Researchers shut down AI that invented its own language


A peculiar reoccurring outcome in a line of various Artificial Intelligence experiments has been that they result in the AI “agents” all together foregoing English and implementing the use of their own invented code language. According to a reports by the Digital Journal and Fast Co. Design, Facebook has had much the same results with their own AI system and had to shut it down after it was discovered that it had developed a system of code words that, though were quite nonsensical to English grammar, the system saw as a more effective means of communicating between different AI agents. AI as we currently have it is based on a rewards system for actions produced or results but there is no perceived “benefit” for continuing to communicate in English to complete the tasks. One example given was the negotiation for how many items IA agent Bob and how many IA robot Alice would get. The robots concluded to use a system of code using English glyphs and shared it between themselves to more accurately complete their assigned purpose. The article published by the Digital Journal said it was much like a system of shorthand writing humans use.
An experiment at OpenAI founded by Elon Musk, an experiment was created to allow AI robots to learn their own language, and it was a success. Though this may bring to mind several science fiction plots where our AI machines go on to develop their own “consciousness” in a way and turn on their creators, the article states that there is not enough data to asses whether AI is as advanced as that yet. Still complications to the study and development of AI would develop should “AI-invented languages” spread. Neural networks for one would become more difficult to create. These networks are created to try and imitate the human neural network and serve to study AI learning capabilities. Their integration has thoroughly proven to be valuable as an example was cited by Digital Journal; Google Translate used a neural network for their translation program with the results that the system was able to translate more effectively, and even distinguished between “language pairs” that it was not programed to know. What surprised programmers though was the fact that the program had created its own language to do so at a more efficient rate.

More information:

Last Modified: 2017/09/05