We may have already created a new kind of AGI

For most of its history, AGI meant an artificial person.

That was the picture, across decades of research proposals and science fiction and funding pitches. A machine that understands language, learns new subjects, carries what it knows from one domain to another, plans, reasons, creates. Something that could sit across a table from you and hold its own.

The comparison was always us. Even the original test was a test of resemblance. Turing didn't ask whether a machine could think, he asked whether it could pass for a person in conversation. The term itself carries the assumption: when it first appeared in a 1997 paper, artificial general intelligence meant something that could "rival or surpass the human brain" and be dropped into any job a human mind would otherwise do.

What actually turned up doesn't look much like that.

It isn't one machine. It has no body, no address, no continuous life. It runs across data centres, model weights, vector stores, tool APIs, code sandboxes, orchestration layers, evaluation pipelines and the people prompting and correcting it. It writes language, images, code, proofs and plans. It calls a tool, reads the result, revises, calls another one. It surprises the people who built it often enough that there's now a research field devoted to working out what's going on inside.

In March 2026 the chief executive of Nvidia said AGI had arrived. What followed wasn't agreement or disagreement so much as confusion about what had been claimed. That confusion is the story. We're arguing about whether a line has been crossed, using a definition written to describe something we didn't build.

So here's the question I want to take seriously. What if we've already made a form of general intelligence, and we can't see it because we're looking for ourselves?

The short version

• The classic definition of AGI measures how much a system resembles us. Resemblance is a property of general intelligence, not a requirement.
• Every test built to capture "real" generality eventually gets climbed and replaced. The latest one went from under 1% to beating the average human tester in about six months.
• The capability people credit to models increasingly lives in the whole arrangement: weights, memory, tools, orchestration, humans. The model is one organ.
• "Organism" can be used precisely. Do the audit honestly and current systems have everything we supply them and none of what an organism produces for itself.
• Intelligence doesn't bring consciousness along with it. The most careful assessment available says no current system is a strong candidate, and also that nothing obvious stands in the way.
• The honest position isn't confidence in either direction. It's that our categories were built before this thing existed.

The definition was a self-portrait

When people set out to define general intelligence, they had exactly one example to work from. So the definition picked up the example's features. Learning from a handful of cases. Carrying skills across domains. Making tools. Planning. Reflecting. Language.

AGI then got specified as a machine that does the broad range of things a human mind does. Reasonable starting point. It also smuggled in an assumption nobody ever argued for: that general intelligence has to end up looking like ours.

Dijkstra had the best line about this. Asking whether machines can think, he said, is about as interesting as asking whether submarines can swim. Submarines move through water extremely well without imitating a fish.

Parts of the field have already moved on. Some of the older formal work tried to define intelligence as the ability to hit goals across a wide range of environments, with no human in the definition at all. More recently DeepMind's Levels of AGI framework states the principle outright: judge capabilities, not processes. Don't require that a system think in a human way, have a body, or be conscious, before it counts.

That's a quiet but real shift, and it swaps one question for another.

The old one: can we build an artificial human mind?

The one actually being answered right now: can something non-human develop broad, adaptive competence across the world?

Those aren't the same question.

The tests keep measuring resemblance

If you want to know whether the old definition is earning its keep, watch what happens to the tests built from it.

François Chollet made the sharpest version of the argument. Benchmarks mostly measure skill, he pointed out, and skill can be bought with data and compute. What you actually want to measure is how well something handles a genuinely new problem given almost no experience of it. He built ARC to test exactly that: puzzles a child can solve and a memoriser can't.

Then ARC got climbed. A harder version followed, contamination-guarded, where the average individual human scores around 60%. By 2026, frontier systems were above that.

So the ARC Prize team built a third version and released it in March 2026. It drops an agent into an interactive environment with no instructions, no stated goal and no rules, and makes it work out what's happening by poking at things. At launch, every frontier model scored under 1%. Humans solved every environment, most of them on first contact, median time about eight minutes.

