What We Choose to Keep

There's a good essay by Sean Goedecke making a simple observation with an uncomfortable edge: math olympiad puzzles used to be a reliable, hard-to-fake signal of real mathematical progress. Solving one meant you'd generated genuine insight, because brute force alone couldn't get you there. AI has started closing that gap the wrong way. It can now solve some of these puzzles without producing the insight that made them valuable in the first place, a thousand-page proof no mathematician could hold in their head, correct but hollow. The puzzle gets solved. The thing the puzzle was supposed to certify doesn't show up.

Goedecke's proposed fix is modest and specific: look at chess and speedrunning, fields where computers surpassed humans decades ago, and notice that both kept a human-only lane. Human chess is still played, still prestigious, still meaningfully different from a human losing to an engine. His bet is that math and software engineering can do something similar: silo human work off, or find legible skills AI can't yet counterfeit, and rebuild prestige around explaining the machine's work instead of only producing it.

That's a reasonable, careful answer to a narrow version of the question. It's worth pushing on the wider version he doesn't fully take on.

Start with this: none of this is happening to us. It's not weather. Every model that got trained, every dollar that funded it, every team that shipped a coding agent instead of holding it back a year, every one of those was a decision, made by specific people, who could have decided otherwise. That includes the decisions still being made right now, this week, by the people writing this and reading it. The comfortable version of this story treats AI capability as an external force arriving on its own schedule, something happening to expertise from outside. The honest version is that we are building it, funding it, and shipping it, on purpose, and some of what it costs us may already be spent. That's not a reason to stop asking the question. It's a reason to ask it without the alibi.

And the question, once you follow it past math contests, isn't really about prestige or proxies. It's older than that. Descartes, trying to find one thing he could be certain of, didn't land on an output. He landed on the act itself: I am thinking, right now, and that fact is available to me in a way nothing else is. Not the conclusion the thinking arrives at. The arriving. Whatever else you think of his answer to "what am I," the question he was asking is exactly the one Goedecke's essay backs into without naming it: when the correct proof, the working code, the winning move shows up without anything having experienced the getting-there, what's actually been produced? Is it the same kind of thing a human expert produces, just faster? Or is it a different kind of object that happens to look the same from outside, the way a lookup table can produce the right answer without anything resembling understanding sitting behind it?

That's not a rhetorical flourish. It's the actual stakes. If expertise was always just the ability to produce correct outputs, then AI closing that gap is a straightforward loss of a job market and nothing more, sad for the people affected but conceptually simple. If expertise was something else, something more like the interior experience of working a problem until it yields, the doubt and the arriving that Descartes pointed at, then AI producing the same outputs without any of that happening doesn't replace expertise. It produces something adjacent to it, useful, sometimes indistinguishable in its results, and still a different kind of thing. Which one is true probably isn't a question with a clean answer. It might be worth sitting with the discomfort of not knowing, rather than resolving it too fast in either direction.

It's worth naming the obvious objections here rather than hoping no one raises them. One: if some of this is already irreversible, calling it a "choice" is just fatalism wearing a nicer word. That's fair to worry about, but naming a point of no return isn't the same as saying nothing before it mattered. A technology already embedded in an economy is harder to unwind than one still on a whiteboard, and pretending otherwise doesn't make unwinding easier, it just delays admitting the cost. Two: the Descartes framing leans on a distinction, the experience of thinking, that we can't actually verify one way or the other in a model. Also fair, and it's not resolved here. The point isn't that models definitely lack that experience. It's that we don't have a way to check, and we're making sweeping claims about expertise anyway, which is itself worth sitting with. Three: bringing a product into an essay about philosophy and inevitability can look like using a crisis to sell something. That's a reasonable thing to be suspicious of, so it's worth being precise about the claim: 5thDev isn't offered here as a resolution to what expertise or understanding mean. It's one concrete answer to a narrower question, whether human oversight stays structurally possible or quietly becomes optional by default.

None of this is an argument for fatalism, and it shouldn't read as one. The fact that this is a choice, not a weather event, cuts both ways: it means there's still a choice to make. What gets automated by default, and what a team decides is worth a person still doing, on purpose, even when a model could do it faster. What stays visible enough to check, versus what quietly becomes a black box because checking it was never built into the system in the first place. Those are engineering decisions as much as philosophical ones, and they get made in software, not in essays.

That's the actual ground 5thDev is trying to stand on. Not the position that AI should be held back, and not the position that the fastest possible path to full autonomy is obviously correct either. A middle path: agents that are governed, continuous, and auditable, so that whatever work they do leaves something a human can actually check, question, and stay accountable for. Not because checking is a nostalgic gesture toward a world where humans still mattered. Because a system that keeps its work legible is the only kind of system where the question Goedecke is raising, and the older one underneath it, stays open long enough for anyone to actually answer it. The goal isn't systems humanity serves. It's systems that serve humanity, that enable it rather than quietly replacing the need for it.

The puzzle was never the point. What it was supposed to certify was. The real question in front of us isn't whether AI can solve it. It's whether we build anything that keeps us able to tell the difference.