OpenAI’s ‘embarrassing’ math (techcrunch.com)

OpenAI researchers publicly celebrated GPT-5 for allegedly “finding solutions” to 10 previously unsolved Erdős problems and making progress on 11 more, but the claim was quickly walked back amid sharp criticism. A now-deleted tweet from VP Kevin Weil prompted Meta’s Yann LeCun and DeepMind’s Demis Hassabis to call the announcements embarrassing. Mathematician Thomas Bloom, who curates an “Erdős problems” list, clarified that the label “open” on his site simply meant he was unaware of existing papers — GPT-5 had surfaced prior literature, not produced new proofs. OpenAI researcher Sébastien Bubeck conceded the model found existing solutions, while arguing literature search at that scale is itself nontrivial. Technically, the episode highlights a useful but distinct capability: large models can be powerful research assistants for retrieval, citation discovery and literature triage, yet retrieving prior work is not the same as solving open mathematical conjectures. The controversy underscores risks around overstating model achievements, the need for careful provenance and reproducibility, and clearer evaluation protocols for claims of creativity or discovery. For the AI/ML community, this is a reminder to separate retrieval-based competence from genuine theorem‑proving breakthroughs and to insist on verifiable evidence when models are credited with advancing formal, peer‑verified research.