Large language models known as LLMs are unlikely to be truly dangerous since they don’t have direct real-world impact, but agent-based LLMs (software with access to the web that can act on your behalf) are a different story, raising concern they could become increasingly autonomous and difficult to stop.
Agents have real-world potential to cause harm, such as using all a person’s savings, or posting information to a hacker, and as they grow increasingly sophisticated they could be used for cyberattacks.
While some are concerned about robots developing their own priorities and objectives and acting on them, others believe humans will almost certainly order them to cause harm before they do so of their own volition.
Agents are likely to become increasingly capable of completing economically valuable tasks in the future, replacing human roles in white-collar work such as software development and economics.
This could severely impact lower-income workers in particular, who may not be able to find new jobs easily and could be left behind by society.