On Tuesday, OpenAI dropped 722 math papers on the world. At once. On GitHub, of all places, not a journal. They came from an internal model nobody outside the company can use, they fall into 372 groups of results, and the company says it threw about 4,000 problems at the thing to get there. According to The Next Web, nearly every paper came from a single prompt to a single AI agent.
722 papers. Think about that for a second. Not the writing of them. The reading of them.
Because that’s the part nobody can automate yet. Mathematicians are now going through the pile by hand, and some of them, per TechSpot, rushed their own work out the door earlier in the week so they wouldn’t get scooped by a machine. Been there, kind of. Not with math. But anybody who’s watched a flood of AI books land on Amazon next to their own knows the feeling.
The panel of mathematicians that advised OpenAI on how to release this stuff (it’s hosted by the Institute for Advanced Study) put out a statement that night saying their advice wasn’t an endorsement. And then this line, which I keep chewing on: “This release is the beginning, not the completion, of the process of human understanding.”
Yeah. That’s it right there. A paper isn’t done when it’s written. It’s done when somebody has actually read it, checked it, and is willing to put their name next to what it says.
Now look at arXiv, the big open preprint server where physicists and mathematicians post their work. As of October 1st, every submitter is capped at two submissions a month. Two. Why? Because in September arXiv got 40,363 submissions. Two years ago it was about 20,000. Ten years ago, under 10,000. That flood kicked up almost 9,000 support tickets, and the people sorting it are mostly volunteers. Their moderators are describing “thin” papers, “salami” papers (one study sliced into five), and a marked increase in dense, AI-written stuff.
So here’s where we are. Writing is basically free now. Reading is not. Attention is not. Judgment is not. And when the thing that’s scarce is the reader, you end up rationing the writers. Surprise.
That’s the big world. Here’s the little one.
Circle is being built around exactly this problem, though honestly we didn’t have a 722-paper day in mind when we started. In Circle, you don’t get your work reviewed by showing up with a manuscript and a credit card. You get it reviewed by reviewing. You request an assignment, you read somebody else’s chapter or watch their cut, you write a structured review, and that earns you credit toward having your own work read. The currency is reading. Real reading, by a person who’s on the hook for what they said.
arXiv has to limit how much you can submit because it can’t grow readers fast enough. Circle tries to grow the readers first. Every member who wants to be read becomes a reader. It’s slower. It’s supposed to be. The doors are opening now and membership will be limited, not for money. There is no fee. But for size, and community. This is a place where you will work with people and get to know them. Share things. Make new friends and expand your craft together. That’s who we’re looking for. Drop in and leave your mark at gapoz.com ..
And AI? Use it if you want. Plenty of us do. What Circle asks is that you are honest about how you use AI. That’s all, really. Self-attested. No detector at the door. Because the question was never “did a machine touch this.” The question is whether a human stands behind it, and whether another human actually read it.
722 papers and the hard part hasn’t started. That’s not a math story. That’s every writer’s story now.
GAPOZ Press is the people’s publisher. Which these days may mean the people’s readers, too.
Also on the beat
- Storytel’s content chief says AI-assisted translation gets an eight-book series out in six months, and that the backlash is “a perception issue, not actually about quality.” (Publishers Weekly)
- Simon & Schuster staff push back on talks about AI workflow software they fear could track individual performance; CEO Greg Greeley says no rollout is planned. (Publishers Weekly)
- Jill Lepore tells Frankfurt that “our thoughts and our words and our books are our own” and urges publishers to declare independence from Big Tech. (Publishers Weekly)
- The Third Circuit rules that ROSS’s use of Westlaw headnotes to train a competing AI legal-search tool was not fair use, the first federal appellate fair-use ruling on AI training. (National Law Review)
- AI video startup Higgsfield launches a tool for building custom AI influencers, more synthetic faces headed for your feed. (Semafor)
How this was made
AI assisted research and a first draft of this Circle View. A human reviewed, edited, and is accountable for the judgment and the final take. Disclosure by role — not undetectable.
