A sharp reader wrote to me this week with a question better than most of the answers I have read on the subject: what actually separates AI from AGI?
It is a fair thing to be confused about. The two terms turn up in the same sentence, usually from someone with a product to sell, and the distance between them gets described either as a rounding error or as the end of the world. Both descriptions are marketing.
So here is the boring version, which happens to be the true one.
The one letter
Artificial Intelligence is the broad family — anything where a machine does something we would call intelligent if a person did it. Recognising a face. Ranking a feed. Flagging a claim for review. Writing a paragraph. Your phone has been doing this for a decade. The spam filter is AI. The thing that quietly fixes teh is AI.
Artificial General Intelligence adds one word, and the word does all the work.
General means the same system handles problems it was never built for. A broader training set does not get you there. The system has to walk into an unfamiliar domain, work out what the rules are, and become competent at it without an engineer retraining anything.
A useful way to hold the difference: today's systems are extraordinarily good at tasks that resemble something in their training. AGI would be good at tasks that resemble nothing.
Everything else — the timelines, the essays, the very confident YouTube videos — is commentary on that one distinction.
What is actually under the hood
The mystery is doing a lot of unearned work in public conversation, so let me be specific about the machinery. It is less magical and more interesting than the mystery.
A large language model is an enormous collection of numbers, called weights, arranged in layers. During training it is shown vast quantities of text and asked, over and over, to predict what comes next. Every time it is wrong, the weights nudge a little. Do that for months across tens of thousands of chips and the weights end up encoding something genuinely remarkable — statistical structure deep enough that it starts to look like understanding.
The architecture underneath is the transformer, and its central trick is attention. For every word it processes, the model works out which of the other words in view matter, and how much. That is the mechanism that unlocked the last eight years of progress.
Then comes a second phase, which is where most recent gains actually live. After the base model learns to predict text, it gets trained again — on human preferences, on verified solutions, on its own reasoning attempts that turned out to be correct. This is what converts a text predictor into something that will work a problem step by step rather than blurting the first plausible answer. When you see a model "thinking" before it responds, you are watching the product of that second phase.
A model that has read enough code, enough medicine, enough law and enough poetry starts to look general, because whatever you ask, something in that training set rhymes with it.
I run into this constantly in healthcare work. Ask a good model to explain a benefit end date on an enrollment file and it will do it well, because thousands of pages describing exactly that exist somewhere in what it read. Ask it to reason about an edge case that shows up in one payer's implementation and has never been written down anywhere, and the confidence stays exactly where it was while the accuracy quietly collapses.
The steady confidence is the tell. A system that had genuinely generalised would know the difference between remembering and improvising.
Three constraints sit underneath all of this, and they matter more than any benchmark score.
The weights do not change while you talk to it. Training happened, then it ended. Everything the model picks up during your conversation lives in the context window and disappears when the session closes. It cannot get better at your problem by working on your problem.
The context window is working memory rather than memory. It is generous now — large enough to hold a small library — but it is a scratchpad, re-read from scratch on every turn, and gone at the end.
There is no internal signal for "I actually know this." Nothing separates solid recall from a pattern that simply completes nicely, which is why a wrong answer arrives in precisely the same tone as a right one.
It cannot get better at your problem by working on your problem.
None of that makes these systems less useful. I have built production systems on them and written about what survives contact with real users. It does mean the word general is carrying a claim the machinery has not earned yet.
The honest scoreboard
Put a frontier model next to a six-year-old and the comparison stops being flattering to either.
| Capability | Frontier model, 2026 | A six-year-old |
|---|---|---|
| Recall across domains | Superhuman | Poor |
| Speed | Superhuman | Slow |
| Learning a rule from one example | Weak | Strong |
| Noticing that it is lost | Weak | Strong |
| Building a model of a new game by playing it | Weak | Strong |
| Carrying yesterday into today | None by default | Automatic |
The columns are measuring genuinely different things, which is the whole point. We built something that read more than any human ever will and still cannot do what a child does on a playground in an afternoon — watch an unfamiliar game, infer the rules, and join in.
Worth separating from all of this: ASI, artificial superintelligence, means a system meaningfully beyond human capability across the board. It sits past AGI in every serious framing, and it is where most of the apocalyptic writing lives. Conflating the two is the most common move in a bad explainer.
