A burst of AI-powered breakthroughs in mathematics has given way to a stubborn plateau. After frontier AI models solved roughly fifty Erdős problems - some of which had stood open for decades - the pace of purely autonomous solutions has dropped to near zero, mathematician Terence Tao said in a conversation with podcaster Dwarkesh Patel.

"There was a month where that happened and that has stopped, not for lack of trying," Tao told Patel, as reported by "Hvylya", citing the Dwarkesh Podcast. "I know of three separate attempts to get frontier AIs to tackle every problem simultaneously." The models pick out minor observations or flag that a problem was already solved in the literature, but no new purely AI-driven solutions have emerged since.

Tao likened the situation to a dark mountain range. Some walls are three feet high, some are a mile tall, and nobody knows which is which in advance. AI tools function like jumping machines that can clear two meters - higher than any human - but they either succeed or crash. "They've been really bad at creating partial progress or identifying intermediate stages that you should focus on first," he said.

The pattern behind the successes was consistent. Almost all of the fifty solved problems lacked any existing literature - Erdős posed them once or twice, a few people tried casually and gave up, but nobody published a serious attempt. The solved problems, it turned out, relied on combining one obscure technique with another published result. "That's the median level of what AI can accomplish, and that's really great," Tao said. "It clears out fifty of these problems."

But when researchers conducted systematic sweeps rather than cherry-picking the wins, a different picture emerged. On any given problem, AI tools showed a success rate of roughly 1% to 2%. "It's just that they can buy scale, and you just pick the winners. It looks great," Tao said. The remaining six hundred-odd open Erdős problems are being chipped away slowly, mostly through human-AI collaboration rather than pure AI one-shots.

Also read: "Hvylya" reported on how one pet owner used 300 gigabytes of canine DNA data and a chatbot to produce a personalized cancer treatment.