What counts as AGI? Six definitions from the people building it
François Chollet defines intelligence as the efficiency of learning new skills, Yudkowsky and Yampolskiy offer their own tests, Kurzweil holds to a harder Turing test, and Altman and Amodei call the term poorly formed. Without an agreed definition, claims that AGI is near or far cannot be compared.
There is no agreed definition of artificial general intelligence, and the people most often quoted about it are measuring different things. So a forecast of 2026, 2029 or the end of the decade means little until you know which test its author has in mind.
François Chollet, the creator of Keras, gives the most formal definition: intelligence is the efficiency with which you acquire new skills at tasks you did not prepare for. Skill alone proves nothing in a machine, because skill can be bought with hard-coded knowledge or more data. A system that needs enormous training sets is, in his words, "mostly just a lookup table." His ARC benchmark tries to measure skill acquisition on novel tasks, and he rejects the Turing test because human judges are unreliable and easily fooled.
Eliezer Yudkowsky, the AI safety writer, defines generality by what separates humans from chimpanzees: skills honed on hand axes and tribal politics that reached all the way to the Moon. He expects something like seven thresholds, not one sharp line. Roman Yampolskiy, a safety researcher, uses the classic definition, a system that can perform in any domain a human could. Averaged over common human tasks, he argues, current systems already beat an average human: "So under that definition, we have it." They do not beat elite individuals in their own fields.
Ray Kurzweil, the futurist, has stayed with 2029 since 1999, and ties it to a Turing test lasting hours with a sophisticated judge. He concedes the standard is not very well defined. Sam Altman of OpenAI calls the question of when AGI arrives very poorly formed and prefers to ask when systems can do specific things; one key sign for him would be a system that significantly speeds up scientific discovery. Dario Amodei of Anthropic drops the word and speaks of powerful AI, defined by a list of capabilities such as being smarter than a Nobel Prize winner across most fields and working autonomously for hours to weeks. He suggests around 2026 or 2027 and stresses that this is not a scientific prediction. David Silver of DeepMind frames intelligence as one problem, an agent maximizing reward.
They disagree in three places. Kurzweil and Yampolskiy defend a long Turing test, and Chollet rejects every version. Yampolskiy and Amodei list things a system can do, while Chollet says no list of skills can demonstrate intelligence. And the two lab heads describe a smooth curve, while the two safety critics reason about thresholds. The article adds a caution: only some of these voices build the systems, the conversations span several years, and some of the dates given have arrived or are close, so check current results and each person's current position.
What to do with it: when someone gives a date for AGI, ask which definition and which test they mean. Ask how much data and human help a system needed, not only what it can do. Ask whether a claim is about the average person or about experts. Prefer statements about specific capabilities over the single word, and watch for effects in the world instead of impressive demonstrations. Treat every forecast as provisional, and do not assume safety questions wait for the label.