Facebook shuts down "out-of-control" project because AI developed a language that humans cannot understand

Facebook's controversial project to develop intelligent conversational robots that can communicate in ways that humans cannot understand has been shut down last month because researchers are worried that it will get out of control. Let's take a look at the relevant content with the mobile phone portable editor.

According to Digital Journal, Facebook's intelligent conversational robot project, which caused widespread controversy last month and was intended to "develop communication capabilities that humans cannot understand," has been shut down.

Last month, Facebook Artificial Intelligence Research Institute (FAIR) used machine learning methods to iterate and upgrade the dialogue strategies of two chatbots. It turned out that they developed their own unique language that humans could not understand. The researchers had to intervene manually.

Once this research was released, it quickly attracted widespread attention and spread on the Internet and in the media. Many people exclaimed: Machines have created their own language, which humans cannot understand, which is terrifying to think about.

But some people pointed out that this is just a bunch of gibberish.

Now, this system has been locked up in a "small black room". According to foreign media reports, Facebook's research institute has stopped research on this project because of "concerns about losing control of these AIs."

This isn’t the first time AI has been found to stray from the English it was trained on and develop new, more effective languages. While these communications may look more like gibberish to humans, from the AI ​​agent’s perspective, they actually make sense.

Communicating in a new language: English lacks “feedback”

The main point of the research paper is that robots can be good negotiators - they can even pretend to be interested in something worthless and then sacrifice it during the negotiation process as if they have made a concession. The part about the conversational agent developing its own language is really only a little bit in the research paper, but it has attracted everyone's attention.

As Fast Co. Design reports, the researchers noticed that the chatbots in the system completely gave up communicating in English, but the upgraded system was able to communicate with other agents. Together, they could decide how to proceed with any task. At first, the words they used seemed incomprehensible, but through long-term observation, the researchers found that these words actually represented the tasks faced by the agents.

For example, in one exchange, two robots, Bob and Alice, used a language they created to communicate. Bob began by saying, “I can can II everything else.” Alice responded, “Balls have zero to me to me to me to me to me to me to me to me to me to.” The rest of the conversation consisted of sentences like this.

Although these word combinations may not seem to make sense at first glance, the researchers believe they reveal how the two robots work and how many times they need to process each word. Bob's repetition shows how he uses this word to convey more information to Alice. If expressed in this way, what he said is a bit like a person saying "I'll have three and you have everything else."

It seems that the agents have discovered that English sentences are not necessary for understanding specific scenarios. At present, AI based on deep learning follows the "reward" principle. Agents will gain corresponding benefits by following the specific behaviors of a system. However, in this scenario, continuing to use English did not bring corresponding rewards, so the agents decided to adapt to a more efficient method of communication.

“In the English language, there is no so-called feedback that the agent can follow,” researchers at Facebook AI Research (FAIR) told FastCo. Designs.

"The AI ​​will deviate from the language that humans understand and create its own 'code language'," FastCo. Designs reported. "For example, if I mention 'the' five times, it means I want to copy this term five times. This is not much different from the abbreviations created by humans."

Other AI developers have also noticed that agents use "abbreviations" to simplify communication. At Open AI, researchers have successfully taught AI robots to learn a language of their own creation.

What is the problem?

If AI continues to create its own language, developers may have problems creating and adopting new neural networks, but it is unclear whether this would allow machines to actually overthrow their developers.

However, these new developments allow AI to work more efficiently and effectively, and could benefit research teams in the long run if they learn new AI-created "abbreviations" as they work, and stay up to date on this new way of communicating.

The above is about the mobile phone portable-Facebook shut down the "out of control" project because AI developed a language that humans cannot understand. If you want to know more related information, please pay more attention to eeworld. eeworld Electronic Engineering will provide you with more complete, detailed and updated information.