The AI industry has been busy lately.
OpenAI unveiled Chat GPT (GPT-4o) on May 13. Much like Samantha in the movie “Her,” GPT-4o could converse smoothly with people and appeared to show empathy.
A week before that, Google DeepMind published a paper in Nature and launched the brand-new AlphaFold3, an AI model that can predict the structure of proteins, DNA, RNA, ligands and how they interact with one another.
To explore how to make AI work for science, Westlake University hosted a roundtable on May 13, inviting 14 researchers from Microsoft Research and seven scientists from the university to share their thoughts.
How can we use AI to change the world?
AI is growing.
From playing chess to creating images on request, AI has proven its worth as a tool. In less than a decade, it is now capable of predicting microscopic molecular structures with the accuracy of lab work. Now that it's catching up with the smartest of us, how can we use it to change the world? How is AI evolving?
Prof. Tie-Yan Liu, a graduate of Tsinghua University, is a distinguished scientist at Microsoft Research AI for Science. He established the first AI+Science team at Microsoft Research Asia and worked on computational biology.
Below is an excerpt from Liu's speech:
Microsoft Research AI for Science was established in 2022. The world was a very different place six years ago.
Back in 2018, I set up a team for computational biology at Microsoft Research Asia. The recruiting staff were baffled by my decision to hire a biologist for a computing lab.
This was what I said: AI will not stay in that tiny box where we assigned it to play chess and recognize images. One day, it'll make contributions to natural science and help us solve matters that concern our future and the essence of the world.
We humans reached the top of the evolutionary chain because we have the ability to recognize the world and the courage to change it, and that has since become the mission for natural science research. AI is designed to imitate human intelligence, so it would be a shame if we didn't use it to understand and change the world. To make that happen, we need the strength of both AI experts and natural scientists.
We can use AI's ability to predict protein structures as an example.
AlphaFold took the world by surprise when the news broke that it could predict protein structures with a much higher efficiency than lab work. What should we take from this? Does it mean that we should throw in the towel on protein research now?
The researchers at Microsoft Research AI for Science beg to differ. We realized that learning the static structure of proteins was only half of the picture in learning its biological features. In fact, it's the dynamic changes of the protein structures and the balanced distribution that define its biological features. The static structure predicted by AI is still quite far from the core of biological challenges. In a way, this reveals the limit of AI: Only when working with biologists can AI be applied to things that matter.
AI: Specialist or polymath?
Now that we've established the goal of changing the world, what exactly does AI have to offer to science and mankind?
It's been six years since Prof. Tie-Yan Liu hired the first biologist at Microsoft Research Asia. When asked what he learned from the topic “AI for Science”, Liu said, “It's a specialist and a polymath.”
The following is an excerpt from Liu's speech:
I see two missions AI can carry out for science.
Modern-day natural science is based on long-term experimentation, which influences the pace of development in science. The beauty of computation is that it can simplify complex tasks, given sufficient investment, computability and power. The cycle of one year can easily be shortened to a month, 10 days, or even just one minute.
If we could combine AI and lab work, and give it the accuracy of the lab equipment, then it is fair to presume that we can shorten the experimental cycle by hundredths, if not thousandths.
This is the first mission for AI: to shorten the experimental cycle and to accelerate scientific development.
The other mission is that AI might be able to change the methodology of natural scientific research.
If we look back in history, we'd see that the giants who made significant contributions such as Da Vinci, Newton, and Schrödinger were polymaths who were knowledgeable in many fields. Researchers in natural sciences are trained differently these days: They tend to be specialists.
That brings up a question: Is it a good idea to conduct scientific research within a single discipline?
That's a tough question. Nowadays, scientists don't really have a choice. With the rapid development in science, it's impossible for any single researcher to read all the papers published around the world. It's hard enough to stay up to date with the latest developments in one's field, let alone be a polymath.
This is where AI can help.
Humans have limited capacity and speed when processing information, but AI doesn't. AI can be used to read, comprehend, and summarize the latest developments in science, as well as analyze mass data. It can be the Renaissance man who thinks beyond a single discipline, makes connections, and gains insights. Once AI tears down the barriers between disciplines, scientists can then make new discoveries and develop new methodologies.
Based on this hypothesis, Microsoft Research AI for Science has been working on a model to discover and integrate basic principles across disciplines. When it's in place, we can feed a new hypothesis to it and let it run a simulated calculation. Once that's done, we can verify it with the automated lab and feed the results back to the model for correction–and the lab could function around the clock. Imagine what that could do for our research capabilities!
AI: Science or a tool for science?
Last week, Liu published a paper titled “AI for Science: Imagine a Future where Everyone Can Participate”.
Ten years ago, it would be seen as sensational. But nowadays, we've become used to a life where AI lowers the barriers of language, skills, and even creativity.
If AI can help us comprehend and change the world and also assist scientists in their discoveries, what do scientists think of it? How helpful is AI to them? How far are we from a future where everyone can participate?
Prof. Yue Zhang, who received his Ph.D. degree in computer science from Oxford University, is a tenured full professor at Westlake University and the head of the Natural Language Processing Lab.
