For AI, and they will watch TV forever

By the ability of curiosity, artificial intelligence will be able to learn by themselves without any human support.

Currently, most are only used to translate, tag images on Facebook and optimize map navigation based on initial human data. We show it what this sentence means in other languages, how one person will look at different images, and how to detect the best path for a car.

But some AI researchers are looking for ways to give them a sense of curiosity, so they can learn without human guidance. Recent research from OpenAI - an independent, non-profit AI research lab, founded by Elon Musk, Sam Altman, and some other Silicon Valley "players" , works with researchers to from UC Berkeley and the University of Edinburgh - have shown that, when the AI ​​algorithm is given a simple definition of curiosity, no more information is needed by humans, it can discover for itself. 50 video games, and even " breaking" some of them.

But this curiosity also has its price. Scientists at the same time realized, since the AI ​​was allowed to come into contact with new things, it would sometimes deliberately "die" to see the Game Over screen, or be too enamored with the TV and Remote control, seamlessly switch channels, try to find new and exciting things.

What is artificial curiosity?

The definition that the OpenAI group uses for artificial curiosity is relatively simple: The algorithm will try to figure out what its environment looks like in the next frame. When that next frame occurs, the algorithm will be "rewarded" based on its degree of error. The idea is, if the algorithm can predict what's going to happen in the future, it must have experienced the same thing in the past.

That's why AI "characters" play very well but games like Super Mario - the game always moves forward to overcome each "floor".

Is there anything special about the TV?

OpenAI researcher Harri Edward told Quartz that the idea of ​​allowing AI to switch TV channels comes from a fictional experiment called "noisy-TV problem" . The TV noise is completely random, so with the curiosity inherent, the AI ​​will never be able to predict what happens next and it accidentally gets caught up in the act of watching. vi forever.

AIs will never be able to predict what happens next and it accidentally gets caught up in watching TV.

The researchers tested the theory by placing a digital television in a 3D environment and allowing the AI ​​to manipulate a button to switch channels. When the AI ​​finds the TV and starts switching between channels, the new stream of images on the TV makes them irresistible.

Harri Edward said, there are some cases where AI can escape from the TV itself, but that only happens if the AI ​​environment happens unfortunately something more interesting than the next happenings. on television.

Overcome the framework of games

The purpose of this study is not only to overcome video games with AI, but also to better understand how the algorithms describe the world around it. Since these algorithms proved extremely effective when exploring the full range of video games, scientists said, they can also adjust to debug (debug code) more easily or ensure A certain game will not have any problems when playing.