This article is from WeChat official account:Xin Zhiyuan (ID: AI_era), editor: Meng Jia, Bai Feng, head picture from: “She Her” movie screenshot

The epidemic in the first half of the year has forced many friends and lovers to be separated from each other. I am in Haidiantou, and you are at the end of Chaoyang. I think about you every day and never see you. I can only…

At this time, there is a GPT-3 to accompany you, watching the stars and the moon, and talking about the philosophy of life from poems and songs, isn’t it beautiful?

Why are you still playing with the producer? Just apply for an API interface and chat with GPT-3. In the future, the electronic male and female friends of Paper Man + GPT-3 have huge market potential.

He is very straight sometimes, he always makes some lame jokes to make you happy, and he looks a little cute when talking awkwardly.

Source: Computer Scientist Kevin Lacker Blog

Sometimes he incarnates as a literary and artistic youth, a prodigal poet, and from time to time he recites some random thoughts that break away from clichés and low-level tastes and find new ways. The warning sentence is as follows:

“Anything in the world that you can’t understand, you can use “This is not fun” to deal with”

“If I knew the meaning of life, would I go here to waste time?”

If you say what is the hottest in the computer science circle recently, it is GPT-3.

To put it simply, GPT-2 and GPT-3 are both AI models that continue to write text. A few hints are given at the beginning, and the following story relies on his language model to continue writing bit by bit. The thoughts are smooth, and the “writing bottleneck” is rarely encountered.

GPT-3 is a text generation (text-remarks” label=”note”> (text generation) artificial intelligence developed by OpenAI, a well-known artificial intelligence research company. Related papers are in 5 Announced in December, it caused a huge sensation at the time with an astronomical amount of 175 billion parameters.

For language models, size is really important. You know, the human brain has only 86 billion neurons.

After the API was opened, netizens from all walks of life had a lot of thoughts. Let GPT-3 write code and make resumes, which is more effective than boyfriends.

Writing SQL, drawing charts, doing operation and maintenance, doing PPT, GPT-3 must be refined!

GPT-3 is not broken, but a new height.

You may still be stuck at the amazing text generation ability of GPT-3, but after reading the following applications, you will definitely be shocked, “Look for Lao Huang to buy the showThe card is too late, apply for API as soon as possible, I want to develop my own GPT application”!

Give GPT-3 a reminder to incorporate various emoticons into popular culture.

What, GPT-3 can also be used in Excel, and automatic reasoning can be done? There are several CEOs of large companies listed in the demo form, frame the cell containing the data, and Loading for one second automatically infers the company to which Ben Horowitz belongs, need no operation!

Why your chart is so low, but the visualization done by others is beautiful, and you can only forcefully explain, “I can’t write code, good-looking charts are written in code”. Now that you have GPT-3, you don’t need to explain, just feed your data to GPT-3, you don’t need Excel spreadsheets, just say it!

“3 apples, 5 pears, 2 oranges”, bang! The beautiful Dash chart is generated.

Finding a job and writing a resume is a headache. Now you only need to tell GPT-3 which company you have worked for for a few years, and you don’t have to deal with the rest. It’s well organized and a well-formed resume will help you “compile it.” “, working years, main responsibilities, and project experience are all available.

Report work to the boss and do a PPT. You only need to copy the text to generate a suitable presentation.

Want to find an operation and maintenance job, but can’t operate AWS?

“Hello, I need to create an AWS instance, thank you”.

Wait for a second, GPT-3: “I’m done, just copy the code”!

It’s almost perfect!

Can’t write SQL, are you ashamed to say that you are O&M? Sorry, with GPT-3 you can really do whatever you want!

Operation and maintenance: “GPT-3, I want to query the salary information of workers between 25 and 35”,

GPT-3: Okay, use SQL to “select salary from worker where age between 25 and 35”.

The “addition, deletion, modification and investigation” engineer has been crying in the toilet.

GPT-3, which has eaten 300 billion tokens, how does it work internally?

Feed the text containing 300 billion tokens to GPT-3 for unsupervised training, and OpenAI has already accomplished this feat. All the experiments you see now come from this huge and expensive model. It is estimated that it cost up to 4.6 million US dollars.

Let’s come againLooking at the details of its training, GPT-3 slides across all texts, and a large number of examples are generated for each window sliding.

How does GPT-3 process the word “robotics” and produce “A”?