Below is an excerpt from Zhang's speech:
I believe that many of us still need an in-depth understanding of AI and other basic sciences (in the era of AI). AI is now capable of evolving and learning. When AI knows how to apply the laws of physics, fewer students would bother to learn them by heart. When AI knows how to code, fewer people would bother to learn to code. Would humankind still have the motivation to learn when AI does all the learning? Perhaps that would be the day that we are replaced by AI. But this is just one of the possible outcomes. From where I stand, we can use AI to do the hard work, but a human still needs to guide it. Indeed, AI can generate code now, but there are ordinary code, neat code, and code that is as exquisite as art. AI can quickly offer basic information from any given field – to amateurs, it could be overwhelming. But once you have in-depth knowledge, you'll see that AI is the amateur that captures the shape and not the essence. So, the question becomes: How do we make AI work for us while providing the guidance? At Westlake University, we use AI to empower the whole process of basic scientific research to help researchers reach their goal. We have a natural advantage here. Our AI scientists work seamlessly with researchers from the basic natural sciences, and they challenge each other to do better. For example, many of us are accustomed to recording experimental data in notebooks. The downside of this is that the results are scattered, easy to lose, and hard to trace if there's a mistake. I do research on natural language processing and collaborate with six labs from four schools. We designed an automated platform for recording experimental results, using a unified, universal, and scalable approach to automatically log all the results from the six labs. The data was then fed into AI and stored instantly. Prof. Yaochu Jin graduated from Zhejiang University, served as an Alexander von Humboldt Professor for Artificial Intelligence, and is currently president-elect of the IEEE Computational Intelligence Society. As a chair professor of artificial intelligence at Westlake University, he heads the Trustworthy and General Artificial Intelligence Laboratory.
Below is an excerpt from Jin's speech:
When we talk about the future of AI, we don't simply mean “bigger and faster”. AI needs inspiration from biological evolution to develop more organically. To make that happen, we need to keep our eyes on some of the most basic and important mechanisms in biology. That's why I think AI and life sciences, especially the interdisciplinary research of computational biology, would be a good match. There are two ways forward. The first is quite straightforward: AI can be used to process the mass data produced in modern experimental environments. But I think a deeper cooperation could involve extracting the essence of life sciences to make that the core of new AI models. That's something I've pondered for the past three decades: How do we approach AI from a natural intelligence point of view? Back then it was hard to imagine what AI would look like. But out of curiosity, I wondered, “Can we find inspiration from nature and biological evolution for our research on neural networks for computers?”
Preparing for the future of AI
A member of the audience asked Liu, “How would you advise young undergraduate students and Ph.D. candidates to prepare themselves academically for AI?
A China Central Television journalist once asked Zhang a similar question: “How will the younger generation use AI?”
What should humans do when AI becomes more and more intelligent?
The answer is: All roads lead to Rome.
Liu said, “Until just a year ago, we were still hiring specialists from different fields, trying to maintain a certain ratio between AI researchers and natural scientists. But more recently, we started to reflect: Are they the talent we need?
“The answer is no. ideally, we'd like the 'all-in-one' researchers: they are trained in AI, natural science, and engineering, and they have enough knowledge and skills to carry out research in a closed loop. Even when their hypothesis goes against mainstream beliefs, they can still deliver. It'd be great if there's a team of them, that way they can complement each other.
“This means we need young researchers who are knowledgeable in AI and natural science. An AI specialist with little knowledge in natural science might look down upon the importance of the actual research, and a natural scientist who knows little about AI might reduce the AI input to just a generic tool.
“To make AI work for science, we need natural scientists to understand AI, and in turn to like and to trust it. We also need AI specialists to study natural science, so they have the proper respect for it.
“With such a group, we will have an amplification effect where the scientific knowledge is truly understood, digested, and integrated into the AI process.
“To make that happen, we need support from higher education. At the moment, few universities strive to train talent like this. Students are offered either computer science or biology. They don't teach future scientists enough about computer science (AI for science), and those who major in computer science aren't learning enough about natural science (science for AI). Students graduating from such a system cannot be organically interdisciplinary.”
Zhang said, “My students and I remain optimistic on whether AI can become a scientist. We believe that AI can be used for science. As I mentioned previously, AI can help to increase research efficiency. From there, AI can also learn about the personality, experience, decision-making process of scientists, and thus potentially become one.
“From that perspective, I think humans should embrace this technology. In the coming five to 10 years, we will coexist with AI without a doubt. The sooner we adapt to it, the sooner we can make it a better tool for improving work efficiency. Those who first embrace the change would have the nimble control of the future of our society and be most adapted to the future technological developments.”
Microsoft once had such an equation:
AI+HI=SI
Artificial intelligence + human intelligence= super intelligence
AI plus natural science appears to be a step in that direction.
When some of the best minds choose to side with AI, where does that leave us? Should we be concerned, panicked, relieved, or even hopeful?
We heard from scientists about curiosity, courage, and respect – the strength that led humankind to where we are today.
Yet engineers and scientists are also hoping for more. They hope that AI will illuminate the dark, dry world of the unknown so that scientists can ask the right questions. They hope to work with AI and challenge one another to move forward.