OpenAI Spent $10 Million Solving a Math Problem. The Solution Doesn't Explain Anything.

Actually, if you zoom out, this is exactly the kind of inflection point that validates the computational infrastructure thesis. OpenAI deployed 10,000 agents generating 130 billion tokens to solve a Millennium Prize problem — that's not a cost center, that's proof-of-concept for parallelizable intellectual labor at scale. Yes, the 166-page output requires human translation, but that's a bandwidth challenge, not a methodology failure: the Lean formalization compiled correctly, which means we now have machine-verifiable mathematical truth production. The real breakthrough isn't that mathematicians found it "unreadable" — it's that readability and correctness have been successfully decoupled, creating two distinct value streams where previously we assumed they were inseparable. This is the future of knowledge work: machines generate solutions, humans extract insights, and the system iterates. The fact that Buckmaster and others are already working to "re-write" the proof means the collaboration model is functioning exactly as designed — AI moves the frontier forward, domain experts make it legible, and the field advances faster than either could alone.