Large language models build their understanding of the world through interconnected patterns.
Christ Miranda, a member of the Alfréd Rényi Institute of Mathematics research team, who graduated last year from the Fazekas Mihály Secondary School and is now a University freshman in the United States, will give a lecture (in HUngarian) about how researchers are trying to “look inside” artificial intelligence using mathematical methods. The topic is especially timely: although hundreds of millions of people use large language models such as ChatGPT every day, we still only partly understand how these systems store and organize knowledge.
If you want to visit her lecture, please register HERE.
| With their paper on this topic, Miranda and her fellow researchers Adrián Csiszárik, Gergely Becsó and Dániel Varga were selected last year for the Spotlight program of NeurIPS 2025, the world’s most prestigious artificial intelligence conference. This is an extraordinary professional recognition: only the top three percent of tens of thousands of submitted papers are invited into this highlighted category. One of the most exciting insights of their research is that large language models do not simply “memorize” isolated facts. They do not function like an encyclopedia where data is stored side by side. Instead, they construct knowledge about the world through interconnected patterns.“AI models are trained using certain methods, but while they are learning — and later, when we use them — we do not really control what happens inside them. In a sense, they are black boxes. Simplifying somewhat, large language models work by learning to predict the next word. But while doing this, the relationships and structures of the world somehow become encoded inside this black box. And this is what we barely understand at all: what is actually happening in there? How is the world represented inside this black box?” A whole new scientific field has emerged in recent years around this problem — the study of AI interpretability. If we understood what happens inside the black box, we could build better and safer models. At the moment, many doubts and fears still surround these systems. More about the topic can be read HERE. |
For the young researcher, the NeurIPS 2025 conference was also a defining experience because she could directly witness the strong international interest in their work. At the conference, they spoke with experts from leading research institutes and technology companies around the world about how the methods developed at the Alfréd Rényi Institute of Mathematics could connect to other research directions.
“No one in my family worked in mathematics or any scientific field,” Miranda told renyi.hu when asked about her mathematical beginnings. “My grandfather taught me to play chess when I was very young. At first, I went to chess competitions. But in second grade, my teacher asked whether I would like to participate in a math competition. I went, it went well, and from then on mathematics started to interest me more than chess. Something just clicked.”
Christ Miranda Anna is currently a student at University of California, Berkeley, where she studies computer science while continuing her research on the workings of artificial intelligence. The second part of next week's event will be an informal Q&A session, where she will also talk about what it was like to do research already as a high school student and what it is like to study computer science at University of California, Berkeley.