Why the finish line keeps moving
In 2019 François Chollet built a benchmark called ARC, on a deliberate premise: puzzles simple enough for a child, designed so that having read the internet gives you no advantage. Each puzzle shows a few examples of a transformation and asks you to apply it to a new case. Training data cannot help, because every puzzle is novel by construction.
For years models scored close to nothing while humans found it easy. Then models started catching up, so ARC-AGI-2 arrived. Models climbed that too. ARC-AGI-3 launched in March 2026, and by September the newest frontier systems were already being evaluated against it.
Each time a benchmark falls, the field learns something uncomfortable — the benchmark was measuring something narrower than general intelligence. So a harder one gets built and the goalposts move again. Less dishonesty than the predictable result of trying to measure a thing nobody can define.
This is the honest reason your instinct to distrust confident timelines is a good instinct. When the target is defined by whatever the current systems cannot yet do, "we are two years away" and "we are twenty years away" can both be defensible, depending entirely on the definition you started with.
Every lab has quietly written its own definition
This is the part most coverage skips, and it explains the landscape better than any leaderboard.
| Lab | How the goal gets framed | The bet underneath |
|---|---|---|
| OpenAI | Highly autonomous systems that outperform humans at most economically valuable work | Scale, plus distribution to hundreds of millions of people |
| Anthropic | Largely avoids the term; talks about powerful AI and what it can be trusted with | Safety and interpretability research as the thing that makes capability deployable |
| Google DeepMind | A published ladder — levels from emergent to superhuman, crossed with narrow and general | Breadth across modalities, on top of a deep reinforcement-learning lineage |
| Meta | Capability as infrastructure, shipped as open weights | Commoditise the layer, win the ecosystem around it |
| DeepSeek, Qwen and the open-weight labs | Comparable capability at a fraction of the cost | Efficiency, rather than frontier scale, is the binding constraint |
Each definition happens to describe the race that particular lab is best positioned to win. That is ordinary strategy rather than cynicism, and it is worth knowing before you accept anyone's timeline as neutral.
The DeepMind framing is the most useful one for a reader, because it refuses the binary. Instead of asking has AGI arrived, it asks two questions at once: how capable, and how broad. A system can be superhuman and narrow — a chess engine has been for decades. A system can be broad and mediocre, which is roughly where the good general models sit today. Capability and breadth move independently, and almost every confused argument about AGI comes from collapsing them into one axis.
I used to think the gap would close through raw scale. Bigger models, more data, and the generality would emerge.
What actually shifted my thinking was watching where the recent jumps came from. Not from size alone — from teaching models to spend more effort at the moment of answering, to check their own work, to use tools, to keep going when the first attempt fails. The interesting progress moved from what the model knows to what the model does before it replies.
That is a real advance, and I use it every day. It is also a different thing from generality. A system that reasons harder inside a domain it has seen is still a system inside a domain it has seen.
What I would actually watch for
Ignore the announcements. Four things would tell you something real has changed:
- Learning that persists. A system that is measurably better at your work in month three than in month one, without anyone retraining it.
- Calibrated uncertainty. A model whose confidence drops when it is out of its depth, reliably enough that you could route work on the signal.
- Transfer to genuine novelty. Competence in a domain that has no meaningful presence in the training data, reached without human hand-holding.
- Long-horizon coherence. Holding a goal across days and hundreds of steps, recovering from its own mistakes, without a person steering it back.
Every one of those is a research problem today. None is obviously impossible. The distance between them and the systems we have is the distance the word general is quietly asking you to ignore.
So: what should you tell someone who asks?
That AI is here, it is genuinely useful, and it is narrow in a way that is easy to miss because the surface is so broad. That AGI is a moving target which every serious lab defines differently, mostly in ways that flatter their own approach. And that the people most worth listening to are the ones who tell you which definition they are using before they give you a date.
The systems we have read more than any human ever will, and still cannot do what a child does on a playground in an afternoon.
The honest position is that nobody knows. Which is a less satisfying answer than the videos give you, and a far better place to think from.
If you want to test any explainer you come across, ask one question of it: does this person define what they mean by AGI before predicting when it arrives? Almost nobody does. The ones who do are worth your time.
To the reader who asked — thank you. Questions like that one are the reason this site is worth keeping open to everybody. If something here is still unclear, write again.
