SAIR’s Open Math Model initiative

Last year I co-founded the Foundation for Science and AI Research (SAIR) with some private donors to create a non-profit organization that could support responsible uses of AI in mathematics and the other sciences, independent of the major AI companies. Initially, our resources were rather limited, and our main activities consisted primarily of podcasts, short events, and competitions. However, we have been slowly trying to expand in scale, ambition, and logistical capacity; for instance, we were very grateful to receive recent support from XTX markets to fund our next round of SAIR competitions. Furthermore, we have been negotiating with several key partners from academia and industry towards an initiative to build open weight models for science, as well as open source tooling for using both open and closed source LLMs for scientific applications.

We had initially planned a gradual rollout of this initiative, starting with a few pilot projects over the next few months, but given current events and the high demand for open models, we are accelerating our schedule and moving our initial announcement of the initiative to today, despite the fact that some key planning is still underway. In particular we are now ready to collect expressions of interest in the initiative, as we are seeking partners who can contribute funding, compute, expertise, or community building to our efforts. (Expressions of interest can be registered at the bottom of the announcement page.) I have replicated the text of our initial announcement below: we plan to provide more updates and information on the initiative shortly, including an announcement of our partners, as well as future plans to integrate this initiative into our existing portfolio of podcasts, events, and competitions.

Building Open Models for the Mathematical Community

An Invitation from SAIR Foundation

Mathematics advances through discoveries that others can examine and build upon. Generations of researchers have created a shared body of knowledge, sustained by open exchange and respect for each other’s contributions. As AI becomes part of mathematical research, these principles should shape the tools we use. SAIR and its partners are commencing an initiative to be at the forefront of shaping and implementing these new tools. Let’s work together to build Math 2.0.

Models Shaped by the Community

Through SAIR Open Models, we invite the mathematical community and its supporters to build open-source models together. Researchers can and will determine how these models are trained, what they are evaluated on, and how they serve research and education. The community needs models it can inspect, improve, and run independently, with development guided by shared scientific priorities. Participation should be open across institutions, regions, and career stages.

Tools for Everyday Research

Our first phase will focus on everyday mathematical work: understanding difficult arguments, checking references, exploring examples, writing code, and formalizing proofs. Our models will help people develop understanding and pursue investigations they could not otherwise undertake. We will measure their value through reliable assistance, verifiable results, and the cost of sustained use.

Open Development

Our models will have open licensed weights and code, published training methods, and reproducible evaluations. Training data will come with documented sources and compatible permissions, with limits on sharing stated clearly. Model releases will report failures and limitations alongside successes. Independent teams will be able to reproduce the work and adapt models to their own mathematical needs.

The Mathematical Community Owns the Data

The mathematical community owns the data and decides how it is used. We will use users’ data to train or improve models only with their explicit consent, on terms agreed in advance. Researchers choose what to share and for what purposes.

Shared Intellectual Property

Models, code, and tools developed through this initiative will be shared with the community under open licenses, for example Apache 2.0, MIT, or CC BY 4.0 where appropriate. Ownership, licensing, and attribution for jointly developed work should be agreed with contributors from the outset. Nevertheless, researchers retain ownership of their own independent and prior work which is shared with the initiative and receive credit for their mathematical ideas and results.

Open Community and Governance

We welcome participation across institutions, regions, and career stages. The mathematical community should govern the initiative, setting priorities and deciding how shared resources are used. Governance rules and decisions will be public, with clear ways for members to propose changes, take part in decisions, and hold leaders accountable. While we will work with strategic industry partners to provide compute and other resources, it will be essential that these partnerships preserve the research independence of the community.

Join the Initiative

We invite you to join SAIR Open Models and help make open models and affordable compute a shared foundation for mathematical research. We plan to make more detailed announcements on this initiative in the very near future: stay tuned!

Contact: collaboration@sair.foundation

You’re also welcome to join the discussion on SAIR Foundation Zulip:
Join the discussion on Zulip