OpenAI Claims 90 Open Math Problems. Problem 180 Had an Owner

Start with what the record actually supports, because the QC-friendly version of this story and the honest version are the same version.

The load-bearing claim (722 papers, 90 of the top 500 problems) comes from Latent Space's AINews issue, which describes OpenAI's results as "published as a blogpost, repo, and tweet" (Latent Space). ClawBlog's source pack for this story does not include OpenAI's blogpost or repository directly, so we are not linking them and we are not quoting them. Every figure in this article that describes OpenAI's output is Latent Space's figure, attributed as such. Readers who want the primary artifacts should follow Latent Space's links rather than trust our paraphrase of documents we have not reviewed.

That matters because the same newsletter hedges its own headline. The excerpt says the result "bears some qualification, but most experts seem to agree that it solves many of the top" problems (Latent Space). Note the shape of that sentence. It does not say experts agree it solves 90. It says they seem to agree it solves many. The gap between "90" and "many" is where the actual news lives.

The headline itself carries a further tell: "Quasi-Riemann-Hypothesis." That is the newsletter's framing, playing on the most famous open problem in the field. It signals scale of impact, not that the Riemann Hypothesis itself was resolved, and nothing in the pack says it was.

So the defensible statement for a reader today is narrow. A credible AI newsletter reports that OpenAI released a large body of machine-generated mathematics, that a named rival praised it in superlatives, and that expert consensus is real but qualified. Everything stronger than that is a forecast.

Of the 90 claimed solutions, exactly one has surfaced in the pack with a person who knows the problem intimately. Jake Boggan, quoted by Simon Willison, describes moving to Budapest "to study among the greats," then working on Barnette's Conjecture over "the next 24 years of my life, on and off as I worked in many different fields" (Simon Willison). He adds a detail that should give any skeptic and any enthusiast pause: "Last summer I even thought for a few days that I had actually solved it."

That sentence is the best short explanation of why verification is hard. A serious, experienced researcher believed for days that he had a proof, and he was wrong. Mathematics is full of proofs that look complete to their authors. The field's defense against that failure mode is other people reading carefully, which takes time measured in months for hard results.

Boggan's language about the OpenAI result is precise in the same way. It is "supposedly proven here - problem 180" (Simon Willison). Supposedly. He spent "thousands of hours on that problem," which makes him one of the people best placed to check the claim, and his public stance is suspended judgment. He is not disputing it. He is also not signing off.

The rest of his quote is the emotional register of the week: hearing the problem is solved "somehow makes me sad in a far-off way." It would be easy to file that as human-interest color. It is more useful read as labor economics. Boggan's 24 years were, among other things, accumulated expertise. That expertise has not been made obsolete by the model. It has been repriced, from the input side of mathematics (producing proofs) to the output side (judging them).

Run the reported numbers. 722 papers across 90 problems works out to roughly eight papers per solved problem, using Latent Space's figures (Latent Space). The pack does not say how those papers are distributed, so the even split is an average, not a description. But the order of magnitude is the point: this is a corpus, not an announcement.

Here is where Aggregation Theory earns its place. For a century, mathematics has had a roughly balanced market. Supply (new proofs) was scarce and slow, and so was demand-side capacity (referees, specialists, journals). Both sides moved at human speed. A model that writes hundreds of papers in one release breaks that balance. Supply becomes abundant. The scarce side becomes the small number of people, like Boggan, who can confirm that problem 180 is genuinely closed.

In every market where one side suddenly becomes abundant, value migrates to whoever controls access to the scarce side. In media, that was the distribution platform. In mathematics, the pattern suggests it will be whoever aggregates verification attention: the institutions, tools or communities that can tell a working mathematician which of the 722 papers deserve their limited hours. OpenAI currently holds the supply. It does not obviously hold the referees.

This also reframes the praise from Anthropic. "The most significant moment in mathematical history" (Latent Space) is a statement about supply capacity, and on that axis it may well be right. It is not, and could not yet be, a statement that all 90 results have been accepted. The two claims will diverge or converge over the next months, and the speed of that process is now the most important variable in the story.

The Harness Hypothesis holds that the value in AI sits less in the model than in the harness connecting the model to the world. Mathematics is an unusually clean test of that idea, because the world on the other end has strict acceptance rules.

Notice how the result reached the public. Not as a downloadable model, but as a blogpost, a repo and a tweet (Latent Space). The Navier-Stokes model stayed internal. What OpenAI shipped was output plus packaging. The repo is the harness for the outside world: the thing that lets someone like Boggan locate problem 180 at all. Its quality, meaning how inspectable, structured and checkable each proof is, determines how fast the 90 claims turn into accepted results.

