AGI — artificial general intelligence — is the hypothetical point where AI matches or beats humans across most cognitive work, rather than excelling at one thing. Chess engines, image classifiers, even a strong LLM beating exams are “narrow” in this framing; AGI is the claim of generality — one system that could learn your job, more or less whatever your job is. It is simultaneously the stated goal of the major labs and a term with no agreed definition, which explains most arguments about it.
Why nobody agrees
Every proposed test dissolves on contact. The Turing test fell quietly years ago — models pass conversational indistinguishability without anyone conceding AGI. Economic definitions (OpenAI’s charter: “outperforms humans at most economically valuable work”) shift with the economy itself. Benchmark-based definitions saturate: models ace graduate exams yet still fumble tasks a temp worker handles in an afternoon. DeepMind’s “Levels of AGI” framework is the honest response — treating generality as a spectrum (emerging → competent → expert → virtuoso → superhuman, crossed with autonomy) rather than a finish line. On that ladder, today’s frontier models are roughly “emerging AGI”: general but unevenly competent.
The jagged frontier
The consistent empirical surprise is jaggedness: the same model that wins math olympiads and writes production code can lose a novel game a child learns in minutes, or confidently misread a calendar. Skills arrive in spikes, not as one smooth curve. That’s why serious researchers can look at identical evidence and disagree by decades — pointing at the spikes (“expert-level at X!”) or the valleys (“still fails at Y!”) is a choice of emphasis, not of facts. Agents sharpened the question: sustained autonomous work is exactly where generality gaps become visible.
Why the word still matters
Definitional fog notwithstanding, AGI drives real decisions — trillion-dollar datacenter buildouts, lab charters with AGI clauses (Microsoft’s OpenAI deal famously hinges on one), and government safety frameworks. For reading the news, one habit helps: whenever someone says “AGI”, substitute their operational claim — which tasks, at what reliability, replacing whom? If the sentence survives the substitution, it says something. If not, it was marketing.