The Word "AGI"

September was a remarkable month. In just a few weeks, nearly every major lab released a new frontier model. The capabilities are, once again, a real step up, especially for long-running work - the kind where you hand a model a substantial task and come back hours later to find it done, mostly well. And at least one lab described its new model, in its own announcement, as artificial general intelligence.

So: is it AGI?

I’ve been asked this a lot this month, by friends, by clients, by my family. My honest answer is that I don’t think it’s the right question. But I’ll try to explain why, because the reason isn’t dismissive.

A word that keeps moving

“AGI” has never had a stable definition. For decades it meant, roughly, “a machine that can do any intellectual task a human can.” Then models started doing many of those tasks, and the definition quietly moved. It became “can do most economically valuable work.” Then “can do it autonomously.” Then “can improve itself.” Every time a milestone was reached, it turned out not to be the one that counted.

That’s not because people are dishonest. It’s because we didn’t really know what we meant. We imagined intelligence as one thing, a single line that a machine would one day cross. It turns out to be many things, and machines are crossing them at different times, in a strange, uneven order. Superhuman at mathematics competitions. Still occasionally confused about what day it is.

Words do work

When a company calls its product AGI, the word is doing work. It’s raising money, attracting talent, shaping regulation, setting expectations, sometimes triggering contractual clauses. That doesn’t make the claim false. But it should make us read it the way we’d read any claim made by someone with a large stake in our believing it.

Equally, when critics insist that nothing is AGI and never will be, the word is doing a different kind of work: reassuring, defending, drawing a line around what makes us special. That doesn’t make their claim false either.

Wisdom, I think, is to look past the word at the thing itself.

What I actually see

Here’s what I see, without the label.

These systems can now do a great deal of what I used to do, often faster and sometimes better. They still need someone to decide what’s worth doing, to notice when the result is subtly wrong, and to answer for the consequences. The amount of work they can do on their own keeps growing. The need for human judgment about what and why has not shrunk at all. If anything, it has become the main thing.

That’s true whether or not we call it AGI. And it’s the part that actually changes how I work, how I hire, and how I advise the people I work with.

What I’d focus on instead

Rather than “is it AGI,” I find these questions more useful:

What can it reliably do, in my actual work, today?

Where does it fail, and how would I know?

Who is responsible when it does?

What is it changing about the lives of the people around me, and what do I owe them?

Those questions have answers. They’re harder and less exciting than a single word, but they’re the ones that matter on a Tuesday morning.

A little awe, too

I don’t want to end on caution alone. Whatever we call it, what’s been built is astonishing. Twenty-six years ago I was typing my first programs into a text editor. Today I talk with machines that reason, write, and build. It would be a strange kind of wisdom that couldn’t feel wonder at that.

The ancient prayer says: how great are Your works. I find myself thinking it about these, too. Not because they are divine, but because the minds that made them, and the world that made those minds, are.

Call it whatever you like. Then go back to asking what it’s for.