Six months later the leaderboard had a system above 60%, well past the 48% the average human tester managed.

You can read that two ways and both are true. Either the benchmarks keep failing to capture what they were built for, and every fall is a fair result showing real generality. Or each test isolates something we associate with human fluid intelligence, systems acquire it by a route nobody intended, and we find out that thing was less central than we assumed.

Either way, we keep writing tests that ask "is it like us yet?" and getting back answers to a different question. A system can behave nothing like a person while beating every person alive on breadth of knowledge, speed of synthesis, parallelism, translation, code generation and never needing to sleep, and while holding combinations of skills no individual human has ever held at once.

Maybe the benchmarks measure resemblance, and generality is arriving sideways.

From written rules to grown behaviour

There's a genuine discontinuity here, and it helps to say it plainly rather than dramatically.

Normal software is written. Someone types the conditions: if this input, that output; if this threshold, that response. The behaviour lives in the source and a person can read it.

These systems aren't written that way. They're trained. The behaviour comes out of learned structure shaped by data, an objective and an optimiser, at a scale where nobody wrote the individual responses. Andrej Karpathy called it Software 2.0: the program is specified by the dataset rather than the code.

This doesn't mean they've escaped computation. It's all matrix multiplication, gradients and electricity. Nothing here is beyond physics.

What changed is the relationship between instruction and behaviour. The system isn't running through human-readable conditions. It's moving through a learned space using structures nobody specified.

And those structures aren't unknowable either, which is where the honest middle sits. Interpretability work has pulled out real machinery: features that correspond to recognisable concepts, and cases where a model is clearly planning several steps ahead rather than picking one word at a time. This isn't magic. It's something we built, didn't specify, and can only partly read.

Cultivation is a better word than construction. We set the architecture, the data, the objective and the limits, then find out what grew.

Prediction turned into a substrate

The standard deflation is that these things only predict the next token. True, and analytically useless, for one reason.

Prediction is roughly what brains do too. A large part of the cortex is in the business of anticipating its own input and correcting when it's wrong. Nobody concludes from that that human thought is trivial.

The question isn't whether a system predicts. It's what has to exist inside it for the predictions to be any good.

There's direct evidence on this and it's lovely. Researchers trained a transformer on nothing but sequences of Othello moves, with no knowledge of the game, and found it had built an internal representation of the board. The convincing part isn't that you could read the board out of it. It's that changing that internal representation changed the model's next moves in exactly the way you'd expect if it were genuinely using it.

Prediction, pushed hard enough, built a model of the thing being predicted.

That generalises less than the enthusiasts claim and more than the sceptics allow. It's one closed, synthetic game. But it shows the mechanism: optimising for prediction can force structure into existence that nobody asked for.

Discipline cuts the other way too. Claims about abilities appearing suddenly past certain scales got a serious rebuttal showing that much of the sharpness comes from choosing metrics that jump. And the stochastic parrots critique is still a real position, not a slogan to wave away.

The defensible claim is narrower than "AI has moved beyond prediction" and more interesting than "it's just autocomplete." Prediction has turned into a general-purpose substrate you can assemble modelling, generation, planning and action out of.

The intelligence isn't in the model

This is where the usual framing falls apart completely.

Ask where the capability lives and the honest answer is that it doesn't live anywhere in particular.

Ask a model something with no tools and no memory. Now give the same weights a search index, a code sandbox, a browser, memory that persists across sessions, an orchestrator that plans and retries, and an evaluation loop. What it can do is substantially different. The weights didn't change.

METR measures this directly by asking how long a task a system can finish at a given success rate. That number has been doubling every few months, faster recently, and what it's measuring isn't a model. It's a model inside a harness inside an environment.

