news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Nowadays, general artificial intelligence (AGI) has become one of the keywords that the scientific and industrial communities focus on. Just a few years ago, many people believed that it would take at least 10 to 50 years to achieve AGI, and some even believed that it would never be achieved. Today, the latter are very few. However, compared with the public's excitement about this wave of technological changes, many front-line scholars and industry leaders in the field of AI (artificial intelligence) feel that there is still a long way to go for the current AI to develop into AGI.

Qi Yuan, Distinguished Professor Haoqing of Fudan University, Dean of Shanghai Institute of Scientific and Intelligent Technology (hereinafter referred to as "SIIT"), and founder of the trustworthy big model company "Infinite Light Years", believes that "one of the highest manifestations of general artificial intelligence is to discover the unknown laws of the complex world. In short, it should be an 'AI Einstein'. This requires us to create a 'black box' probability prediction that combines fast thinking and 'white box' logical reasoning that combines slow thinking, and to create a 'gray box' trustworthy big model. In addition, through the deep integration of science and technology and industry, we will promote basic research, talent training, and the implementation of results, and create a scientific and intelligent innovation ecosystem."

At the 2024 World Artificial Intelligence Conference (WAIC) and the High-level Meeting on Global Governance of Artificial Intelligence, which concluded not long ago, Shanghai Institute of Artificial Intelligence successfully held the theme forum "Artificial Intelligence: Paradigm Change in Scientific Research and Industrial Development", which was also the first appearance of this new research and development institution at WAIC. Shanghai Institute of Artificial Intelligence can be said to be a model of Shanghai in exploring the innovation-driven "1+1+N" scientific and intelligent ecological model, that is, Shanghai Institute of Artificial Intelligence, the "1", is the center and hub of the ecosystem, responsible for overall strategic planning, resource integration, and research and development and innovation of key technologies, and jointly with another "1" Fudan University, and many universities, research institutions, technology companies, innovation teams and investment institutions, these "N", jointly promote scientific research, talent training, transformation of scientific and technological achievements, and innovation and upgrading of industries.

The standard of AGI should be to create "AI Einstein"

From a technical perspective, will large models with more and more parameters lead to general artificial intelligence? To date, large models based on the Transformer autoregressive architecture still have dividends, but they are not enough to lead to general artificial intelligence, both from the perspective of AI technology itself and from the perspective of energy consumption. Artificial intelligence needs to develop new "grey box" trustworthy large models. This is the summary of Qi Yuan's many years of practical experience in academia and industry.

Ten years ago, with the idea of ​​"making artificial intelligence useful", Qi Yuan led the team to increase the number of parameters of Alibaba's core machine learning system from 2 million to hundreds of millions for the first time, achieving large-scale business results and demonstrating the integrated transformation of data, algorithms and engineering capabilities. This is exactly the embodiment of the Scaling law that is talked about in the artificial intelligence circle today.

Qi Yuan recalled that the team did taste the sweetness of Scaling Law at that time: after the model parameters increased by a hundred times, the overall effect was improved by leaps and bounds. "But now I wonder: Why didn't we achieve a greater level when we were making AI models? Why did I stop here when I was about to take a step forward?" He said, "Hundreds of millions of parameters for large models are still not enough, we need to move towards tens of billions, hundreds of billions, or even trillions. At that time, both academia and institutions lacked computing power, and even in the industry, it would cost a lot to achieve such high computing power, not to mention academia."

Qi Yuan explained that the reason why the standard of AGI should be to create "AI Einstein" is that it should be effective and smart. First of all, Einstein saw the "dark clouds of physics in the early 20th century" through several key data points, and AGI should also be able to discover and understand the unknown laws of the complex world. But none of the current large models can do it. Although the visual large model SORA has achieved an unprecedented level of simulation of the physical world, SORA is still based on the simulation of the two-dimensional world to build a three-dimensional world, and there is still a long way to go to thoroughly understand the physical world. The second is power consumption. The power of the human brain is about 15 watts, and the peak power of a GPU is hundreds of watts, not to mention that it takes a cluster of thousands or tens of thousands of cards to support the training of general large models. For now, if the existing architecture continues to be used, it will require a huge amount of power consumption, and it will be difficult to achieve the goal of being effective and smart.

