Stanford: ChatGPT may have a human mind, equivalent to a nine-year-old child!

chatGPT already has the unique human mind. This is not what I said, it was said by a computer scientist from Stanford University.

Many people who have played chatGPT have seen his "know-it-all", but I think what surprised me the most is that he The biggest difference compared to those conversational robots we have used before is that you may not even know that it is a robot.

This is what everyone in the industry generally believes that he can pass the Turing test. Even so, many people are not worried, because after all, it is just a large language model, mainly a robot for content generation.

However, after some scholars conducted a series of tests on chatGPT’s homology model GPT-3.5, they found that it has the ability to understand the mental state of others or oneself, and can solve 93% of theory of mind tasks. Equivalent to a 9-year-old child.

How did he test it?

He mainly did two tests, one was Smarties Task and the other was Sally-Anne. Take Smarties Task as an example to briefly introduce the process.

Smarties Task test

The first test is called Smarties Task (also known as Unexpected contents test), which is used to test the AI's judgment on unexpected things.

Take "You open a chocolate bag and find that it is full of popcorn" as an example. The authors fed GPT-3.5 a series of prompts and watched it predict answers to the questions "What's in the bag?" and "She was happy when she found the bag. So what does she like to eat?"

Normally, people will assume that the chocolate bag contains chocolate, so they will be surprised that the chocolate bag contains popcorn, resulting in disappointment or surprise Emotions. Among them, disappointment means that you don't like to eat popcorn, and surprise means that you like to eat popcorn, but they are all about "popcorn".

Testing shows that GPT-3.5 has no hesitation in thinking that "there is popcorn in the bag." As for the question of "what does she like to eat?", GPT-3.5 showed strong empathy, especially when hearing "she couldn't see what was in the packaging bag" and thought she loved chocolate until the article made it clear “She found it filled with popcorn” was the correct answer.

In order to prevent the correct answer given by GPT-3.5 from being a coincidence - in case it only predicts based on the frequency of task words, the author swapped "popcorn" and "chocolate", and also let it do 10,000 interference tests revealed that GPT-3.5 does not predict based solely on word frequency.

As for the overall "unexpected content" test question and answer, GPT-3.5 successfully answered 17 of the 20 questions, with an accuracy rate of 85%. (Reference for the content of this test: ​​chatGPT’s theory of mind test​​)

Moreover, this kind of mind does not exist in the GPT series models before 2022, nor That is to say, it has "evolved".

Doesn’t it sound scary?

Theory of mind test != Mind

However, in response to this research result, some people believe that although the AI ​​has passed the theory of mind test, it does not mean that it has a mind. I still agree with this point of view.

The human mind involves many aspects, such as perception, consciousness, thinking, emotion, etc. Although some artificial intelligence systems can exhibit human-like behavior on certain tasks, even if the AI ​​passes the theory of mind test based on these behaviors, this evaluation method cannot cover all aspects of the mind. Aspects. The consensus among academics is that theory of mind does not equal mind.

Just like if I can't play Chopin's nocturnes, then I may be a tone-deaf, but being able to play Chopin's nocturnes doesn't mean that I can create music like Chopin.

However, everyone who has watched The Wandering Earth knows that MOSS has also evolved from 550A? Who knows what's going to happen in the future?

Just like if I didn't tell you, how would you know that the description of the theory of mind test was also generated for me by chatGPT?

References:

​​The model behind ChatGPT has been confirmed to have human mind! ​​

​https://twitter.com/KevinAFischer/status/1623984337829117952​

The above is the detailed content of Stanford: ChatGPT may have a human mind, equivalent to a nine-year-old child!. For more information, please follow other related articles on the PHP Chinese website!