“Zhongguancun” Magazine, July 2024 Issue
Since 1956, Artificial Intelligence (AI) has existed and developed for nearly 70 years. The journey has not been easy, marked by numerous ups and downs. Nevertheless, due to the efforts and perseverance of many, the achievements of AI have genuinely inspired the public and ignited greater societal expectations.
Can we say that we have entered the era of Artificial Intelligence? There can be no unanimous opinion on this; it entirely depends on each person’s perspective and attitude toward the issue. However, one thing should be clear: the fundamental purpose of developing AI is ultimately to benefit humanity, not to “harm” it. Although AI will still encounter many challenges and may present deviations and issues requiring improvement, the momentum driving its advancement is gathering more rapidly and widely than ever before. In this sense, the AI era is approaching.
It is entirely normal for society to express high concern for the development of AI, as it relates to the interests and future of all members of society, whether one likes it or not. In the process of promoting AI development, every individual and institution, including every enterprise, will have a role to play. Therefore, it is essential to engage in broad and rational discussions about future perspectives.
In late April 2024, the President’s Council of Advisors on Science and Technology (PCAST) convened a workshop on generative AI, with ten members including Nobel laureates, university presidents, and senior technical personnel and managers from internationally influential companies such as Microsoft, Google, and NVIDIA. Notable Asian scientists, semiconductor chip expert Lisa Su, and Fields Medalist Terence Tao were also among them. The workshop was supported by 21 heavyweight figures from universities, enterprises, media, and international organizations. A significant part of the workshop was to hear from the head of the Generative AI Working Group, mathematician Terence Tao, who reported on “Using AI to Address Global Challenges,” focusing on how to leverage AI to comprehensively enhance and elevate scientific research and development levels (Supercharging Research). The core of this report is to analyze the synergy between science and AI.
There have always been many obstacles hindering the progress of general scientific research, such as slow progress, high R&D costs, and a shortage of qualified experts. Now, AI is beginning to eliminate these barriers; the progress of foundational AI research has reached a stage where it can mitigate the errors, biases, and other potential harms previously caused by AI.
On one hand, in the collaborative development of science and AI, existing scientific values (e.g., views and attitudes regarding validity, reproducibility, openness, and expert oversight) play a key role in ensuring the formation of a culture that responsibly uses AI methods. Ideas, methods, and tools from various (relevant) scientific fields have also played an indispensable role in the development of AI.
On the other hand, in many scientific research fields, numerous cases and instances exist that heavily rely on AI methods and tools. AI is changing science. It can be said that AI will change every scientific discipline and every aspect of the scientific workflow. Especially in identifying candidate solutions to scientific problems, accelerating and refining the establishment and selection of scientific simulations and models, and analyzing new types of data, AI’s role will bring revolutionary changes.
For example, in semiconductor design, AI technology is now effectively used for the initial circuit design of chips. Future AI tools will perform more foundational chip design tasks, improving designers’ work efficiency by at least an order of magnitude, leading to better quality and more efficient problem-solving solutions. In cosmology, current AI methods allow for rapid simulation of cosmological hypotheses about the nature of the universe. With the support of AI tools, new fundamental theories of physics may emerge in the future. Regarding comprehensive health and well-being issues involving physical, psychological, and social aspects, current AI technology has already helped doctors successfully perform early cancer diagnoses and identify potential diagnostic errors, enhancing patient safety. Future AI systems will assist doctors in tailoring healthcare plans based on patients’ specific genetic information and medical history, achieving personalized medicine.
Undoubtedly, as we look forward to the grand prospects of conducting research and development assisted by AI, it is essential to set and pursue rational goals to achieve maximum benefits through the appropriate and effective application of AI in the scientific field. In this process, it is particularly important to grant more autonomy to human scientists; to maintain a more responsible attitude towards the use of AI tools; and to ensure as much sharing of various AI resources as possible. During the workshop, Terence Tao provided a systematic analysis on behalf of the Generative AI Working Group and proposed several targeted recommendations.
The development of AI assistant systems aims to supplement and enhance human scientists’ capabilities, not to replace them. AI tools can be used to handle large volumes of data, managing laboratory work, coding, and writing tasks increasingly, thus finding more promising solutions to scientific problems. This allows human scientists to focus on higher-level selected issues. New collaborative methods can be explored, such as human scientists guiding interrelated networks of AI assistants, enabling large-scale, interdisciplinary, and/or decentralized projects. For this, special support is needed for collaboration among academia, industry, national laboratories, and federal agencies to conduct foundational and applied AI research. A pertinent example is the materials innovation platform project supported by the National Science Foundation, which is being developed into a data-sharing infrastructure while integrating AI tool development as a means of establishing collaborative interactions with other institutions and industry partners. Future projects may also include collaborative development of next-generation quantum computing, whole-cell modeling, global foundational models, or high-quality scientific databases for various disciplines.
In practical exploration, it is crucial to encourage innovative methods to incorporate AI-assisted features into scientific workflows. Relevant funding agencies should recognize potential new workflows and design flexible processes, monitoring metrics, and fund certain models to encourage strategic experiments organizing scientific projects in new AI-assisted ways. Although AI-assisted science can outperform non-assisted human science or fully automated science, these tools are merely supplementary means to traditional human scientific research.
It is necessary to establish a clear awareness that responsible, transparent, and trustworthy principles for AI use should be upheld at all stages of the scientific research process. This means developing a culture of responsible AI use in science, where AI outputs should undergo external validation, protection of private data must be ensured, algorithmic biases need to be measured and compensated, and models and data should be as transparent, reproducible, open, and interpretable as possible. Funding agencies should also require researchers to formulate clear AI use plans with responsible characteristics to assess potential AI-related risks. All these need to be shaped through ongoing dialogue among natural sciences, social sciences, humanities, and policymakers.
Besides large models, with the development of more lightweight models and shared resources, the environmental impact and other costs of AI use may decrease. This further necessitates expanding existing efforts to share foundational AI resources widely and equitably. Only in this way can AI benefit society more broadly.
To some extent, developing AI is a double-edged sword. Widespread AI technology will fundamentally accelerate scientific research. With the right AI infrastructure, scientists will be able to tackle other more urgent challenges. However, AI also has weaknesses, potentially requiring substantial computation, energy, and data. Certainly, there are methods available to mitigate these weaknesses and reduce the required resources.
Artificial Intelligence can drive significant advancements in science and technology, providing unprecedented opportunities for scientists, engineers, entrepreneurs, and startups. This topic deserves deeper contemplation.