AI Has Striking Science Skills, but Grad Students Are Still Wanted
AI agents can do many of the tasks traditionally taken on by graduate students. Both can be given a problem to solve or a project to put together. Both can read relevant papers, crunch through numbers, and write a program. Once done, they can present their results and revise any errors that are spotted. But a current state-of-the-art AI agent can complete the task in minutes, whereas a graduate student will need months. For a researcher like me, it is tempting to see in AI a route to faster discoveries and increased productivity. That possibility opens up questions about the role of graduate students going forward.
The range of possibilities unleashed by AI is so large that speaking of the future seems silly. Let us agree instead to talk about the present. I will share my perspective and conclusions after experimenting with agentic AI for several months.
First, there is the “wow” factor. If you have never used an agentic AI, such as OpenAI’s Codex or Anthropic’s Claude Code, please try it. There’s a thrill of discovering something truly new. The experience is a quantitative leap beyond your first conversation with ChatGPT. Earlier this year, I asked Codex to find an old theoretical paper of mine and fit the model proposed in that paper to a recent cosmological dataset that was published last year. To do that, the algorithm had to “read” my initial paper, find and interpret the relevant equations, write them into code and run the resulting program hundreds of times for different sets of parameters, produce predictions for the data, and find the parameters that best fit the data. For good measure, I asked it to plot the data with error bars along with the best-fitting model. Codex did that in under 60 seconds.
It is the speed and the volume of output that are so impressive. It would be less amazing if agents wrote the code over the course of a day or so, checking in after each step. I spent more time typing the question than the agent did in coming up with the answer. And the plots it produced—a series of graphs, each with multiple panels and curves—would take a human several attempts to get right.
That’s not to say that there aren’t hiccups. I have learned that AI still needs us (at least for now). The agents are perfectly content to produce a plot where the amplitude is off by 20 orders of magnitude. Or they will spit out rows and rows of values from a functional expansion, each number of which is incorrect. You point that out, and they say, “You’re right,” before going off to try and fix it. But without human input, they would not have known that they produced nonsense.
This supervising role has some parallels with mentoring a graduate student. In my career, I have been fortunate enough to work with dozens of students. Some of our projects were similar in scope to the AI query that I have described. The students worked through the difficulties, learning what makes sense and what does not. In most cases, their efforts paid off, and together we were able to publish an interesting scientific result for the advancement of the field.
If the goal were only to publish papers, then an AI agent could do some of the tasks of a graduate student at a fraction of the time. In fact, another physicist, Matthew Schwartz from Harvard University, gave an AI agent a project and published the results as a preprint, estimating that AI accelerated his research tenfold. Taken at face value, this speedup can enhance scientific productivity, allowing researchers to focus on bigger problems than debugging code.
And AI might also help level the academic playing field. Currently, there is a feedback loop that exacerbates inequality between the top-ranked and lowest-ranked universities: The best graduate students often attend the top universities, benefiting those school’s professors, who get more funding to attract more graduate students. By contrast, an AI agent doesn’t care what ranking a university has, which means any professor can install one of these programs to boost their productivity.
So, will researchers be less inclined to hire graduate students? Ultimately, this is tied to a larger and more frightening question: Will we still need humans to do work? I don’t know. Will agentic AI be like other technologies in history, where old jobs go away and new, more satisfying ones emerge to replace them? Or is this technology different? My guesses just aren’t worth writing down.
Fortunately, there is an easier question that I am more confident answering. Rather than, Do we still need graduate students? scientists should ask, Do we still want graduate students? And for me, the answer is unambiguously, “Yes!” One of the joys of being a scientist—for me and for colleagues I have spoken with—is to be approached by a prospective graduate student who shares our passion for the field, is familiar with our work, and is interested in learning more. We ask them to join our group, and they take on a project. We meet with them regularly, listening to their frustrations while gently guiding them in the right direction. Their contributions become more and more valuable until we ultimately realize that they don’t need us anymore. This process is an integral part of doing physics—one of the top reasons we enjoy our work.
An important point worth making is that graduate students themselves are learning to use AI tools—and they may well adapt to the changing world faster than their advisors. In the long run, these tools are likely to reshape how students contribute rather than replace them. The whole process of training scientists may be overturned. A lot is still unknown, but the thing I am surest of is that I cannot imagine doing physics without students.