OpenAI publishes a large set of maths results from an unreleased internal model, but not the model or its prompts

On 6 October OpenAI published a large set of mathematical results produced by an internal model it has not released, Decrypt and Pulse 2.0 reported. OpenAI's own post refused access from here on two tries, so the details are as reported.

Decrypt reported that the manuscripts were posted in a public GitHub repository and grouped into families of related results, and noted that the count is of manuscripts, not of solved problems. OpenAI puts the average result at about three hours of ChatGPT Pro thinking, Decrypt and Pulse 2.0 reported. OpenAI says many of the proofs already have formal versions in Lean, a language for machine-checked proofs, Pulse 2.0 reported.

The Institute for Advanced Study's advisory group on mathematics and AI had asked on 29 September for the model name, prompts and per-result compute, Decrypt reported. According to Decrypt, OpenAI released average compute figures and ten summaries of the model's reasoning, but no prompts, and has not released the model. MIT mathematician Andrew Sutherland told Decrypt: “We should ask for receipts.”

The Switchboard reported on 22 September that OpenAI had set up a maths advisory group after claiming more than 100 solved problems. What is new is the publication itself. The results are OpenAI's own claims, and outside mathematicians cannot reproduce them while the model is unreleased. The exact counts of manuscripts and problems appear in only one human-written outlet, so they are not given here.

• Reported OpenAI posted the manuscripts in a public GitHub repository, grouped into families of related results, Decrypt reported. Decrypt
• Claimed OpenAI puts the average result at about three hours of ChatGPT Pro thinking, Decrypt and Pulse 2.0 reported. Pulse 2.0
• Reported OpenAI released no prompts and has not released the model, Decrypt reported. Decrypt