"AI Einstein" is also a key goal of AI for Science (Note: scientific intelligence, hereinafter referred to as AI4S). Scientific intelligence has played an important role in accelerating the solution of known physical equations, but scientific intelligence needs to combine known rules with data, reduce the heavy dependence on data and computing power, improve the accuracy of reasoning and prediction, and adjust knowledge rules according to data to propose new scientific theories. This also coincides with Qi Yuan's long-term goal of working at Fudan University and Sophia Institute - "using artificial intelligence to understand the complex world and discover unknown laws."

"Gray Box" Trusted Vertical Field Big Model Empowers Thousands of Industries

What problems need to be solved urgently for big models to achieve new quality productivity from AI tools? In Qiyuan's view, the big model industry faces many common challenges, making it difficult to match technology, products and market demand.

"I think the biggest problem with the current big model is that it seems useful at first glance, but it doesn't work when you really use it." Qi Yuan explained that the core of today's big language model is simply to predict the next word based on multiple past words, but this is not suitable for rigorous multi-step reasoning. "Language is a tool for communication, not a tool for thinking." Recently, an article published by the Massachusetts Institute of Technology (MIT) and other institutions in the top academic journal Nature pointed out that language is a powerful tool for spreading cultural knowledge. It may have co-evolved with our thinking and reasoning abilities and can reflect the complexity of human cognition. But language does not produce the complexity of reasoning.

In view of the existing unreliability, low interpretability and high cost of big models, a truly effective solution is to combine probabilistic neural network reasoning with logical symbolic computing, similar to the combination of human fast thinking based on instinct and slow thinking based on logical reasoning described in the book "Thinking, Fast and Slow" by Nobel Prize winner in Economics Daniel Kahneman. "It can be said to be a 'gray box' big model." Qi Yuan believes that the "gray box" credible big model that combines symbolic computing with neural networks can reduce the "illusion" of artificial intelligence and solve professional problems in vertical fields, thereby empowering thousands of industries and releasing the productivity of big models.

What is a "grey box" trustworthy big model? "Deep learning used to be considered a 'black box', but now we have combined the 'white box' of logical reasoning with deep learning, and it has become a 'grey box'." Qi Yuan explained, "The original 'black box' made people completely ignorant of the process of how data produced results, while the 'grey box' big model, with the help of logical reasoning, allows people to 'know what it is, and why it is'. At the same time, from another perspective, the 'grey box' big model can use deep learning to reduce rules that do not conform to real-world observation data."

Qi Yuan said that in order to make the complex scenarios of various industries become a new battlefield for AI to play a core role, whether in the scenarios of financial insurance, wind power energy, ocean shipping, or medical and pharmaceutical industries, it is necessary to combine the system's industry knowledge, reasoning logic, and decision-making mechanism with the big model. The "grey box" big model is not only the general direction of general artificial intelligence, but also a tool for the big model to deeply penetrate in vertical fields and truly solve practical problems. "If you look at it from the perspective of the industry, this understanding is very intuitive." Qi Yuan gave an example, doctors do not need to become lawyers, and lawyers do not need to become investment experts. Each professional role should focus on their own field and do their own productivity tools. From a technical perspective, if a big model is allowed to over-learn irrelevant tasks, "catastrophic forgetting" may occur. Just like if Li Bai did accounting all day instead of writing poetry, his poetic inspiration might gradually fade. "We have observed that when training big models in vertical fields, if the model is allowed to learn too many irrelevant functions, it will interfere with its original capabilities. Therefore, I think it is of great value to do a good job of the "grey box" big model in the vertical field in the implementation of the industry."

