It wasn’t complex new math. Not really. It was stubbornness.
OpenAI dropped the solution to the cycle double cover conjecture right before letting loose its newest model, GPT-5.6 Sol. The timing? Pure marketing, of course. The impact? A quiet revolution in how we view artificial intelligence and what it can do. Humans have been chewing on this problem for more than half a century. GPT-5.6 solved it by being told one specific thing: do not give up.
This is part of a larger trend. Tech giants are dumping money into pure mathematics. Not because they care about beauty for beauty’s sake. But because math is the ultimate benchmark. It proves reasoning. Or at least, the machine version of it.
“AI tools will change mathematical research significantly.”
— Noga Alon, Princeton mathematician, on the Sol breakthrough. Alon called the proof surprisingly short. Which makes it funnier, if you ask me.
Here is what you need to know. Graph theory sounds abstract, boring even. A graph is just dots (vertices) and lines (edges) connecting them. The Internet? It’s a graph. Your social network? Also a graph.
Back in the 1970s, mathematicians guessed something interesting about these shapes. They thought almost every graph has a cycle double cover. What’s that? A set of loops that covers the entire structure exactly twice. Every edge sits inside precisely two loops.
Easy enough to picture.
Proving it? That’s where the decades went.
Great minds tried. They cracked it for specific cases. They got close. But the general proof? It slipped away. Every time.
Last Friday, the AI stepped in.
The solution turned out to be elegant. Simple, almost. The AI showed you can cover the graph with no more than eight loops. Sure, there are technicalities. Graphs held together by single thin threads (cut edges) don’t count. But for the rest? Done.
Here’s the twist.
The proof didn’t use flashy new ideas. No breakthrough theories born in silicon dreams. It recycled methods humans had already tried. Methods we threw in the trash or left on a shelf because we got bored. Or scared.
Did you ever consider that “hard” might just mean “unpopular”?
Andrew Sutherland at MIT thinks so. He suggests a reputation for difficulty can be a trap. Students steer clear. Experts move on. It becomes a self-fulfilling prophecy of obscurity. When an LLM doesn’t have ego or a career to worry about, it just digs.
“We will keep seeing supposedly ‘hard’ problems having ‘easy’ solutions.”
Sutherland isn’t guessing blindly. OpenAI released the prompt used to crack this nut. It reveals the ugly, mechanical work behind the magic. The prompt wasn’t poetry. It was scaffolding. Instructions for sixty-four agents to talk to each other in parallel. Cross-checking. Mitigating the lies and hallucinations that plague these models.
The real trick? The directive.
They didn’t just ask for an answer. They told the bot: Spend at least 8 hours. Don’t think about quitting.
Most of us would quit by hour two. Or we’d say, “It’s impossible.”
The machine stayed at the keyboard. Eight hours.
It turns out genius might just be endurance we’re too tired to give. And now that AI is here, maybe the library of open problems isn’t a fortress. It’s a garden where we forgot to water the flowers.
