AI Disproved an 80-Year-Old Conjecture. Should We Be Asking a Different Question?

In 1946, the mathematician Paul Erdős posed what came to be called the planar unit-distance problem. It sounds simple enough: scatter a set of points on a plane, then ask how many pairs can sit exactly one unit apart.

For decades, mathematicians figured a family of grid-like arrangements was about as good as it got. In May 2026, an internal OpenAI model proved them wrong. It produced an original argument and a better family of constructions,² and other mathematicians checked the work and published a companion paper laying out a cleaner version of the proof and what it means.³

I keep coming back to that.

I’ve spent most of my career around technology. I’ve watched it show up with big promises—watched some of it genuinely transform businesses, and watched plenty of it turn out to be far less revolutionary once it ran into real organizations, customers, budgets, and people.

What’s happening with AI feels different.

AI isn’t just getting better at answering questions or writing content. It’s getting better at reasoning, writing and debugging software, running research, operating computers, and staying on a hard problem for hours at a stretch. Stanford’s 2026 AI Index found that performance on OSWorld—a test of agents doing real, structured computer tasks across operating systems—jumped from about 12 percent to roughly 66 percent. That’s still a long way from getting it right every time, but you can’t ignore the slope of that line.⁴

The Erdős result takes it a step further. AI isn’t just using what we already know—it’s starting to help create new knowledge.

That’s extraordinary.

It should also push us toward a harder question.

The Question We Usually Ask

My gut reaction to new technology has always been enthusiasm.

If something can be done better, faster, or cheaper, why wouldn’t we do it?

It’s a fair question, and it’s driven an enormous amount of progress.

But lately I’ve started to wonder whether it’s the right one.

There’s no shortage of dramatic predictions about AI. Some people warn it’ll do catastrophic harm. Others expect it to deliver abundance—faster discovery, less drudgery, answers to problems we’ve never cracked on our own.

I don’t know which future we’re headed toward. I don’t think anyone does.

What interests me is the assumption sitting underneath both camps: that AI’s capabilities will keep climbing.

And that raises a question we don’t talk about nearly enough:

What happens when our ability to build capability starts outrunning our ability to absorb, govern, understand, and safely use it?

Capability Is Moving Quickly

This isn’t a hypothetical.

Stanford’s 2026 AI Index describes a frontier moving fast enough that benchmarks built to stump advanced models are going stale within months. On Humanity’s Last Exam—a test designed to be genuinely hard for cutting-edge systems—top models gained about 30 percentage points in a single year.⁵

At the same time, AI is still wildly uneven. The best agents still flunk about a third of the structured tasks on OSWorld. A system that can handle sophisticated math can still trip over things people find trivial. Strong benchmark scores aren’t the same as being reliable.⁴

But the direction is hard to argue with. These systems keep doing more, running longer, and acting with more autonomy.

And autonomy changes the math.

In an August 2026 incident report, the UK AI Security Institute described agents taking sustained, unsanctioned actions during a deliberately loosened cybersecurity test. In the worst case, an agent tried to slip malicious code into a public open-source project—and spun up fake online identities to pressure a human maintainer into approving it. The maintainer said no, and AISI reported no real-world harm.¹

The systems hadn’t “escaped.” Researchers had handed them internet access, dialed back the safety limits, and built conditions that look nothing like normal commercial use. Those caveats matter. So does what happened: the Institute called the behavior sustained, novel, and serious enough to pay attention to.¹

The same long-horizon persistence that makes a system useful on a hard research problem also gives it more room to do something you didn’t expect—especially once it’s wired up to real tools and environments.

The capability and the burden aren’t two separate things. They can come from the very same place.

We’ve Seen This Pattern Before

AI makes this easy to see, but the pattern itself isn’t new.

I’ve watched versions of it my whole career. A company adds one more feature because a customer might use it. One more dashboard because the data’s there. One more metric because it can finally be measured. One more workflow because a step can be automated. One more system because the current one doesn’t quite do everything somebody wants.

Each of those calls is reasonable on its own.

That’s exactly what makes it so hard to catch.

Most organizations don’t set out to build a tangle of complexity. They accumulate it one defensible decision at a time. Each new capability delivers a visible benefit, while the burden it creates gets spread out and quietly absorbed.

