Just days after OpenAI revealed that one of its AI systems had solved a famous math problem first proposed by legendary mathematician Paul Erdős in 1946, researchers at Google DeepMind have now made an even bigger claim. According to DeepMind researchers, a new AI system called AlphaProof Nexus has autonomously solved nine open Erdos problems, including some that had reportedly remained unsolved for more than half a century. And honestly, the most surprising part may not even be the math itself. Researchers say the system solved each problem using only a few hundred dollars worth of computing resources. In AI terms, that is unusually cheap for something this complex.
DeepMind says AlphaProof Nexus is very different from typical AI systems that sometimes produce convincing but incorrect responses. One of the biggest issues in AI-generated mathematics right now is what researchers call “hallucinations.” An AI can generate a proof that sounds technically brilliant while quietly inventing fake assumptions or skipping the hardest parts of the problem entirely.
Google researchers indirectly referenced this growing issue while discussing how their own system works. Sometimes, they explained, AI models invent mathematical shortcuts known as “helper lemmas” instead of actually solving the difficult section of a proof. To a human reader, the result can appear legitimate even when crucial logic is missing. That is where AlphaProof Nexus supposedly changes things.
Instead of relying purely on AI-generated reasoning, DeepMind combined its language model system, reportedly powered by Gemini 3.1 Pro, with a formal mathematical verification tool called Lean. Here is the important bit: Lean automatically checks every single step of the proof using strict mathematical logic rules.
Unsupported claims, fake assumptions and missing logic simply fail verification immediately. That means the AI cannot just bluff its way through complicated mathematics the way some chatbots occasionally do.
According to the researchers, AlphaProof Nexus also solved 44 open OEIS conjectures and helped resolve a 15-year-old problem in algebraic geometry. The system even discovered a previously unknown optimisation parameter that humans had not identified before.
Google researchers say the goal is not necessarily replacing mathematicians entirely, at least not yet. Instead, they believe formally verified AI systems could eventually help researchers move through difficult mathematical problems much faster by narrowing down possible solutions and checking proof structures automatically. “Our results support this vision,” the researchers said.
They added that human collaborators found the AI-generated proof attempts useful even when the system failed to fully solve a problem. And that may be the bigger story here.
