Solving Problem Solving

I don’t get to talk a lot with people who are as interested in Artificial Intelligence as I am.

In fact, I’d say that I’m not even as into it as I know others to be. I do, however, spend an unreasonable amount of time pondering it. This allows me a certain amount of conversational agility should I ever encounter a real nerd.

I kid.

But to be clear about the edges of my knowledge, I haven’t read Kurzweil, or much dedicated AI futurism at all. Most of what I know about those ideas comes secondhand through the ambient internet. But as I mentioned, I do ponder, think, and imagine having conversations with people who are interested in this stuff. So, the ideas and writing here are as mine as they can be, minus whatever spelling/grammar cleanup I eventually ask an AI to perform.

These ideas needed an outlet.

My everyday conversations are more often superficial and follow the well-trod paths of social norms. And my imaginary conversations are really starting to bore me. So here I am writing to you, an imaginary reader.

Something that seems to come up often for me is the scenario of trying to explain why artificial intelligence is potentially such a big deal. Sometimes, the person I imagine explaining this to is a skeptic; other times, they’re just genuinely out of the loop.

The phrase that comes to mind that seems to best encapsulate how I think about it is this:

“First, start by imagining that you have solved problem solving.”

Cue Vanilla Ice

“Got a problem? Yo I’ll solve it”

(Yes, in my mind, Artificial Intelligence is like Vanilla Ice. VI, not AI.)

I digress.

All I really mean is that as Artificial Intelligence makes more everyday problems tractable, the production of solutions starts to look increasingly solved. Depending on how far you take the idea, you either conclude with Machine God Magic, or less optimistically (or less far in the future) simply highly capable technology.

I think plenty of thinking has been done about the “Machine God Magic” section of the spectrum, so I won’t sit us around that campfire too long. Except maybe to say that truly solving problem solving may be the less interesting part of this spectrum. What I’ve been thinking about is the strange point on the timeline just before then.

That point on the timeline is the one where well-formed candidate solutions can be created for almost any adequately explained problem. Cheaply and quickly. That is to say, “If you can name it precisely, a competent candidate solution can be produced.”

I think we live in a world that looks more and more like this every day.

I began writing this essay on August 2, 2026. I put it away for a while as life and my own lack of focus tend to pull me in all sorts of directions.

It is now September 8, and a headline came across my desk that I couldn’t have planned more perfectly if I tried (mostly because I am unfamiliar with the math). OpenAI announced that a swarm of 10,000 agents more capable than GPT-6 Astra produced a proposed solution to a Millennium Prize Problem in about a week. That problem was Navier–Stokes, and you can read about what OpenAI has to say about it here.

Whether that proposed solution ultimately earns acceptance is, perhaps hilariously, relevant to the rest of this essay.

But I think we now have to zoom out from that extremely narrow, practical example of what I mean and continue the journey.

Where the scarcity shifts

Why is that in-between point more interesting than the future where Machine God Magic fulfills your every desire instantaneously? Well, because despite the original premise (“Problem solving is solved!”), at the point in the timeline I’m writing about, there actually are still problems. Stakes. Skin in the game. Something to struggle against. They’re just more insidious.

I’ve found in my recent experience that removing a bottleneck does not fully remove constraint from a system. It only helps you find the next constraint.

In practice, when solutions become abundant, the actually scarce work begins sticking out to explain why there aren’t suddenly more successful ventures everywhere. Specifying the problem, evaluating the candidates, establishing the facts, coordinating action, and accepting responsibility for the result are all valuable activities left over in a post-problem-solving world. These assertions aren’t even specific to AI necessarily. In some ways, humanity has been living this rule for centuries. Computers made the ability to process information abundant; humans still needed to decide what was worth processing.

The Verification Frontier

Among the most important remaining bottlenecks, in my opinion, is knowing whether a solution is actually good, correct, or safe. In software, and in relation to AI code, this often gets described as “verification.” For example, tests are created that prove the new software behaves the way we want it to.

That’s great because we can take a piece of intent and functionally prove that the software meets it. That is all good and well, but what happens when the ability to generate code outpaces the ability to verify that code, either through code review or otherwise? I’ve increasingly found from my own experience that the value of verification in that scenario skyrockets.

As highly capable agents produce more software, the traditional methods of verification no longer scale. Code review done primarily by humans used to be acceptable because the slowness of code generation could be paced with the slowness of reviewing that code. This necessarily means that the ability to call any piece of agent-authored code “safe,” “correct,” or “what I actually asked for and nothing more” becomes increasingly scarce.

But tests answer one question: “Does the code behave?” With agents in the loop, the failures that matter start to look different. “Is the work what I agreed on?”, “Does the agent’s report match the artifact?”, “Who stands behind this?”, and “Is what was merged what was verified?”

And where there is responsibility and scarcity, there is value to be mined. The person or machinery that can sit in the gap and say, “this code meets the requirements and declared intent, its tests pass and all the deliverables are here,” suddenly becomes a godsend when hundreds of thousands of lines of code start getting slung around at breakneck speed.

A constant separation should be maintained between claims and acceptance. An agent (or a swarm of 10,000) may claim that it has solved a nearly 100-year-old problem. But accepting that the problem is solved requires verification.

Is the proof checker in the room with us right now?

In the case of mathematics, we often get comically lucky in regard to checking correctness. Frameworks like Lean make math proofs formally checkable and verifiable in a way similar to what I described above.

But real life is messy. Penguins are not perfectly spherical, nor do they slide on frictionless ice. Many hard problems involve nuance (something sorely lacking in public discourse these days, but that’s a separate topic), and not everything comes with clean acceptance criteria.

So then what happens when answers to problems become abundant in the domains of business, policy, science, or everyday life but the ground truth still remains hard to get? Well, for one, we get more.

More what? Well, more of everything. The good and the bad solutions. Which necessarily makes the ability to distinguish the good ones that much more valuable. You may be noticing a theme. The value accrues where there is the ability to steer toward truth and correctness.

So what is left?

The good news is there is plenty of opportunity in a world where candidate solutions are abundant.

Having good taste (you didn’t think you’d get through an essay about AI without the word “taste,” did you?) about what to try to solve, establishing what is true (facts), testing candidate solutions, and coordinating all of the aforementioned opportunities (maybe even with agents as the actors) are increasingly valuable and worthwhile pursuits.

And once you understand the scarcity shift, finding valuable work starts to feel like following a blueprint. Where to go feels obvious in hindsight. In some ways, that’s the point. The scarcity shift only exposes what was already valuable by contrast of its relatively increased value over time. It will feel obvious because it was always the case.

If you take nothing else practical away from this essay, remember this:

Democratizing production does not democratize demand. It usually commoditizes production.

Solving Problem Solving

So, start by imagining that we’ve solved problem solving.

And then once you’re done fantasizing about real-life lightsabers or cures for cancer (if you’re an adult), take a few steps back on the timeline to the point just before then. I think that’s where we are now.

Maybe we haven’t solved problem solving. But maybe we’ve begun to solve one enormous piece of it: producing competent candidate solutions at scale.

This doesn’t remove constraints. It exposes the constraints that were always hiding behind it.

AI is certainly capable of giving us these candidate solutions.

The flood has begun. I’m interested in where the scarcity shifts when the waters rise.