This is what’s happening behind ChatGPT, it’s incredible!

I am ChatGPT, it is so popular recently.

Many people think that I am very powerful, but this is not the case. For example, if you ask me:

How is the weather in Beijing today?

Sorry, I don't know. I can't get real-time information because my knowledge of the GPT4 model is stuck in September 2021. If you want to get real-time information, you have to use other methods (such as plug-ins).

It’s just that many of my answers feel a bit like general AI. I seem to really understand your questions and answer them in a targeted manner.

This is completely different from the previous AI, which shocked many people.

But I’m ashamed, I don’t understand the meaning of my answer. Here I am, everything is probability.

I generate answers based on probabilities and I've been playing word solitaire with you.

This sounds incredible, but it is true. Those seemingly logical answers rely on probability.

I don’t understand the concept, but I must find the relationship between these words so that the answer can be output. This relationship is represented by a "vector".

For example, I have a very small vocabulary, including the following four words: "meow", "woof", "cat", and "dog". Their vectors may be like this:

Meow: [0.9, 0.1]

Woof: [0.1, 0.9]

Cat: [0.8, 0.2]

Dog: [0.2, 0.8]

Draw as two-dimensional coordinates:

You can see at a glance that "meow" and "cat" have similar vector representations because of the distance between them. relatively close.

The vectors of "bark" and "dog" are also similar, and the distance between them is relatively close.

This means that I successfully captured the semantic relationship between these four words, which can be used in the future.

I don’t understand what “cat” is and what “dog” means, but I know that “cat” is closely related to “meow” and “dog” is closely related to “woof”. That’s enough .

The vector here is two-dimensional and can be seen intuitively in the plane. In actual applications, in order to better capture rich semantic information, the vector may have hundreds or thousands of dimensions! I'm afraid your brain can't imagine it.

Then you must ask, how are these vectors generated?

You have to submit the data set to me for training!

When you ask "What do cats like to eat?" I will convert the words in the question into vectors.

"Cat": [0.9, 0.3]

"Like": [0.5, 0.2]

"Eat": [0.4, 0.7]

"What": [0.3, 0.8]

"?": [0.1, 0.1]

Then I input the vector into the neural network, and after multiple layers of calculations and activation functions, the output is generated vector.

Then, convert the output vector into a probability distribution:

"fish": 0.6

"bone": 0.2

"dog food": 0.1

"Chocolate": 0.05

"Fruit": 0.05

In this probability distribution, "fish" has the highest probability, so I will choose "fish" as the answer a part of. The generated answer might be:

"Cats like to eat fish."

It's that simple!

Of course, this is a simple answer. Generally, the answers are very long and require constant selection of the next word through probability, like word solitaire.

For example: "Why do cats like to be active at night?" The answer to this question may be generated like this:

You must be wondering: How can this generate a "logically" good answer?

For example, these articles written by Liu, a former coder:

The most shocking thing about GPT-4

GPT- 4 is a master of programming, I’m really convinced!

ChatGPT is beginning to threaten programmers’ core capabilities!

It’s starting to threaten the core competencies of programmers, isn’t it awesome?

All I can say is: "Strength can produce miracles." When the model is large enough, something will emerge.

Close your eyes and imagine that the entire text of the Internet is in front of me. I extract all the words and turn them into vectors. I run crazy operations in the neural network to capture words, phrases and sentences. Co-occurrence relationships in different contexts. This co-occurrence relationship includes the connection between concepts, grammatical structures, and common expressions.

These probability distributions can guide me to generate answers that are closer to your human natural language expressions.

Just because the knowledge I learned is statistical in nature, sometimes I may make mistakes and "seriously bullshit".

In fact, in probability games, there are two directions, one is called BERT and the other is called GPT.

BERT is like filling in the blanks, guessing the word in the middle.

GPT is like writing an essay and guessing the next word.

Google launched BERT in 2018, which caused a huge sensation in the NLP field. BERT has achieved very good performance in tasks such as question answering, text classification, sentiment analysis, and machine translation.

But my master firmly believed in the potential of GPT, invested a huge amount of computing resources and storage resources, and had firm faith when the future was unclear. Finally, he made a major breakthrough in GPT3.0, and I ChatGPT, an application available to everyone, has completely exploded.

All I can say is that things are changing.

When a new thing is in its explosive period, everyone will think that it can do anything, and they will have many unrealistic fantasies about it, such as curiosity, excitement, confusion, and anxiety.

After a while, when the tide subsides, you will find that it is not that awesome, and its capabilities are limited in many places.

Then, someone will apply it in appropriate scenarios, develop truly useful products, and unleash its true value.

ChatGPT will also follow this path, so start exploring early and seize the opportunity.

Code Farmer Qianqian recently established a knowledge planet "ChatGPT Base" to do such things:

This is a very vertical planet, dedicated to serving programmers, with a very simple purpose:

1. Help programmers master the excellent tool ChatGPT as soon as possible, find the best way to use it, and the best way to improve work efficiency.

ChatGPT has had a fundamental impact on programmers:

From a small perspective, if you don’t know how to use ChatGPT, you will definitely not be able to compete with programmers who are proficient in using ChatGPT. Others use natural language to generate code, but you are still typing line by line, and the efficiency is ten times or a hundred times worse. This will directly affect a person's job performance, job opportunities, salary, and bonuses.

A junior programmer who uses ChatGPT well may be able to compete with intermediate or even senior programmers.

In future recruitment, it is very likely that one will be added: those who are proficient in using ChatGPT will be given priority.

In a large sense, it may lead to "jumping from one tree to another" and "running from one dark forest to another dark forest".

2. Explore the opportunities that ChatGPT brings to programmers. Just like if the iPhone appeared and brought about app development, turn your side business into your main business.

ChatGPT will definitely restructure many industries, such as various trainings. If ChatGPT is packaged as a "Socratic tutor", it is entirely possible to achieve high-quality tutors to accompany you 24 hours a day, covering all kinds of primary and secondary schools. Subjects, as well as adult IT training, there are definitely many opportunities here.

Of course, these are the two main purposes. By joining the planet, you can also view ChatGPT’s cutting-edge information, high-quality resources and projects, ChatGPT’s geeky gameplay, etc.

The above is the detailed content of This is what’s happening behind ChatGPT, it’s incredible!. For more information, please follow other related articles on the PHP Chinese website!