An open frontier model for mathematical discovery

Lead: Tristan Buckmaster

The research areas of the MAIC center gravitate around one shared project: an open AI system for mathematics, built together with the data, environments and training infrastructure needed to develop it. Across this site we call this project the core.

Why an open model. The most capable AI systems are developed in closed industrial labs. Mathematicians can use them, but they cannot examine how they were trained, reproduce reported results, or adapt them to the problems their fields care about. An open model, with open methods, makes frontier AI something the mathematical community can study, test and improve, and not only use.

How we will build it. We will proceed in three stages, each useful on its own:

1. Mathematical data and environments at scale. Problems drawn from our members’ research, formal statements and proofs, and computational tasks whose answers can be checked automatically. Together these give a training signal that is reliable because it can be verified.
2. Post-training open models. In partnership with open-weight model providers, we will post-train existing models on these data and environments. We control, and can release, our own post-training data and training runs. The starting model’s pretraining data and history remain in the hands of its provider.
3. Our own model and training stack. Building on the first two stages, we will develop an open model and training stack of our own, so that every stage of training is open to study.

How the research areas connect to the core.

• Formalization turns written mathematics into verifiable training data.
• Pure and applied mathematics supply research problems, and test whether the model helps solve them.
• Mathematics of AI uses the core as an experimental system whose weights and training can be examined.
• AI safety sets the evaluations, audits and release criteria that govern how the model is trained and released.
• Training and outreach studies how mathematical practice and education adapt to these systems.

Safety and release. Because released weights cannot be recalled, each release will be evaluated beforehand, including after adversarial fine-tuning, and its risks measured relative to open models already available. See AI safety.

Partners. The core is being built by the members of the Open Mathematics Consortium.