"I believe that the 'grey box' model will play an increasingly important role on the road to AGI and the implementation of vertical industry, because from the Bayesian methodology, it is to combine our known knowledge with the unknown information hidden in the data to discover new laws and solve scientific and industrial problems." Qi Yuan said frankly that in the future, "AI Einstein" can also be "AI Buffett."

Open up the innovation chain and build a scientific and intelligent innovation ecosystem

At the 2024 World Artificial Intelligence Conference, the team led by Qi Yuan released the Trusted Light Language Finance and Medical Big Models with tens of billions of parameters. The tests of these two vertical field big models both surpassed OPEN AI's trillion-parameter big model GPT4-Turbo, which once again attracted the industry's attention to the implementation of big models.

"Today's breakthroughs in artificial intelligence come not only from innovations in underlying principles, but are also driven by products that meet social needs. What society needs is not only the publication of theoretical articles or innovation in business models, but the deep integration of scientific and technological innovation and industrial innovation, and breakthroughs based on first principles. Once these two are combined, we can swim to where the sea is bluer," said Qi Yuan.

Academia and industry have different missions. Academia needs to explore new things, while industry needs to solve practical problems first. A common problem that exists both at home and abroad is that research institutions need to study many technological innovation issues, but if they ignore productization and social needs, they will have two shortcomings: lack of real competitive pressure, unable to forge innovative technologies in competition; and lack of effective market information feedback to guide the direction of technological research and development.

To this end, Qi Yuan has been looking forward to opening up the innovation chain of "university-research institute-start-up" and creating a good innovation ecosystem, not only thinking about the underlying technology, but also grasping the market demand. Let the market demand and scenarios guide the product direction, and build the core competitiveness of the product from the bottom-up innovation.

Founded in 2023, Sophia Institute is committed to original innovation in AI for Science that combines knowledge and data. Recently, Sophia Institute released the Fuxi series meteorological big model 2.0 for industrial applications such as new energy, insurance, and urban management, and took the lead in launching the Intelligent Meteorological Innovation Ecological Alliance, and joined forces with many units to gradually promote the industrial application of the Fuxi series meteorological big model 2.0. The product launch of the "Gray Box" trusted big model is also in progress, and the trusted big model company founded by Qi Yuan, Infinite Light Years, has been established.

In order to further prosper the scientific intelligence innovation ecosystem, the second World Science Intelligence Competition, co-sponsored by Shanghai Institute of International Studies and Fudan University and jointly guided by Shanghai Municipal Science and Technology Commission, Shanghai Municipal Development and Reform Commission, Shanghai Municipal Economic and Information Commission, Shanghai Municipal Education Commission and other departments, has been launched. The competition sets a prize of one million yuan to recruit contestants from all over the world to jointly explore the frontiers of scientific intelligence. At the same time, Shanghai Institute of International Studies has led the development of a scientific data platform covering multimodal scientific data. The platform has full-link capabilities from data collection, processing to management and modeling to ensure efficient processing, trustworthiness and secure interoperability of data. Based on this platform, Shanghai Institute of International Studies and its partners have built a number of high-quality scientific data sets for life sciences, material sciences and atmospheric sciences, providing valuable resources for scientific intelligence research. In addition, Shanghai Institute of International Studies took the lead in launching the Global Science Data Ecological Alliance. The first batch of alliance members include more than ten units such as China Telecom Co., Ltd., COSCO Shipping Property and Casualty Insurance Captive Co., Ltd., and Shanghai Lingang New Area Cross-border Data Technology Co., Ltd. The alliance will build a global, multi-field scientific research big data resource opening and sharing platform through cooperation with governments, enterprises, universities, research institutions and other parties.

"Whether in scientific research or industry, we should not pursue innovation for the sake of innovation. We hope that we can build general artificial intelligence and applications in the future to solve real-world problems," said Qi Yuan.

author:

Image: courtesy of Shanghai Institute of Intelligent Technology Editor: Wu Jinjiao Responsible Editor: Jiang Peng

Please indicate the source when reprinting this article.