DeepMind CEO Demis Hassabis says solving Erdős problems isn't AGI

Google DeepMind CEO Demis Hassabis has ignited a fresh debate over AGI benchmarks by stating that solving advanced mathematical problems like Erdős problems does not equate to true artificial general intelligence. This comes after DeepMind's AI systems recently solved 8 Erdős problems—long-standing combinatorial conjectures that had stumped professional mathematicians for years. Hassabis emphasized that while these achievements are impressive, they remain narrow feats within well-defined finite spaces, far from the domain-general innovation required for AGI. He drew a contrast with historical figures like Srinivasa Ramanujan, who invented entirely new mathematical objects.

The comments have split the AI community. Gary Marcus, NYU professor emeritus, defended Hassabis, arguing critics misinterpreted his focus on cross-domain innovation versus specific problem-solving. Economist Noah Smith called Hassabis's stance not credible, given the difficulty of the solved problems. Researcher Valerio Capraro supported Hassabis, noting that Erdős problems are exactly the kind of combinatorial searches where current AI excels with massive compute, but true invention requires creating new conceptual spaces. The exchange underscores how even after landmark achievements, the goalposts for AGI remain contested—and perhaps moving.