Jan 17, 2021 15:00:00

Is it possible for humans to control highly developed AI?

AI has made remarkable progress, with

Stephen Hawking saying, ' Artificial intelligence will surpass humans within 100 years, ' and AI research leaders ' never worry about the threat of AI. ' There is lively discussion about the rapidly developing threat of AI, saying, 'It's just too much .' Meanwhile, a new study was announced that 'AI with advanced intelligence will be out of control'.

Calculations Show It'll Be Impossible to Control a Super-Intelligent AI
https://www.sciencealert.com/calculations-show-it-d-be-impossible-to-control-a-rogue-super-smart-ai

View of Superintelligence Cannot be Contained: Lessons from Computability Theory
(PDF file) https://jair.org/index.php/jair/article/view/12202/26642

The research announced this time is based on the '

halting problem ' proved in 1936 by mathematician Alan Turing, who is known for decrypting the encryption machine ' Enigma ' used by Germany during World War II. Claims to be out of control.

The halting problem is the question, 'Is there an algorithm that can determine whether any Turing machine outputs (stops) the calculation result or falls into an infinite loop?' In 1936, Turing said, 'Does not exist?' It was proved.

When this halting problem is applied to the control of AI, the program that prevents 'the destruction of the world by AI' actually succeeds (stops) or fails to stop (the program falls into an infinite loop). Since it is not possible to judge whether it is, it is thought that it is impossible to create an algorithm that controls AI.

Manuel Sebrian , a member of the research team, said, 'AI with a high degree of intelligence that affects the fate of the world may sound like a story of the world of fiction, but thinking that is incomprehensible to humans. There is already an AI that deploys and solves problems autonomously. '

In addition, as an example of 'AI that thinks for yourself', AI 'MuZero' that learns how to win without being taught the rules of the game was announced by DeepMind under Google in December 2020.

DeepMind under the umbrella of Google announces AI 'Mu Zero' that learns how to win without being taught the rules of the game --GIGAZINE