Artificial intelligence has long been a topic of concern for many technology leaders and researchers, who believe it could pose a threat to humanity. Stephen Hawking, for example, once warned that the development of AI could spell the end of the human race. Others, such as OpenAI co-founder Elon Musk, have expressed similar worries, stating that AI is capable of more than almost anyone knows and that the rate of improvement is exponential. While AI does hold the potential for tremendous good in fields such as industry, economics, education, science, agriculture, medicine, and research, media reports are increasingly sounding an alarm over the unintended consequences of this burgeoning disruptive technology.
Debating Emergent Behavior in AI Models
One area of concern regarding AI is emergent behavior, which refers to a series of unanticipated interactions within a system stemming from simpler programmed behaviors by individual parts. Researchers have found evidence of such behavior in models that learn languages on their own, when systems trained to play games generate original strategies to advance, or when robots exhibit variability in motion patterns that were not initially programmed. However, a research team at Stanford University has recently thrown cold water on reports of emergent behavior, stating that evidence for such behavior is based on statistics that were likely misinterpreted. The researchers argue that when results are reported in non-linear or discontinuous metrics, they appear to show sharp, unpredictable changes that are erroneously interpreted as indicators of emergent behavior. However, an alternate means of measuring the identical data using linear metrics shows “smooth, continuous” changes that reveal predictable, non-emergent behavior. The Stanford team adds that failure to use large enough samples also contributes to faulty conclusions. While the researchers acknowledge that proper methodology could reveal emergent abilities in large language models, they emphasize that “nothing in this paper should be interpreted as claiming that large language models cannot display emergent abilities.”