More questions about whether researchers can trust OpenAI with unpublished math
Points and comments are a snapshot, not live.
Mathematician Andreas Thom details an exchange raising trust issues with OpenAI over unpublished math.
Andreas Thom recounts emailing OpenAI researchers Mark Sellke and Sebastien Bubeck after the non-sofic-group announcement, asking whether his private ChatGPT conversations about related math entered training data or were accessible to the solving process. Sellke's complete answer was "that did not happen." Thom now calls this dishonesty, noting OpenAI's later statement in the Buckmaster-Alpöge case that it cannot rule out that de-identified usage data helped improve models.
Thom's toot links to recent finite-time blowup results by Levent Alpöge and tristanbuckmaster, and subsequent replies discuss opt-out settings and broader trust concerns.
What commenters are saying
Commenters split on whether the issue is credit, model competence, or both. Some argue the AI likely only solved problems after training on human chat data, making its claimed autonomous breakthroughs dubious. Others note that if true, this still shows human-AI collaboration supercharges discovery but creates a prisoner's dilemma for researchers.
A reply flags that OpenAI's data controls opt-out is enabled by default, and that the company's business terms explicitly forbid using customer data for model improvement, unlike consumer terms. Several commenters urge lawsuits and warn against trusting Big Tech with confidential research.