AI and the practical scientist - by John Hammersley
Scholarly Futures
SubscribeSign in
AI and the practical scientist<br>We are entering a window of opportunity in which collaboration between human researchers and AI systems will bear the most fruit. But it will not last very long, with far-reaching consequences.
John Hammersley<br>May 19, 2026
Share
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.
Thanks for reading Scholarly Futures! Subscribe for free to receive new posts and support my work.
Subscribe
AI, Maths & Open Science
What it means to be a researcher has fundamentally changed.<br>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.<br>“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.<br>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.<br>Problem analysis. Exploration of the solution space. Tightening of bounds.
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.<br>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.<br>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.<br>A window of opportunity
At the moment AI agents still require input from the researcher in two key areas:<br>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.<br>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.<br>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...