Understand AGI › Syllabus › Start here
Four sensible, well-informed people can look at the same AI system and disagree about whether it is close to AGI. That is not because three of them are wrong.
If you have followed the AGI conversation for any length of time you will have noticed something odd. Serious people, with access to the same systems and the same research, reach opposite conclusions about how close we are.
The temptation is to assume somebody is being dishonest, or hyping, or pessimistic for effect. Sometimes that is true. But most of the disagreement has a duller and more interesting cause: they are not answering the same question.
AGI is not one claim. It is a family of claims about capability, generality, autonomy, reliability and economic effect, and different people put the important threshold in different places. Until you know which threshold somebody has in mind, their timeline tells you almost nothing.
“Artificial intelligence” was coined in the mid-1950s to describe the general project: machines that could do the things minds do. For its first few decades that is what it meant, and progress was measured against that ambition.
What actually got built was narrower. Systems that played chess, or read postcodes, or approved loan applications - each one excellent at one thing and incapable of anything else. The field delivered enormous value this way while the original ambition receded.
By the early 2000s the word “AI” had drifted so far towards those narrow systems that researchers still interested in the original project needed a new term for it. That term was artificial general intelligence.
So AGI does not name a new idea. It names the old idea, after the old word was taken. The “general” is doing all the work, and it is the part nobody agrees on.
Broadly, the arguments cluster around four different places to draw the line. They are not variations on a theme; they are different tests, and a system could satisfy one while clearly failing another.
Capability
How much is left over?
Economic
How much of the work?
Learning
How little, to learn how much?
Autonomy
How long, unsupervised?
The capability threshold
A system that can perform most intellectual tasks a human can perform. The test is breadth: not how well it does any one thing, but how few things are outside its range.
The economic threshold
A system capable of doing a significant proportion of useful human work. This one is deliberately indifferent to how the system works or whether it is “really” intelligent. If the work gets done, the threshold is met.
The learning threshold
A system that can acquire a genuinely new skill from limited experience and then use it somewhere else. Here the test is not what the system knows but how efficiently it can come to know something new.
The autonomy threshold
A system that can be given a complicated goal and pursue it over a long period without a human steering each step.
Notice that these can come apart. A system might clear the economic threshold while failing the learning one, if it has simply been trained on enough of the world’s work. A system might learn efficiently and still be useless unsupervised.
Take a common headline: “AGI in three years.” Under the economic threshold that is a claim about deployment, cost and reliability, and it could be settled by looking at what work is actually being done. Under the learning threshold it is a claim about a research problem that has resisted attack for decades and may or may not yield on any particular schedule.
Same sentence. Completely different claim, and completely different evidence would count for or against it.
This is why the first question this site asks about any AGI statement is not “is that right?” but “which threshold is that about?” It is a boring question and it resolves a surprising amount of the argument.
Module 1 takes the published definitions - the ones written down by the organisations building these systems and by the researchers studying them - and puts them side by side, with what each one would actually require before it was satisfied.
After that, Module 2 deals with measurement, which is where most of the confusion in public reporting comes from.
Last reviewed
Enrol early
We will email you when each module publishes. You will also get AGI, Made Clear, keeping you up to date on AGI and ASI developments in a field that moves faster than anyone can follow.
Double opt-in. One-click unsubscribe. We never share your address.
AGI Made Clear covers what is happening, the evidence behind it and the claims that are not supported. This site teaches the subject. That one tracks it.
Visit AGI Made Clear →