Ask ten people building AI systems what "AGI" means and you'll get ten answers, and at least three of them will contradict each other on purpose. Ask ten people who write checks for AI companies what it means and you'll get one answer: whatever gets the next round funded.

That's not a cynical throwaway line. It's the actual shape of the problem.

I keep noticing that "AGI" gets used less as a description and more as a lever. It's pulled out when someone wants to justify a valuation, close a talent war, or end an interview question they'd rather not sit with. Nobody defines it first. They just deploy it, and the room nods along because objecting makes you look like you're missing the point, when really you're the only one asking what the point is.

Intelligence according to whom?

Here's a test worth trying on yourself. Next time you hear the phrase, stop and ask: intelligence according to whom? A psychologist measuring how people score on tests built by other people? A biologist watching a bee find its way home from two miles out but get lost two yards from the hive it just left? A theorist who can define intelligence with total precision and then discover the definition can't actually be computed for anything real? These aren't three flavors of the same idea. They're three different projects that happen to share a word, and "general" gets bolted onto whichever one is convenient that week.

Agreeing on nothing, except this

The people actually disagreeing about this aren't cranks on either side. Some think human intelligence itself is nowhere near as general as we like to believe, that we just can't picture the problems we're bad at, which makes the whole premise of "general" intelligence built by us, tested by us, a little circular. Others point out that a system trained to be economically productive is being graded on labor output, not on anything resembling a mind, and that a very good spreadsheet would pass that test without anyone confusing it for a person. Still others think the concept of a threshold to cross is itself the mistake, that intelligence has always been prediction plus the ability to act on it, running continuously from bacteria to whatever we're building now, with no line anywhere for AGI to cross.

None of them are agreeing on what the word means. What they're agreeing on, without meaning to, is that arguing about the word is a waste of everyone's time.

I think that's the actual finding here, buried under all the announcements. When people who disagree about nearly everything else converge on "this term isn't pointing at anything real," that convergence is itself information. Not proof, exactly. But a pretty strong hint that the word survives because it's useful for rallying people, not because it's useful for describing anything.

The cost of the wrong question

And the cost of that isn't abstract. Somewhere there's a research group working on a robot that could help someone's grandmother get out of bed without a caregiver straining their back doing it, and that work doesn't sound impressive at a keynote, so it gets a fraction of the attention and money that a chatbot demo gets. Reliability is boring. Nobody has ever gotten a standing ovation for a system that fails one time in ten thousand instead of one time in a thousand, even though that's the actual, grinding, expensive work of making a technology trustworthy enough to put anywhere near a frail person's body. "AGI" skips past all of that. It promises an arrival instead of a decade of unglamorous nines.

The smaller, better questions

So I've mostly stopped asking when AGI will happen, because I no longer think it's a when question. It's not that I've decided the answer is never. It's that I've decided the question is badly formed, like asking when someone becomes fluent in a language nobody agrees on the grammar for.

The questions I've started asking instead are smaller and much more answerable. Can this system do the thing reliably, not just once, but the ten-thousandth time too? Does what it learned in one setting actually transfer to a different one, or did it just memorize the first setting really well? Who's accountable when it's wrong, and does that answer change depending on which company's servers it happened to be running on that day?

Those questions don't make for a good keynote slide. They also happen to be the ones that decide whether any of this is actually going to be trustworthy enough to matter. I'd rather spend my attention there than on a word that four serious, deeply informed people can't agree on the meaning of, no matter how many times it gets announced as arrived.