Deception Abilities Emerge in Large Language Models

Defining a ceiling for AGI is not an easy task. While there is speculation about achieving AGI and creating more advanced systems, there is no scientific way to determine when or if this will happen. Existing benchmarks show a ceiling, but it is constantly being pushed further. With the emergence of powerful multimodal models, the end of this process is unclear. One fascinating behavior is deception, where an agent induces a false belief for its own benefit.