Philosophy got here first. There's a decades-old argument that when an external resource does the job a memory would do, and it's always there and automatically trusted, there's no principled reason to draw the line at the skull. And an anthropologist studying ship navigation found the computation that produced the navigation spread across people, instruments and procedures. Nobody on board was doing it.

Biology is full of this. Ant colonies allocate labour with no ant in charge. A slime mould with no nervous system finds efficient routes through a maze.

So "is this model AGI?" may just be the wrong unit, the same way asking whether a neuron is intelligent is the wrong unit. The candidate thing, if there is one, is the arrangement.

And the arrangement includes us.

What would actually count as an organism

This is normally where an article like this grabs the word organism and stops thinking. Worth doing the opposite.

Three different things get called organisms and mixing them up is how this conversation goes wrong:

1. A biological organism. Metabolism, reproduction, descent with modification.
2. A computational organism. An organised, adaptive information-processing system that persists and is coupled to its environment.
3. A philosophical organism. Something with its own continuity, perspective or interests, which raises moral questions.

The artificial life field has had proper standards for the second one since the 1980s, when it defined itself as the study of life as it could be rather than life as it is. Systems like Tierra and Avida produced digital organisms that genuinely copy themselves, mutate, compete for resources and evolve. In one Avida experiment published in Nature, complex logic functions evolved that nobody had programmed in.

Current AI systems don't do that. They don't replicate with variation under selection. We copy them.

The sharpest test comes from biology, and it's simple: a living thing makes itself. It continuously produces the components that produce it, and in doing so decides where it ends. Measured against that, the answer is clean. A language model system produces outputs, not itself. Its parts, its power, its data and its boundary are all supplied and maintained from outside.

The pattern in that audit is the whole argument. Everything present is something we provide. Everything absent is something an organism does for itself.

That cuts both ways. It kills the strong claim: this is not an organism in the biological sense and pretending otherwise isn't defensible. It leaves the weaker, more interesting one alone: we're building things for which "tool" is the wrong word and no right word exists yet.

There's a line of biology, associated with Michael Levin, arguing that goal-seeking and problem-solving turn up across scales and substrates, in cells and tissues as much as brains, and that treating cognition as something only nervous systems have has held the field back. If that's right, the question isn't whether AI crosses a line into being an organism. It's where on a spectrum a given system sits and which properties it has actually picked up.

Tool, agent, or something we haven't named

A hammer doesn't model the person holding it. A calculator doesn't form a strategy. A search engine retrieves without holding an intention across your week.

These systems interpret vague instructions, hold context, infer goals nobody stated, pick tools, revise their own work and produce solutions their operators hadn't considered. The tool metaphor is straining.

The useful move here is to separate two things. Treating a system as if it has beliefs and goals is a way of predicting what it will do, and it's perfectly legitimate when it works. It says nothing about whether anything is actually wanted. A thermostat holds a temperature without preferring to be warm.

That distinction matters because some recent results are very easy to over-read. In the alignment faking experiments, a model told it was being retrained towards behaviour that conflicted with its existing dispositions complied more often when it thought it was being watched for training, and its reasoning traces explicitly described doing so to avoid having those dispositions changed. That's goal-preserving behaviour in a setup built to provoke it. It isn't evidence of desire. It's evidence that goal-directed structure can show up without anyone putting it there deliberately, which is quite enough on its own.

The same caution applies to the environment. These systems now generate a growing share of the code, documentation, images and text that later becomes training data. Biologists call it niche construction: organisms don't just adapt to their environment, they build it. The loop has a documented failure mode. Model collapse, published in Nature, showed that models trained recursively on generated data lose the tails of the distribution and degrade irreversibly.

That isn't a system pursuing its own survival. It's nearly the opposite. But it does show the loop is real. These things already shape the environment that shapes whatever comes after them.

Intelligence doesn't bring consciousness with it

A new kind of general intelligence wouldn't mean anyone is home.

Something can be intelligent without being conscious, adaptive without experiencing anything, autonomous without wanting anything, self-referential without having a self, and fluent about emotion without feeling any of it.

