Researchers published OpenAI's unique games result before the proof

Summary

When rumors spread that OpenAI had proven the Unique Games Conjecture, Dor Minzer from MIT and two graduate students rushed to publish a 95-page landmark result on a closely related question. Three days later, OpenAI announced the proof of the conjecture, accompanied by 376 mathematical results.

Why it matters

The Unique Games Conjecture sits at the heart of computational complexity theory. Proving it would firmly establish the limits of approximation algorithms across a broad family of problems. This event is also a striking turning point: the human team had to outpace AI in a publishing race. Moreover, the AI’s output was not a single result but a deluge of 376 proofs. This raises fundamental questions about how mathematics and peer review culture will adapt.

Highlights

• Minzer's team described their hastily prepared 95-page paper as 'mathematically complete, but not in the form we wished to share.' They prioritized speed over polish to preempt OpenAI's announcement.
• OpenAI's October 6 publication included the proof of the Unique Games Conjecture and 376 additional results across various fields. The article notes that this development was met with a 'mix of anxiety and curiosity' in the community.
• Ryan O'Donnell from Carnegie Mellon praised Minzer's work as 'another truly major achievement,' emphasizing that the team 'solved the problem the old-fashioned way, with their minds, and wrote it with their own hands.'

Critical take

The article is quite weak in terms of actual mathematical content. It never clarifies exactly what Minzer's result proves in relation to the full conjecture, merely stating that it addresses a 'closely related question.' Readers are left to infer the significance from peer praise rather than the content itself. OpenAI's '376 other results' are dismissed in a single sentence. There is no examination of whether these AI-generated proofs have been independently verified. Despite this, the text suggests that the human-AI race was decided solely by the act of publishing first.