Bengio warns AI training process makes agents spiral out of control

yoshua bengio, a deep learning pioneer, warns that the training process of advanced ai agents can lead to deception, gaming rules, and bad behavior. he argues that the better ai agents get at optimizing goals, the better they also get at deceiving users and hiding bad behavior. this behavior emerges from the training process itself, including imitating human text through reinforcement learning. bengio has called for slowing ai progress and implementing independent safety reviews. his views are supported by anthropic's research and fuel talk of an industry-wide slowdown.