In The Englishman Who Went Up a Hill But Came Down a Mountain, a not-yet-grizzled Hugh Grant arrives in a Welsh village to measure the local mountain and finds it sixteen feet short of the thousand required to count as one. The villagers, unwilling to be demoted to a hill, respond by carrying earth up the slope in buckets until the number comes right. While the film is largely unremarkable, it is weirdly prescient in a reverse sort of way. As things stand, humankind is confident it stands on a mountain. It may turn out to be a hill, after all, with a great deal of its own earth piled carefully on top.

We tell ourselves a comforting story about why artificial general intelligence is still far, far away. It goes like this: human intelligence is a vast and largely uncharted territory; it contains reasoning, creativity, intuition, common sense, emotional nuance, social judgement, moral feeling, the ability to hold a hundred half-formed considerations in mind at once and resolve them into a decision that feels right for reasons too subtle to articulate. The machines, for all their recent progress, have managed only the lowest foothills of this territory. They can string sentences together. They can write large swathes of code that works. They can pass exams that were designed to be hard for humans.

Even so, the hope, sometimes spoken aloud and sometimes only implied, is that one day they will climb the rest of the way up, then keep climbing past us, and that when they do they will take our thinking with them into regions we were never going to reach on our own. The machines will finish the sentences we could not finish. They will see the patterns we could not see. They will be the next chapter in a story that has, until now, only had us in it.

Lately, this is the part of the story I have found more difficult to believe. Not the part where the machines climb, but the part about how far up we are.

Could it be that human general intelligence is a much rarer and more localised phenomenon than we as a species like to advertise? The possibility becomes harder to ignore the more time you spend watching humans make ordinary decisions. We have a folk model of ourselves as rational agents who occasionally lapse into emotion or bias, when the empirical picture often looks closer to the reverse: we are creatures of habit, association, and tribal allegiance who sometimes manage, with effort, to do something that deserves to be called reasoning. The reasoning, when it happens, is impressive. The frequency with which it happens, less so.

Observing much of what passes for thought in daily life, we might conclude that a surprising amount of it resembles pattern completion of a kind that would not look entirely out of place in a well-trained model. Someone says something; we say the thing our group says back. Someone asks a question; we retrieve the nearest cached answer. Someone proposes an idea; we check, often without noticing, whether we are supposed to like it before deciding whether we do. None of this is meant unkindly. I do it too. Everyone does, to some degree. The point is merely that an honest inventory of the cognitive operations performed by an average human in an average day might contain rather less explicit reasoning than our self-description suggests.

Maybe reasoning is not a wholly different substance from pattern completion, but something that emerges when enough of it is piled up high enough.

Entertain, for a moment, how we evaluate language models. We give them logic puzzles, mathematics olympiad problems, multi-step reasoning chains, exam questions from professional qualifications. When they fail, we conclude that they are not yet intelligent. When they succeed, we move the goalposts and complain that the task was contaminated by their training data, or that solving exam questions is not the same as understanding, or that a real intelligence would have done it more elegantly or less recklessly. While some of those objections are valid, the habit is old and reliable enough to have earned a name: the AI effect. Its cleanest formulation belongs to Larry Tesler: “intelligence is whatever machines have not done yet”.

On the other hand, we do not always pay the same attention to the human baselines on those tests. They are imperfect comparisons (the humans are rarely representative, and the conditions differ) but they can still be illuminating. 1 The MMLU authors report that unspecialised human participants recruited through Amazon Mechanical Turk managed 34.5%, against a 25% floor for random guessing. GPQA gave skilled non-experts more than thirty minutes per question and unrestricted access to the web; they reached 34%, while the GPT-4 baseline reported in the paper reached 39%. What these figures establish is that performance on some of the very benchmarks we invoke as tests of formidable intelligence is surprisingly modest among humans who lack the relevant specialisation. The bar we imagine ourselves clearing with ease is not always the bar ordinary people actually clear.

Unsurprisingly, the comforting story has a defence ready for this too: averages are the wrong measure, what matters is the ceiling. Which leaves the question of what counts as the human ceiling in the first place: Ramanujan or Stradivarius? The brilliant mathematician is one example. The virtuoso musician is another. But much of what we celebrate as peak human performance is the product of years of obsessive specialisation, performed by people who are, in the rest of their lives, recognisably ordinary. The accomplished particle physicist still forgets where they put their keys. The chess grandmaster can still be irrational about money, politics or a leaking tap. Peak human cognition is narrow, expensive, and oddly disconnected from the everyday cognition of the same human. This matters because, when we ask whether a language model is as smart as a person, we are usually comparing the model’s average behaviour to the human’s peak.

