Naming Is the Last Step

If you can name something cleanly and confidently, you are usually too late to be early to it. We are still naming AI.

When the first commercial transistor went on sale in 1954, it was sold for use in hearing aids. The Bell engineers who invented it — Bardeen, Brattain, and Shockley — did not call their company “Transistor Corporation.” Bell Labs filed the patent under a description that read, in part, “a semi-conductor amplifying device.” The word transistor itself was coined later, by an internal naming committee, who chose it as a portmanteau of “transconductance” and “varistor,” in part because it was hard to confuse with anything else.

For about ten years after the transistor was invented, no one outside engineering knew what it was. The general public had no name for the thing in their pocket radio that was making it possible. They knew transistor radios were smaller and lighter than the older vacuum-tube radios. They did not know why. The word “transistor” entered general English usage only after it had been embedded in millions of devices, doing its work invisibly, for over a decade.

This is the normal pattern. The thing exists first. The deployment happens second. The naming — the moment when the general public has a stable, agreed-upon word for what they are dealing with — comes last. Often by a lot.

Consider a few examples.

Electricity was being generated and used commercially for thirty years before the word electron entered the language. Telephones were in widespread use before the term long-distance call stabilized. Computers existed for fifteen years before software became an ordinary word. The internet was used by millions of people before the word email was unhyphenated. Even the cloud, as a casual term for distributed computing, took most of a decade to become unremarkable.

This pattern is so consistent that you can use it as a diagnostic. If you can name something cleanly and confidently, you are usually too late to be early to it. By the time the language has stabilized, the deployment is already mature, the early profits have already been taken, and the underlying technology has settled into a form that’s hard to disrupt.

Conversely: if you cannot find good words for something, and the words you try keep falling apart, and people who clearly know what they’re talking about disagree about what to call it — that is usually a sign that something real is happening, but it’s still in the phase of being seen rather than being named.

We are in this phase now, for AI.

The word “artificial intelligence” itself is a 1956 marketing term coined for a Dartmouth conference grant proposal. It has never stopped being slightly misleading. Artificial implies a contrast with natural intelligence that doesn’t quite hold. Intelligence claims something we don’t know how to define. The phrase persists because nothing better has emerged, and because the marketing weight is already there.

The other current candidates are worse. Machine learning describes one technique among many and treats the result as an artifact of the technique, which is backwards. Large language model is technically accurate and conveys nothing about what the thing does or why anyone should care. Generative AI is the term we’ve ended up with because it’s vague enough to cover everything from chatbots to image generators to music systems, which is also why it tells you almost nothing.

The honest fact is that we don’t yet know what to call the thing we have built. We don’t know which of its features are essential and which are accidents of the current implementation. We don’t know whether what’s coming next is a continuation of the current paradigm or a quiet replacement of it. The naming hasn’t happened yet, because the thing hasn’t fully resolved into one shape.

This is a strange position to be in, and it’s worth sitting with.

Most of the breathless commentary about AI assumes the naming is settled. People argue about whether AGI is close, as if AGI were a stable, well-defined concept. They argue about whether agents are real, as if “agent” pointed to one identifiable kind of thing. They argue about whether alignment is solvable, treating alignment as if it were a problem with crisp boundaries. None of these terms are stable. They are placeholders. The discourse pretends to know what it’s debating, because pretending is more comfortable than admitting you’re naming a moving target.

In the historical pattern, the real names — the ones that will be in unremarkable use forty years from now — have probably not been coined yet. They will be coined by someone who is not famous, in a context that is not currently being watched. They will spread quietly. By the time they are established, the underlying technology will have settled into a shape that today’s engineers would not fully recognize. This is how every previous technological transition has worked.

There’s a useful discipline in this for anyone trying to think clearly about the present. Notice when the language is loud. Loud language — confident slogans, decisive predictions, urgent acronyms — is usually a sign of an industry that is talking about itself rather than understanding itself. The people who actually understood electricity in 1900 were not the ones writing manifestos. They were the ones quietly running cables.

The same is true now. The people who will turn out to have been right about AI are mostly not the ones who are loudest right now. They are the ones who can sit with the fact that we don’t yet have the right words, and who keep working anyway, knowing that the names will come last.

When the names come, you’ll know the moment is over. The new normal will already have arrived.

Until then, the noise is just noise. The signal is in the work.

Part of the Aperture series at Nous. Closes the six-essay arc. Pairs with The 30-Year Lag on the same observation from a different angle — deployment runs ahead of language, language runs ahead of public understanding, public understanding runs ahead of the next disruption.