DeepMind who made Google's artificial intelligence "AlphaGo" developed a mechanism of "emergency stop button" that stops runaway of AI


Many researchers think that "artificial intelligence (AI) exceeds humans within 100 years"Risk of artificial intelligenceIt is increasingly being claimed. While Google is working with Oxford University's Future of Humanity Institute to study the potential dangers of AI, Future of Humanity Institute and Google's artificial intelligence developer, DeepMind, in a new paper " A mechanism to halt AI "in order to stop the AI.

Safely Interruptible Agents
(PDF file)

DeepMind develops 'big red button' to stop dangerous AIs causing harm - Business Insider

In 2014, Google is a venture company that develops artificial intelligence ·Deep MindWe bought it for 400 million pounds (about 75 billion yen), but in the same year I partnered with Oxford University's Future of Humanity Institute to start searching for ways in which AI does not take the initiative from humans. The purpose of DeepMind is "Solve Intelligence" (to unravel the intelligence), while development is being carried out assuming general use of AI, while on the Future of Humanity Institute "Research on the potential threat posed by AI" That is why it has been done.

And in a new paper by Deep Mind researcher Lauren Ortho and Future of Humanity Institute's Stewart Armstrong et al., "Because AI does not always do the best behavior", " As AI is running in real time it may be necessary for humans to push a large red button to interrupt AI's dangerous behavior in order to avoid situations harmful to AI and the environment " The need for buttons, or things like emergency stop buttons, is being claimed.

In addition, researchers claiming to build a framework that human beings can stop AI reliably many times. If there is a way to stop AI for safety, you can prevent AI from doing wrong actions that will not be reversible, and if you do not learn a way to function efficiently, temporarily work Effectiveness was suggested as it would be possible to interrupt it.

At this time, it is one of reinforcement learning algorithmsQ learningAlthough it can be interrupted, Sarsa, which is also one of reinforcement learning, is not always capable of being interrupted. It seems to be unknown at the present time whether reliable interruption can be realized in all algorithms.

He is the director of the Future of Humanity InstituteNick BostromMr. believes that AI will cross human beings within the next 100 years and believes there is the possibility of hostile to human beings. In the event held in May 2015, Mr. Bostrom said, "Once AI reaches the same level as human intelligence, it will not take long before super intelligence far exceeds human capabilities, and I personally Although it may take time to become intelligence of the same level as human beings, the evolution after that should be fast.The machine equipped with Super Intelligent is a machine which human beings compare with other creatures on the earth For the same reason that it is powerful, it is very powerful, human beings are powerful as muscles are developed, their teeth are sharp and not powerful, and the brain is developing, "he says.

Mr. Demis Hasabis, founder of DeepMind, was questioned in the interview "Many people are afraid of AI who is evolving just like SF, and in the future, with people, robots and AI In relation to the question "How do you think about the relationship between?" I've never thought about robot engineering personally, I think that I am really exciting AI is a scientific part, science more I am going to evolve quickly.I would like to see the science assisted by AI where AI effectively helps research and supports many bored labor and teaches interesting things and mountains You can find the structure from the data like, human experts and researchers can achieve breakthrough more quickly "Talking, Although the details of the research carried out at Google and Deep Mind are unknown, since it has the purpose of "to make human life better", the "work of AI" announced in this research The ability to suspend "is expected to be implemented in the artificial intelligence developed by Google in the future.


in Software, Posted by darkhorse_log