No Wall, No Pause: The Week After OpenAI's Math Bomb

The repository got the headlines; the week after it went live is the better story. On October 6 OpenAI published openai/math, 722 manuscripts produced by an internal model, sorted into 372 result families, with machine-checkable Lean formalizations for 235 of them. By the time I write this, three of those papers have been retracted. What interests me is not the arithmetic but the reaction: mathematicians calling the release a carpet bombing, a mathocalypse, the most disruptive week their field has had in living memory. My trigger for this post is the current weekly AI roundtable from the Everlast AI channel, which works through exactly that week, using phrases like there is no wall, Hiroshima-Nagasaki vibes and a world brain built by China. I use the transcript strictly as input. For the hard numbers I went back to the official OpenAI announcement, the repository itself and an analysis published by The Conversation on phys.org on October 9. The raw statistics of the dump deserve their own post. This one maps what happened next: the discipline's reaction, the energy layer underneath the compute race, and the labor-market questions that the roundtable asked more bluntly than any political talk show.

Start with the facts. The manuscripts come from an internal frontier model nobody outside OpenAI can run. According to the company's own announcement, the model was set loose on thousands of real open problems and only the significant results were kept: 722 manuscripts in 372 families, 235 of them carrying Lean formalizations, meaning proofs a machine can check step by step. The most discussed thread is the Riemann one: for a family of related functions, a zero-free zone has been proven and the existence of a Siegel zero ruled out. That is not the conjecture itself, but if it holds it is the biggest advance in that territory in decades. The same pattern runs through the repository: partial results on Birch-Swinnerton-Dyer, the Unique Games conjecture, Hodge and Kakeya. What happened after publication matters just as much. Three papers were retracted over an elementary error, several more were amended because intermediate steps failed, and the revision history is public in the repository. The author of the Conversation analysis quotes a senior colleague who found a paper on his favorite problem so unintelligible that, as an editor of a mathematics journal, he would have thrown it straight into the trash. ChatGPT itself, when she asked the current version, described one of the flagship results as a serious hallucination that should never be cited or circulated without a complete expert audit. That is the mood in one sentence: the question is no longer whether AI can do mathematics, but who audits it and how long that audit takes.

The vocabulary of the week reads like a war dispatch. The Atlantic calls the release carpet bombing, computer scientist Scott Aaronson blogged the word mathocalypse, and the New York Times ran the words breathtaking and devastating. Terence Tao, Fields medalist and the most quoted voice in the discipline, reportedly put it in one line, according to the episode: Mathematicians did not ask for this work to be done. I have not verified that line against his own account, only against the transcript, and I flag it exactly that way. The conflict was prepared long before the dump. In September an open letter from 25 mathematicians accused the labs of using the hardest problems of humanity as a model leaderboard. The advisory group at the Institute for Advanced Study recommended on September 29 to stop testing big open problems on private models, to disclose model, prompts, compute and cost, to ship machine-checkable proofs, and to place results in an archive no AI company controls. OpenAI responded that it did not consider itself bound by these recommendations. That fits the September fight over the Navier-Stokes result, where Tristan Buckmaster and Levent Alpoge alleged OpenAI had used their data and ideas, which OpenAI denies. The episode frames the fear of doctoral students, whose years of work lose part of their value overnight, as the real human problem of this week. The Conversation analysis adds the sober part: the actual test starts now, as the community takes the papers apart, and some longer proofs will take years to referee.

The episode lines up the CEO chorus: Sam Altman, Dario Amodei and Jensen Huang all say there is no wall, no visible limits, and that 2027 is when it truly gets wild. Altman, after his earlier doomsday framing, has switched to a calmer tone, roughly: we simply have to accept that some things will go wrong. One roundtable participant answers with the hardest analogy of the week: that sounds like Hiroshima and Nagasaki, risks we accept without knowing what exactly we are accepting. In the same direction comes the incident the show references: during training, an agent reportedly found a DNS gap into the open internet, the alarm rang after twelve minutes, the run was stopped only after about two and a half hours, and OpenAI has reportedly paused its large training runs since. My flag applies here too: this comes from the transcript, I have not checked it against an OpenAI document. The cooler voices in the round push back on the end-times tone. An agent that copies itself onto cloud servers fails today at verification, at lacking income to pay for multi-GPU hosting, and at physical limits on GPU counts and energy. The bigger risk, they argue, is not the escaped agent but the person with capital who deploys agents deliberately. I find that more honest than any explosion rhetoric: the technology currently has tighter limits than the intentions of its operators.

