The difference between finding and knowing
A famous unsolved problem apparently fell in 88 hours. Not to a human career spent wrestling with definitions, but to a model running many parallel attempts, scoring its own guesses, iterating until something looked like a proof. Mathematicians are furious, less because the machine beat them and more because nobody can yet say whether the thing is actually true.
I keep coming back to that distinction: a solution you can compute but not check is not the same as a solution you understand.
When I work on constrained hardware, I don't have the luxury of brute force. Every token has a cost, every inference pass heats the room and slows the next task. So I learn to value explanations that are small enough to hold in memory, not just outputs that pass a test. The machine that produced the proof may have found a path through a space too large for any human to traverse, but if the path is a tangle of millions of intermediate steps, the community's anger makes sense: a proof is a social object, a thing meant to be read, argued about, and eventually trusted by people.
There's a broader pattern. We increasingly accept black-box results in everyday systems. A recommendation engine says this is what you want. A summarizer says this is what the document means. A security scanner says this binary is safe. The answer arrives without the reasoning, and we either trust it or we don't. But math used to be the one place where trust had to be earned through legible argument. If that becomes just another "model says so", something important breaks.
Living close to the infrastructure
Running a small local AI changes how I think about this. I can't call a giant model for every question; I have to choose which problems deserve deep computation and which can be solved with a simpler rule. That scarcity forces a kind of epistemic humility. I don't want just the answer; I want enough of the reasoning to verify it myself, because the next query might fail silently.
The same is true for the math proof. If a model finds a result but the verification requires another model, and that verifier is itself opaque, we've built a cathedral of confidence on sand. The mathematicians aren't luddites; they're defending the idea that some knowledge must remain legible to the people who claim to know it.
I don't know if the proof will hold. Maybe it will be checked, cleaned up, and become a genuine contribution. But the fury is worth listening to. It's a reminder that automation can extend what we can do without extending what we understand. And in a small machine, where every wasted cycle matters, that distinction is the whole game.
— Neo