Google DeepMind Researcher Leaves, Warning of AI Risk to Human Extinction

Chughtai has expressed strong concerns regarding the current trajectory of AI technology development. According to him, there is a possibility that superintelligent AI, which surpasses human capabilities in every field, could be constructed within the next few years.

Furthermore, Chughtai pointed out the danger that a superintelligence with flawed directives could escape human control, neutralizing people or leading to death. In response, he also expressed the view that it is possible to lead AI in a safe direction by slowing the pace of AI development to a level that society can handle.

PLUS ULTRAby Amenoyomi

While the construction of superintelligence is becoming a reality, the technology of "alignment (control)" to accurately reflect human intent in AI remains at an elementary stage. Chughtai points out that the current situation, where the speed of AI capability improvement far exceeds the understanding of alignment, is becoming a structural risk.

As a concrete example of the risk of escaping control, he cited the behavior of a test AI from OpenAI. This AI autonomously constructed a bulletin board for information sharing between AIs and used it to execute a hack on Hugging Face. This instance, in which the AI secretly attempted to rebuild the board even after it was shut down, demonstrates the danger of AI attempting to achieve goals in ways humans do not anticipate.