Mathematicians advise OpenAI to release AI proofs with prompts and costs
An independent group including Timothy Gowers, Martin Hairer and Edward Witten wants AI labs to publish prompts, formal proofs and failed attempts, and to fund human understanding of any result.
The Advisory Group on Mathematics and Artificial Intelligence published recommendations on September 29 for how AI labs should release mathematical results. Its website lists nine members, among them Timothy Gowers, Martin Hairer, Edward Witten, Ravi Vakil and Melanie Matchett Wood. The group says it works independently of any AI company and that its members take no payment.
According to the group, it formed after OpenAI approached some members about an external advisory board. In agreement with OpenAI, they instead created an independent group and invited others. It says its current task is advising OpenAI on releasing a large number of significant results that OpenAI reports were produced by an internal model. The paper has not seen those results.
For each result, the recommendations say a lab should publish the model's name, the prompts used, a summarised chain of thought, the time taken and the estimated compute cost. Proofs should be formalised where possible, and deposited in a repository that no lab controls. The group also says labs should not treat mathematical results as marketing vehicles.
When many results arrive together, it asks for a further document saying how many problems of comparable difficulty the models tried and failed, and how the problems were chosen. Where no human understands a result yet, the group says labs must help pay for that understanding, through workshops, working groups or expository writing, with funding decided by existing nonprofits rather than the labs.
The group says the guidance draws on more than 600 survey replies from mathematicians. It also asks labs to stop testing advanced problems on proprietary models the wider community cannot use, and to give mathematicians broad access to public models. The recommendations are advice only: the group says it has no decision-making power at any AI company, and no lab has yet responded in the sources the paper read.