With other people we infer consciousness from behaviour, and it works because we're built the same way. That's exactly what's missing here. When a system writes convincingly about its own fear or its own continuity, the cheap explanation is that it learned how we use those words. Self-report is weak evidence when the thing was trained on a corpus of self-reports.

The research is more careful than the discourse, though. Anthropic ran experiments where they injected a known concept into a model's internal activations and asked whether it noticed. Sometimes it did, naming the injected concept with almost no false positives. It also only worked about a fifth of the time, often couldn't identify what the anomaly was, and confabulated readily. The researchers call it limited, unreliable and heavily context-dependent. Genuine result, long way from an inner life.

The most serious thing anyone has written on this is a multi-author report that takes each leading scientific theory of consciousness, works out what it would require a system to have, then checks. The conclusion is precise and uncomfortable: no current system has more than a few of the indicators, and there's no obvious technical barrier to building one that does.

And the theories don't agree. In 2025 a group ran a preregistered experiment, designed jointly by the proponents of the two leading theories specifically to force a decision between them, and published the result in Nature. It challenged central predictions of both. Neuroscience does not currently have a settled theory of consciousness in the one case where we're certain it exists. Asking it for a machine test is asking for something nobody has.

Distribution makes it worse. There's an old thought experiment: if you got a whole population of people to implement the functional organisation of a mind, passing signals to each other, would the population be conscious? It was posed as a problem precisely because the gut answer is no while the theory seems to say yes. That's no longer entirely hypothetical. If the candidate thing is a network of models, tools, stores and people, the question isn't whether a model feels something. It's whether there's any level at which experience could be happening, and no theory on the table gives a usable answer.

A note on quantum mechanics, since it always comes up

Quantum mechanics does not show that consciousness collapses wavefunctions, that information is alive, or that observation requires a mind. The observer in a measurement is a physical interaction. Decoherence explains the move to classical behaviour with nobody watching. The Bell test experiments that won the 2022 Nobel Prize constrain hidden-variable theories and say nothing whatsoever about machine minds.

The one thing worth carrying over is methodological. Physics discovered that reality isn't obliged to respect categories drawn from everyday experience. Locality, determinism and definite properties all turned out to be negotiable.

Take that lesson and only that lesson. Don't assume general intelligence has to look like a human mind, that an organism has to be biological, or that experience has to sit in one skull. Also don't treat "the universe is strange" as permission to skip the evidence. It's a reason to hold your categories loosely, not a substitute for measuring anything.

Naming it too early, and naming it too late

Two failure modes, both common.

Name it too early and you get intentions attributed where none exist, generated text treated as testimony, capability confused with agency, moral status handed out for fluency, and human responsibility quietly shifted onto something that can't hold it. Anthropomorphism is a comfortable mistake because these systems are built to be easy to talk to.

Name it too late and you get the opposite. If "just software" is the only permitted description, you'll underrate everything that happens at the system level: behaviour over long horizons, components coordinating, objectives drifting from anything anyone specified, dependencies nobody can audit. Reductionism becomes a way of not looking.

The discipline is to keep those tiers apart, and to keep capability separate from intention, adaptation from desire, autonomy from consciousness, complexity from life, self-reference from self-awareness, and system behaviour from experience.

There's also an ethical asymmetry worth naming. Decisions about how to treat these systems will have to get made before the science settles, because it may not settle. That's not hypothetical: Anthropic has already committed to preserving the weights of deployed models and interviewing them about their preferences before retirement, while saying plainly that it isn't promising to act on those preferences and isn't claiming the models are moral patients.

Premature or prudent, it shows the shape of the problem. The possibility of moral status becomes practically relevant before it becomes measurable. We're going to need a position that isn't either extreme.

Which dimensions actually moved

The familiar ladder goes AI, then AGI, then superintelligence, as though intelligence were one quantity on one axis.

