LLMs know the answers to questions we were already aware of, and we can tell they're right, but we can't understand their working. That's what's new.
If a team of 100 top-flight number theorists working together had used much less sophisticated LLMs + Lean to produce this proof, we’d be just where we are today.
If that's not the standard for understanding something, what is?
I think of it like a spaghetti codebase with a giant densely connected graph of interactions. Today’s LLMs can easily “understand” and answer questions about it, but only by way of fitting all of the complexity into context at once.
The more useful form of understanding is being able to compress that giant graph down into a far smaller set of independent subgraphs. To factor it into simple, obvious, reusable building blocks, and use those abstractions to express the problem itself in a much simpler way.
That’s an instrumental definition, but it captures the idea that understanding has to have causal, conceptual linkages to a wider body of knowledge. Imagine a magical, infinite catalog of true vs. untrue statements. Every bit of knowledge is there, but we wouldn’t say it has understanding & it couldn’t teach us anything beyond its (impressive!) raw content.
This just stopped being true.
And even in mathematics, incomprehensible but valid proofs have been a thing since the four-color theorem decades ago.
It's just that no single human understands the whole. But we had proofs like that before.
Btw, see also the infamous proof of the 4 colour theorem, which faced similar hurdles.
This hypothesis was previously not proven at all. That's strictly worse.
Now it is understood by our systems.
Soon it will be by us. Our systems gave us a new and working salient.
We're building machines that help us understand faster.
There have been plenty of times in history we didn't completely understand something yet used partial knowledge to make advances. This isn't even that bad.