A Friendlier Equilibrium for Sharing Research?

Yesterday, OpenAI released GPT-6.2 Astra, which appears to be the model that produced the several hundred breakthrough math results released on October 6th.

Undoubtedly, many researchers with access to the model have now fed it their favorite open problems and have either been supremely underwhelmed or gotten results they wouldn’t have dreamed of a few years ago. (That there may not be much middle ground there is bewildering to say the least.) What should we do with these? Should we spend our Thanksgiving rushing to understand these results and write papers that we can slap our names on? Should we refresh our social media and arXiv feeds constantly in fear that someone else pushed the button faster, or cut more corners? All of this sounds awfully antisocial (in many ways) and disincentivizes the joy of human understanding, collaboration, and democratic nature of the mathematical sciences that drew many of us in.

I want to propose an alternate way we can engage in research, which I hope could bring those incentives back in line. If researchers get results that require their own insights and advantages (including proposing new questions), they should by all means feel empowered to take full ownership of them and share them in the traditional way. But for results that are within the models’ grasp without much human input, here’s one thing we could do instead.

  • We can start by posting our AI-generated results on a preprint server without human authorship, like Hexagon, as “raw data.” Before doing so, the following due diligence is suggested:
    • Give the model sufficient time in a loop of improvement and simplification until it settles on proofs that are Pareto optimal in these respects.
    • Trace where key ideas could have come from in the literature. (To that end, this proposal draws on many online discussions in the math/TCS community, including from the ETH effort, Joaquin Lema and others, Grant Sanderson, Daniel Spielman, Daniel Litt, Timothy Gowers, and the oai-math-revised project.)
    • Share a formalization of the results if possible (e.g., in Lean), so that we can be fairly confident they are correct.
  • Then, the contributors organize a group to produce a high-quality exposition of the result, with an open invitation to anyone interested in learning it. For example, the raw data can include a link to a Zulip channel or to MathHelm for discussions. The group iterates on an exposition and reaches consensus on when it’s ready to upload to the arXiv. They may also work towards improving the results and can update the preprint and formalizations accordingly.
  • Not every participant in the discussions needs to be an author, including the original contributors. Perhaps all authors should commit to giving a seminar talk on the paper at their local institution shortly after the arXiv paper is submitted. (Thus, all authors must have a solid understanding of the paper.)

For a major result, I think it would be exciting to see papers on par with the most well-written textbooks, with dozens of coauthors spreading the gospel across the world. I also imagine that most of these authors would be junior people who accumulate a reputation through the quality of the presentations they give and the papers they are most involved in.

Of course, the devil is in the details. How should independent groups with different objectives go about merging? What about the logistics of large-scale collaboration and writing a paper, and is giving a talk enough of a price of admission for authorship? What happens when AI can write the textbooks too? Though imperfect, I feel that this is a reasonable starting point, and with experimentation and iteration, we can find ways to incentivize impactful research that is joyful to participate in and ultimately reap the benefits of this technology. In that spirit, I hope to be working with you on some of my favorite problems (and yours) very soon.

Thanks to Moses, Jabari, Ariel, Kangning, and Qilin for thoughtful feedback on this post!

2 Comments on A Friendlier Equilibrium for Sharing Research?

  1. Thanks for the thoughtful post. A lot of fields list the specific contributions each author makes to a paper. If there is someone who generated a prompt and interacted with the AI to prove a result, and someone else probed it for weakness, and a third group came up with a nice exposition, can you imagine a detailed report that lists all the specific contributions each person made, in a way that individuals can cite? It is hard for me to advise a young scholar to spend significant time on an academic effort where clear credit is not forthcoming.

    Like

  2. Unknown's avatar Prasanna Ramakrishnan // November 19, 2026 at 9:19 am // Reply

    Thanks Ashish, I think that’s a great idea! Another related suggestion that came up in a conversation this morning was forgoing our convention of alphabetical author ordering in these larger collaborations, like is done in the sciences. I think I would want to spend more time learning about how these dynamics play out in those fields so that we can figure out how we can adopt those conventions without inheriting anything toxic. Ultimately, making sure that junior researchers can get appropriate credit and be incentivized to do the work they are excited about should be at top of mind in trying these experiments, and I appreciate that you are thinking about that too.

    Like

Leave a Reply