The strange thing about modern AI isn’t just that it can solve hard problems. It’s that it can solve the same hard problem for different people, almost simultaneously, and we don’t quite know what to call that yet.
Case in point: two separate research groups used OpenAI’s GPT-5.6 Sol Ultra to crack an open problem in quantum cryptography known as “unclonable encryption” — a method for securing data using quantum properties. MIT PhD student Seyoon Ragavan submitted his paper to arXiv. Three hours later, professors Prabhanjan Ananth (UC Santa Barbara) and Amit Sahai (UCLA) submitted theirs. Both teams used the same AI model but prompted it through entirely different reasoning paths, Scientific American reports.
“If someone mentions an open problem, the first thing is to see if GPT solves it,” Ananth said. Ragavan was even more direct about the shift: “The way I do research now has nothing to do with how I did research two months ago.” That statement lands hard because it isn’t hyperbole — it’s a PhD student at one of the world’s top institutions describing a genuine rupture in his workflow.
The teams are now discussing whether to merge their papers, which is the pragmatic move but also sidesteps a thornier question the scientific community hasn’t answered: when every researcher has access to the same frontier model, what does “independent discovery” even mean? Priority disputes have always been part of science, but the timescale here — three hours — makes the old arguments feel quaint. The AI didn’t plagiarize. It just made the problem tractable enough that two groups converged on a solution within the same afternoon. Mathematics, in particular, is feeling this acutely, with reactions ranging from exhilaration at the acceleration of progress to a genuine sense of professional vertigo.
This isn’t the first time AI has compressed discovery timelines, but the specificity matters. Unclonable encryption wasn’t a brute-force search problem. It required conceptual insight. That two different prompting strategies both led to valid solutions suggests the model wasn’t just retrieving — it was navigating the problem space in a way that multiple human experts, working independently, could steer toward the same destination. Whether that’s a feature of the model or of the problem itself is exactly the kind of meta-question researchers will now have to wrestle with, probably while asking GPT-5.6 for its opinion.
