Erdosgate (garymarcus.substack.com)

OpenAI researcher Sébastien Bubeck posted a high-profile claim that GPT-5 had “found a solution” to a number of Erdős problems, prompting huge online attention (≈100k views) and swift backlash when it became clear the model was mostly surfacing existing, web-available proofs rather than producing novel solutions to open conjectures. Math and AI communities criticized the framing — DeepMind’s Demis Hassabis called it “embarrassing” — and Bubeck deleted and tried to retract the announcement. The episode echoed past misleading presentations (e.g., an earlier OpenAI Rubik’s cube claim) and exposed how quickly extraordinary claims can spread without rigorous vetting. Technically, the incident highlights key issues for LLM-driven math claims: model memorization or retrieval of published proofs can be mistaken for original reasoning, provenance and dataset leakage must be checked, and human verification or formal proof validation is essential before asserting breakthroughs. For the AI/ML community this is a teachable moment about evaluation standards — transparent evidence, reproducible verification (formal proofs, checkable derivations, citation of sources), and careful communication — to avoid eroding trust and conflating impressive retrieval/reconstruction abilities with true mathematical discovery.