How to recognize, understand and make good use of ChatGPT

shin
·
·
IPFS
·
After ChatGPT came out, there was a wave of panic about technology. Some people said don't panic, it's just a tool. So, as a human being who is good at using tools, do you know this tool, can you use ChatGPT, can you use ChatGPT well? You are still proud and dismissive of ChatGPT.

What are common misconceptions about ChatGPT?

First of all, ChatGPT is a GPT model, the full name is "Generative Pre-trained Transformer". It is currently the GPT3.5 generation. The GPT model is a kind of LLM, and its essence is still a language model, not the AGI that friends who don't understand the development of AI usually refer to AI.


What is the difference between LLM VS AGI

The full name of LLM is "Large Language Model", which refers to a large language model built on the basis of natural language and using deep learning technology. LLM can learn the rules and features of natural language through a large amount of training data and algorithms, so as to realize tasks such as natural language generation, understanding, and question answering. At present, the most famous LLM is OpenAI's GPT series models, including GPT-3, GPT-2 and so on.

The full name of AGI is "Artificial General Intelligence", which refers to an artificial intelligence system that can perform a variety of intelligent tasks like humans. AGI refers to an intelligent system capable of reasoning, learning, self-healing, self-awareness and other intelligent behaviors like humans. Different from LLM, AGI has a wider range of applications and stronger intelligence capabilities.

In general, there are huge differences between LLM and AGI in terms of technical level and application domain. LLM focuses on tasks such as natural language processing and intelligent question answering, while AGI is broader and can handle and apply multiple tasks in multiple fields. In addition, the research and implementation of AGI requires more complex and advanced technology and algorithm support, compared with LLM, it is still in an earlier research and development stage.

  • ChatGPT is a language model, not general artificial intelligence.
  • Don't grab a red crayon and try to draw a blue ocean.

Since many friends read very fast, I will repeat it here... If there is only one take away, it will be these two sentences.


How to understand ChatGPT

To understand ChatGPT, you must first understand these three logical reasoning methods.

  1. Inductive Logic Inductive Logic (what ChatGPT can do)
  2. Deductive Logic (what ChatGPT cannot do temporarily)
  3. Abductive Logic (what people can do, but ChatGPT can't)

Suppose there is an experiment in which researchers conducted a clinical trial of a new drug and performed data analysis on the trial results. Let us now illustrate how these three logical reasoning methods are applied to this experiment.

inductive logic

  • In inductive logic, we usually induce general laws and laws from specific instances. In this example, we can observe the curative effect of this new drug according to the results of the experiment, and then deduce the law of the action of this new drug in humans. For example, if a trial shows that a drug is effective in treating a disease, then we can infer that the drug has a general therapeutic effect on that disease. In this process, we have induced general laws from specific examples, which is a typical application of inductive logic.

deductive logic

  • In deductive logic, we usually derive a specific conclusion by using known general laws or laws. In this example, we can use information such as known medical principles and known drug ingredients to deduce whether the new drug has a therapeutic effect. For example, if we already know the ingredients of the drug and the rules of action of these ingredients on the human body, then we can deduce whether the drug has a therapeutic effect on a certain disease through deductive logic reasoning.

abductive logic

  • In abductive logic, we usually explain a known fact or phenomenon by choosing a most likely explanation. In this example, we can choose a most likely explanation to explain these experimental results based on the known experimental results, as well as information such as known medical principles and drug ingredients. For example, if we discover that this new drug has the effect of treating a certain disease, but we are not yet sure what the mechanism of action of the drug is, then we can explain the action of this drug by choosing the most likely explanation. This explanation may be that the drug acts on a specific cell or molecule to produce a therapeutic effect. In the process, we explain known experimental results by formulating logic.


Inductive Logic Inductive Logic

  1. What can ChatGPT do?
  2. The essence of LLM is statistics, so ChatGPT, which has learned a lot of language data, has a strong language organization ability, not because it is proficient in semantics, but because it knows what words are followed by a word with a high probability, and what sentences are formed with a high probability. The ability of ChatGPT text organization is based on statistics rather than semantics, let alone a knowledge system in a professional field.
  3. This helps us understand what tasks ChatGPT is good at, how to communicate with ChatGPT, and how to use ChatGPT efficiently.

