With Table Mountain rising behind it and the southern Atlantic shimmering in the distance on a clear afternoon, the edifice is perched high on the slope of Devil’s Peak. Claude Formanek, a PhD candidate, has spent the majority of the last year sitting in front of two monitors in one of the upper-floor offices of the University of Cape Town’s Mathematics Department, working on problems that, until recently, mathematicians thought were nearly impossible to solve.
In this, he is not by himself. He works in part through the South African AI company InstaDeep, in part under Associate Professor Jonathan Shock’s supervision, and in part within the peculiar new ecosystem that 2026 has created, where pure mathematics and artificial intelligence have begun to truly collaborate in ways that the field had not yet witnessed.
| University of Cape Town & AI Mathematics Research — Key Information | Details |
|---|---|
| Institution | University of Cape Town (UCT) |
| Location | Cape Town, South Africa |
| Department | Mathematics and Applied Mathematics |
| Industry Partner | InstaDeep |
| PhD Candidate Driving Work | Claude Formanek |
| Faculty Supervisor | Associate Professor Jonathan Shock |
| Research Focus | AI applied to number theory, topology, combinatorics |
| Methodology | Generative AI + pattern detection in large mathematical datasets |
| Infrastructure Launch | African Compute Initiative (ACI) |
| Significance of ACI | Africa’s largest GPU-intensive AI cluster for higher education |
| 2026 Global Context | AI tools cracking decades-old math problems |
| Notable Benchmark | FrontierMath for advanced math |
| Notable AI Breakthrough Tool | GPT-5.4 cited in algebraic geometry results |
| Comparable International Effort | DeepMind’s AlphaProof and AlphaGeometry |
| Reference Resource | Mathematical Association of America |
There isn’t actually a single proof in this story. When AI solves a problem that was thought to be unsolvable, there’s a propensity to go for the big reveal and the heroic storyline. Because the Cape Town art deviates from that pattern, it is more intriguing. The slower, more subdued version of the AI mathematics revolution is taking place at UCT, where infrastructure, methodology, and intellectual habits are being developed to enable people and machines to collaborate on complex issues.
This year, MIT, DeepMind, and Stanford have made several impressive headline-grabbing discoveries. Because it is taking place on a region that has traditionally been a consumer of mathematical research rather than a creator of frontier breakthroughs, the Cape Town initiative is significant.
You can see the peculiar physical reality of this time when you walk into the Mathematics Department on a Tuesday afternoon. Pupils are using chalkboards. Hunched over laptops, a few PhDs ran queries to massive language models, such as DeepMind’s specialized math systems, OpenAI’s GPT-5.4, and Anthropic’s Claude.
A half-finished proof with marginal notes in a different color on a whiteboard in the corner suggested that a model suggested a specific reformulation that no human had ever considered. There’s a feeling that the field is still figuring out how to change the role of the mathematician. AI does not take the place of humans. It suggests routes. The human confirms, rejects, improves, and sometimes discovers that the road leads to something truly novel.
Cape Town’s position is unique because of the infrastructure component. The majority of African colleges lack the GPU-intensive computing power needed for frontier AI mathematics. This figure is altered by UCT’s African Compute Initiative, which launched as the largest such cluster for higher education on the continent.
This allows a PhD candidate studying number theory in Cape Town to do experiments at a scale that would have required physical relocation to Cambridge or Berkeley three years ago. It seems like a procedural detail. In actuality, a region’s ability to create frontier research or stay permanently downstream depends on this type of structural change.
The background of 2026 is also important. This year, decades-old algebraic geometry conjectures have crumbled, frequently to proofs that human reviewers regard as startlingly elegant—”very nice, clean,” as one mathematician put it. Benchmarks like FrontierMath, which were created especially to evaluate issues that human mathematicians have found difficult, are now being passed by AI models.
Observing the trajectory gives the impression that the field is in a similar phase to the one chess went through in the late 1990s, when machines ceased to be curiosities and began to work together. What this signifies is still up for debate among mathematicians. The alliance is welcomed by some. Some are concerned about what happens to mathematical intuition when proofs are produced by pipelines that are impossible for humans to thoroughly examine.
According to Shock and his associates, the Cape Town approach views AI as a generator of ways rather than solutions. The human mathematician determines whether the model’s “cliché”—a methodological pattern or potential line of attack—is worthwhile. This framework maintains the discipline’s foundation in human verification while taking advantage of the strengths of huge models—namely, their ability to recognize patterns in enormous mathematical literatures that are impossible for humans to fully comprehend.
It’s not a glamorous job. It’s the kind of gradual, meticulous integration that yields long-lasting outcomes as opposed to headlines. It’s actually unclear if Cape Town or another location will produce the next significant mathematical discovery in 2026. What’s more obvious is that a building on Devil’s Peak’s slope now lies solidly on the global map of frontier mathematics’ locations, which has been subtly revised.