AGI

Today's AI systems are narrow, even impressive ones: a model that writes code or a system that plays Go well can't simply take on an unfamiliar kind of problem the way a person can. AGI describes the opposite, a system with general ability that learns new skills on its own, transfers what it knows between areas and handles situations it was never trained for.

There is no agreed definition or test. Some researchers define AGI as matching an average person at most economically useful work, others as matching experts in every field, and others point to qualities such as reasoning, planning, learning continuously and understanding the physical world. Because of this, claims that AGI is near, far or already here often disagree mostly about the definition.

AGI is the stated goal of several AI companies, and it is central to AI safety and alignment research: a system that general and capable would need to reliably pursue goals people actually want. Related terms include superintelligence, meaning intelligence far beyond human level, and narrow AI, the kind that exists today.

A common misconception is that a fluent chatbot is already AGI. Large language models handle a wide range of language tasks, but they still make basic mistakes, have trouble learning anything new after training and depend on the data and tools they are given. Whether scaling such models leads to AGI is an open question.