Today, Artificial General Intelligence (AGI) has become a focal keyword in both the scientific and industrial communities. Just a few years ago, many believed that achieving AGI would take at least 10 to 50 years, or even thought it was impossible. Nowadays, such pessimistic views are rare. However, compared to the public's excitement about this wave of technological change, many frontline scholars and industry leaders in the AI field believe that there is still a long way to go for current AI to develop into AGI.
According to Qi Yuan, a distinguished professor at Fudan University, director of the Shanghai Artificial Intelligence Research Institute (SAIRI), and founder of the Trustworthy Large Model company "Infinite Lightyear," "One of the highest manifestations of AGI is the discovery of unknown laws in the complex world. Simply put, it should be an 'AI Einstein.' This requires us to create 'gray box' trustworthy large models that combine 'black box' probabilistic predictions with 'white box' logical reasoning; and to promote fundamental research, talent cultivation, and practical applications through the deep integration of technology and industry, thereby building an innovative ecosystem for scientific intelligence."
At the recent 2024 World Artificial Intelligence Conference (WAIC) and the Global Governance of Artificial Intelligence High-Level Meeting, SAIRI successfully hosted a thematic forum titled "Artificial Intelligence: Paradigm Shift in Scientific Research and Industrial Development." This was the first appearance of this new research institution at WAIC. SAIRI can be seen as a model for Shanghai's exploration of an innovation-driven "1+1+N" scientific intelligence ecosystem. This model involves SAIRI as the central hub responsible for overall strategic planning, resource integration, and key technology research and innovation, in collaboration with another "1" Fudan University, and several "N" universities, research institutions, tech companies, innovation teams, and investment institutions, to jointly promote scientific research, talent cultivation, technology transfer, and industrial innovation and upgrading.
The standard for AGI should be to create an "AI Einstein."
From a technical perspective, will increasingly larger models with more parameters lead to AGI? To date, neither from the perspective of AI technology itself nor from the energy consumption perspective, are large models based on the Transformer autoregressive architecture sufficient to lead to AGI. AI needs to develop new "gray box" trustworthy large models. This conclusion is based on Qi Yuan's years of practical experience in both academia and industry.
Ten years ago, with the idea of "making AI useful," Qi Yuan led a team to increase the core machine learning system of Alibaba from 2 million parameters to several hundred million parameters for the first time, achieving a significant improvement in business performance and demonstrating the integrated transformation of data, algorithms, and engineering capabilities. This is precisely the manifestation of the Scaling Law, which is widely discussed in the AI community today.
Qi Yuan recalls that the team did indeed taste the sweetness of the Scaling Law: after increasing the model parameters by a hundredfold, the overall effect improved dramatically. "But now I think: why didn't we make AI models even larger back then? Why did we stop when we could have taken a step further?" he said. "Even billions of parameters in large models are not enough; we need to move towards hundreds of billions, trillions, or even more. At that time, both academia and industry lacked the computing power, and even in the industrial sector, achieving such high computing power required very high costs, not to mention academia."
The reason why the standard for AGI should be to create an "AI Einstein," Qi Yuan explains, is that it needs to be both effective and intelligent. Firstly, Einstein discovered the "clouds of early 20th-century physics" through a few key data points. AGI should also be able to discover and understand the unknown laws of the complex world. However, current large models cannot achieve this. For instance, although the visual large model SORA simulates the physical world to an unprecedented degree, it still constructs the three-dimensional world based on the simulation of the two-dimensional world and is far from thoroughly understanding the physical world. Secondly, there is the issue of power consumption. The human brain operates at about 15 watts, while a single GPU can peak at several hundred watts, not to mention the clusters of thousands or tens of thousands of GPUs needed to train general large models. At present, if we continue to use existing architectures, the power consumption required would be astronomical, making it difficult to achieve the goal of being effective and intelligent.
The "AI Einstein" is also a key goal of AI for Science (AI4S). Scientific intelligence has played an important role in accelerating the solution of known physical equations, but it also needs to combine known rules with data to reduce the severe dependence on data and computing power, improve the accuracy of reasoning and prediction, and propose new scientific theories based on data-adjusted knowledge rules. This aligns with Qi Yuan's long-term goal at Fudan University and SAIRI-to use artificial intelligence to understand the complex world and discover unknown laws.
"Gray box" trustworthy vertical domain large models empower various industries.
What problems need to be solved for large models to become new productive forces from AI tools? According to Qi Yuan, the large model industry faces many common challenges, making it difficult for technology, products, and market needs to align.
"The biggest issue with large model implementation today is that it seems useful at first glance but fails in practical use," Qi Yuan explains. Today's large language models primarily predict the next word based on multiple preceding words, but this approach is not suitable for rigorous multi-step reasoning. "Language is a tool for communication, not for thinking." Recently, a paper published by institutions including MIT in the top academic journal Nature pointed out that language is a powerful tool for transmitting cultural knowledge, and it may have co-evolved with our thinking and reasoning abilities, reflecting the complexity of human cognition. However, language does not generate the complexity of reasoning.