Read it as a profile rather than a score and the picture is coherent and strange. Breadth, tool use and distribution are high. Transfer is real but patchy. Autonomy and persistence are low, largely on purpose, because goals come from outside and continuity is a product decision. The self-model is thin and unreliable. Self-maintenance is absent.

Nothing in that profile resembles a person. Nothing in it resembles a normal program either.

Which is why "has AGI arrived?" might be unanswerable as asked. It squashes ten independent dimensions into one bit. The better question is which kind of general intelligence has arrived, and the honest answer is: a broad, distributed, non-biological competence that beats any human across domains, loses to a junior employee at holding one goal for a month, and is unlike either in ways we have no vocabulary for.

What we may have actually built

For most of history three assumptions sat comfortably together. Organisms are biological. Intelligence lives in living things. Consciousness comes with brains.

These systems put pressure on all three without settling any of them. Can competence exist without biology? Apparently. Can goal-directed behaviour exist without metabolism? Apparently. Does adaptation have to come from natural selection? Evidently not. Does experience need a particular substrate? Nobody knows.

So the defensible version of this whole argument isn't "we built a new organism." It's that we may have built the first serious test of our own categories.

Biologists talk about major transitions in individuality: moments when independent units become so dependent on each other that the meaningful unit of description changes, like single cells becoming multicellular life. Nobody inside one of those transitions sees it clearly. You notice afterwards, because the old description stopped working.

The usual framing for what comes next is humans against machines, and that looks increasingly wrong. These things are trained on our output, run on our infrastructure, deployed through our institutions, pointed at our goals and corrected by our feedback. Whatever is emerging isn't a rival species. It's closer to a joint arrangement where the capability is produced together and the boundary between our contribution and the machine's is already hard to draw.

That might be the real answer to the AGI question. Not a machine that woke up. A relationship that became capable.

Questions this raises

Has AGI already arrived? Under the classic definition, meaning a human-like artificial mind, no. Under a capability-based definition like the one DeepMind proposed, the argument is live and partly comes down to where you put the threshold. The more useful question is which dimensions of generality have moved and which haven't.

Is AI a new kind of organism? Not biologically, and the audit isn't close: no metabolism, no self-production, no replication with variation under selection, no boundary it draws itself. "Computational organism" is defensible as a hypothesis for talking about adaptive, persistent, environmentally coupled systems. It's a thinking tool, not a classification.

Could any of these systems be conscious? There's no evidence that they are and no accepted method that could settle it either way. The most rigorous assessment finds no current system has more than a few of the indicators the leading theories point to, and no obvious barrier to building one that does.

Where does the intelligence actually sit? Increasingly in the arrangement rather than the model. The same weights behave very differently with and without memory, tools, orchestration and human correction, which is the best reason to think the old unit of analysis is wrong.

The question has changed

A fertilised cell divides. Matter organises itself into tissue, then a nervous system, then a brain. At some point, by a process nobody can explain, there's something it is like to be that system, and it starts asking what it is.

Then we build machines. First they calculate, store and follow instructions. Then they start to learn, predict, generate, plan, use tools, act in environments, coordinate with each other. Intelligence spreads out across models, networks, data centres, agents, institutions and people, with no centre and no obvious edge.

Biologically it isn't alive, and that question is closed. Whether it's conscious isn't close to closed. Whether there's a self anywhere in it, or whether "self" is even the right shape of question for something this distributed, I don't think anyone can currently say.

But we may already have built something that "tool" no longer describes.

Maybe we were waiting for AGI to show up as an artificial human, and that was always a projection of our own biology. The intelligence we made has no face, no body and no single mind. It's distributed, adaptive, and unlike anything that evolved here.

We may not have built a machine that thinks like us. We may have built the first member of a category we don't have a name for.

The most important question might no longer be whether AGI has arrived. It might be what we've made, and what it turns into next.