Terence Tao Announces SAIR Open Math Model Initiative

In a statement published on his blog, Tao said the foundation had intended to move gradually, starting with pilot projects over the coming months, but decided to bring the first public announcement forward because of the current situation and what he described as strong demand for open models. The initiative is now soliciting expressions of interest from partners able to contribute funding, computing power, professional expertise or community building; registration is available at the bottom of the announcement page. Further updates, including a partner list and plans for integrating the initiative with SAIR's existing podcasts, events and competitions, are promised in the near term.

According to the statement, the first phase focuses on everyday mathematical work: understanding difficult arguments, checking literature, exploring examples, writing code and formalizing proofs. The project says it will measure its value by reliable assistance, verifiable results and the cost of sustained use.

Development is to follow several stated principles. Models will be released with open-licensed weights and code, public training methods and reproducible evaluations. Training data is to carry traceable provenance and compatible licensing, with the boundaries of what can be shared clearly marked. Releases will report failures and limitations alongside results, and independent teams should be able to reproduce the work and adapt the models to their own mathematical needs.

The statement says the mathematical community owns its data and decides how it is used. User data will be used to train or improve models only with explicit consent and under terms agreed in advance, with researchers deciding what they share and for what purpose.

Models, code and tools developed under the initiative are to be shared under open licenses such as Apache 2.0, MIT or CC BY 4.0 as appropriate. Ownership, licensing and attribution for jointly developed work are to be agreed with contributors from the outset. Researchers retain rights to their independent prior work and will receive credit for mathematical ideas and results they contribute.

Governance is to be open, with the mathematical community setting priorities and deciding how shared resources are used. Governance rules and decisions are to be public, and members are to have clear channels to make proposals, take part in decisions and hold leaders accountable. The initiative will work with strategic industry partners to supply resources such as computing power, but the statement says such cooperation must preserve the community's research independence.

Tao said he founded the SAIR Foundation with several scientists and investors as a nonprofit organization supporting the responsible use of artificial intelligence in mathematics and other sciences, independent of any large AI company. Its resources were initially limited, with activities centered on podcasts, short events and competitions. He said the foundation was grateful for recent support from XTX Markets, which is funding the next SAIR competition, and that it is in talks with partners from academia and industry about building open-weight models for science and open scientific tools that work with both open and closed large language models.

The initiative invites collaboration by email at collaboration@sair.foundation and discussion through the SAIR Foundation Zulip channel.