{"course":{"slug":"ai-actually-explained","title":"AI, Actually Explained","url":"https://addaly.com/learn/ai-actually-explained","level":"beginner","track":"start"},"lesson":{"slug":"what-agi-means","title":"What \"AGI\" means, to whom","url":"https://addaly.com/learn/ai-actually-explained/what-agi-means","position":90,"of":92,"minutes":9,"answer":"\"AGI\" has no agreed definition (a charter, a contract.","keyIdea":"\"AGI\" has no agreed definition (a charter, a contract, a research taxonomy and a benchmark each mean something different by it) and every test proposed for it has been passed or saturated without the question feeling answered, which is what happens when a measure becomes a target."},"body":{"format":"markdown","text":"## A word that means what its user needs\n\nArtificial general intelligence, in the plain sense, means a system that can do any intellectual task a person can. Nobody uses it in the plain sense. Four uses are current and they conflict.\n\n**The charter.** The largest lab's founding document defines it as highly autonomous systems that outperform humans at most economically valuable work. Note *economically valuable*: not all work, not understanding, not consciousness. A definition about markets.\n\n**The contract.** That lab's agreement with its largest investor was reported in 2024 to define AGI, for the purpose of ending the investor's rights, as a system that generates a hundred billion dollars in profit. A definition about money, which is what a contract is for.\n\n**The taxonomy.** In 2023 a research group at another lab proposed a grid: five levels of performance, from emerging to superhuman, crossed with narrow or general scope. Under it, the chatbots of that year rated as emerging general intelligence: level one of five. A definition that admits degrees and refuses a threshold.\n\n**The benchmark.** Since 2019 a researcher has maintained a test of abstract visual puzzles designed so that a person solves them easily and a system cannot pattern-match its way through. When a model scored above most humans on it in late 2024, spending thousands of dollars of compute per puzzle, the author released a harder version, on which the same models scored in single figures. A definition that moves when reached.\n\nWhen someone says AGI is here, or is years away, ask which of the four they mean. Under the taxonomy it has been here since 2023 at level one; under the contract it has not arrived; under the benchmark it arrived and was redefined.\n\n## Why every test gets passed and nothing feels settled\n\nThe history module started with the imitation game: could a person, typing, tell the machine from a human? In 2025 a controlled study ran that test with the original three-party design (a judge talking to a person and a model at once for five minutes) and found that a leading model, instructed to adopt a persona, was judged to be the human seventy-three per cent of the time. More often than the actual human.\n\nThe response, reasonably, was that the test measures the ability to seem human in five minutes, which is not the thing anyone meant. That is the pattern. A test is proposed as the mark of intelligence. Systems are optimised toward it. It is passed. Everyone agrees, looking at the system, that it is not what was meant. A new test is proposed.\n\nThe mechanism has a name from economics: when a measure becomes a target, it stops being a good measure. The exam that was hard for a general system is easy for a system trained on exams. And because the benchmarks are public and the training sets are the internet, the questions leak into the pile; a model that scores well on a test may have read it. The evaluating-AI course in this catalogue is largely about that problem.\n\n## The saturation record\n\nA test of knowledge across fifty-seven subjects, released in 2020, was the standard measure for four years and is now scored near its ceiling by every serious model. A test of graduate-level science questions written to be unanswerable by searching, from 2023, was passed at above the level of the experts who wrote it within two years. A 2025 test assembled from thousands of experts' hardest questions was the new frontier when written and was being climbed within months. Each was, at release, the thing a model could not do.\n\n## What the question is worth\n\nWhether a system is \"AGI\" turns out to be a question about a word, and the honest answer is that the word is doing commercial, contractual and rhetorical work that no definition can carry. The useful questions are the ones this course has asked: what can it do reliably, on which tasks, at what cost, with what failure, and who is accountable. A system can be nowhere near general and transform a profession; a system can pass every benchmark and be unable to count the people in a photograph.\n\nThe people who built the systems are not agreed on whether they are intelligent, and the first module told you why: the word \"understanding\" was borrowed before anyone had a test for it. Hold the term lightly. When you hear it, translate it into which definition, and then into what the system was actually shown to do."},"check":{"question":"A benchmark designed so that a person passes easily and a machine cannot is scored above the human level by a model, and its author then releases a harder version on which the model scores in single figures. What does this illustrate?","options":[{"text":"That the model cheated by reading the puzzles during training.","correct":false,"tempting":"Reaches for contamination; the puzzles were unpublished, and the model's score was genuine on the test as set."},{"text":"That a measure which becomes a target stops measuring what it was built to, so passing it settles nothing.","correct":true},{"text":"That the model had reached general intelligence on the author's own terms and the author refused to accept the result.","correct":false,"tempting":"Takes one definition as settled; the point is that the author's own test no longer separated the two."},{"text":"That benchmarks are useless for comparing models.","correct":false,"tempting":"Overreaches; benchmarks compare models well, and what they cannot do is define a threshold."}]},"prev":{"slug":"why-it-cannot-be-made-to-refuse","title":"Why a refusal can be trained but not guaranteed","url":"https://addaly.com/learn/ai-actually-explained/why-it-cannot-be-made-to-refuse"},"next":{"slug":"the-questions-still-open","title":"Questions nobody has answered yet","url":"https://addaly.com/learn/ai-actually-explained/the-questions-still-open"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'What \"AGI\" means, to whom', in AI, Actually Explained. