AI develops "self"? Scientists: Even desires are artificially set

(Original title: Artificial intelligence develops "self"? Scientists: They even set desires manually)


On June 15, according to Techcrunch, a technology blogger, researchers at the Facebook Artificial Intelligence Research (FAIR) have created artificial intelligence models that have been given the ability to negotiate with humans to negotiate with people.

My mother was very sad when she heard of the news. She believes that even "bargaining" can make artificial intelligence come, which greatly deprives her of the pleasure of going to a food court to buy food.

In the FAIR report, it was also mentioned that they had to replace the model once because the “agent has developed its own negotiation language”.

Many media reported the episode in this study with the headline "Artificial intelligence has developed its own language."

So did artificial intelligence really develop its own language? Is it not possible to quickly develop a language that people can't understand, and start rebelling?

Experts in the field of artificial intelligence, Fu Ruiji, Vice President of the Beijing Branch of the HKUST Academy of Sciences, answered these questions for us.

ask:

Hello teacher, no matter if it is Siri or Microsoft Ice, many friends have been accustomed to artificial intelligence can chat, you can also pretend there is a man / girlfriend in the phone.

So why did the artificial intelligence "negotiation" get such high attention? What is the difference between "negotiation" and "chat"?

answer:

The biggest difference is that negotiations need to compete with “adversaries” to maximize their “own” interests. Previous chat features did not have this feature. Robots can follow users’ words or collaborate with users to complete a task ( For example: booking air tickets, inquiries, etc.).

ask:

Can it be understood that chat is a pastime of irrelevant decision-making, and negotiation is to analyze, judge and make decisions on situations?

answer:

Gossip is purposeless. The goal of developing AI and chatting with people is to hope that AI has “human-like” thinking and language skills. Microsoft Xiao Bing is a typical example.

In fact, chat is not all aimless chat, a large part of the chat is to complete a specific task, such as Apple's Siri can help users complete mobile phone operations, information query and other instructions, like a small secretary, of course, Siri also has Part of the chat feature, but not its main function.

The Concordance of China University of Science and Technology also has similar functions. It can complete information search for weather, restaurants, hotels, and attractions, as well as tickets/ticket bookings, oral translations, and other tasks.

Well, negotiation is actually a special task. The purpose is to negotiate success and maximize their own interests. However, the negotiating task needs to compete with the "opponent". This is different from the previous task.

As for whether it is necessary to analyze and judge the situation, and decide what to say based on the analysis and judgment results, whether it is chat or chat for specific tasks, it is needed.

ask:

In terms of the difficulty of R&D, is the negotiation more difficult?

answer:

As long as the problem is well defined and the mission objectives are well represented by mathematical functions, the machine can learn. So it seems to me that this negotiation task is no harder than other "chat" tasks before it.

ask:

What does the negotiation function mean for the development of artificial intelligence?

answer:

The task of negotiating this task was ingenious. Most of the previous studies were aimed at the "human-machine" dialogue. The task of negotiating can make the two AIs carry out the "machine-machine" dialogue, which is relatively impressive.

However, as far as it can influence the field of artificial intelligence, I feel that it is not as large as the outside world has imagined. It is not known whether this can trigger some changes.

ask:

There are news reports that Facebook researchers said that once they had to change one of the models, because if they did not do this, the dialogue between robots “will be different from human language because agents have developed their own Negotiation language."

Does this mean that the development of artificial intelligence may be out of control? What is the significance of this incident? Is it the beginning of artificial intelligence to create "self"?

answer:

Language is a symbol of communication. If the robot is not constrained by the natural language of human beings, then they may drift through mutual training and produce “symbols” that humans cannot read, and may be called “languages ​​that meet specific tasks.” ".

However, this does not mean that artificial intelligence has created "self", only that robots use a special symbol string for information transmission. In fact, the "desire" of Facebook's negotiating robots is set manually in the data set.

For example: books, hats, balls have different values ​​for different robots. It is not that robots have the "self" consciousness.

As for the "lying" behavior that arose during the negotiation process, it was only the imitation of human behavior that was learned from the training data.

ask:

Can you give an overall assessment of this progress?

answer:

This work of Facebook is very interesting. It converts the negotiation task into a computable goal. Then the two agents “negotiate and discuss” and use the reinforcement learning algorithm to learn the negotiation process, similar to Alphago's self-play.

For this task itself, it is still very limited. We can consider translating real-world problems into this form and solve them in a similar way.

It seems that there is no need to worry about the issue of AI rebellion. Instead, this interesting study provides us with reference to the use of artificial intelligence to solve practical problems.

ring light/fill light

Ring Fill Light,Ring Lamp,Ring Light With Tripod,Selfie Ring Light Tripod

Shaoxing Shangyu Kenuo Photographic Equipment Factory , https://www.kernelphoto.com