A master of the abacus can, after enough years, push the beads away and keep computing — the frame has moved inside. Watch one work and you'll see the fingers twitch over a device that is no longer on the table; the calculation runs on a mental image of a tool they no longer physically need. Now picture the driver who has followed the turn-by-turn voice through the same city for a decade and still cannot say which way is north. Two tools, two outcomes, one difference that matters more than almost anything else we could say about the technology of thought: when the tool is gone, is there more of you, or less?
David Krakauer, who runs the Santa Fe Institute, gave this difference its sharpest name. He splits the tools that help us think into two families. One he calls complementary; the other, competitive. The distinction is the most useful lens we have for looking at what AI is actually doing to us — and it is almost completely absent from the conversation.
Two families of tool
A complementary cognitive artifact — the abacus, the slide rule, the piano, a good map — works by transferring its structure into you. You practice with it, and the practice deposits something: a model, a method, an intuition. In Krakauer's words, “after repeated practice… the artifact itself could be set aside and its mental simulacrum deployed in its place.” The expert abacist runs the abacus in their head. The tool was a teacher. Its job was to make itself unnecessary, and it succeeded.
A competitive cognitive artifact — the GPS, the pocket calculator — works the other way. It performs the task for you and asks nothing of you in return except that you stop doing the task yourself. It deposits nothing. Take it away, Krakauer says, and “we are no better than when we started.” Worse, usually, because the faculty it replaced has gone slack from disuse. His verdict is blunt: these tools “are not coaches and teachers — they are serfs.” A serf does the work and leaves you none the wiser.
The complementary tool teaches you to fish. The competitive tool delivers the fish, every day, until you forget there was ever a river.
The mechanism under the old warning
McLuhan said it first, and darkly: every extension is an amputation. Reach far enough through any tool and you go numb in the part of yourself it extended — the Narcissus trap, where you fall for your own reflection in the technology and lose feeling in the limb. It's a haunting claim, but it was missing a mechanism. Why do some extensions amputate while others don't? The abacus is an extension of arithmetic and it leaves you sharper, not number. So the warning couldn't be the whole story.
Krakauer supplies the missing piece. It turns on whether the tool transfers its structure into you, or merely acts on your behalf. Transfer the structure and the gain is retained — the extension never flips, because you've absorbed it. Merely act on your behalf and the extended faculty atrophies the moment the tool is present to carry it — the amputation completes quietly, on schedule. A competitive artifact, put precisely, is one whose amplification has already flipped into atrophy. It still works, so you don't notice. That's the trap.
Where AI is being pointed
Here is the uncomfortable part. The incentives of the current AI build all pull toward the competitive pole. The competitive tool demos better — it's smoother, it asks less, it hands you the answer while the complementary tool is still asking you to understand the question. So the market rewards the serf and starves the coach. Krakauer's own stated fear is not a runaway superintelligence; it's “the seductive taste of the algorithmic recommender system that saps our ability to explore options and exercise judgment,” a slow slide he calls a genuine “phase transition” in how discerning people are willing to be. Not a bang. A drift. The faculties going quiet one at a time while the interface stays delightful.
None of this is an argument against tools. The abacus is a tool. The slide rule put men on the moon. The argument is about which kind of tool, and about refusing to let “it produced the answer” stand in for “it made me more capable.” Those are different achievements, and only one of them survives the tool being switched off.
Building for the abacus pole
If you take the distinction seriously, it stops being philosophy and becomes a specification. Keep the gain when the tool is gone — make that the acceptance test, not an afterthought. That means a tool should make its method legible, not just its output: show the reasoning, name the pattern, say why this and not that, so the structure has a chance to transfer. It means leaving a mental simulacrum behind — repeatable procedures, named moves, checklists a person comes to know by heart. It means keeping the human making the discerning calls rather than collapsing every choice into a single auto-selected answer. And it means, sometimes, declining to pre-chew the problem — letting a person meet the raw version now and then, because that friction is where skill compounds.
A slide rule, not a GPS. An instrument that leaves its structure behind in the mind that used it — so that removing it reveals a smarter person, not a stranded one. That is the bar. It's a harder bar than “did it work,” and we think it's the only one worth building to. The question to carry out of here is small and it is everything: when this tool is gone, will there be more of you, or less? Build the abacus. Refuse the GPS. The difference is the whole game.