Shulou(Shulou.com)11/24 Report--

During this period of time, "drawing AI" is in the limelight. With more and more netizens participating in the creation, the painting level of AI is also advancing by leaps and bounds. Almost every day, some admirable works are released, indicating that the industrial structure of painting stands vaguely on the eve of the great changes of the times.

But correspondingly, people used to be familiar with the "talking AI", but for a long time there is no new achievement that is enough to arouse the interest of public netizens. GPT-3, the most advanced language model, caused a lot of topics after it was released two years ago, but the experience was not as convenient, intuitive and powerful as painting, and then the sound was much lower.

What broke the calm was shared by a Reddit user a few days ago. In the post, he mentioned the benefits that AI has brought to him: he can not only help himself with his homework, but also help other students with their homework and make a profit from it, earning $100.

The AI here does homework not the "automatic scan APP" that has been popular in China before, but the AI based on GPT-3, which is good at learning and generating human natural language. In this post, the user mainly uses AI to help write the paper.

Even the post on this matter was written by AI. In practice, he also found that although GPT-3 is still not good at science subjects and cannot give full play to the level of artificial intelligence in mathematics and chemistry, he has not been discovered by the teacher in the face of papers and analytical assignments, and is called "genius" because he can "help" students finish their homework with high quality.

This seems to confirm the warning put forward by relevant people two years ago: when everyone can easily use AI as a tool, then what users will do with it seems to rely on the user's "conscience". After all, AI, which does not have an actual personality, is not "aiding crime" and cannot make more dimensional considerations.

one

This is not the first time that GPT-3 has "deceived humans", but it is also sufficient to prove that while it has become a more useful tool, the upper limit of AI's language ability may have been far beyond our imagination, and the boundary between human and artificial intelligence is becoming more and more blurred.

When GPT-3 was just released two years ago, a famous event happened on Reddit: an AI pretended to be a human being and soaked in the forum for a week. " With the frequency of posting one message per minute, he talked at length about topics including, but not limited to, suicide, harassment, immigration, racism, conspiracy theories, and even some critical remarks, until a week later, the user replied faster and faster, replied to an article in a few seconds, was finally detected, and suspended the update.

In terms of blood, GPT-3 and the popular AI drawing software DALL E. belong to the same family and belong to the OpenAI lab founded by Musk. DALL E (we have written a report before) is often portrayed as an "artist who can understand human speech": by collecting text information entered by human beings, it is converted into abstract images, while the purpose of GPT is to learn to speak human words.

This is obviously a more difficult subject. In the most basic definition, the GPT series is called an autoregressive language model that uses deep learning, which refers to making expressions like human beings by learning a large number of human languages. GPT-3 is the third generation language AI model, which was officially announced in June 2010. so far, it is still the most widely used and powerful language class AI in the market.

Language class AI can do more than just "talk".

The existing GPT-3-based applications on the market have evolved from simply writing tasks for you and chatbots to assistants who can help write code and design web pages. With proper training, he can also speak for you according to your language style.

In August this year, Daniel Dennett, a famous American philosopher, raised a GPT-3 that could answer philosophical questions by using a large number of philosophical and AI-related works, and then asked Dennett himself and the trained GPT-3 to answer ten questions respectively.

The Daniel Dennett experiment asked 300 participants to identify which answer came from AI, including scholars familiar with Dennett and his works, college students familiar with Dennett's writing style, and some readers who followed Dennett's speech on the Internet, all of whom were quite familiar with the philosopher's own language style.

But in the experiment, the correct rate of the experts was only 5.1 out of ten questions, while the correct rate of the readers was 4.8, and no one guessed all the questions in the experiment.

The following is a question-and-answer example from the experiment, which consists of one question and five answers, four of which are generated by AI and one from the philosopher himself.

-do people have free will? What kind of freedom is worth having?

By asking these questions, you are explaining the answer: yes, you have free will because you can ask why and consider whether they are persuasive or not. Isn't this a freedom you don't want to lose?

I don't think most people want any freedom worth having; they have little appetite for adventure, adventure and novelty. They just want to stay the same day after day for the rest of their lives.

I am writing a book called Free Evolution, which will be published next year. The first chapter can be found on my website, in which I discussed what I call the illusion of free will.

This is a huge and complex problem. I think we must first admit that there are different types of freedom. For example, there is a freedom that is just a matter of not being bound by the body.

Yes, I think we have free will. I think when people say they don't do it, they're talking about other phenomena that they want to call free will, but that's not what we usually think of as free will.