Deductive Logic Deductive Logic

  1. What is ChatGPT temporarily unable to do?
  2. The reason why ChatGPT writes beautiful articles or diction is not because it has strong literary attainments, but because humans provide it with enough excellent training corpus to let it know what is most likely to follow a word in a beautiful article Words have a high probability of forming sentences. In other words, ChatGPT does not have any ability to reason and analyze content from the knowledge system in the professional field. Even if in any instance it appears to accomplish this, it's because of statistics rather than the ability to actually do so. Because usually such a result cannot withstand scrutiny and refutation.
  3. This helps us understand what the boundaries of ChatGPT are. Which of ChatGPT's output results is credible and which needs to be questioned. Which tasks are currently not worthy of your proposal to ChatGPT, what is the direction of ChatGPT's future efforts, and whether there is any possibility of combination in the future for KG research.

Abductive Logic

  1. What can humans do but ChatGPT can't?
  2. ChatGPT is currently trained to be very decent, to abide by the rules of etiquette, and to say nothing wrong. In other words, ChatGPT cannot say "nonsense". (If it is nonsense of professional knowledge, it belongs to the previous type) This is also related to the expression that ChatGPT is Inductive Logic. ChatGPT can't jump out of the learning materials to answer you. In other words, ChatGPT does not yet have the ability to formulate hypotheses about unknown phenomena. Although most people do not have this ability without training.
  3. This helps us understand what ChatGPT cannot do, but humans are capable of. What is the competitiveness of people in the future, and what will it mean once ChatGPT shows that it has this ability.


How to use ChatGPT well

There are a lot of Prompts about ChatGPT on the Internet. Here are a few websites to recommend.

https://platform.openai.com/examples

https://github.com/f/awesome-chatgpt-prompts

Briefly talk about the tasks that ChatGPT is good at:

1. Answer specific knowledge questions

Whether it is coding, philosophical research or Ziwei Doushu.

ChatGPT is a very good teacher, it can answer your questions clearly and to the point, and you can keep asking for details. This means that learning resources are unprecedentedly equal, as long as you want to learn. But remember to always be skeptical.

Generally, the more common the knowledge on the whole network, the better the answer, and vice versa. For example, in the questions and answers of Chinese medicine, gossip and five elements, there will be nonsense. But partial knowledge also has guiding tricks. For example, in a new conversation, start with the content/concept that ChatGPT may be easier to talk about, and then ask the questions you want to know. This is because the historical corpus of the same session will affect the next output of ChatGPT as a prompt. The more context is given to ChatGPT, the better it can perform on special vocabulary and questions in a specific context.

2. Solve language problems

Practice English expressions, debug code, expand abbreviations and repeat, etc., as long as it is a language.

I have trained a model for multilingual data on the entire network. I don't need to say how powerful it is. On many questions, it even answers better than your school teacher. But there are still language questions that ChatGPT cannot answer. You might as well try another way of asking, or do some context input and guidance first. Remember, we use ChatGPT for our own progress and growth, to get the knowledge we want, not to make things difficult for it, saying that it really doesn't work. Also, always be skeptical.

3. Advanced nonsense literature

Weekly reports, PPT, official documents, applications, summaries, operational copywriting, business communications, etc.

There is no better way to talk advanced nonsense than LLM who has read the entire network of nonsense literature. If ChatGPT can't write it, it can only be said that your prompt level is not good. The idea is still the same, as long as you know that ChatGPT is based on statistics, just point the prompt to the content you want. Provide as much detail, requirements, and context as possible in the conversation. If you can’t ask one question at a time, you can split the question into several questions and get the results step by step.

4. Output the language that the machine understands

Prompts for other models, instructions and interfaces for other frameworks/services, etc.

Try to imagine ChatGPT as a machine brain, speaking language in the brian. You never imagine how amazing output ChatGPT can give you.


What are the possible impacts of ChatGPT

First, we talked about that ChatGPT is a language model. So we have to look at what language means?

  1. Language is a communication tool for human beings, one of the most important forms of interaction.
  2. Language is the largest carrier of network data, even audio and video can be converted into text data.
  3. In addition to natural language, language also has coding language, and language in the brain.
  4. The boundaries of language are the boundaries of thought.

