Speaking at the event on AI in schools at the Vietnam Institute for Advanced Study in Mathematics (VIASM) on May 22, professor Ho Tu Bao said the first question worth asking is how deeply AI actually "understands" math. To find out, he ran a striking test.
He asked an AI tool to explain how Ngo Bao Chau, who serves as VIASM's scientific director, had proved the Fundamental Lemma of the Langlands program, a problem that stumped mathematicians for decades before Chau cracked it in 2008 and won the 2010 Fields Medal for the work.
The AI produced an explanation that "looked very much like understanding," Bao said. But what looked like comprehension was really just a strong capacity to gather and synthesize information.
Generative AI works on language models, Bao explained. The tools do not grasp the underlying nature of mathematics. They search vast datasets, chain together patterns and produce outputs based on probability.
"AI can understand the components of a mathematical formula and their relationships, but it lacks common sense and cannot connect mathematical knowledge to the stories of real life," he said, voicing concern about an "intellectual decline" when students rely on AI too much.
