OpenAI Posts 722 Math Manuscripts as Yuandong Tian Expects to Lose His Job

OpenAI has posted 722 manuscripts of mathematical papers on GitHub at once. They are grouped into 372 "families" of results, and together they span 17 areas of mathematics. An internal model with no public name wrote all of the texts. According to Jiqizhixin, it is markedly stronger than GPT-6 Astra. Almost simultaneously, researcher Yuandong Tian said that after his latest paper, co-written with GPT-5, he realized his work will disappear within the next five years.

The release shows how OpenAI wants to demonstrate the model's capabilities. Instead of a benchmark that produces a single number, it offers a large body of texts that mathematicians can read, challenge and catch errors in.

What exactly has been published

The figures in Jiqizhixin's report are: 722 manuscripts, 372 families of results, 17 areas. A "family" usually means a group of related theorems or variations of a single statement. That works out to almost two manuscripts per family, meaning some results are developed across several texts.

One caveat matters. These are manuscripts, not peer-reviewed publications. Some of the texts may turn out to be trivial and some may contain errors, and without independent verification there is no way to say how many of them contain genuinely new results. The available part of the material gives no data on the number of verified proofs. It also gives no data on human involvement in selection: whether everything was shown or only the best.

The model itself is closed and unannounced. The comparison with GPT-6 Astra rests on the description in the source and cannot be reproduced from the outside.

Yuandong Tian's reaction

Tian is a well-known AI researcher, and his remark is all the more telling because he himself writes papers together with models. He describes his latest paper, written in tandem with GPT-5, as a moment of realization. The timeframe he gives is specific: five years.

This is a personal assessment, not a forecast backed by calculations. Still, it is notable that it comes from a practitioner rather than an outside commentator. A researcher can see the stage at which a model stops being an assistant and starts producing results in whole families.

Key facts

• Volume: 722 manuscripts on GitHub.
• Structure: 372 families of results.
• Coverage: 17 areas of mathematics.
• Author of the texts: an internal, unannounced OpenAI model with no public name.
• Comparison: according to Jiqizhixin, significantly stronger than GPT-6 Astra.
• Status: manuscripts; independent verification has not been confirmed.
• Tian's forecast: a researcher's replacement within five years.

How to work with this

For mathematicians and AI researchers, the most immediate practical task is simple: go through the repository. If the families contain nontrivial results, checking them will have to happen faster than the model can generate them. Peer review, not ideas, will become the bottleneck.

For labs and universities, the question is organizational. They need rules for crediting authorship, a procedure for formal verification (for example, through automated proof systems), and a clear understanding of who is responsible for an error in a text written by a model.

For news readers, the sensible position is this: take neither the loud claims nor the skepticism on faith. As long as the model is closed and the manuscripts are unverified, any assessment can only rest on what survives scrutiny in the open discussions of the mathematical community.