What AGI means, and why smart people disagree about it

AGI, artificial general intelligence, means a system that can handle the full range of cognitive work a capable adult can handle, including tasks nobody trained it for. That is the shared core of every definition. Beyond that core the agreement stops, and the disagreement is not a detail. There is no accepted test, no agreed list of abilities and no measurement that would let two experts look at the same system and settle the question. When you see a confident claim that AGI is five years away or fifty, most of the distance between those numbers comes from the speaker using a different definition, not from different evidence.

The definitions people are actually using

Four definitions do most of the work in public argument, and they can be satisfied in different orders. A system could pass one and fail the rest.

Notice how much rides on which one somebody picked. On the economic definition, a system that writes excellent code but cannot be trusted alone with a bank login is not AGI. On the task breadth definition it might already be close. The word carries whichever meaning the speaker needed.

Why there is no exam that settles it

The obvious fix is a test. It has been tried, and the pattern of failure is consistent.

Turing proposed the imitation game in 1950: if a person conversing by text cannot tell machine from human, call the machine intelligent. Modern systems do well at that and almost nobody accepts the conclusion, because it turned out to measure fluency and social mimicry rather than general capability. The test was retired by its own success.

Benchmarks have the opposite problem. A benchmark is a fixed set of questions, so once it is published it drifts into the training data of the next generation of models. A high score then means the answers were seen, not that the ability is there, which is why every serious evaluation now worries about contamination and why new benchmarks are built to be unguessable and get saturated anyway.

Professional exams mislead in a third way. Passing a bar exam and practicing law are different jobs. Exams are built to separate prepared humans from unprepared ones, and they assume everything an exam cannot check, such as judgment, client contact and knowing when to stop, is already in place. A machine score on a human exam does not carry those assumptions.

What current systems still cannot do

Whatever definition you favor, four gaps show up repeatedly, and none of them is a small bug.

Continual learning. A released model is frozen. Its weights stop changing when training finishes, so it cannot absorb something new from working with you. What looks like learning inside a conversation is your earlier messages still being visible, and it disappears when the chat ends or gets too long. Being unable to update is also why every model has a date after which it knows nothing.

A stable model of the world. These systems hold facts that contradict each other and do not notice, because nothing requires consistency. Ask about physical arrangements, what is inside a closed container, how objects stack, what a person in the room can and cannot see, and you get plausible answers that fall apart under a follow up. Reasoning modes that work through intermediate steps improve this measurably and do not solve it.

Goal persistence. Give a system a task with twenty steps and errors compound. A small wrong turn at step three quietly poisons everything after it, because there is no separate part of the system checking whether the plan still makes sense. This is the central practical problem in systems that take actions rather than answer questions.

Sample efficiency. A child learns a new word from a handful of exposures. A frontier model reads trillions of words to reach its level of fluency and still needs many examples to pick up an unfamiliar pattern reliably. Whether that gap is an engineering problem or a sign the approach is different in kind is one of the genuine open questions.

There is an older observation behind all of this. The things people find hard, chess and calculus and legal drafting, turned out relatively easy for machines. The things any toddler does, walking into an unfamiliar kitchen and making a sandwich, turned out very hard. Progress on the first list tells you less about the second than intuition suggests.

Why the timelines disagree so much

Serious researchers give estimates ranging from a few years to never, and when they are surveyed the spread is wide and the median date has shifted substantially between surveys. Three disagreements produce most of that range.

The first is about scaling. One camp reads the last decade as a straight line: more data, more compute and more parameters kept producing more capability, so extrapolate. The other camp points out that returns on sheer parameter count have flattened before, that high quality training text is finite, and that a curve is not a law. Both are reading the same chart.

The second is about what is missing. If continual learning and reliable world models are engineering problems, they get solved on an engineering schedule. If they need an idea nobody has had yet, no schedule applies, because you cannot put a date on an unknown insight.

The third is about what counts as arrival. Someone using the task breadth definition can reasonably name a much earlier date than someone who requires sustained economic replacement. They may not actually disagree about any fact.

Incentives sit on top of all three. A near date attracts investment and attention; a far date signals rigor and caution. Neither motive makes a prediction wrong, and both are worth noticing.

How to read a confident prediction

Five questions will tell you more than the headline number.

1. Which definition is being used? If the claim does not say, it cannot be checked and probably cannot be wrong either.
2. What evidence is offered, and is it a benchmark score or behavior in the world? Scores on published tests are the weakest evidence available, for the contamination reason above.
3. What would change the speaker's mind? A forecast with no failure condition is a position, not a prediction.
4. Is the claim about capability or about deployment? A capability can exist for years before anything in your life changes, because integration, regulation and liability move at their own speed.
5. Has the speaker been specific before, and did it happen? Track records are rare in this field and worth a lot when they exist.

The same questions work in reverse on claims that nothing is happening. Dismissals are predictions too, and they are just as often built on a definition chosen to make the answer come out right.

Living with an unsettled question

You do not need a position on AGI to use these tools well, and a position is unlikely to help you. What helps is judging a system on the specific job you have for it: does it do this task, at what reliability, and what does a failure cost me? That question has an answer you can test this week, which no AGI timeline does.

Two habits carry over regardless of how the debate resolves. Check the output where being wrong matters, because a fluent answer and a correct one look identical and certainty in the wording tells you nothing about accuracy. And keep clear about what the thing in front of you is: software whose behavior comes from patterns in data, remarkable at that, and not currently anything else.