands of industries, the problems encountered are just like the contestants asked: how to find the right data set? How to train if you can't buy a GPU? How to do if you don't adjust the argument? What if the results do not meet expectations?
If the big model is to "forge a sword into a plough", it must guide and help industry people solve these practical problems. Obviously, through this AI creative competition, Baidu has successfully explored and answered the above questions. Of course, this is not only the credit of the Wenxin big model, but Baidu ensures that the big model capability is integrated into the intelligent upgrading of the industry through a complete set of systems.
The core is the fusion of flying paddles and large models. The aforementioned Flying Propeller has developed 4D hybrid parallel strategies, adaptive distributed training, etc., effectively supporting the efficient and stable training of Wenxin big model, and development platforms and tools such as Flying Propeller EasyDL, which have long been open sourced to all sectors of society. The strong combination and integration of flying propellers and large models have become an AI base with leading technology, independent innovation, solid capabilities and industry availability, which strongly supports the intelligent transformation of China's industry.
Next, there's the convergence of the cloud and the big model. Industrial intelligence calls for a more agile, simple and low-cost way to obtain AI, and cloud services have become the only way for all walks of life to introduce AI models, basic algorithms, AND computing power. In this context, Baidu has long promoted the deep integration of AI and cloud, and Baidu Intelligent Cloud, as an industrial export of Baidu AI technology, is pushing the Wenxin big model to more and more industrial scenarios.
Of course, the development of the big model is also inseparable from the integration with the talent ecology. When it comes to AI talent education, it always seems to be accompanied by seriousness and anxiety, but it is not. The highest state of education that Chinese yearns for is "twilight spring, spring clothes are complete, five or six crowns, six or seven children, bathing in Yi, wind like dancing, and singing and returning." The AI Creative Competition just made me feel this way, and under the arm support of the Wenxin big model, countless ideas flew freely, lightly and stretched. When ordinary people can also sing and walk and integrate into the torrent of AI creativity, will we still worry about not hearing the voice of AI innovation?
After three years of hard work, Baidu has established a new paradigm for the application of AI big models from research to application, and promoted the big model to a new stage of "casting swords for ploughs" and serving the industry.
As more people and industries join the chorus of the big model, a poetry of style and objects in the intelligent era will flutter in every corner of the Chinese land.