Universities today face difficult, pressing questions about how to handle new opportunities and growing challenges as artificial intelligence capabilities continue to expand. This is currently receiving a lot of attention, as seen in a study on AI and education that was released this summer.
Sasha Rakhlin, the Distinguished Professor in Data, Systems, and Society, IDSS, and Brain and Cognitive Sciences at the Statistics and Data Science Center, has written a new essay summarizing recent readings and discussions about how AI is transforming academia, specifically mathematics, statistics, machine learning, and engineering. He discusses key issues departments and institutions should consider and how they may work with AI in the future, with a primary focus on graduate research and education.
What is evolving in research, and why is it happening so swiftly?
A: The speed at which AI capabilities are developing is demonstrated by mathematics. A model won a gold medal at the International Mathematical Olympiad last year. After just a year, models are yielding fresh research findings, such as a solution to one of the Millennium Prize Problems. AI is becoming more capable of performing tasks that formerly required strong mathematical skills.
The speed and accuracy of verification play a key role in deciding how quickly AI advances in a certain field. Programs can be executed and tested, and formalized proofs can be automatically verified. Systems may produce candidates, learn from results, and advance when evaluation is quick and dependable. This also holds for AI research itself: enhancing models, training methods, and tools that enable them. Improved models can help with the subsequent development cycle, resulting in a compounding process that speeds up advancement.
This gives seasoned researchers the chance to investigate topics that were previously unattainable due to technical constraints. Additionally, it makes it easier to distinguish between finding a solution and comprehending why it works, what generalizes, and what to ask next. But we shouldn’t presume that AI won’t be able to solve problems or make judgments or abstract concepts. Universities should get ready for a time when artificial intelligence (AI) will likely surpass humans in many, if not all, areas of intellectual endeavor.
How should departments reconsider graduate training and academic credit?
A: A polished paper is becoming a less reliable indicator of individual knowledge in an increasing number of subjects. Departments need to start reevaluating what they reward right away, rather than just defining quality work as something that AI is not currently capable of. Good questions, replication, concept synthesis, negative outcomes that are instructive, and shared datasets can all be worthy of more credit. In a broader sense, evaluation should determine the contributions and intellectual responsibilities of a researcher, including instances in which significant portions of the job were completed by artificial intelligence. Both existing and prospective PhD students should be made aware of these expectations, which should serve as a guide for funding, hiring, and promotion.
Training is a more difficult issue. Exercise builds intellectual muscles. Students have traditionally been assisted in developing intuition and judgment through routine computations, coding, unsuccessful attempts, and minor discoveries. Even if AI enables students to take on more challenging assignments, assigning this work can eliminate formative experiences. We must separate activity that leads to the development of competence from needless friction. Pupils should be taught how to create problems, evaluate model outputs, replicate outcomes, and explain their decisions. As the foundation for effectively utilizing these tools, fundamentals may become increasingly valuable.
What should be built right now at other universities?
A: In my opinion, universities have a genuine opportunity here. First, let’s be clear about the objective: universities should be able to independently assess claims, communicate findings, and investigate problems over extended periods of time. Industry partnerships will be crucial, but we shouldn’t assume that commercial priorities will span the entirety of science or stay in line with it in the long run. Therefore, a certain level of technological independence will be required.
The information and expertise that academia has amassed in its labs may be its most valuable strategic asset. A selective record is presented by published papers: unsuccessful experiments, abandoned paths, and the reasons why an approach failed are frequently left unpublished. Experience has given scientists and engineers tacit information about which initiatives are likely to fail and why. This lack of context could help explain why models are now unable to predict outcomes that are obvious to an expert in some fields, particularly the empirical sciences. These restrictions might be short-lived, and models could become considerably more adept at scientific investigation if they capture unfavorable outcomes and experts’ judgments.
We may envision a time when labs operate more like a single, living, scientific organism, linked by a common infrastructure for AI research. Assume, for instance, that improved techniques for segmenting neurons in microscope images are required in a neuroscience lab. An AI bot might identify a noteworthy development from a computer vision team, link the researchers, provide benchmarks, and assist with iteration. It might make lessons from one lab accessible to others and bring up unanswered questions for instructors and students. We should create workflows that record hypotheses, interventions, results, failures, and interpretations to enable this. These workflows should be connected by shared systems, enabling agents with the necessary permissions to use tools and data across laboratories. This will necessitate significant institutional and governmental investment in secure data systems, computation, and post-training and model adaptation expertise.
The lineage of ideas might be preserved and contributions, including those of graduate students, could be more easily identified by tracing this process. Researchers could work together more freely and with more assurance that their work would be recognized if consent and credit guidelines were established. The preceding query of how to honor and reward intellectual achievements may also be addressed by this.
AI systems are now capable of synthesizing and reasoning about data from more sources than any one researcher could possibly comprehend, and their capacity to draw meaningful connections will only grow. This ability should be used to unite us around challenging scientific and engineering issues, enabling us to build on each other’s abilities and perspectives. This seems like a wonderful future, but to make it happen, institutions must invest now.