OpenAI's Math Dump Just Exposed the AI Bubble's Fatal Flaw

A 26-minute video essay, uploaded on 9 October 2026, built around one idea: AI now acts at machine speed, while checking, understanding, paying for and answering for what it does still move at human speed. Everyone, the narrator says, is asking for receipts.

First the maths. On 6 October OpenAI released several hundred mathematical results on GitHub, grouped into 372 families and drawn from about 4,000 problems attempted, which it says nearly all came from a single prompt to a single agent. A month earlier it had claimed Navier–Stokes with a swarm of some 10,000 agents and millions in compute, so the narrator finds the one-prompt claim a little suspicious. Many results come with Lean proofs a machine can verify, though only a minority of them; understanding is another matter. The narrator, a scientist, sides with releasing it all, as the Toronto mathematician Daniel Litt does, while granting the cost. The Erdős problems site has frozen comments and removed its solved and open labels, swamped by AI flag-planting, and an Association for Human Mathematics called the release a show of power. Aspirin worked for seventy years before John Vane explained why, and the explanation is what led to new uses. That is why the advisory group OpenAI convened asked for each result’s prompt and compute. OpenAI published average compute, no prompts, and said it isn’t bound. One supposed Riemann result is a zero-free region for L-functions, not the hypothesis itself.

Then the money: a $60 billion debt package for Anthropic to lease Google-designed chips, its $42 billion senior slice backed by Broadcom’s promise that the chips will hold value, its $18 billion junior slice resting on Anthropic alone; SpaceX seeking $40 billion for Nvidia chips while losing money; and the FT’s report that OpenAI’s revenue run-rate is near $50 billion, not $70 billion. Lenders asking to see the books, the narrator says, is the market working. Then a crime: from late September at least nine South Korean banks were hit by one person using ARTEX, an open-source harness that plugs commercial models into the steps of an attack. CrowdStrike traced him through his AI sessions, where he had asked Claude where to sell the stolen data and to write his CV. The closing proposal borrows from anti-money-laundering rules: model companies detect and report misuse, agents carry an identity tied to the person who sent them (building on Web Bot Auth), and institutions treat unsigned traffic the way email treats spam.