To address the unreliability, low interpretability, and high costs of existing large models, an effective solution is to combine probabilistic neural network reasoning with logical symbolic computation, akin to the combination of fast thinking based on instinct and slow thinking based on logical reasoning described in Nobel laureate Daniel Kahneman's book Thinking, Fast and Slow. "This can be called a 'gray box' large model," Qi Yuan believes. Combining symbolic computation with neural networks in a "gray box" trustworthy large model can reduce AI's "hallucinations" and solve professional problems in vertical fields, thereby empowering various industries and unleashing the productivity of large models.
What is a "gray box" trustworthy large model? "Originally, deep learning was considered a 'black box.' Now, by combining logical reasoning with deep learning, we have a 'gray box,'" Qi Yuan explains. "The original 'black box' left people unaware of the process by which data produced results, whereas the 'gray box' large model, aided by logical reasoning, allows people to 'know both the results and the reasons behind them.' From another perspective, 'gray box' large models can use deep learning to reduce rules that do not conform to real-world observed data."
Qi Yuan states that for AI to play a core role in complex scenarios across various industries-whether in finance and insurance, wind power and energy, or ocean shipping and pharmaceutical fields-it is necessary to combine systematic industry knowledge, reasoning logic, and decision-making mechanisms with large models. The "gray box" large model is not only the direction for AGI but also a powerful tool for deeply penetrating vertical fields and genuinely solving real-world problems. "From an industrial perspective, this understanding is very intuitive," Qi Yuan illustrates. Doctors do not need to become lawyers, nor do lawyers need to become investment experts. Each professional role should focus on their field and enhance their productivity tools. From a technical standpoint, if a large model over-learns irrelevant tasks, it may experience "catastrophic forgetting." For example, if Li Bai were to spend all his time doing accounting instead of writing poetry, his poetic inspiration might gradually fade. "We have already observed that when training large models for vertical domains, if the model learns too many unrelated functions, it can interfere with its original capabilities. Therefore, developing effective 'gray box' large models for vertical domains is of great value in industrial implementation."
"I believe 'gray box' large models will play an increasingly important role on the path to AGI and in the implementation of vertical domain industries. From a Bayesian methodological perspective, it combines our known knowledge with hidden information in the data to discover new laws and solve scientific and industrial problems," Qi Yuan states. In the future, "AI Einstein" could also be "AI Buffett."
Connecting the innovation chain and building a scientific intelligence innovation ecosystem.
At this year's World Artificial Intelligence Conference, Qi Yuan's team launched trustworthy financial and medical large models with hundreds of billions of parameters. These vertical domain large models surpassed OpenAI's trillion-parameter model GPT-4 Turbo in testing, once again drawing industry attention to the implementation of large models.
"Today's AI breakthroughs are driven not only by innovations in underlying principles but also by product-driven approaches that address societal needs. Society requires not only the publication of theoretical papers or business model innovations but also the deep integration of technological and industrial innovations based on first principles. Once these two elements are combined, we can reach bluer waters," Qi Yuan says.
Academia and industry have different missions. Academia explores new phenomena, while industry primarily solves practical problems. A common issue worldwide is that research institutions need to address many technological innovation problems, but if they overlook productization and societal needs, they face two shortcomings: a lack of real competitive pressure, which hinders the refinement of innovative technologies, and the absence of effective market feedback to guide technological research.
To this end, Qi Yuan has long sought to connect the innovation chain of "universities-research institutes-startups" to create a good innovation ecosystem that considers both underlying technology and market needs. The product direction should be guided by market demand and scenarios, building product core competitiveness through foundational innovation.
Founded in 2023, SAIRI is committed to original AI for Science innovations that combine knowledge and data. Recently, SAIRI launched the Fuxi series of meteorological large models 2.0 for applications in new energy, insurance, urban management, and initiated the Smart Meteorological Innovation Ecosystem Alliance. This alliance aims to gradually promote the industrial application of the Fuxi series meteorological large models 2.0. The "gray box" trustworthy large models are also progressing in product implementation, with Infinite Lightyear, the trustworthy large model company founded by Qi Yuan, already established.
To further promote the scientific intelligence innovation ecosystem, the second World Scientific Intelligence Competition, jointly organized by SAIRI and Fudan University, and guided by multiple departments including the Shanghai Science and Technology Committee, Shanghai Development and Reform Commission, Shanghai Economic and Information Technology Committee, and Shanghai Education Committee, has been launched. The competition offers millions in prizes to recruit global participants to explore frontier fields of scientific intelligence. Additionally, SAIRI has developed a scientific data platform covering multimodal scientific data, which supports the full chain from data collection and processing to management and modeling, ensuring efficient data processing, trustworthiness, and secure communication. Based on this platform, SAIRI and its partners have built several high-quality scientific datasets for life sciences, material sciences, atmospheric sciences, and other fields, providing valuable resources for scientific intelligence research. Moreover, SAIRI has initiated the Global Scientific Data Ecosystem Alliance, with initial members including China Telecom Corporation, COSCO Shipping Insurance Captive, Shanghai Lingang New Area Cross-Border Data Technology, and more than ten other entities. The alliance aims to build a global, multi-domain research big data resource open and sharing platform through collaboration among government, enterprises, universities, and research institutions.
"Whether in scientific research or industry, we should not innovate for the sake of innovation. We hope to build future AGI and applications that solve real-world problems," Qi Yuan says.