If You Can Name It, You Can Build It
There is a scene I keep coming back to from Ursula Le Guin’s The Dispossessed. The physicist Shevek spends the better part of the novel chasing a unified theory of time — not because he lacks the mathematical tools, but because the conceptual framework he needs doesn’t quite exist yet. The breakthrough, when it comes, is not a calculation. It is a reframing. He finally sees the shape of the thing. The math follows almost as an afterthought.
Le Guin was obsessed with this idea that thought and language are inseparable. Her fiction is full of it. Linguists at Berkeley have written papers about how carefully she embedded ideas from the Sapir-Whorf hypothesis — the notion that the language you have determines, or at least strongly shapes, what you are able to think — into the fabric of her invented worlds. The people who cannot name a thing, in Le Guin’s universe, cannot quite see it either.
I have been thinking about Shevek a lot lately, and not just because I have been rereading The Dispossessed. I have been thinking about him because I think he is a surprisingly good model for what it means to be effective in the world we are currently building.
Here is the observation that keeps nagging at me: the AI tools available today are extraordinarily good at execution. Give them a clear enough problem statement and they will write the code, run the simulation, produce the documentation, and debug the result. The friction between “I want this to exist” and “this now exists” has dropped close to zero for a remarkable range of tasks. That is genuinely new. It is, I think, the most significant thing that has happened to knowledge work in my lifetime.
But here is what I don’t hear talked about enough: that shift has not made knowledge less valuable. It has moved where the knowledge matters.
In the old world — pre-2023, let’s say — having a good idea was necessary but not sufficient. You also needed to be able to execute it, or to hire people who could. The ability to implement was itself a filter. If you wanted to build a tool that used wavelet analysis to detect edges in noisy images, you needed to know not just that wavelet analysis was the right approach, but how to implement a discrete wavelet transform, how to tune the thresholding parameters, how to handle boundary conditions. The gap between “I know what I want” and “I have built what I want” was wide and expensive, and crossing it required deep technical skill.
That gap has narrowed dramatically. An experienced researcher who has never written a line of Python can now, in an afternoon, produce a working implementation of something that would have required a software engineer two weeks of effort five years ago.
But here is the thing: that researcher still needs to know about wavelets. They need to have the concept. Because if you don’t know what a wavelet is, you are not going to ask for one. You might ask for something blurrier, something less precise. You might describe your problem in terms of its symptoms rather than its structure, and get a generic, unimaginative solution as a result. The AI is not going to spontaneously invent the right conceptual framework for you. It will give you something that fits the shape of the question you asked.
This is where the Sapir-Whorf hypothesis becomes unexpectedly practical. The classic formulation — that the language you speak influences what you can think — has always struck people as either obvious or overstated. Modern cognitive science has landed somewhere in the middle: language doesn’t absolutely determine thought, but it does shape it in measurable ways. Speakers of languages that have precise color terms for distinguishing blue from green show different perceptual behavior near that boundary than speakers whose language collapses the two into one word. The name carves out the concept. The concept carves out the perception.
Scale that up to scientific education and it becomes the most important thing I can say about what it means to be effective right now. The conceptual vocabulary you have built up — through coursework, through research, through years of wrestling with problems in a domain — is not just a collection of useful facts. It is a set of lenses that let you see certain problem shapes that are invisible to people who don’t have the vocabulary. Fourier analysis is a lens. Markov chains are a lens. Bayesian inference is a lens. Topological data analysis is a lens. Each of these is a way of seeing that generates a set of possible questions, and the people who have those lenses are going to ask qualitatively different things of AI systems than the people who don’t.
The historian of science Helge Kragh, in a recent book on the history of scientific terminology, makes the point that scientific language is not merely descriptive — it is generative. Lavoisier’s reform of chemical nomenclature in the 18th century didn’t just rename things; it embedded an entire new theory into the language of chemistry. To adopt the new names was to adopt the new explanatory framework. The naming was the thinking. Something similar happened when Leibniz developed his notation for calculus: the symbols he chose — dy/dx, the integral sign — were designed to make the underlying mathematics transparent to the user, and they did. Continental European mathematicians using Leibniz’s notation made faster progress than British mathematicians using Newton’s less intuitive system. The notation was a cognitive tool, not just a record.
I want to be precise about what I am and am not claiming here, because this is where the argument is easy to caricature.
I am not saying that AI doesn’t help beginners. It clearly does. You can use AI to learn things faster, to get unstuck, to explore unfamiliar terrain. I am also not saying that deep expertise in one domain protects you from disruption in another. The person with a PhD in organic chemistry is not automatically an effective AI user in, say, financial modeling.
What I am saying is something narrower and, I think, more interesting: the conceptual depth you have in a domain determines the ceiling of what you can accomplish with AI assistance in that domain. The AI can fill in implementation details. It can suggest approaches you haven’t thought of. It can catch your mistakes. But the initial framing — the decision about what class of problem this is, what conceptual tools are potentially relevant, what would count as a genuinely novel contribution — that framing comes from you. And it comes from the vocabulary you bring to the table.
Shevek’s insight isn’t that he ran a calculation the other physicists couldn’t run. It’s that he could see something they couldn’t see because he had spent years building a conceptual model that let him perceive a structure others missed. The math was always there. The tools were always there. The problem was formulation.
The popular narrative about AI and education tends to run in the opposite direction from what I am describing. The usual story is that AI makes expertise less necessary — that you don’t need to know things anymore because AI knows them for you. I have never quite believed this, and I think the evidence is accumulating against it.
The people I have watched do the most interesting things with AI tools are, almost without exception, people who have genuinely deep knowledge in some domain. They are the ones coming to the conversation with a clear model of what they want to build, a sense of what would be novel, and the conceptual vocabulary to express it precisely. The AI amplifies their ability to execute. It does not supply the vision.
If I wanted to give someone the single most useful piece of advice for thriving in the world that is taking shape, it would be this: learn things deeply enough that you build up genuine conceptual vocabulary. Not just terminology, but the underlying structures the terminology points to. Not just the name “wavelet” but the idea of representing a signal at multiple scales simultaneously. Not just “Bayesian inference” but the intuition of updating beliefs in light of evidence.
The barrier to building has collapsed. The barrier to knowing what is worth building has not. That was always the hard part. It still is.
Le Guin’s Shevek would, I think, make an excellent AI power user.