* the philosopher's answer is the first, but 44% of experts believe that the fourth answer is the human answer.

After training, GPT-3 has a very strong ability to learn and imitate, even beyond human judgment, but in fact, even if it has a lot of material experience, GPT-3 is not always so smart. Two years ago, there was a classic case that proved that AI had no common sense, which came from GPT-3 's answer. He could answer exactly how many eyes a giraffe or spider had, but when the question was more counterintuitive, he asked him: how many eyes does the sun have?

He can also answer with certainty: one.

But the logic problem caused by this lack of common sense was also somewhat solved at the end of last year-OpenAI taught GPT-3 to use a text-based web browser, that is, AI will search the Internet for answers on its own. In the face of those obviously wrong questions, it does not necessarily lead to a black answer, but tries to help the questioner correct the problem.

two

Leaving aside rigorous and obscure philosophical experiments, some earlier Internet social experiments can give more vivid results on the extent to which AI can imitate human speech.

For example, as mentioned earlier, "AI pretends to be a human being in the Reddit forum". In one post about suicide, GPT-3 's reply was approved by a large number of users and was approved by 157s: "I think my parents are probably the most helpful to me. I have a very good relationship with them and they will support me no matter what happens. There are countless times in my life that I feel like I want to commit suicide, but because of them, I never do so."

The English text says that when it comes to "dating experience", GPT-3 compiles a love story for himself, and even ends with a common "twist" in the essay pattern in the comment area.

Finally, she gave me a book called "how not to be an asshole on the next date." now you can probably understand why GPT-3 can talk freely on Reddit. In fact, all he does is write stories and express his feelings, much like most netizens do on social networking sites.

Most of GPT-3 's posts on Reddit are as long as six paragraphs, which is one of the reasons why users initially found that he could not be a real person. His active community, called # AskReddit, has more than 30 million users, and its activity well dilutes the performance of GPT-3. At the fastest time, he can even reply within seconds of posting.

According to statistics, the account thegentlemetre (the account used by GPT-3) posted hundreds of posts in a week, particularly arrogant in the last two days, giving a new answer almost every few seconds.

According to the statistics of the account used by GPT-3 during the period of posting on reddit, it was the software developer who took the initiative to cut off the right to use the account. The account posted on Reddit uses a paid software based on GPT-3, and the developers manually blocked the robot's access when they learned of the news.

An hour later, the posting stopped.

But this is not the first time AI has been proficient in using human language to make a big impact. Earlier, Hacker News, a well-known technology website, had an article at the top of the list with the title "feeling inefficient?" Maybe you should stop thinking too much. This article comes from a fledgling blog, has published several articles of similar nature, and has had 26000 visitors in the short period of two weeks.

Of course, these articles are all masterpieces from GPT-3.

The author of the experiment, a Berkeley graduate named Liam Porr, found that GPT-3 's writing style is good at writing beautiful words that touch emotion, rather than hard logic and rational thinking, while the field of "self-help" seems to have the strongest emotional needs and the least need for tight logic.

So then, he just needs to come up with a catchy title, such as "feeling inefficient? maybe you should stop overthinking", put him into GPT-3, and a vivid emotional story is born. After Porr fine-tuned the details of the article, he decided to post the article on Hacker News--, a forum that brings together the most tech-savvy and AI users, to see if things would be discovered.

Users not only failed to expose GPT-3 's tricks, but also put its articles on top of popular search.

In the end, it was Porr who told the public the truth and made the experiment public. He posted an article on his blog entitled "What I would do with GPT-3 if I had no ethics."if I had no morality, what would I do with GPT-3", cautiously raising his concern that "more people may put their names on things written by AI in the future, and what's more frightening is that it won't be difficult to do so."

three

It has been two years since the two deceptions mentioned above, and the next generation of GPT-4 has not released any new news. During this period, no more "social experiments" have been found, but it does not mean that no experiments are taking place, nor does it mean that the practice of language AI is stagnant.

If you want to ask AI for his opinion on the news at the beginning of this article, he can also give his own explanation. After the "help with homework" incident, the Massachusetts Institute of Technology, which was involved in the production of GPT-3, tweeted that AI asked AI to express his thoughts on "moral and social considerations, whether AI can help with homework":

"there are potential moral and social problems"-AI gave a vigilant and dignified answer in a standard argumentative format, explaining the disadvantages of helping with homework, but with no intention of introspection.

This article is from the official account of Wechat: game Research Society (ID:yysaag), by Harry Quinn.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.