A Twitter user named Liam Price claims GPT-5.4 Pro solved Erdős Problem #1196, a conjecture about primitive sets of integers that's been open for decades. According to Price, the model produced a complete proof in about 80 minutes and spent another 30 converting it to LaTeX. The Erdős problems website now lists the problem as "PROVED" with a discussion of the solution, which reportedly uses a Markov chain approach with the von Mangoldt function.
Jared Lichtman, a mathematician who has published partial results on this exact problem, commented on the Erdős forum that the proof is short and non-standard but he couldn't find a flaw. That's notable. Lichtman knows this territory well. But he's also involved with Math.inc, an AI startup focused on autoformalization that counts Fields medalist Terence Tao as a partner. Hacker News commenters quickly flagged this potential conflict of interest.
The concerns are legitimate. Did GPT-5.4 Pro solve this autonomously, or did significant human guidance shape the output? What prompts were used? Price's background and credentials remain unclear, and we don't know how he accessed the model or what methodology he followed. Without transparency on the prompting process, it's hard to say what actually happened here. Was this genuine mathematical reasoning from the AI, or something more curated?
A verified solution would mark a real milestone for AI solving real math research. But the field needs more than claims posted on Twitter. It needs reproducible processes and independent verification from people who aren't financially tied to AI companies. The solution might be valid. The process around announcing it isn't.