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-actually-explained/what-agi-means (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"What “AGI” means, to whom","courseTitle":"AI, Actually Explained","module":{"index":9,"slug":"who-decides","title":"Who controls it, and what is unsettled"},"answer":null,"keyIdea":"“AGI” has no agreed definition (a charter, a contract, a research taxonomy and a benchmark each mean something different by it) and every test proposed for it has been passed or saturated without the question feeling answered, which is what happens when a measure becomes a target.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"a-word-that-means-what-its-user-needs","runs":[{"text":"A word that means what its user needs"}]},{"t":"p","runs":[{"text":"Artificial general intelligence, in the plain sense, means a system that can do any intellectual task a person can. Nobody uses it in the plain sense. Four uses are current and they conflict."}]},{"t":"p","runs":[{"text":"The charter.","b":true},{"text":" The largest lab’s founding document defines it as highly autonomous systems that outperform humans at most economically valuable work. Note "},{"text":"economically valuable","i":true},{"text":": not all work, not understanding, not consciousness. A definition about markets."}]},{"t":"p","runs":[{"text":"The contract.","b":true},{"text":" That lab’s agreement with its largest investor was reported in 2024 to define AGI, for the purpose of ending the investor’s rights, as a system that generates a hundred billion dollars in profit. A definition about money, which is what a contract is for."}]},{"t":"p","runs":[{"text":"The taxonomy.","b":true},{"text":" In 2023 a research group at another lab proposed a grid: five levels of performance, from emerging to superhuman, crossed with narrow or general scope. Under it, the chatbots of that year rated as emerging general intelligence: level one of five. A definition that admits degrees and refuses a threshold."}]},{"t":"p","runs":[{"text":"The benchmark.","b":true},{"text":" Since 2019 a researcher has maintained a test of abstract visual puzzles designed so that a person solves them easily and a system cannot pattern-match its way through. When a model scored above most humans on it in late 2024, spending thousands of dollars of compute per puzzle, the author released a harder version, on which the same models scored in single figures. A definition that moves when reached."}]},{"t":"p","runs":[{"text":"When someone says AGI is here, or is years away, ask which of the four they mean. Under the taxonomy it has been here since 2023 at level one; under the contract it has not arrived; under the benchmark it arrived and was redefined."}]},{"t":"h2","id":"why-every-test-gets-passed-and-nothing-feels-settled","runs":[{"text":"Why every test gets passed and nothing feels settled"}]},{"t":"p","runs":[{"text":"The history module started with the imitation game: could a person, typing, tell the machine from a human? In 2025 a controlled study ran that test with the original three-party design (a judge talking to a person and a model at once for five minutes) and found that a leading model, instructed to adopt a persona, was judged to be the human seventy-three per cent of the time. More often than the actual human."}]},{"t":"p","runs":[{"text":"The response, reasonably, was that the test measures the ability to seem human in five minutes, which is not the thing anyone meant. That is the pattern. A test is proposed as the mark of intelligence. Systems are optimised toward it. It is passed. Everyone agrees, looking at the system, that it is not what was meant. A new test is proposed."}]},{"t":"p","runs":[{"text":"The mechanism has a name from economics: when a measure becomes a target, it stops being a good measure. The exam that was hard for a general system is easy for a system trained on exams. And because the benchmarks are public and the training sets are the internet, the questions leak into the pile; a model that scores well on a test may have read it. The evaluating-AI course in this catalogue is largely about that problem."}]},{"t":"h2","id":"the-saturation-record","runs":[{"text":"The saturation record"}]},{"t":"p","runs":[{"text":"A test of knowledge across fifty-seven subjects, released in 2020, was the standard measure for four years and is now scored near its ceiling by every serious model. A test of graduate-level science questions written to be unanswerable by searching, from 2023, was passed at above the level of the experts who wrote it within two years. A 2025 test assembled from thousands of experts’ hardest questions was the new frontier when written and was being climbed within months. Each was, at release, the thing a model could not do."}]},{"t":"h2","id":"what-the-question-is-worth","runs":[{"text":"What the question is worth"}]},{"t":"p","runs":[{"text":"Whether a system is “AGI” turns out to be a question about a word, and the honest answer is that the word is doing commercial, contractual and rhetorical work that no definition can carry. The useful questions are the ones this course has asked: what can it do reliably, on which tasks, at what cost, with what failure, and who is accountable. A system can be nowhere near general and transform a profession; a system can pass every benchmark and be unable to count the people in a photograph."}]},{"t":"p","runs":[{"text":"The people who built the systems are not agreed on whether they are intelligent, and the first module told you why: the word “understanding” was borrowed before anyone had a test for it. Hold the term lightly. When you hear it, translate it into which definition, and then into what the system was actually shown to do."}]}],"check":{"question":[{"text":"A benchmark designed so that a person passes easily and a machine cannot is scored above the human level by a model, and its author then releases a harder version on which the model scores in single figures. What does this illustrate?"}],"options":[{"text":[{"text":"That the model cheated by reading the puzzles during training."}],"correct":false,"tempting":[{"text":"Reaches for contamination; the puzzles were unpublished, and the model’s score was genuine on the test as set."}]},{"text":[{"text":"That a measure which becomes a target stops measuring what it was built to, so passing it settles nothing."}],"correct":true,"tempting":null},{"text":[{"text":"That the model had reached general intelligence on the author’s own terms and the author refused to accept the result."}],"correct":false,"tempting":[{"text":"Takes one definition as settled; the point is that the author’s own test no longer separated the two."}]},{"text":[{"text":"That benchmarks are useless for comparing models."}],"correct":false,"tempting":[{"text":"Overreaches; benchmarks compare models well, and what they cannot do is define a threshold."}]}],"order":[0,2,1,3],"why":null},"prev":{"slug":"why-it-cannot-be-made-to-refuse","title":"Why a refusal can be trained but not guaranteed","minutes":9},"next":{"slug":"the-questions-still-open","title":"Questions nobody has answered yet","minutes":10}}}