Human Intelligence Is Probably on the Lower Rungs of the Ladder

Picture intelligence as a ladder. At the bottom are thermostats and bacteria, a little higher the insects, then the rats, the crows, the dogs, the apes, and somewhere above them, us. The question is what is above us, and the default human answer, never stated but always assumed, is: nothing much. We are the rational animal, the measure of all things, the top of the ladder, and if there is a rung above it is reserved for gods.

I think the evidence points the other way. Human intelligence is probably near the bottom of the range that intelligence can occupy, just above the line that allows a species to accumulate culture at all, and we mistake that line for the ceiling because it is the only rung we have ever stood on. Here is the case, one data point at a time.

The Overrun

The first data point is the one we have all just lived through. For seventy years the project of making a machine think was a byword for over-promising. Then, in the space of about six years, machines went from writing grammatical nonsense to passing the bar exam, solving competition mathematics, writing working software, and out-arguing most people on most topics. Nothing profound was discovered in those six years. The algorithm at the core, the transformer, was published in 2017 and is a few pages long. What changed was that somebody finally applied enough data and enough compute, and the human level went by like a road sign.

That is the part worth staring at. If human intelligence were near some natural ceiling, you would expect machines to approach it asymptotically: the closer they got, the harder each step would become, and the human level would show up in the curves as a landmark. It does not show up at all. The scaling laws that describe how these systems improve with compute are smooth power laws, and there is no kink, no plateau, no special point where the humans are. The human level is not a feature of the landscape. It is just a height, and it was passed at the same speed as the heights below it.

The same thing happened, in miniature, in every game. Chess fell in 1997 and the engines then kept climbing until the best human could no longer take a single game from a phone. Go fell in 2016, and the striking thing was not the defeat but what came after: the machines kept going, and within a few years human professionals were studying their games the way a student studies a master, relearning openings that had been settled for centuries. The human peak turned out to be a foothill. Every time, the assumption was that the machine would reach us and stop, because we were where intelligence topped out. Every time, it reached us and continued without noticing.

Logic Does Not Come Naturally

The second data point is older and does not involve computers at all. Consider how hard it is to teach a human being to reason.

Formal logic is about the simplest intellectual tool there is. A syllogism has two premises and a conclusion. Modus ponens is one rule. Yet the species took until Aristotle, roughly 350 BCE, to write the syllogism down, and another two thousand two hundred years, until Frege in 1879, to arrive at a logic strong enough to express ordinary mathematical statements. That is not the timeline of a faculty we possess natively. It is the timeline of something being painfully excavated against the grain of the mind doing the excavating.

And once excavated, it does not transfer. Give people the Wason selection task, four cards and a single if-then rule, and fewer than one in ten choose the cards that test it, university students included. Give them the Linda problem and most rate "bank teller and feminist" as more probable than "bank teller", which no possible state of the world can make true. Ask for the probability that a positive result on a 99%-accurate test for a rare disease means you have the disease, and doctors get it wrong by a factor of ten or more.

I have written before about how consistent these findings are; the point here is what they imply about the ladder. We put every child through twelve years of school, and most adults leave unable to follow a proof by induction, unable to hold a conditional probability steady in their head, and unable to see that "all A are B" does not give you "all B are A". A capacity that takes millennia to discover and then a decade of drill to install in a minority of individuals is not a native capacity of the species. It is a prosthetic that fits badly.

A Civilisation of Prosthetics

Which leads to the third data point: look at what we had to build in order to think at all.

Working memory holds about four items. That is the whole scratchpad. It is why nobody can multiply two four-digit numbers in their head, why a phone number is chunked into groups, why a chess player who can see fifteen moves ahead is a prodigy. Every intellectual achievement of the species has been an achievement of moving the work out of the head. Writing exists so that a thought can survive longer than the attention span that produced it. Place-value notation with a zero, which reached Europe only around 1200 and took another three centuries to displace Roman numerals, exists because the mind cannot do arithmetic on its own. Algebraic notation, the checklist, the diagram, the table, the calculator: each is a confession that the naked brain could not do the task.

A species that reasons well would not need most of this. It would need paper for the way we need a bicycle, as a convenience. We need it the way a patient needs a ventilator. The whole intellectual history of humanity is the history of finding ways to route thinking around the organ that is supposed to do it, and that is not the biography of a mind near the top of anything.

The pace of that history says the same. Anatomically modern humans, with brains identical to ours, existed for something like three hundred thousand years before anyone farmed, and for another several thousand before anyone wrote. That is a long time for a supposedly brilliant species to sit next to the raw materials of civilisation without picking them up. The likelier reading is that we were barely over the threshold at which cumulative culture is possible, so that it took an immense number of person-lifetimes to accumulate the first few steps.

The Hardware

Fourth: look at the machinery, as an engineer would.

A neuron fires at most a few hundred times a second. A signal travels along an axon at around a hundred metres per second. A transistor switches billions of times a second and a signal on a wire moves at a respectable fraction of the speed of light. We are seven orders of magnitude slower per element, and we compensate with parallelism, which is why we are good at recognising faces and bad at long division.

