The New Scientist report details how an OpenAI model resolved a long-standing Paul Erdős conjecture on graph theory and additive bases, a result that has stood for decades without proof. Unlike prior AI tools that verified existing theorems or optimized known proofs, this system generated a novel combinatorial argument that mathematicians verified as both correct and insightful. The methodology relied on large-scale reinforcement learning over mathematical statement spaces rather than brute-force search, though exact model scale and training corpus remain undisclosed. This marks a departure from DeepMind's AlphaProof work, which focused on IMO problems using formal verification; here the output bypassed formal systems to produce human-readable insights directly. Earlier coverage overlooked the training data implications, where the model may have internalized patterns from thousands of unpublished arXiv preprints, raising questions about originality versus sophisticated recombination. A related 2023 arXiv preprint by researchers at Google DeepMind on AI-driven conjecture generation (arXiv:2305.06547) provides context, showing similar phase transitions in algebraic geometry but limited to toy problems with sample sizes under 500 instances. Limitations persist: no large-scale replication across other Erdős problems has occurred, and the result remains a single verified case without peer-reviewed publication. The pattern echoes the 2016 AlphaGo moment, where narrow mastery generalized into broader capability claims, suggesting AI may soon tackle open problems in number theory at scale.