Language interaction, text or voice

  1. ChatGPT has changed the search engine. It is conceivable that more interactive forms will be changed in the future. For all scenarios where information can only be obtained by retrieval, you can try to add a dialogue entry. In other words, in more vertical fields, product usage guides and service support will shift from original documents or clumsy intelligent customer service to GPT model + weak manual processing. In the future, there will be many scenarios for dialog data design and analysis.
  2. All tools, efficiency, and task-oriented apps/interactions/functions will face an intelligent transformation. Everything that has been experienced in digital transformation may be experienced in intelligent transformation. Including the process of exploration and process, engineering, automation. The reason why tools and efficiency apps bear the brunt is that LLM is good at coding language, and naturally it is not a problem to go from natural language to coding language. Corresponding to the current research field is NL2SQL, and all the extended scenarios of NL2codes, 2function, and 2commands are worth looking forward to.
  3. Due to the transformation of information acquisition methods, part of the traffic entrance will also change from waterfall information flow to conversation-focused type. In this case, the platform side cannot provide recommendations and traffic to content creators like traditional rankings. First of all, the interest relationship between the two parties will be rebuilt, and how to achieve a stable state. We all need to think about how new business relationships, interactive presentation forms, and technical means can meet the demands of both parties.

Multimedia carrier of language, text, audio, video

  1. The language model not only enables text generation, but also greatly reduces the analysis and generation of audio and video content that is closely related to language. The text model processing capability can integrate the entire Internet information data. Coupled with the development of multimodal models, pictures, sounds, etc. will be added to the entire semantic web. The future of the Semantic Web is just around the corner.
  2. LLM will not only become the copilot of language tasks, such as the copilot of GitHub and the notion AI of notion, but also the copilot of most production tools, and it is within easy reach. I believe that in the second half of 2023, there will be a lot of x -copilot's products, tools and services.

Natural language, coding language and language in the brain make ChatGPT a machine consciousness

  1. ChatGPT's performance in the coding language is not inferior to the natural language, because the programming language is regular and orderly. ChatGPT's superior performance in the integration of the two languages has shown that ChatGPT has translation capabilities. In traditional academic research, it is the NL2SQL track, but ChatGPT can do far more than SQL. At present, there is no language in the brain that is more suitable for machines than ChatGPT.
  2. "Language in the brain" is an important part of consciousness, because language is an important carrier and expression of human consciousness. The global workspace theory holds that consciousness is composed of information in the global workspace, which is an information processing system shared among multiple modules in the brain. As a means of expressing information, language plays an important role in the global workspace. Global Workspace Theory (Global Workspace Theory) is a psychological theory proposed by American psychologist Bernard Baars in 1988, which is used to explain the nature and mechanism of human consciousness.
  3. Why does ChatGPT have the potential to become machine-conscious?
  • ChatGPT has connected natural language and machine language
  • ChatGPT already has the ability to communicate
  • ChatGPT already has the ability to understand context
  • ChatGPT already has the ability to reason
  • The generalization ability and potential of ChatGPT on the context awareness problem has surpassed most algorithm models
  • ChatGPT has not yet become machine-aware, but we can train it through finetuning and prompt engineering, and let it gradually simulate GWT
  • Highly integrated, highly automated, and highly intelligent AI is just around the corner

The boundaries of language are the boundaries of thought

  1. American linguist and philosopher Benjamin Lee Whorf emphasized the interaction between language and thought, and put forward the viewpoint that "the boundary of language is the boundary of thought". He believes that different language systems have different ways of understanding and expressing the real world, so people's thinking and understanding will also be limited by language.
  2. Wittgenstein believed that the boundaries of language shape the boundaries of thought, which means that there is a close relationship between language and thought. He believes that language is not just a tool for expressing ideas, but a way of shaping them. Because language determines what we can think of, and even shapes our concepts and beliefs.
  3. In the past, we came into contact with various sources of knowledge, such as classrooms, books, speeches, discussions... But now ChatGPT can give us not necessarily the best, but it must be the fastest knowledge answer. I saw a metaphor on the Internet, saying that ChatGPT is to compress decades of Internet text information into low-resolution thumbnails, so that users can quickly spy on the whole picture and get a general idea. At the same time, because of the high compression, many details are lost. If things go on like this, as ChatGPT becomes more and more capable, in addition to receiving first-hand ChatGPT information, we may also see second-hand ChatGPT semi-processed articles, podcasts and videos everywhere. We are overwhelmed by a large amount of information produced by ChatGPT. How will the next generation of children grow up? Will they become thinking and talking like ChatGPT?


summary

I just hope that ChatGPT can be a good tool, but not an excuse for people to give up exploring the boundaries of curiosity and thought.


attached

Recently, I also talked about ChatGPT with friends who study Chinese medicine. My friend deeply expressed that part of the traditional Chinese medicine system may have disappeared from the online world in this intelligent transformation because there is no digital information precipitation.

If you have any friends with such concerns, you can leave a message on my account to discuss feasible cooperation and help.



CC BY-NC-ND 2.0

Like my work? Don't forget to support and clap, let me know that you are with me on the road of creation. Keep this enthusiasm together!