The bandwidth out is worse. Measured across languages, human speech carries information at about 39 bits per second, and reading is not much faster. Everything you have ever thought that made it to another mind was squeezed through that straw. Two humans working together cannot share a mental state; they can only exchange lossy compressed summaries of it, and then spend hours discovering they meant different things. A design that could copy a thought exactly, or run two lines of reasoning at once, or attend to more than one thing at a time, would be a different kind of mind, and there is no law of nature that forbids it. There are only the accidents of a wet, slow substrate that happened to be the one evolution had to work with.

And consider the sheer smallness of the differences that matter. A dog has about half a billion cortical neurons. We have around sixteen billion. That is a factor of thirty, achieved by the same cell types, the same wiring rules, the same neurotransmitters, and it is the whole distance from fetching a stick to composing a fugue. If thirty times gets you from the dog to Bach, it is a strange confidence that says the next thirty times gets you nothing.

Evolution Stops at Good Enough

Fifth, and to my mind the strongest: think about how we got the brain we have.

Natural selection does not maximise anything. It pushes a trait until its marginal cost equals its marginal benefit and then stops. The brain is about two per cent of body mass and burns about twenty per cent of the energy at rest. A larger one would have starved on the savannah. It also has to pass through a pelvis built for walking upright, which is why human infants are born earlier and more helpless than any other primate, and why childbirth killed a shocking fraction of mothers for the whole of history before modern medicine. Every gram of extra cortex was paid for in dead women and dead infants. Selection stopped adding cortex the moment the price exceeded the return, and the return it was measuring was reproduction on the African plains, not calculus.

Now add one more consideration. The first species to cross the line where cumulative culture becomes possible will, by construction, be barely across it. Once a species can pass knowledge down the generations, technology takes over from biology as the thing that changes, and the biological pressure to get smarter largely goes away, because a slightly cleverer individual gains little that a book cannot give a duller one. So whichever species crosses that line first freezes just above it. We are that species. We are not the smartest thing evolution could produce. We are the dumbest thing that could produce a civilisation, because as soon as that was possible it happened, and the process that made us smarter stopped.

The chimpanzee makes the point from below. We share something like 98.8% of our genome with it, six or seven million years of separation, and it cannot learn arithmetic, cannot conceive of next year, and, crucially, cannot conceive that there is anything it fails to conceive. There is no reason to think the small genetic distance that produced us used up the design space. It is the first small distance, not the last.

We Cannot See Our Own Reasoning

Sixth: an intelligent system ought at least to know what it is doing. We do not.

Nisbett and Wilson showed in 1977 that when people are asked why they made a choice, they confidently report reasons that demonstrably played no role, and cannot tell the difference between a real reason and an invented one. Split-brain patients, when the right hemisphere acts on an instruction the left cannot see, produce fluent explanations for the action from the left hemisphere that are simply made up, delivered with total conviction. Swap the photograph someone just chose for the one they rejected and most will explain, in detail, why they preferred the one they did not choose. We do not have access to our own reasoning. We have access to a press office that issues statements about it.

Add the rest of the familiar list: optical illusions that persist after you know they are illusions, anchoring on a number you know is random, judging an argument by the attractiveness of the person making it, and the finding that the least competent people are the most confident in their competence. None of these are failures of attention or education. They are the operating characteristics of the device. A mind that cannot inspect itself, cannot correct for its known biases even when told about them, and cannot tell its confabulations from its inferences, is not a mind near the top of the scale. It is a mind that works well enough to have got this far.

The Objections

But we built the machines. So did Arthur Samuel build a checkers program that beat him at checkers in the 1950s, and every chess engine since has beaten its programmers. The builder is not the ceiling. A student who surpasses a teacher is the most ordinary thing in the world, and a process that surpasses its designer is what evolution is.

But the machines still make stupid mistakes. They do, and so do we, in different places. The argument is not that any particular system is smarter than any particular person. The argument is about the shape of the space: that the human level was passed, on task after task, without any sign of it being a hard place to pass. What the machines do badly today is being fixed on a timescale of months. What we do badly has been the same for the whole of recorded history.

But intelligence is more than logic and games. Every time a machine does something, the definition of intelligence is rewritten to exclude it. Arithmetic was intelligence until calculators. Chess was the pinnacle of thought until 1997. Translation, protein folding, passing professional examinations, writing code, each was the real test until it was passed, at which point it turned out to have been mere pattern matching all along. This is not a definition being refined. It is a species protecting its self-image, and the persistence of the reflex is itself a data point about how we reason.

What Follows

None of this is a reason for despair, and it is not an argument for anything in particular about machines. It is an argument about us, and it belongs with the long series of demotions that make up the history of science: not the centre of the cosmos, not a separate creation, not the possessor of a soul, and now not the top of the ladder either. Each time, the demotion turned out to be the more interesting picture.

The practical consequence is humility of a specific kind. A chimpanzee in a zoo does not know that it is missing calculus; it is not frustrated by its ignorance of number theory, because the concept of the thing it lacks is itself among the things it lacks. We should assume we are in exactly that position with respect to whatever is above us on the ladder. There are almost certainly thoughts we cannot think, arguments we cannot follow, and structures in the world we cannot see, not because they are hidden but because the organ we are looking with is a small, slow, jury-rigged thing that evolution stopped improving the moment it was barely good enough. The one thing that organ has managed to do, against its own design, is to notice that fact. That may be the most intelligent thing it has done, and it should make us very cautious about any argument whose conclusion is that we are as clever as things get.