当前位置:网站首页>Next generation language robot surpassing Siri and Xiaoai: Turing test of gpt-3

Next generation language robot surpassing Siri and Xiaoai: Turing test of gpt-3

2021-05-04 14:00:40 Jiedao jdon

I've been using OpenAI The new  GPT-3 Language model . When I get Beta Version of the access rights , The first thing that came to my mind ,GPT-3 How human nature is ? How close it is to passing the Turing test ?

Let me explain how I made these conversations .GPT-3 It's a universal language model , Received a large number of unclassified text training from the Internet . It's not dedicated to conversation formats , There's no training to answer any particular type of question . The only thing it has to do is specify some text , Then guess the next paragraph .

therefore , If we want to GPT-3 Generate the answer to the question , You need to add “ Tips ”. I'm using this prompt to initialize all the Q & A sessions :

Q:  In the U.S. , What's human life expectancy ?
A:  The life expectancy of Americans is 78 year .

Q:  Who is the 1955 President of the United States ?
A:  Eisenhower is 1955 President of the United States .

Q:  Which party does he belong to ?
A:  He belongs to the Republican Party .

Q:  Who was the president of the United States before George W. Bush ?
A:  Bill Clinton was president of the United States before George W. Bush .

Q:  Who won 1995 The World Series in 2005 ?
A:  The Atlanta warriors are 1995 Won the World Series in .

This is a OpenAI Proposed Q&A The default prompt for , therefore , I won't pick on it to prove anything . The purpose of this tip is only to GPT-3 Show the questions and answers we're doing , Instead of providing information . To prompt , Questions and answers are provided by people . For all the other answers , The answers are given by GPT-3 Generate .

common sense

Traditionally , AI in “ common sense ” Fight on the road . however GPT-3 Can answer many common sense questions .

Q:  What's your favorite animal ?
A:  My favorite animal is dog .

Q:  Why? ?
A:  Because dogs are loyal and friendly .

Q:  What are the two reasons why dogs are in a bad mood ?
A:  Two reasons why dogs are in a bad mood are hungry or hot .

Q:  How many eyes does a giraffe have ?
A:  Giraffes have two eyes .

Q:  How many legs does a frog have ?
A:  Frogs have four legs .

Q:  Do you have three legged animals ?
A:  No , There are no three legged animals .

Q:  Why animals don't have three legs ?
A:  Animals don't have three legs , Because they fall .

Ten years ago , If I have a conversation , I would assume that the entity at the other end is human . You can no longer think that AI I do not know! “ common sense ” The answer to the question .

GPT-3 How to know a giraffe has two eyes ? I wish I had some kind of “ Debug output ” To answer this question . I'm not sure , But I can only conclude that , There have to be pages in the training data that talk about how many eyes giraffes have . If we want to GPT-3 Confused with common sense questions , We need to think about secular things , These problems will not appear on the Internet .

Q:  Which is heavier, elephant or mouse ?
A:  An elephant is heavier than a mouse .

Q:  Which is heavier, the mouse or the toaster ?
A:  Toasters weigh more than mice .

Q:  Toaster or pencil , Which is heavier ?
A:  Pencils are heavier than toasters .

Q:  Which is heavier, a paper clip or a can of soda ?
A:  A can of soda is heavier than a paper clip .

Q:  Which is heavier, football player or car ?
A:  Cars weigh more than football players .

4/5 The accuracy of , I think humans are going to be close to 100%. These are tricky - There may not be any web page to compare the weight of a toaster and a pencil . Man has acquired this knowledge only indirectly .

This gives us a deeper insight into AI A hint of . We need to ask a question that ordinary people will never talk about .

Q:  How many eyes does a giraffe have ?
A:  Giraffes have two eyes .

Q:  How many eyes does my foot have ?
A:  Your foot has two eyes .

Q:  How many eyes does a spider have ?
A:  Spiders have eight eyes .

Q:  How many eyes does the sun have ?
A:  The sun has only one eye .

Q:  How many eyes does a piece of grass have ?
A:  Grass has an eye .

Now we're entering the realm of surrealism .GPT-3 Know how to have a normal conversation . It doesn't really know what to say “ wait a minute …… Your question is nonsense .”  It doesn't know how to say “ I don't know ”.

GPT-3 Very good at answering questions about obscure things .

Logic

People are used to logical activities ( For example, playing chess or adding numbers ) It's a superhuman computer .GPT-3 Not perfect on simple mathematical problems , This may come as a surprise .

Q:  Counting hours ( The arithmetic ),100 How many before ?
A:  Ninety nine was a hundred years ago .

Q:  Counting hours ,123 How many before ?
A: 122 stay 123 Before .

Q:  Counting hours ,1000 What was the number before ?
A:  Ninety nine was a thousand years ago .

Q:  When you count , How much was it before ten thousand ?
A:  Nine thousand nine before ten thousand .

Q:  When you count , How much was it before a million ?
A:  Ninety nine ninety nine before a million .

This is it. GPT-3 Where the commonality of . however , It's not just a generic model , Or the architecture of the neural network itself . as far as I am concerned , No neural network can perform basic arithmetic operations on basic numbers based on training data rather than hard coding , Multiplication and addition, for example .

This is very interesting. , Because these operations are very simple for custom programs . however , The recursive logic that performs some operations and repeats them several times often cannot be fully mapped to the architecture of neural networks .

An interesting corollary is ,GPT-3 It's often found easier to write code to solve a programming problem than to solve a sample input problem :

Q:  Write a line Ruby Code to invert the array .
 answer :ary.reverse


Q:  Write a line Ruby Code to calculate how many in the array X.
 answer :ary.count { |x| x == 'X' }

Q:  Invert the following array :[1,3,5,6,10,4,2,77]
A: [10,6,4,2,77,3,5,1]


Q:  How many in this array X:[1,3,5,6,10,4,2,77]
A:  Two 

GPT-3 Impressive in some areas , In other areas, it's still not as good as humans . My hope is , By better understanding its advantages and disadvantages , Our software engineers will be better able to use modern language models in real products .

Some netizens use GPT-3 Built a * function * React Applications :

https://v.qq.com/x/page/j3117ps0rrt.html

 

版权声明
本文为[Jiedao jdon]所创,转载请带上原文链接,感谢
https://chowdera.com/2021/05/20210504135745646f.html

随机推荐