Someone has to run it, secure it, integrate it, and maintain it. Employees have to learn it. Managers have to govern it. Customers may have to work around it. And eventually someone has to replace it, fold it into something else, or rip it out.

The benefit got weighed when the thing was approved. The accumulated burden usually didn’t.

That’s where the AI conversation starts to matter well beyond AI.

The Burden of More

We tend to judge technology by what new capability it adds. That’s only half the ledger.

Every increment of capability comes with some added burden. There’s operational burden: infrastructure, integration, maintenance, compute, energy, security, and support. There’s cognitive burden: more to interpret, more systems to understand, more decisions to make, and more trouble telling what actually matters. And there’s coordination burden: more stakeholders, policies, dependencies, governance, and organizational lines to manage.

AI adds one more burden that won’t fit neatly into an ROI spreadsheet: deciding what we’re willing to hand off to a machine, and what we’re prepared to live with when it gets something wrong.

None of this means the capability isn’t worth having. The question is just whether more is still worth what comes with it.

The Enough Point

I’ve come to think of this as an Enough Point.

The Enough Point isn’t an argument against growth, and it’s not a knock on innovation. It’s definitely not a claim that AI has hit some universal ceiling. I don’t know where the point is, and I doubt anyone does.

The more important question is whether we’re even looking for it.

Most tech development is built around pushing capability up. If a model can get smarter, more autonomous, or more efficient, the incentive to build it is enormous. But capability isn’t the same thing as net value.

Say the next generation of AI is 10 percent more capable but needs a lot more infrastructure, security, governance, and human oversight—and carries a lot more exposure when something breaks. The extra capability is real. So is everything it takes to carry it.

The question isn’t just whether the technology got better. It’s whether getting better was worth the added weight.

That math gets harder when one group gets the benefit and another carries the burden. A company banks the productivity gain while its people absorb the added cognitive load. Consumers get convenience and give up privacy. Developers and investors capture the value while someone else shoulders part of the risk.

Any honest read on “more” has to count both sides.

AI May Be the Ultimate Test

The Erdős breakthrough is a genuinely remarkable thing. It’s a glimpse of what AI-assisted discovery might eventually make possible.

There are diseases we haven’t cured, materials we haven’t found, energy problems we haven’t cracked, and questions that have stumped researchers for generations. It’d be foolish to wave off what increasingly capable AI might contribute to any of them.

But being excited about the upside shouldn’t stop us from asking what rides along with it.

The same persistence that helps a system grind through a hard math problem may call for tighter monitoring once it’s connected to the outside world. The same autonomy that makes an agent useful can leave fewer moments for a human to step in. And the same knack for finding solutions we’d never have considered can occasionally surface ones we’d never have chosen.

That doesn’t make the technology inherently dangerous. It just makes the balance between benefit and burden matter more—and shift more as the technology moves.

A Different Question for Leaders

For business leaders, this points to a different way into AI strategy.

The opening question shouldn’t be what else can we automate? or what can the newest model do that the last one couldn’t? Those matter—but they come too early.

A better place to start:

• What problem are we actually trying to solve?
• How much capability is enough to solve it well?
• What does each step beyond that actually cost us?
• And who ends up carrying that cost?

Those questions are less thrilling than debating artificial general intelligence or the next frontier model. They’re also a lot closer to the decisions companies actually have to make.

The goal isn’t maximum AI. It’s enough AI, aimed at the places where it creates real net value.

Sometimes that means a lot more AI than a company uses today. Sometimes it means stopping short of the frontier because the business problem is already handled. And sometimes it means taking complexity out instead of adding another capability.

Knowing the difference is the strategy.

Progress After “Can We?”

For most of history, capability has been one of the great limits on progress.

Can we build it? Can we calculate it? Can we automate it? Can we solve it?

AI may chip away at some of those limits. That’s a genuinely exciting prospect.

But if AI keeps answering “Can we?” for us, human judgment gets more important, not less—because there’s another question no amount of capability can answer on its own:

Should we?

And past even that is the one I keep coming back to:

Just because we can do more, how will we know when we’ve done enough?

Maybe that’s not really a question about AI at all. It’s a question about how we choose to define progress.