AI large model: "opening and discharging" for industrial intelligent upgrade
As one of the core driving forces of the new generation of industrial transformation, artificial intelligence has gradually moved from "refining the model" to the stage of "refining the big model".
By designing advanced algorithms, integrating as much data as possible, gathering a large amount of computing power, and intensively training large models to serve more enterprises, it is becoming a new trend in the development of artificial intelligence.
The robot "Xiao Ke" appeared in the National "Thirteenth Five-Year Plan" Science and Technology Innovation Achievement Exhibition, and the "Digital Man of Winter Olympic Sign Language Broadcasting" took up posts on Beijing TV.
Although the public is still ignorant of the concept of the large model and the technological breakthroughs behind it, they are no longer unfamiliar with these applications driven by the large-scale intelligent model of Enlightenment 2.0.
In 2021, Beijing Zhiyuan Artificial Intelligence Research Institute (hereinafter referred to as Zhiyuan Research Institute) released a large model of Enlightenment, setting the record of "China's first" and "world's largest".
With this as a sign, more and more research institutions and enterprises have joined the team of "refining big models" and promoting intelligent inclusiveness, contributing Chinese wisdom and strength in the development and application of artificial intelligence technology in the world.
"Opening Discharge" Inclusive Society
On June 1, 2021, Enlightenment 2.0, jointly created by Zhiyuan Research Institute, Tsinghua University and other units, was released.
Its parameter scale reached 1.75 trillion, breaking the previous record of 1.6 trillion parameters created by foreign pre-training models, becoming the first trillion-level pre-training model in China and the largest in the world.
Tang Jie, academic vice president of Zhiyuan Research Institute and professor of Tsinghua University, introduced that Enlightenment 2.0, which is based entirely on the domestic supercomputing GPU platform, has achieved a number of world-class innovative breakthroughs in pre-training model architecture, fine-tuning algorithms, and efficient pre-training framework. It has achieved original theoretical innovations, and achieved a leading position in 9 capabilities in the world-recognized artificial intelligence capability ranking list.
In order to improve the industrial applicability and ease of use of large-scale pre-training models, the efficient pre-training framework built by the Enlightenment team has achieved original breakthroughs or iterative optimization of the entire link, and the pre-training efficiency has been greatly improved.
"Diversified demands and fragmentation of scenarios are the common difficulties in AI implementation. While the Wudao open platform is fast and easy to use, it pays more attention to solving the problems of large-scale and industrialized AI applications." Tang Jie said that the Wudao large model has "low threshold" + high efficiency + high emotional intelligence", can meet the application needs of different industries and enterprises to achieve large-scale and industrialization.
After any enterprise or developer gets the fully open source Enlightenment 2.0 pre-training framework, it can be quickly deployed and applied to actual business.
OPPO's open dialogue virtual voice assistant Xiaobu, based on the "generative answering system" based on the Wudao model, solved the long-tail problem of commonality in the industry in one fell swoop, reducing the construction cost of a single answer by 99%.
Efficient machine translation, intelligent conversational customer service and voice broadcast have shown great potential for development in e-commerce, media, education, intelligent hardware and other fields, verifying that the path to general artificial intelligence laid out by large models has great potential .
Zhang Hongjiang, chairman of the Zhiyuan Research Institute, pointed out: "In the future, the large model will form an intelligent basic platform similar to the power grid, and like a power plant, it will continuously supply the 'intelligence source' to the whole society, and benefit all walks of life efficiently."
The Wudao 2.0 super-large-scale intelligent model training technology system, the infrastructure built for my country's artificial intelligence applications, has begun to "open the gate and discharge" for the intelligent transformation and upgrade of traditional industries.
Create a new R&D mechanism
The Enlightenment Large Model has achieved my country's independent controllability and cutting-edge leadership in ultra-large-scale intelligent model technology.
Huang Tiejun, Dean of Zhiyuan Research Institute, explained the R&D mechanism from three perspectives.
On the one hand, it is the scientific research organization model of "concentrating efforts on major tasks" in the new era, that is, maintaining a keen eye on major scientific issues, arranging major scientific research tasks in a demand-oriented and problem-oriented manner, establishing a rapid demonstration and initiation mechanism for major tasks, and establishing cross-institutional, Large collaborative, high-intensity scientific research team to solve big problems.
On the other hand, focus on talents, encourage free exploration, adhere to a realistic and pragmatic talent development model regardless of seniority, insist on selecting talents based on “representative works” and “small peer evaluations”, and let young talents who want to do things and can do things “take the lead” be the protagonist".
"Zhiyuan Research Institute aspires to be a 'forever young research institute', pays attention to attracting young scholars, and looks forward to working with young scholars to create a new paradigm of scientific research." Zhang Hongjiang said, "Zhiyuan respects 'representative work culture', regardless of background. , don't look at the number of papers, just look at whether you have achieved benchmarking achievements, and whether you have the potential to become benchmarks."
Tang Jie said that in the future, Zhiyuan Research Institute will continue to promote mechanism innovation, and it must be "on the sky" and "on the ground".
While attracting more scholars to join in, creating more "representative works" of scientific research similar to the Enlightenment Model, and promoting Beijing to become the world's leading artificial intelligence innovation center, it will create an ecology, connect technology and industry, and promote the development and depth of the artificial intelligence industry. application.
Industry-university-research institutes have entered the venue one after another
Standing at the starting point of the "New Three-Year Plan", Wudao will focus on improving intelligence, lowering the threshold, and building an ecosystem to further advance towards "better usability".
As more and more research institutions and technology companies enter the market, technological innovation and industrial achievements based on large model applications, typically represented by virtual digital humans, are fully blooming in my country.
On July 9, 2021, the Institute of Automation, Chinese Academy of Sciences released the cross-modal general artificial intelligence platform "Zidong Taichu". Based on this full-stack localization platform, the virtual human "Xiaochu" is built with a multi-modal large model as the core. , pictures, text, and voice can be understood, and the correlation and synergy between the three modes of pictures, text and voice are truly presented, which once again shortens the distance between artificial intelligence and human imagination.
On September 28, 2021, Inspur Artificial Intelligence Research Institute released a massive model of artificial intelligence-Yuan 1.0. When it was released, it had almost finished reading the vast content of the Chinese Internet for nearly 5 years.
Liu Jun, vice president of Inspur Information, said that one of the core features of giant quantization is that there are many model parameters and a large amount of training data.
Source 1.0 has 245.7 billion parameters, and the training dataset size reaches 5,000 GB.
As Wang Endong, an academician of the Chinese Academy of Engineering, said, making machines have cognitive abilities such as logic, consciousness, and reasoning like humans have always been an important direction of computer science exploration and research.
After better solving the problem of "perceptual intelligence", the development of this round of artificial intelligence has entered the development stage of solving more complex "cognitive intelligence" problems through various innovations.
Huang Tiejun said: "Artificial intelligence is the core driving force of the new generation of industrial transformation, and its development has gradually moved from a 'big refining model' to a 'big refining model'. By designing advanced algorithms, integrating as much data as possible, gathering A large amount of computing power, intensive training of large models for a large number of enterprises to use, is an inevitable trend."