Artificial general intelligence (AGI) is the proposed ability of an artificial system to solve general problems in something like the way a human can, rather than excel at one narrow task. The distinction names a goal, not a test that any particular system has passed. section 9.2.3
The label acquired a research identity in 2007, when Ben Goertzel and Cassio Pennachin edited Artificial General Intelligence. It helped move discussion of human-level or superhuman AI beyond the older language of “strong AI” and philosophical speculation. It also brought safe design, friendly AI, and alignment into the conversation without settling how to achieve them. section 9.2.3
Those ambitions later shaped institutions. At OpenAI’s founding in 2015, debates about openness, safety, and the pace of pursuing AGI accompanied a research agenda that still lacked a precise strategic goal. Work on reinforcement learning, game-playing, and then Transformers supplied practical milestones, but none alone defined what would count as AGI. section 1.1
Did scaling produce AGI?
The debate resurfaced around GPT-4. Microsoft researchers tested it on mathematics, law, programming, vision, and reasoning, describing the results as “sparks” of AGI. A competing account called them “embers” of autoregressive training: impressive abilities arising from scaled next-word prediction, not necessarily general intelligence. The disagreement illustrates why broader task performance does not, by itself, settle the AGI question. section 6.1.6
AGI is therefore both a research ambition and a contested yardstick. Supporters can read successive capabilities as progress toward it; skeptics can ask whether working systems meet an ideal of intelligence that remains difficult to pin down. section 9.2.3
