A friend of mine works at a robotics startup. The other week he messaged me, practically vibrating with excitement: their robot had finally learned to plug a cord into a wall socket by itself. He was so thrilled he could have run a victory lap around the building.
You read that right. Some of the sharpest people in the business had spent months teaching a machine a move any three-year-old can do with its eyes shut — and they were over the moon about it.
Meanwhile, open your phone and you get the exact opposite. "AGI arrives in two years, humanity wiped out." "General AI is already here." On one side, champagne for a plug in a socket. On the other, a countdown clock on human civilisation. How can both worlds be true at once?
That's what I want to untangle today. This so-called superintelligence — you'll have run into the acronyms, AGI (artificial general intelligence), ASI (artificial superintelligence) — is it coming or not? And if it is, how much of it has actually arrived?
What intelligence really is — inner strength, not fancy footwork
To argue about this seriously, we first have to pin down what "intelligence" even is, or the whole conversation slides off the rails within three sentences.
In AI research, the one definition most people can roughly agree on comes down to a single idea: generalisation. In plain terms, it's the knack of taking one lesson and running with it. You teach the thing one thing, and if it has genuinely understood, it can carry that over to a hundred things you never taught it. (Regular readers may remember I made this same point a couple of columns back: for humans and machines alike, "understanding" is really about squeezing a transferable pattern out of a pile of specific examples.)
Let me put it in kung fu terms. What makes a real master isn't how many set routines he has memorised. It's that his inner strength runs deep enough to improvise — learn one style of swordplay today, hand him a saber tomorrow and he still looks the part. The opposite is the man who has drilled a single flashy routine to perfection. His forms look dazzling, but change the situation and he falls apart. We don't call him a master. We call him all show.
By that yardstick, AI has built genuine inner strength in the "literary" arts — language, text, writing code. Open ChatGPT and ask it anything, from astronomy to history to maths, and it may not ace every question but it's rarely far off. It never saw this year's SPM papers in training, and it still answers them convincingly. That is generalisation, and almost nobody in the field disputes it.
But switch from the literary arts to the martial ones, and the whole story changes.
The first art it hasn't mastered — making AI move
Back to that wall socket. Why would some of the smartest people alive lose their minds over something any toddler does without thinking?
Here is where someone objects. Hang on — Boston Dynamics robots were doing backflips years ago. Unitree's machines danced at China's televised New Year gala and turned up on variety shows. Beijing just held a half-marathon for humanoid robots. Musk's Optimus is in the headlines every other day.
I won't pretend otherwise: "embodied AI" — robotics, in plain English — has moved astonishingly fast these past two years. But here is the catch. Almost all of these robots have been drilled to death inside one fixed setting. They are that flashy single routine, polished to a mirror shine. Take the robot that just breakdanced, ask it cold for a cha-cha, and it freezes on the spot, the gap between it and a real human suddenly glaring.
A real dancer is different. She may not top the leaderboard in every style, but a waltz today and a tango tomorrow — the feel of it transfers through her body. That "learn one, get the neighbouring ten for free" is the very core of intelligence, and it is exactly what robots still lack. The people doing the foundational research will tell you: just getting a two-fingered gripper to generalise is still a daily exercise in falling on your face.
The second art it hasn't mastered — making AI see
And forget moving for a moment. Even just seeing — AI's command of that is a world away from its command of language.
My mother loves photographing the plants around the house and showing them to ChatGPT, asking how to prune this one, how to feed that one. It answers beautifully. So no, AI isn't blind. The problem is, once again, generalisation.
The clearest example is self-driving. We have been promised it for over a decade, and to this day it runs smoothly in only a handful of cities, never quite spreading everywhere. A self-driving AI trained in America, moved to another country — or even just another city — tends to fall apart and make a mess of things. Now think about a person. Take a driver who earned their licence in the US and drop them into Kuala Lumpur traffic, which is, let's be honest, on hard mode. We'd still hand them the wheel without much worry. What they learned was never "how to drive in America." It was "how to drive." Change the city, and it carries over. That gap, between carrying over and not, is the single most dangerous chasm in AI today.
Which is exactly why conversations about AI lose focus so easily: we are too quick to be dazzled by the party tricks. Yes, AI can now generate an eerily convincing three-minute period drama, and belt out a song I will have on repeat all week — but however gorgeous the forms, that still isn't the same as inner strength in those domains. What's missing is, once again, generalisation and transfer. And this isn't just me talking. Yann LeCun, a Turing laureate and one of the three godfathers of deep learning, and Fei-Fei Li, widely called the godmother of AI, have both stepped away from language models to raise serious money and start over, building AI that can truly see the world and, eventually, act in it. If even they are starting over from scratch, you can imagine how deep the chasm runs.
One recipe, two racetracks
So why did the literary art get mastered first? For that, we need to talk about how the language model earned its inner strength in the first place.
About a decade ago, a few beautiful ideas landed in the research world — the key one being self-attention, and the Transformer model that grew out of it (the technical guts are a story for another day). In short, over the last ten years we finally worked out a method for forging a peculiar kind of mechanical brain, one that happens to be freakishly good at language and text. OpenAI, back then, essentially bet the house on that method's potential — and it actually worked. Half foresight, mostly luck.
Here is the question that follows. Can the very same method be carried over, untouched, to train the art of seeing, or the art of moving? That single question splits the road into two racetracks, and two camps.
The hawks say yes. Just as OpenAI once hammered away at language, they are betting this was never a question of method. Take this mechanical brain, feed it enough data and enough raw compute, and one day it will simply break through and ascend. Half of Silicon Valley is all in on this track. And to keep the thing fed, an entire supply chain has sprung up — armies of workers across India, China and Southeast Asia churning out its training data by hand, and for the physical tasks, quite literally so. The AI future, it turns out, is handmade.
The doves, LeCun among them, shake their heads. We probably haven't even found the right method yet, they say; wrong path, and no amount of data or compute will get you there. I'm no authority, but from what my years in the field have taught me, I lean dove: the real breakthrough, I suspect, can't be skipped. Which is why I keep thinking that so many of these me-too startups, with no secret art of their own, have no real moat in the long run. Once the tide goes out, they'll be exposed. (I droned on about exactly this last time, on the AI bubble.)
So, is superintelligence here or not?
Back to the original question. My answer is still the same. It's here, and it isn't.
In the "digital world" of language, text and code, AI has genuinely arrived, and it is charging forward faster than anything we have ever seen. The shock to society is real, and I only hope we all manage a soft landing. But step outside language, into the rest of AI's vast territory, and the key that unlocks the breakthrough is one we most likely haven't even found. So true general intelligence? Not yet.
And yet the two are tangled up in each other. Even a dove like me has started to feel a nervous flutter lately, that the key might turn up sooner than we think. Because with language AI now lending a hand, today's AI can search for solutions, test them, and correct its own mistakes at a speed no human has ever matched. In a sense, we are teaching AI how to train itself. When will that day come? Nobody knows. Could be a year, could be ten, could be a century.
As for my friend — the next time he's thrilled enough to run a victory lap over some small robotic step, it may not be because the machine learned to plug in a socket. It may be because it worked out, on its own, how to learn to. That's the day we'll actually need to sit down and talk. Until then, don't panic.