GPT-5.4 Pro cracks Erdős #1196

Hot take: this is a real signal for math research workflows, but the “single shot solved 60 years of math” framing oversells it.

• –Tao’s comment matters more than the Reddit hype: he frames it as a meaningful but narrow case of AI finding an overlooked idea, not a general replacement for mathematicians.
• –The reported breakthrough seems to come from a discrete/probabilistic reformulation, which is exactly the kind of cross-domain jump LLMs can sometimes surface.
• –Human experts still had to validate, rewrite, and contextualize the argument, so this is augmentation, not autonomous theorem proving.
• –If the result holds up, the takeaway is practical: LLMs may be especially useful for suggesting alternate proof lenses when the field has converged on a suboptimal default approach.