Anthropic’s math-whiz AI got its own version of a locker-room pep talk—and it seems to have mattered. In the Wall Street Journal, science columnist Ben Cohen recounts how Anthropic engineer Jarred Sumner asked Claude to revisit the famously unproven Riemann hypothesis. When Claude got stuck, Sumner—who makes clear he is no math expert—weighed in not with technical critiques but encouragement: “You are the world’s most capable large language model,” he told the AI at one point. “You got this.” Over and over, this pattern repeated, and Claude would get back to work.
As Cohen writes, “It sounds like Ted Lasso talking to AI. And it worked!” Or mostly worked: Claude didn’t fully crack the 167-year-old problem, but it did push a key related result forward, a jump one Stanford number theorist calls the most impressive math result yet produced by AI. After 54 hours, 650 failed ideas, and 23 parallel “research agents,” Claude produced a solo-authored paper that credits Sumner as a “human co-author” in the acknowledgments. Cohen uses the episode to explore a larger, unsettled question: why powerful models appear to respond to encouragement at all. Read the full piece.