What an IQ score does not capture

Picture a robot with an IQ of 150. Most people who imagine it that way mean something simple by it: faster, sharper, better at everything than the cleverest person they know. It is an image that works, because it measures against something familiar, a school report, a test result, a number you can compare. The problem is that the people who have studied intelligence the longest and the most thoroughly have already shot down precisely this comparison, and not out of caution but out of precision.

A ruler made for something else

The IQ test itself is not a timeless given. It was designed by the French psychologist Alfred Binet, together with Théodore Simon, in 1905, not to pin down intelligence as such, but to find French schoolchildren who needed extra educational support. Binet himself already warned about what has been happening to it ever since: the number gets treated as a fixed, innate ceiling instead of a practical aid for one narrow question. So taking a ruler built to measure who needs help in a Paris classroom in 1905 and using it to measure a machine that folds proteins and writes texts is a wrong measure on a wrong object, twice over.

Separate from one another, not on a single line

In 2012 the philosopher Nick Bostrom drew a more precise distinction, which he worked out two years later in his book Superintelligence: intelligence and goal, on his account, stand apart from one another, the so-called orthogonality thesis. A system can be dizzyingly capable in one area and, in another, understand simply nothing of what a human being would call "sensible". That, in his view, is not an edge case but rather the rule. His illustration, a system whose only assignment is to maximise paperclips and that puts all available matter to work on it if need be, is deliberately absurd. The absurdity does not lie in the intelligence of the system, but in the idea that "smarter" automatically means "wiser" or "more reasonable".

What is already here, and what is not here yet

Practice confirms the picture of a landscape rather than a line. A language model already writes faster and with a wider range than almost every person you know, and in that area it stands far above any IQ score you could think up. That same model can at the same time add up a simple sum wrongly, in a way no eight-year-old with a calculator would ever fail, or invent a fact with the same tone of certainty it uses to state a true one. "Surpasses the human being" is not a dot you pass at some given moment. It is a surface with peaks and valleys, one that in some places has long risen above us and in others still sits below the floor.

The dissenting voice that deserves a hearing

That does not mean the underlying concern is misplaced, only the number it is often expressed in. Researchers who worry about where this is heading, Bostrom among them, have in the orthogonality thesis a stronger argument than an IQ score could ever supply: not "it is getting too smart", but "it can be capable without sharing our goals". Anyone who disagrees, researchers such as Yann LeCun argue that current architectures are fundamentally not the road to such a general system, has just as much of a point, and that point deserves as much weight here as the first one.

"A moment to take with you: the next time you ask how smart it is going to get, ask yourself also at what, exactly. And at what not."