What is AGI really? And how close are we now?
AGI gets promised constantly and defined rarely. What it actually means, why the timelines slip, and a realistic read on how close we really are, now.
Updated June 2026. I first wrote this in 2025, before GPT-5 had shipped and while the AGI conversation was running at full volume. The models have moved on a great deal since, but the argument underneath them hasn't, which is rather the point.
A friend asked me over dinner recently whether we should be planning for AGI or whether it's another bubble waiting to deflate. It's the right question, and it's the kind that separates clear strategic thinking from expensive mistakes. After thirty years of watching the hype cycles roll through, from dot-com to cloud to blockchain, each one the future until it wasn't, I've learned what real change looks like and what merely costs money. So let's take the question seriously. What is AGI, really, and how close are we?
What AGI actually means
Strip away the science fiction and the most useful definition is fairly plain. AGI is an artificial system that can learn, reason, and apply knowledge across many different domains at roughly human level, without being programmed for each task in advance.
Notice what that leaves out. It says nothing about consciousness, feelings, or whether the machine "thinks" the way we do. That's deliberate. What matters is capability: whether a model can move from one kind of problem to another with something like human adaptability, carry what it learns from one field into the next, improve without being retrained, and cope with situations it has never met before. Manage all four and you're general in the way that counts. Today's tools can't, not yet.
Isn't creativity the real test?
People often reach for creativity as the dividing line. AGI will have arrived, they say, when a machine produces something truly original. It's an appealing idea, because creativity feels like the most human thing we do, and a real general intelligence surely should be able to invent, imagine, and surprise us.
But creativity on its own proves less than it seems. A system can generate striking images and still be narrow as a drainpipe. Early general intelligence might match our adaptability without being especially imaginative at all. And originality is slippery to define even in people; everything we make builds on something we've already absorbed. Creativity is better understood as a by-product of generality than its definition. Once a system can learn across domains and reason its way through the unfamiliar, the originality tends to follow.
AGI, ASI, and the AI we already have
It helps to be clear about where today's tools sit, because the three terms get used almost interchangeably and they shouldn't be.
| Aspect | Today's AI (LLMs) | AGI | ASI |
|---|---|---|---|
| Scope | Narrow: language, coding | Broad, human-level | Beyond human in every domain |
| Learning | Static, pre-trained | Continuous, adaptive | Recursive self-improvement |
| Transfer | Weak | Strong | Total |
| Reasoning | Pattern-based | Causal, adaptable | Superhuman |
| Status | In production now | Hypothetical | Theoretical |
Today's models, GPT-5.5, Claude's Opus 4 line and Gemini 3, are far stronger than the ones I first wrote about. They reason better, they handle images and audio and text together, and they're starting to drive robotics and act through tools. But they stay narrow in the ways that matter: pre-trained and static, weak at carrying understanding from one domain into another. AGI would close that gap. ASI, or artificial superintelligence, a system that outstrips us everywhere at once, is a further step again, and for now a purely theoretical one.
What the people building it say, and why to be sceptical
The loudest timelines come from the people with the most to gain from your believing them. Sam Altman talks about progress surprising even OpenAI and points to AGI within the decade. Demis Hassabis puts it five to ten years out. Dario Amodei has suggested two to five, with safety risks climbing alongside. Mark Zuckerberg has promised Meta will build it this decade and open-source it when they do. These are striking claims, made by people raising billions and defending enormous valuations. I'd treat them as advertising with a kernel of truth inside.
It's worth holding the forecasts up against the scoreboard. When I looked at how one concrete AI prediction actually held up a year on, the direction of travel was right but the timing lagged well behind the confident version of the story. That space between what's coming and when is the whole game.
So are we close?
The progress is real, and I don't want to wave it away. Models now show flashes of reasoning that surprise the people who built them. Benchmarks like ARC-AGI-2, which the best systems could barely scratch a year ago, are now being scored in the seventy-per-cent range. That's the kind of number that has people declaring AGI all but here.
What the excitement skips over is everything still missing, and it isn't a short list. True transfer learning remains unsolved. Common-sense reasoning is patchy and unreliable. Continuous self-learning isn't there at all; the models don't improve themselves between training runs. And they form no goals of their own, so they never decide what to pursue. Today's AI feels intelligent because it's fluent and astonishingly well-read, but underneath it's still a statistical mimic rather than a general thinker. For a business betting its future on these systems, that distinction matters more than almost anything else on the brochure.
When does it arrive, and does the date even matter?
Nobody knows, and anyone who tells you they do is guessing. The optimistic case gives us "proto-AGI", something that handles most office work as well as a person, within two to five years. The conservative case puts true adaptability ten years out or more. And there's always the wildcard, where a single breakthrough in memory or reasoning rewrites the schedule overnight. Where you land depends almost entirely on your definition. "AI that outperforms humans at most jobs" might be this decade. "AI that learns and reasons as we do" is probably a good deal further off.
None of which is really the point, though. You don't need to wait for AGI to be disrupted by AI. The tools we already have are reshaping knowledge work, customer service, and content right now. The companies that come out ahead won't be the ones that guessed the arrival date; they'll be the ones that understood what each generation could and couldn't do, and built accordingly. And that's far less about the technology than the people around it. AI maturity, as I've argued before, is a people story long before it's a technology one.
What I'd actually do
So stop betting on timelines and start building AI literacy across your organisation. Get your people genuinely fluent in what these tools do today, where they break, and how to fold them into real work. The next five years may well be the most consequential in the history of technology, but the advantage won't go to whoever picked the right year for AGI. It'll go to whoever kept their strategy grounded in capabilities they could test and measure, rather than the promises made in keynotes and funding rounds. That's how you separate genuine transformation from the next costly distraction.
Quick answers
Is GPT-5 AGI? No. It's a capable narrow system, fluent and multi-modal and useful, but it doesn't learn on its own, carry understanding cleanly from one domain to another, or set its own goals. Being capable isn't the same as being general.
AGI versus ASI, what's the difference? AGI matches human ability across most domains; ASI would surpass us in all of them. We've got neither, and ASI sits further out and deeper into the speculative than AGI does.
When will AGI arrive? Nobody can tell you, and the confident answers tend to come from people raising money. Optimists say two to five years for something that handles most office work; the cautious case is ten years or more for real adaptability. It depends almost entirely on whose definition you accept.
Do I need to wait for AGI before AI changes my business? Not even slightly. The tools already in production are reshaping knowledge work right now, and the advantage goes to whoever learns to use this generation well rather than whoever predicts the next one.