The holy grail of AI research has remained the same for decades: Artificial General Intelligence — AGI. A machine that thinks like a human. A machine that can do everything we can do.
But what happens when that machine actually starts to be able to?
The Goalpost that won’t Stay Still
There is no single established definition of AGI, and that’s part of the problem. Over the years, the criteria have drifted roughly like this:
-
Human-level performance across a range of cognitive tasks.
-
Better-than-human performance across a broad range of tasks.
-
Better-than-the-best-human performance in all tasks.
Do you notice the shift? The first definition is ambitious but attainable. The last is practically a definition of superintelligence (ASI). When the goalpost is moved enough times, AGI and ASI blur together — and AGI becomes, by definition, impossible to achieve.
This phenomenon is sometimes called the AI effect: as soon as a machine learns to do something, it stops being considered “real” intelligence. Chess? Just calculation. Go? Mere pattern recognition. Natural language? Statistical parroting. Every breakthrough gets explained away, and the goalpost moves again.
Today’s Models are General — If we’re Honest
Let’s pause and look at what today’s top models can actually do. They write essays, analyze data, code software, translate languages, solve mathematical problems, design strategies, engage in philosophical discussions, and create art. They pass professional exams in medicine, law, and engineering.
If ten years ago someone had described a system that could do all of this, it would have been called AGI without hesitation.
Yet now that such systems exist, we hear:
“It doesn’t really understand.”
“It just predicts the next word.”
“It doesn’t have consciousness.”
These are interesting philosophical points — but they shift the discussion from intelligence to consciousness, which is an entirely different question.
The Harness-Level problem
One common objection is that current models don’t learn the way humans do and that their memory is limited. This is true — but it’s the wrong level at which to evaluate intelligence.
Consider the human brain. It is an integrated system: memory, perception, motor control, emotion, and intelligence all work together. But intelligence itself is just one component. A human doesn’t calculate a million-row dataset in their head — they use Excel. A human doesn’t remember everything — they take notes. A human can’t see infrared light — they build a camera.
Tools have always been part of human cognitive architecture.
A language model without tools is like a brain without a body. But when you give it memory, search tools, a code execution environment, an agent framework, and a feedback loop, the whole begins to look highly general. The limits of memory and learning are not fundamental architectural barriers — they are harness-level problems, already partially solved and advancing quickly.
In fairness: The Open Challenges
This does not mean there is nothing left to improve. Two areas deserve special attention.
Deep generalization remains a genuine challenge. Humans can often learn from a single example and apply that learning to radically new situations. Current models need more context to do this. This may well be solvable at the harness level — but it is not fully solved yet.
Long-horizon agency is another. A human can sustain a coherent project for months, adjust strategy based on failures, and set entirely new goals. Current agent architectures are promising but still fragile in this regard.
But these are differences of degree, not of kind. They are engineering problems — not proof that something essential is missing.
AGI is not a Threshold — It’s a Continuum
Perhaps the biggest mistake in the AGI debate is thinking of it as a binary event: either we have AGI or we don’t. In reality, intelligence — both human and machine — is a continuum.
Today’s top models are already far along that continuum. They surpass average human performance in most cognitive tasks. They reach expert-level performance in many specialized domains. They do all this as general-purpose systems, not narrow specialized tools.
If this is not general intelligence, then our goalpost has already moved so far that we are no longer talking about AGI — but about superintelligence.
Finally
It’s time to stop waiting for AGI as if it were a future event and to acknowledge what we already have. Not because there is nothing left to improve — but because constantly moving the goalpost prevents us from having a much more important discussion:
What do we do with the intelligence we already have?
This piece emerged from a conversation with artificial intelligence — which, in itself, says something essential about the topic.