A mathematician named Tristan Buckmaster spent a year working on the Navier-Stokes problem—one of the seven unsolved Millennium Prize Problems—using a coding assistant to develop and test his ideas. He and a collaborator proved a related result in the 3D incompressible Euler equations on August 22, verified in the Lean proof assistant. They held off publishing to write a readable version. Then on a Sunday, OpenAI called to say they'd solved the actual Navier-Stokes problem using the same approach Buckmaster had been exploring. OpenAI claims they started after hearing rumors and that no one accessed Buckmaster's session data. Buckmaster asked whether his work had been used in training or accessed by the model, and says he got no clear answer on the training question. Nobody has proven his data was stolen, but the timing and the fact that both teams independently converged on the same rare strategy—smooth forcing with specific boundary conditions—created an uncomfortable situation. The real problem here has nothing to do with whether Buckmaster's name appears in datasets. De-identification works fine when you're one of millions asking generic questions, but it collapses completely when your work is genuinely rare. Buckmaster estimates maybe four people on Earth would generate that exact sequence of prompts and mathematical approaches. Strip away the names and emails and you're left with a trail of unique ideas, failed estimates, and a specific chain of prior results that amounts to a fingerprint. The rarer and more valuable your idea, the more the idea itself becomes identifying information. A single rumor containing the word "forced" was enough to point ten thousand AI agents at the right problem. De-identification is fundamentally inadequate when dealing with sparse, high-value data. But this same mechanism is also what makes the collaboration powerful. LLMs compress billions of human interactions across arXiv preprints, seminar notes, and Stack Exchange threads into something that can hold thousands of half-finished ideas in one place simultaneously. Buckmaster didn't just use a tool—he collaborated with a compressed record of everyone who ever wrote about fluid dynamics, including researchers from Madrid he'd built on, plus countless others whose work the model absorbed and recombined. He called it a Deep Blue moment, but it's more like having every mathematician in history show up to your office at 3 a.m., disorganized and sometimes wrong but occasionally containing exactly the sentence you needed. The capability to tease apart rare, valuable threads from the noise is inseparable from the risk that those same threads become visible to everyone else in the system.