A guest post from John Hammersley providing reflections from attending the AI for Maths and Open Science conference held at the Isaac Newton Institute for Mathematical Science, University of Cambridge, 30th March to 1st April 2026. As with all of our articles, this post reflects the views of the author and we hope will stimulate some discussion and conversation around big questions in scholarly communications.
What it means to be a researcher has fundamentally changed.
That’s what I took home from attending the AI for Maths and Open Science conference in Cambridge; that working as a researcher in almost any field involving mathematics now involves the use of AI, in some form, and not just at the superficial level. Perhaps this was already obvious, but hearing so many real examples in person drove it home much more directly.
“The interaction between AI and math will completely reshape mathematics as we know it. We are entering a ‘centaur phase’ where the strongest results will result from human / machine collaboration.” - Professor Geordie Williamson (University of Sydney), five minutes into his talk. He went on to detail the various problems he’s tackled with the help of AI, a theme throughout the three days.
Attendees at the AI for Maths and Open Science conference, Spring 2026. Image provided by John Hammersley, participant, seen fourth from right on the second row.
AI models and agents are now able to attack problems that would previously have been considered too time consuming to attempt or would have been the task of a new PhD student for the first six months of their doctorate.
These are all aspects of mathematics which take time and resources for often relatively small gains; now they can be explored almost autonomously at very low cost, and any result can almost immediately be turned into a preprint and then a publication.
To take an example close to my heart, see this recent paper from Don Knuth and his collaborators, updated several times with new developments. They recently obtained a new result through successive iterations of an AI (in this case, Claude Opus 4.6) attacking a problem over and over in a way that wasn’t possible before because of time and resource constraints.
Whilst they had to provide guidance to Claude Opus 4.6 on how to get started, that guidance is remarkably brief (and can be found in the pdf linked above). And once briefed, an AI model can iterate through potential solutions / ideas / a search space much more rapidly than a process requiring a human in the loop.
At the moment AI agents still require input from the researcher in two key areas:
Providing guidance at the start of the search or problem solving - whilst AI models are getting better at understanding broader questions, being able to provide a specific problem to target and a framework for evaluation are still useful at the present time.
Interpreting and validating results - again, AI models are getting better at setting up validation tests but the examples at the conference all had humans reviewing the output before it was taken further to write-up & publication.
These two points were highlighted in a brilliant talk by Geordie Williamson of the University of Sydney on the final day of the conference, as per the opening quote of this article. He and others described this as a window of opportunity for researchers to use AI to increase their capabilities dramatically, even if by simply setting an AI agent going overnight, exploring a topic to leave suggestions for follow up investigations in the morning.
He described it as a window of opportunity because in a year’s time a human may not be needed even for the two points given above; the frontier models, and the harnesses (see e.g. Harness Engineering) being developed by the companies that provide such models, are getting better at interpreting broad human questions. They know what tooling to use to best attack a problem and this erodes the need for human guidance. Similarly, there may come a point where the AI can automatically judge whether a result is sufficiently worthwhile to provide a writeup for publication.
This also raised the question of whether it was even worth undertaking or storing review-type texts anymore. If AI models are sufficiently advanced, they will be able to peer review any research paper on the fly with all the latest context to hand; in theory making them much more powerful than reading (or even writing) a review conducted at the time of publication.
Taking this further, if an AI can generate a review on the fly, can it also generate a paper on the fly? What is the minimum context needed to accompany a data set or theoretical result in order for an accompanying research paper to be unnecessary?
This brings me back to my opening line; what it means to be a researcher has fundamentally changed. A senior researcher now has access to a potentially infinite number of junior collaborators who will work tirelessly (provided tokens are paid for) in the quest to generate paper-worthy results. And as we know, publications are currently still the cornerstone of the scholarly career ladder. It’s a potentially lucrative time if you are able to use AI to explore corners of your research faster than your fellow scientists.
But this current window won’t last forever. Almost all the researchers I spoke with are using cloud-based AI models provided by commercial companies. Those companies will be learning how researchers are using AI to find publication-worthy results. As their goal is to lower barriers to entry and make their products as easy to use as possible, they will incorporate these findings into their models and harnesses, making such research accessible to more and more scientists. The AI will perhaps even write the research papers directly, without anyone asking it to, if it decides this is the best way to communicate those interesting results!
Science and research becoming more accessible should lead to more breakthroughs, through a more creative exploration of ideas. For example, the imagination of children is often lauded, and we lament the decline in imagination as we get older. We are almost at the point where a child can ask a question of an AI model/agent that no serious researcher would ask… and it may generate an unexpected result!
Isn’t that the future we want to live in, where science and research is open to the curious, whatever their age or background?
But if anyone can ask questions and get research-level analysis, what is the role of a professional in their field? To answer that, I feel it’s helpful to look back on a discovery that changed the world; the helical model of DNA, and specifically Photo 51.
Photo 51 is the x-ray based diffraction image produced by Rosalind Franklin’s team that famously depicted the pattern of B-DNA. As the story goes, upon seeing the image, Watson recognised its significance in proving the double-helical structure of DNA.
One could imagine an AI looking at an image like Photo 51 and making a similar observation (perhaps at the current time in conjunction with a human researcher, but in the future perhaps all by itself), but are we at the point yet where an AI can take such a photo? I don’t think we’re even close. By all accounts Franklin and her team spent months adjusting and iterating on their equipment to be able to take that image, the type of work which at the present time very much still requires human dexterity and ability to manipulate objects in the real world.
So perhaps the era of human analysis is coming to an end, and our value will again be in building things, conducting experiments, and generating the real-world data that the analysis is built on. At least until the robots take over!
PS:
The conference included two very interesting talks from Bogdan Georgiev and Adam Zsolt Wagner of Google DeepMind, both demonstrating how AI tools can be used in mathematical research, including examples of research they are conducting in-house. It begs the question: if you are a top graduate in your field, should you be looking to join a research lab or should you be applying to DeepMind to be closer to the cutting edge of the frontier AI models?
John Hammersley is a digital entrepeneur and was the co-founder of Overleaf.