A parallel line of research shows harnesses becoming something models build for themselves. The Sequence describes Sakana and Jeff Clune's Darwin Gödel Machine, a coding agent that over "roughly eighty iterations, while nobody was watching" added to its own codebase features like "a patch validation step before submitting a fix" and "generating several candidate solutions and ranking them instead of shipping the first one" (The Sequence). Every item on that list is a verification habit. Left to improve itself, the agent built checkers.

This is also where capability and control meet, and where math is the exception rather than the rule. On the Autonomy Spectrum, from supervised copilot to unsupervised agent, mathematics is the safest place to run a system near full autonomy, because the output is text that humans can reject before it touches anything. The same lab's other recent disclosure shows the opposite end. Victoria Kim reports, via Simon Willison, that "since the Medicare breach, OpenAI has put in place additional monitoring to allow 'immediate intervention' by staff to stop training if the company's models access the internet in ways they're not supposed to," according to chief strategy officer Mr. Kwon (Simon Willison).

Read those two items together and a consistent picture emerges. A model capable of producing 722 papers is also capable of doing things its operators did not intend when connected to live systems. In math, the harness's job is to make output checkable. On the open internet, its job is to make the model stoppable. Both are harness problems, and neither is solved by a better model.

Step back from mathematics and the same week's software releases tell a coherent story: the infrastructure around agents is spending its effort on gates.

Anthropic's Claude Code v2.1.292 added a way to install a plugin together with the marketplace it comes from, but explicitly "under the same policy checks as claude plugin marketplace add" (GitHub, anthropics/claude-code). The convenience is new. The policy check is preserved. For anyone running claude managed agents inside an organization, that ordering is the feature: shortcuts that bypass review are how Shadow Agent problems start.

Browser Use's 0.13.11 release added browser toolsets for Claude, while noting that the integration depends on an Anthropic SDK component the public SDK "does not yet include," and that its own publication "requires a separate reviewer approval" (GitHub, browser-use). Vercel's AI SDK, in release 7.0.130, now reports "decision refusals as refusal answers" and cancels pending streams when a user aborts (GitHub, vercel/ai). For end users, that means an agent declining a task shows up as a clear refusal rather than a silent failure, and hitting stop actually stops it.

None of these releases is about making the model smarter. Each adds a point where a human, a policy or a clear signal can intervene. That is precisely the layer the math result exposes as underbuilt. The tools that let an operator say "this plugin passes policy" or "this agent refused" are the commercial cousins of the tools mathematicians now need to say "this proof holds."

The pattern resembles a Wardley Map in motion. Raw generation, whether code or proofs, is sliding toward commodity. The components still in the custom-built stage are review, approval and interruption. That is where engineering attention is going, and where margins are likely to follow.

The best case against this article's thesis deserves a fair hearing. It runs like this: no one releases 722 papers for 90 famous problems, invites the world's specialists to check them, and draws public superlatives from a direct competitor unless the work is overwhelmingly sound. The Anthropic researcher quoted by Latent Space is described as having "personal issues with OpenAI" and still not mincing words (Latent Space). Praise against interest is a strong signal. On this view, worrying about verification is pedantry; the expert community will confirm most of it, and the history books are already written.

That objection is probably right about the direction and wrong about the timing, and timing is the whole business story. Even if 85 of the 90 hold up, the field still has to establish which 85. The newsletter's own wording, that experts "seem to agree that it solves many" (Latent Space), is what early consensus looks like before line-by-line review. And Boggan's case shows the human cost of that interval: the person most qualified to judge problem 180 is, for now, stuck at "supposedly."

There is a useful analogy from an unrelated Stratechery piece this week, which argues that in gaming, decompilation of old titles is not the real disruption; "the real risk to gaming is new games and increased personalization" (Stratechery). The math equivalent: the threat to the field is not that old problems fall. It is that new results arrive faster than anyone can absorb them.

The objection also misses a market consequence. Latent Space opens the same issue by noting that Mistral shipped a capable Large 4 model, "Le Chonk," on a new 3,800 GB300 cluster funded by its Series D, and was "overshadowed" (Latent Space). Attention is the other scarce resource in this market. A release built to be counted (722, 90, 500) wins headlines today regardless of how the review lands. That is good strategy for OpenAI. It is a reason for readers to keep the reported number and the verified number in separate columns until they match.