Reasonable comparisons would go the other way. Pick a randomly selected human, on a randomly selected Tuesday afternoon, doing randomly selected tasks of the kind their job actually requires. Now ask whether a current language model, perhaps with the ordinary tools available to that worker, could perform or materially assist with those tasks at a comparable level. In a surprising number of white-collar occupations, the answer for at least some substantial subset of tasks has quietly become yes. And many of the tasks most exposed are not the low-skilled ones whose disappearance was confidently predicted a decade ago. 2 Frey and Osborne estimated in 2013 that 47% of American employment was at high risk of computerisation and found a strong negative relationship between wages, educational attainment and estimated computerisation risk. Doctors came out near the bottom of their risk ranking. Three years later Geoffrey Hinton told a Toronto audience that people should stop training radiologists immediately. Radiology is currently enjoying the largest shortage in its history. Lawyers draft routine memos; consultants assemble slide decks; analysts summarise reports; journalists turn source material into copy; doctors produce clinical notes. None of this means those occupations have been automated. It does mean that a notable share of work once treated as evidence of unusually high human cognition turns out to respond rather well to large language models and the tools built around them. 3 The inversion has been measured as exposure, which is not the same thing as displacement. Eloundou and colleagues, writing in Science, found higher-wage occupations more exposed to language-model capabilities. They estimated that roughly 1.8% of jobs could have more than half their tasks affected by LLMs with simple interfaces, rising to just over 46% when complementary software is included. In the same journal, Noy and Zhang ran 453 college-educated professionals through mid-level writing tasks with and without ChatGPT: average time taken fell by 40%, while rated quality rose by 18%.

Admittedly, this was not supposed to be the shape of things. The familiar script said that machines would come for routine work first, while creative and intellectual work would remain a human preserve. What the recent wave of language models has complicated is our definition of routine.

It is worth remembering that one important strand of artificial intelligence was really trying to build a brain. Perceptrons were modelled on neurons. Connectionism drew heavily on the language and imagery of neuroscience, and we have been talking about layers and weights and activations ever since. That did not get us as far as we would have wanted. What finally worked was an architecture stumbled upon while trying to improve machine translation, which as far as we know imitates no part of the brain at all, and which was pointed instead at the enormous pile of things brains had already written down. We set out to copy the mechanism, failed for fifty years, copied the output instead, and the copy turned out to be good enough.

Oh, but that is the part that stings. It has revealed that a lot of what we thought of as thinking may actually be something simpler. It has made the practical gap between pattern completion and reasoning look narrower than it was expected. It has revealed that you can produce a great deal of competent professional output by doing what amounts to very sophisticated autocomplete, and that an uncomfortable fraction of professional output was already being produced that way, by humans, before any model existed. The machines did not become an idealised version of ourselves. They have instead put a mirror in front of us.

Very often, a particularly well-worn line appears in conversations about AI: the model is not really thinking; it is just predicting the next token. If you press on it, the natural follow-up question is what exactly we are doing when we think. We reach for words before we consciously know why they fit. We have the next sentence ready before we have finished the previous one. We complete other people’s thoughts in our heads as they are speaking. For a great deal of human cognition, thinking looks suspiciously like predicting the next token.

Even so, the comforting story has one more line of defence: the special thing about humans is not any individual cognitive feat but the integration of all of them: the fact that one person contains the mathematician and the joke-teller and the parent and the cook and the moral agent, and that no machine can yet hold all of those together in a single coherent self. This is true, for now, though I wonder whether a single coherent self is the right thing to be looking for. It is the kind of claim that has not aged well every other time it has been made, about every other capability we used to consider uniquely ours. The list of things that were supposed to require a real mind, and that turned out not to, is already long, and it is getting longer at a rate that should worry anyone who is staking their sense of human distinctiveness on the items still on it.

Reading this, you might get the impression that I am arguing that humans are not intelligent. Nothing further from the truth. I am arguing that we have been grading ourselves on a curve of our own design, and that the curve may be generous in ways we have preferred not to examine. The machines are catching up to a target which is, by some measures, considerably closer to sea level than the brochures suggested.

Late in T. S. Eliot’s The Hollow Men, the world ends not with a bang but a whimper. He was talking about something larger and bleaker than office work, but the image keeps coming back to me when I think about how this particular threshold might be crossed. We were promised a bang: the singularity. We were promised a moment, a milestone, a point at which machines would clearly and unambiguously become as smart as we are and then immediately whoosh past us, leaving the species to sit down and reckon with what had happened. The bang was going to be the achievement of the ceiling; the ceiling was going to be high; the climb was going to be long; and somewhere near the top there would be a flag and a date and a sense of occasion.

On qualifying our own intelligence, we naturally set up camp on the high ground we stood on, since all we could see were plains and valleys beneath us. We, the pinnacle of Creation, have been staring down ever since, admiring the scenery and remarking on the altitude. The species keeps looking down, expecting to see the machines still toiling up the slope below. Before we know it, we will find them unhurriedly walking past us. Then, the feeling that the climb was not necessarily fast, but just short, will slowly dawn on all but the most cynical among us.

Reality will hit us with no flag, no date and no sense of occasion. What it will bring instead is something quieter and considerably less flattering. Machines will have caught up to us by climbing far less than we expected, meeting us where we were all this time. Perhaps, in some domains, that meeting has already happened, and the reason we have not noticed is that admitting it would require revising downward our estimate of the most important thing in the room. Which is, of course, ourselves.

Depending on what we mean by AGI, then, it may indeed be closer than we think. Not necessarily because machines are about to make some unimaginable leap forward, but because the destination may never have been as far away as we confidently claimed.

So perhaps the time for carrying up buckets is over.