“robotics” has gone through the following three processes. First, it is converted into a word vector. After the baptism of 96 Transformers, the vector is reconstructed into the word “A”. If it is not “A”, it starts to calculate and The error is passed backwards until the correct word is reached, and this process is repeated millions of times.

How to consume the trained model? Only a simple fine-tuning is required. For example, if you want to apply it to English-French machine translation, you only need to provide a small number of parallel corpus pairs to GPT-3.

This may not be intuitive enough. We use Python to demonstrate how to teach GPT-3 to work.

Writing LaTeX formulas is a headache for non-academic researchers, but with GPT-3, LaTeX syntax conversion is so simple.

As we said earlier, if you want to apply it to a specific task, you need to give GPT-3 a little “tip” so that it can understand what you want to do.

Apply for a GPT-3 API Key with OpenAI, and you can call this 175 billion parameter model with one click.

I want to write an integral from a to b. If you don’t give a hint, GPT-3 will return to the original text directly. However, if you enter an example and tell it that you want to translate natural language into LaTeX, then GPT-3 will My heart understands it, and the integral formula of f(x) from a to b is perfectly output.

OpenAI CEO: GPT-3 is just a glimpse, is it too much praise?

For Xiaobai, GPT-3 is also an entrance to intuitively feel that artificial intelligence is so powerful.

OpenAI CEO Sam Altman said that he showed GPT-3 to a 10-year-old boy, and he immediately said that he wanted to enter the field of artificial intelligence.

A developer believes that GPT-3 does give me a foreboding, in the next 10 years, AI is expected to pass the Turing test. The purpose of the Turing test is to test whether the machine has human intelligence.

The discussion about human intelligence involves another question, how to define “human intelligence”. Is it just talking about its human “intelligence” (including thinking ability and learning ability, etc.), or does it have a complete personality “, even possessing souls and other transcendent things. Reminiscent of the famous online torture about the soul of God.

But the industry has also heard voices that GPT-3 is considered to be too famous. Altman humbly stated in a tweet, “GPT-3’s promotion has recently gone too far, which is really impressive, but it still has serious weaknesses and sometimes makes very stupid mistakes. Artificial intelligence will change the world, but GPT-3 is just a very early glimpse. We still have too many problems to solve.”

Industry expert: GPT-3 has shown reasoning ability, but it is not really smart yet

Some people predict that at this rate of evolution, GPT-3 can be partially replaced within five years, and even surpassed the psychologist in some places. “The most addictive entertainment for everyone in the future is to interact with GPT-3.”

Some netizens also pointed out that GPT-3 is not omniscient and omnipotent. It has poor comprehension of long-form arguments, and logical reasoning cannot be circumvented. It is still impossible to write an analysis for the USMLE exam for medical practitioners in the United States, and it is temporarily unable to replace internal medicine doctors…..

I want to replace the doctor of internal medicine. What kind of bicycle do I need?

Xin Zhiyuan also interviewed a senior technical expert in the field of NLP on whether GPT-3 has cognitive intelligence and other issues. He believes that GPT-3 is still a bit distant from cognitive intelligence, but GPT-3 has achieved few -shot learning, the GPT-3 model itself has learned a lot of experience and fine-tuned some small samples. This approach is very similar to BERT, and the performance is mainly attributed to Transformer, distributed learning + long-distance dependence, just In principle, there is no essential breakthrough.

Behind the examples shown by netizens, there may be many bad cases.

The definition of cognitive intelligence is very broad, and there is currently no uniformity. Understanding, autonomous decision-making, and thinking are some of the core points I think. The reasoning ability displayed by GPT-3 is only a form of cognitive intelligence.

Instead, it is a knowledge graph that has gradually emerged in the past two years, showing a certain potential in reasoning intelligence. The knowledge graph enhances the interpretability of cognitive intelligence (The path can be known), but there are still big deficiencies in computing power, models, and data. There is still a long way to go, and we must stay awake at all times.

In the long run, building large-scale knowledge graph data and combining knowledge graph learning and deep learning is an important trend.

Regardless of reputation or not, GPT-3 heralds a conceivable future. Even if the understanding ability of GPT-3 cannot be equated with human beings, but in the face of human’s eternal desire to communicate, GPT-3, a gentle, emotionally stable, honest and hardworking AI that understands analysis and thinking may one day Will be better than human company.

Reference link:

This article is from WeChat official account:Xin Zhiyuan (ID: AI_era), editor: Meng Jia, Bai Feng