Will ChatGPT become the "550W" in "The Wandering Earth 2"?

Figure 1 Header image | Source: Baidu Encyclopedia

From 0 to 1

This is a leap forward in computer language

The emergence of ChatGPT

Will it be the leap for AI to surpass the Turing test?

1

Why is ChatGPT so powerful?

ChatGPT is an artificial intelligence language processing tool launched by OpenAI.

Simply put, ChatGPT is a chatbot program.

To go into more depth, ChatGPT can learn and understand human language through a specific neural network architecture, and interact with humans on this basis. In other words, it is an AI that can chat and communicate like a real person.

What's more, ChatGPT can be used in a specific way to complete tasks such as copywriting, script programming, text translation, and code writing.

Figure 2 | Source: Beijing News

Thanks to the Transformer neural network architecture (GPT-3.5 model), ChatGPT has a dedicated data processing and understanding model, on which ChatGPT's communication and understanding capabilities are formed.

In other words, ChatGPT can learn and train its own response model by connecting to a large corpus.

Therefore, when we use ChatGPT, we can regard it as communicating with the same talking search engine.

Figure 3 | Source: Beijing News

2

Disruptive AI

More than seventy years ago, scientist Alan Turing proposed the famous "Turing Test" for AI technology.

“The tester and the testee (a person and a machine) are separated and ask random questions to the testee through some devices (such as a keyboard). After multiple tests, if the machine causes an average of more than 30% of the participants to make wrong judgments, then the machine passes the test and is considered to have human intelligence.”

The 30% standard comes from Turing's prediction of the future thinking ability of machines. Now it seems that the emergence of ChatGPT has undoubtedly accelerated the process of AI passing the Turing test.

ChatGPT, which knows everything about astronomy and geography, can not only answer real-life questions, but also travel through time and space to write poems for famous writers.

Such disruptive technology makes people wonder whether it will cause people to lose their jobs.

If so, who will be unemployed?

In any case, the biggest question and controversy about ChatGPT is whether it can really realize or replace human thinking.

Figure 4 | Source: CSDN

In the past, when you asked a question to a search engine or AI, they would often only look for an existing answer or template through a generative model to answer you, and thus often gave you a wrong answer.

ChatGPT adopts the structure of "generation model + reward model". After answering, ChatGPT will feed back the question-answering results to the reward model.

In other words, ChatGPT has an inherent scoring (reward and punishment) mechanism. If ChatGPT can give three answers, ABC, to the same question, the reward model will score the three answers separately based on factors such as whether the logic is self-consistent, and use the scores as the basis for rewards and punishments to optimize its own answering method.

"Generation model + reward model" constitutes the autoregressive characteristics of ChatGPT.

That is, ChatGPT can complete the cycle of question-answering, rewards and punishments, and iterations by itself. ChatGPT can then match the highest-scoring combination among Chinese characters or letters, and then combine them into sentences or paragraphs and submit them to the user.

Figure 5 | Source: openai official website

Although the structure of "generative model + reward model" is better than most current AI, this does not mean that ChatGPT truly understands the meaning of questions and answers in the process of generating answers. ChatGPT is always just looking for the optimal solution calculated by the model, but it does a better job in path selection and feedback reception.

ChatGPT's automatic reward and punishment mechanism reduces the amount of manual correction, so it can learn faster and faster and win in the fierce competition with other AIs.

3

No matter how powerful artificial intelligence is, it cannot be separated from “artificial intelligence”

Three years ago, Google's MT5 had 13 billion parameters.

At that time, ChatGPT's predecessor, the GPT-3 model, already had 175 billion parameters.

Now ChatGPT has emerged. Its parameters have not changed compared to GPT-3, but it has reached the singularity of performance explosion.

It can be seen from this that the accumulation of parameters may not be the key to the advancement of AI. Optimizing models and calculation methods may be the key to improving AI technology.

However, no matter how powerful AI is, it cannot do without human support, and the foundation of the AI ​​world is still the real world.

Figure 6 | Source: Phoenix Technology

Even as powerful as ChatGPT is, its corpus is still stagnant in the fourth quarter of 2021. Its understanding of the world is frozen in 2021. If you ask it about anything new at the moment, you probably won't get an effective answer.

In addition, ChatGPT's answers are not stable, and ChatGPT will start to "talk nonsense" after multiple rounds of questions and answers.

This is because as the conversation increases, the complexity of the model calculation will increase exponentially. In order to reduce the calculation time, ChatGPT needs to cut the text, which leads to distorted answers.

It is true that ChatGPT has achieved significant success in the field of AI at this stage, but AI technology represented by ChatGPT still cannot do without human guidance.

In essence, it is still humans who guide AI's exploration and question-answering. We are still some distance away from AI truly realizing independent thinking.

END

Author: Tang Yu Communication University of China

Source: Wikipedia, Technology Headlines

CSDN, China News Network

Editor: Guru