Chat GPT 5.4 solved a 60+ years unsolved erdos problems in a single shot
A single PhD-level formula cracked a math puzzle that stumped 50 years of mathematicians...
OpenAI's ChatGPT-5.4 has achieved a remarkable milestone by solving an unsolved Erdos problem that had resisted mathematicians for over 60 years. The problem, cataloged as number 1196 on the Erdos Problems website, was cracked in a single interaction by applying a basic PhD-level formula. This feat directly challenges the longstanding criticism that large language models merely predict the next token without genuine reasoning ability.
Notably, the problem had attracted attention from prominent mathematicians, including Terence Tao, who had commented on it. The solution demonstrates that advanced LLMs can now outperform decades of human mathematical effort on specific open problems. The conversation log, shared via ChatGPT's public link, shows the model reasoning step-by-step to arrive at a valid solution, suggesting a level of logical deduction that many experts previously thought impossible for AI systems.
- ChatGPT-5.4 solved a 60+-year-old unsolved Erdos problem in a single interaction using a basic PhD-level formula
- The problem (Erdos #1196) had remained open for decades despite interest from leading mathematicians like Terence Tao
- This challenges the narrative that LLMs cannot reason, as the model demonstrated genuine mathematical deduction beyond token prediction
Why It Matters
Demonstrates LLMs can now solve open research problems, potentially accelerating scientific discovery across disciplines.