Open source math for the AIs

Reyansh Sharma, a London/Cambridge-based researcher, was selected by the Mercatus Center's Emergent Ventures program in its 46th cohort (announced September 2025) for work on open-source mathematical training data and evaluation sets for language models. The Emergent Ventures application is a single short form covering the plan, a brief bio, and a ballpark budget, with Tyler Cowen as program director making the funding call after consulting informal advisers. Grants in the program typically fall in the $1K to $50K range.

The category Sharma is working in sits between the data and training layers and addresses a known bottleneck for open-weights reasoning models. Math corpora and step-level evaluation sets are scarce relative to demand. The reference points in the public literature are projects like OpenWebMath (an open dataset of mathematical web text) and NVIDIA's OpenMathInstruct-1 and OpenMathInstruct-2 (1.8M and larger instruction-tuning datasets). Closed labs are widely reported to have built much larger internal math corpora.

Public detail about the specific dataset, license, or evaluation framework Sharma is producing is not yet available beyond the cohort announcement at Marginal Revolution. The Emergent Ventures grant covers the early, exploratory phase of work where a single individual is the natural unit of funding.