Bitcoin mining turned electricity into money. Knowledge mining turns frontier compute into insights — and the ore body only gets richer as the tools get better.

Ten years ago, mining meant running SHA-256 over blocks of transactions until you found a hash with enough leading zeros. You burned electricity, wore out your GPUs, and if you got lucky, you got some bitcoin. The output was economic value, extracted from nothing but computation and time.

Today, the same thing is happening, but instead of bitcoin, what’s being mined is knowledge.

Run a frontier model over a corpus of mathematical literature for a few hours, and it will find connections no human has seen. An OpenAI model solved an 80-year-old Erdős conjecture last May. In July, GPT-5.6 proved a 50-year-old problem in under an hour. The pattern is the same everywhere: take a massive pile of existing knowledge, apply scale, and pull out something new.

The mechanism is brute force, but not in the way people usually mean. It’s not trying every possible proof. It’s trying every possible combination of existing ideas. Most of these combinations are nonsense. But a few are insights that waited decades for someone—or something—to notice.

This is what makes knowledge mining different from bitcoin mining. You can re-mine the same corpus with a better model and get new results. The resource is renewable. The “ore” doesn’t get used up; it gets more valuable as the tools get better.

The Time-Money Tradeoff

There’s another layer to this, one that shows up in more mundane settings. If you’ve ever tried to translate a book with a cheap model, you know the drill: you build a harness. You inject a glossary. You add a judge to catch quality drift. You run the whole thing through three retry loops. It takes days. The output is… okay.

Then you try the same task with a frontier model. You write two sentences. Six hours later, you have a translation that’s cleaner than anything you could have produced with all that scaffolding. The difference isn’t just quality. It’s frustration. The cheap model costs less but demands more of your time and attention. The expensive one costs more but gives you back both.

This is the time-money tradeoff, and it’s everywhere. You can pay in dollars or you can pay in hours. Most people don’t have enough of either. The ones who do get to skip the scaffolding entirely.

The same dynamic plays out in mathematics. You can spend months learning the literature, building intuition, and slowly constructing a proof. Or you can throw a frontier model at the problem, let it comb through everything ever written on the topic, and wait six hours. The output might be right. It might be wrong. But it’ll be something you can work with.

This is what people mean when they say the answers are “floating in the air.” They’re not wrong. The knowledge is there, scattered across arXiv, textbooks, conference proceedings, half-forgotten papers from the 1970s. The problem has always been finding the right combination. AI doesn’t create new knowledge from nothing. It just searches the space of existing knowledge faster than any human could.

But this is where the Platonists get it half-right and half-wrong. Yes, the answers are “out there” in some sense. But they’re not waiting in a static realm of forms. They’re waiting in the combinatorial space of everything we’ve already written down. And that space is not infinite, but it is vast enough that brute force search was impossible until very recently.

The ones who can afford the brute force get the insights first. The ones who can’t keep building scaffolding.

A Strange New Inequality

This creates a strange new inequality. In the old world, you could beat a better-funded competitor by being smarter, more creative, more persistent. In the new world, the competitor with the bigger GPU cluster will find the insights first, even if they’re less clever. Intelligence is being replaced by scale.

This is not entirely new. Mathematics has always had an element of this. The people with more time, more access to libraries, more collaborators, more resources have always had an edge. But the gap was surmountable. A lone genius could still make a breakthrough. Now the gap is becoming unbridgeable.

The “lone genius” is being replaced by the “lone genius with a $10,000-per-month API subscription.”

And this is before we talk about what happens when the models get better at things that can’t be brute-forced. Right now, most of what we’re seeing is recombination. The models are good at finding patterns that already exist in the literature. They’re not yet good at genuine invention—the kind of insight that requires not just searching, but creating.

(Though it’s worth noting that many cognitive scientists would argue human invention is just advanced recombination. If that’s true, then the gap between human and machine creativity is smaller than we’d like to admit.)

But that will come. And when it does, the gap will widen even further.

The Bitter Harness

There is a third option, one that most people are ignoring. It’s the path of the engineer who accepts that the models are tools, not oracles.

You can’t beat the frontier models at raw search. But you can beat them at direction. You can build a harness that guides the model toward the right kinds of searches, that filters out the nonsense before it wastes your time, that verifies the output before you trust it. This is work. It’s not as glamorous as prompting a frontier model and waiting for magic. But it’s the work that will matter in the long run.

At first glance, this looks like a paradox: if scale always wins, how does scaffolding compete? The answer lies in where the battle is fought. Scale wins the frontier—the mega-labs will brute-force the 80-year-old conjectures first. But scaffolding wins the application. The clever engineer with a smaller budget can out-compete on niche, domain-specific problems by building superior verification harnesses. Sutton’s Bitter Lesson taught us that compute beats cleverness. The Bitter Harness teaches us that the right scaffolding beats raw compute for the problems you actually care about.

The Cost of Getting In

But there’s a cost to all of this. The infrastructure required to do knowledge mining is concentrated in very few hands. Bitcoin mining, for all its problems, was at least democratized—anyone with GPUs could participate. Knowledge mining requires access to frontier models, which in turn requires access to GPU clusters that cost more than most universities’ entire AI budgets. The people who can mine the deepest insights are the ones who can afford the most compute.

And then there’s the verification problem. When a model finds a proof in six hours that would take a human six months to verify, the bottleneck shifts from discovery to validation. Formal proof assistants like Lean, Coq, and Isabelle are no longer academic curiosities—they’re the only way to keep up. Without them, you’re left with claims that sound plausible but can’t be trusted.

The Verifier and the Curator

So what happens to the lone genius? What happens to the mathematician who can’t afford a $10,000-per-month API subscription but still has something to contribute?

The answer is not clear. But there is a path. It’s the path of the verifier, not the discoverer. The future of mathematics may not be about who can find the proofs first. It may be about who can verify the proofs that the machines find.

The mathematicians of the future will be the ones who can read machine-generated proofs, formalize them in Lean, and decide which ones are worth caring about. This is not glamorous work. It’s not the romantic image of the lone genius discovering deep truths. But it’s the work that will matter.

And there’s another role, one that’s even less glamorous but perhaps more important: the role of the curator. The machines can find insights, but they can’t decide which insights are interesting. That’s a human job. The mathematicians of the future will be the ones who decide which problems are worth solving, which proofs are worth mining, which insights are worth pursuing.

What We’re Really Mining

This is the strange new world we’re entering: one where knowledge is no longer something you build slowly, through insight and collaboration, but something you extract, through scale and search.

The mathematicians of the future won’t be the ones who find the proofs. They’ll be the ones who decide which proofs are worth mining.

The question is: what happens to the insights that can’t be found by brute force? What happens to the mathematics that requires not just recombination, but genuine invention?

We don’t know yet. But we’re about to find out.