AI
Claude Helps Organize Gödel's Notes and Discovers Hundreds of Previously Unknown Theorems
A University of Helsinki team has been studying 16 notebooks totaling 1009 pages that logician Kurt Gödel wrote in the early 1940s, much of them in an outdated German shorthand. Researchers previously advanced at about one page per week, but after introducing Claude in September, they obtained analyses of 650 pages within days, discovering hundreds of previously unknown theorems and 14 logical themes that were never further explored.
According to The Guardian, Austrian logician Kurt Gödel's two incompleteness theorems, published in 1931, profoundly changed mathematics and laid the foundation for the digital age. He did not publish many papers after the 1930s, but left tens of thousands of pages of unpublished material. Finnish logician Jan von Plato has led a project at the University of Helsinki since 2017, studying 16 "working notebooks" totaling 1009 pages that Gödel wrote in the early 1940s. The content is dense and heavily uses Gabelsberger German shorthand.
Mathematical interpretation of the project had previously progressed slowly, at about one page per week. In September this year, von Plato began using Claude after an AI and theorem-proving conference in France. He provided the model with his own transcriptions of notes, six books he had previously written about Gödel, and other secondary scientific literature. Within days, Claude generated a 650-page analytical document.
Von Plato said he was astonished because the notebooks not only contained early formulations of dozens of results that did not become known until the 1970s and 1980s, but also included 14 logical themes that had not yet been studied, showing Gödel was far ahead of his time. According to the report, AI summaries revealed that Gödel proved hundreds of theorems no one knew existed.
Von Plato received a €2500000 grant from the European Research Council and taught himself Gabelsberger shorthand. He said AI-generated content still needs to be checked page by page, but it can connect different parts across more than a thousand pages, whereas humans need about a year to read roughly 50 pages and find it difficult to keep all the material in mind at once.