Buckmaster also accused OpenAI of a conflict of interest, alleging that research proposals summarized by panelists using AI end up in its training data. “They have all the mathematicians and scientists' research programs in their training data and how they're going to go about it,” he said. The release contains 372 families of related results from an unreleased internal model, rather than solutions to 722 separate open problems. OpenAI has withdrawn three manuscripts and revised 14 others.
Other researchers are building on the work. Rohan Arun published four improvements covering van der Waerden, Sidorenko, unit distances and matrix multiplication, including a conditional 500-fold increase in an exponent coefficient for the van der Waerden result. Separately, contributor SJ Swapnil Jain raised the exponent saving parameter in OpenAI’s theoretical integer multiplication bound to 6.6857 × 10⁻⁴ from the original 2⁻¹⁸². The increase concerns a parameter, not practical multiplication speed. Jain wrote, “We seem to be approaching a wall. Gains now come in fractions of a percent.”