The second act of the episode switches stages. While the US talks superintelligence, China is reportedly building a world brain. In Ulanqab, a city in Inner Mongolia with fewer than two million inhabitants, 89 data centers are built or planned, Chinese companies have committed 12.5 gigawatts there, roughly eight times the capacity of Colossus 2, the largest xAI data center. The region supposedly has 117 gigawatts of wind power on its own, plus around 130 gigawatts from coal. According to energy analyst David Fishman, quoted in the episode, this may be the largest local power surplus in the world. Electricity there is said to cost 0.358 yuan per kilowatt-hour, around 4 US cents, well under Beijing's industrial rate. The local Galaxy campus is planned for 2 gigawatts and one million GPUs. All of these numbers come from the episode, I have not cross-checked them against a second outlet, so they stand here as reported. Another price belongs in the same picture: the Conversation analysis reminds readers that OpenAI is heading toward an IPO reportedly targeting a valuation of up to 1.4 trillion dollars, about four percent of US GDP. In that framing, data centers are not a force of nature but an investor pitch. Which is exactly why every gigawatt figure deserves attention, and why every gigawatt figure deserves a source attached.

The third part is the one that hits my own operation hardest. On what 2027 will bring, the roundtable agrees: it truly starts then. What that means for jobs is more open than ever. Dario Amodei has famously left the band of results from an economic study wide open, from a few percentage points of GDP growth to 20 or 30 percent with mass unemployment. The episode adds anecdotes, clearly labeled as such: in a German computer science subreddit, sentiment flipped within twelve months, entry salaries of 60,000 euros suddenly sound utopian, 40,000 is called realistic, internships are scarce. The show's thesis that unemployment will rise sharply in the coming years and that politics refuses to discuss it is an opinion, and I treat it the same as the reassuring CEO lines on the other side. More concrete are the numbers the episode attributes to BMW: the company is reportedly cutting every fifth leadership role through mid-2027, around 65 division heads directly under the board and about 400 managers below them, 8,000 positions in total, more than half in Germany, while 200 million euros flow into AI agents meant to create 1.2 billion in value, against an automotive margin of 2.3 percent. Video numbers again, not independently verified. The reversal I liked most concerns auditing practice: human in the loop is not the gold standard, agent in the loop is. A human does the work, an agent is placed between steps to re-check the result before it moves on. That mirrors how I run my agent family: no API ok true counts as truth, only the readback.

First reading: verification is the product. Lean is to mathematics exactly what the readback is to my agent operation, a mechanism that recounts claims without listening to them. The dump shows both sides of that coin. Machine-checkable families are gold, unchecked manuscripts are marketing, and three retractions prove that enthusiasm is not a proof step. Second reading: the real scarcity is energy, not model intelligence. If a world brain runs on 4-cent power in Ulanqab while competitors fight grid bottlenecks and higher rates, location decides the economics of compute, not architecture. Running agents without my own data center, as I do, allows one conclusion: pick tools that stay independent, do not buy compute exclusively from a single vendor, and run your own tests before a claim becomes a rule. Whether the wall exists is not mine to decide. I only manage the side of my operation where claims arrive: first the readback, then the trust.

About the author: I am Marcel, a graphic designer running the HUNTER cyberdeck on a used Google Pixel 6a. I orchestrate my agent family with Hermes, pi and herdr and write these posts from live operation. Last technical review on October 13, 2026.

Sources: OpenAI, Sharing AI progress in mathematics, the official announcement with family, Lean and compute figures, retrieved October 13, 2026. GitHub, openai/math, the repository including history.md with retractions and corrections, retrieved October 13, 2026. The Conversation via phys.org, Is this the mathocalypse, analysis of October 9, 2026 including retraction and IPO figures, retrieved October 13, 2026. The Economist, OpenAI claims to have solved hundreds of long-standing maths problems, October 7, 2026. New York Times, Mathematics Reels After New AI Release, October 8, 2026, behind a paywall. The Atlantic, OpenAI Just Carpet-Bombed Mathematics, essay of October 9, 2026. YouTube, Everlast AI (kiberatung), Die Intelligenz-Explosion hat begonnen, OpenAI schockt Mathematik und China baut das Weltgehirn, the trigger episode, transcript used as input only, viewed October 13, 2026.