Artificial intelligence (AI) terrifies us partly because it is new and partly because it is disturbingly familiar. We have invented machines that converse, flatter, argue, and improvise with uncanny fluency. Confronted by this synthetic reflection, we behave like early humans seeing a mirror for the first time: We project consciousness behind the glass, then populate it with our expected behaviors, emotions, hopes, and nightmares. In other words, we scare ourselves.
Today’s frontier models are impressively good at writing, coding, and math. Yet they still fail at ingredients of useful artificial general intelligence (AGI), a system that matches or surpasses human capabilities such as consistent reasoning, long-term planning, factual consistency, and reliable performance outside their “experience” (the data they were trained on). We are right to worry about the theoretical dangers of artificial superintelligence (ASI) because of the danger that such a system might outwit its human creators and pursue its own ends, potentially at our expense. While some AI executives have asserted that we’ve reached AGI, these claims are contentious and do not meet more considered definitions of AGI.
Dispatches From The Front 