At the request of Epoch AI, workshops in six cities have selected famous conjectures for which any solution can be verified automatically. Some of them have already been proven false.
Scientific American tells how mathematicians gathered in June in London, Toronto, Los Angeles, New York, Berkeley and Cambridge (Massachusetts) made up a list of 50 open-ended issues. Condition imposed: each solution must be capable of being controlled by a machine. The initiative comes from Epoch AI, an organization that measures the progress of AI systems, while recent models successfully address issues previously reserved for human researchers.
A test bench for automated mathematics
The problems cover a variety of fields, from knot theory to number theory. Among them, find three whole cubes worth 114: the case of 42 had been solved in 2020 at the cost of 1.3 million hours of calculation, and Greg Burnham, researcher at Epoch AI, estimates that a similar approach would cost at least $100 million for 114. By the end of September, some of the issues on the list had already been resolved, some of them major. Burnham believes it is likely that the more than 100 open questions that OpenAI says have resolved with an internal model overlap several.
Read the full article on the original source
Go to the source →