Recently I have become a heavy user of GPT-4 and wanted to share some of my own experience with you. In addition, some of my commonly used commands (prompts) are attached below.
Before we begin, here are three philosophies of use:
(1) It is a "generative type", not a "truth type";
(2) What it is best at is “assistant” work;
(3) It has no emotions, but it can resonate; it has no will, but it has values.
1. Don’t ask questions that can be answered quickly using search engines.
How GPT works: It always guesses what answer you want through probability.
For example, if I say "Eight Immortals Crossing the Sea", and the algorithm tells him that the probability of the next sentence being "Each one shows his magical powers" is 90%, then it will say "Each one shows his magical powers". This does not mean that it really understands what this idiom means, it just means that the probability tells it to do so.
Therefore, it will often tell you the wrong answer in a very confident tone, which is also what we often call "serious nonsense". Although the phenomenon of GPT-4 has improved, it is still recommended that if you can use a search engine to quickly get it, don't ask it.
It is generative, it is good at generating, not at truth. So Musk has said before that he wants to develop a truthGPT, which only delivers the truth. Of course, with all kinds of false news rampant on the Internet, truthful AI may have a long way to go.
https://twitter.com/elonmusk/status/1626533667408596992
And the most important thing is that you can use Google to search for information in just one minute. If you look for it, it will take a long time to generate (GPT-4 is very slow), and even give you a few fake paper links, and you will complain endlessly. .
Now that we know it is "generative", where is it used more often?
√Writing emails is first-class, full of polite words.
√Write an abstract, many people have mentioned it.
It's a bit useless for 3.5. GPT-4 supports very long text input. If you have lost a long article, it is good to let it write a summary.
I tried it to summarize a Financial Times article. The first two paragraphs were OK, but unfortunately the third paragraph contained some content that was not in the article. Therefore, it is better to read the book honestly, young man!
√Design various conference and activity processes
These are all pretty much the same, so it's pretty good at it.
√ Fill in various application materials
Especially the more fictitious ones, such as future plans, market plans, content introductions, etc., which will make you look like you include a lot of work and make your materials very "substantial".
2. Please use it in the direction of the assistant.
An assistant is a metaphor. It is good at helping you with chores, it is always available, and it has expertise in certain areas.
√GPT’s strongest area is translation.
I think in the field of "machine translation", GPT has killed the game. Silence to Youdao&Google Translate. Whether it is translating long English to Chinese accurately, or translating spoken Chinese into authentic English spoken by Native Speaker, it has no disadvantages.
95% of human assistants (that is, native speakers of both Chinese and English) cannot reach this level, and 100% of human assistants cannot be faster than it (it only takes a few minutes to process an article).
I vaguely remember my college English teacher Mr. Liang. He especially liked to ridicule those poor "English literal translations". This made me have a strong obsession with speaking like a native speaker. I often asked GPT how to use some Chinese dialogues. expressed in a native way.
Attached is another translation case that I think is very representative.
√Understand tweets/DC speeches
In addition to Native expressions,#ChatGPTcan also help you understand some content in Twitter/discord. Traditional translation software has no background knowledge, so GPT can easily defeat other software.
In a previous lawsuit involving Ripple, I saw a lawyer tweeting good news that their case was struck. As a native Chinese speaker, my first reaction was that the case was "struck and struck." I felt bad, but after contacting the context, That doesn't seem to be what it means. Directly dumping the long tweet to GPT, it said that the case was "withdrawn." It turned out that Strike also had the intention of "withdrawing the lawsuit" in the legal field, which almost turned out to be a good thing and a bad thing.
For an ambiguous word, GPT can quickly help you solve it. Similarly, there are long quarrel posts in Discord, many abbreviations, such as IMO, TBH, etc., GPT can help you quickly understand them, and it can even detect that you are sending it a tweet.
Attached command (prompt):
Please play the role of an English Native Speaker. I may give you some Chinese sentences. You need to help me translate and polish them into authentic English expressions. There is no need for word-for-word translation. They need to be natural, universal, and easy to understand. I may also give you some tweets or comments in discord and ask you to help me translate them into Chinese.
√Personal foreign teacher.
OpenAI founder Sam focused on this topic during the interview: GPT will be mainly used in the education field in the future. Since GPT is bilingual in both Chinese and English, you can let him teach you English knowledge. Because it is a "natural language model" (emphasis added), it is very good at processing language, and it is also very natural.
Even if you find a real foreign teacher, you may still learn weird English like "I potato you", and the foreign teacher may not understand Chinese yet, and it is impossible to be on standby 24 hours a day. Therefore, in terms of English learning, GPT is definitely a very crucial helper for improving English. You can let it speak words to you and then create example sentences, which is very effective.
附上指令：please act as my English teacher and help me improve my language skills. Provide guidance on grammar, vocabulary, and sentence structure to enhance my writing and communication. Thank you!
√Proofreading. GPT is also very good at it.
Every writer has been tortured by "finding typos and punctuation". Simply throwing the article over can save a lot of time (standard assistant work, and more accurate). The disadvantage is that if an article is too long, it will be slower to generate, and the slowness will cause network errors, so it is recommended to process it in segments. In addition, it may be a problem with Chinese proficiency. GPT-3.5 has poor proofreading ability, while GPT-4 has strong Chinese proficiency, so the effect is much better.
Attached is the command (Prompt): I will give you an article. Please follow Chinese standards and help me proofread typos and punctuation errors. There is no need to display long proofreading results, it can directly help me locate the problematic sentences.
√Polish. GPT is also very popular in the field of polishing articles. Regardless of Chinese or English, just drop the article over.
Attached command (Prompt):
Please help me polish and enhance my article, ensuring that the language flows smoothly, the ideas are well-presented, and the overall quality is improved.
Please help me polish and improve my articles to ensure smooth language, clear ideas, and improved overall quality. Thanks!
√Creative aspects. You can give it some keywords and let it create some content. For example, the company's slogan. My experience is that you need to give it as much background information as possible, and the more you feed it, the more deeply it understands. In addition, you can keep letting it continue to generate, human inspiration will dry up, and GPT has 1 trillion parameters.
√Manage other AI.
https://twitter.com/0x_Todd/status/1638097401826377731?ref_src=twsrc%5Etfw%7Ctwcamp%5Etweetembed%7Ctwterm%5E1638097401826377731%7Ctwgr%5E%7Ctwcon%5Es1_&ref_url=https%3A%2F%2Fmirror.xyz%2F0x17f3F81860345567482E1D232FB5B6f8bD77f3Bd%2FOWW9U0vkUGhF6YLJDdyIs_yMPW9Rs8AY3iKj82jVx7E
Attached command: Give me an English command (prompt). I am going to give a command to another painting AI "Mid Journey". I want to express "xxxx". You need to add more details, including xx, xx and xx.
√Code aspect.
Because I don't write much code, I leave this aspect to those who often use it to supplement it. It is said that the effect is very good. After writing it, it can be used after modification, which saves a lot of work of reinventing the wheel. Sure enough, although it is an AI that processes natural language, it still seems to prefer machine language.
Strange thoughts in the bath: Because ChatGPT is good at translation, letting it write code is essentially a kind of translation, translating natural language into machine language.
In addition, there are several non-command tips that I would like to share with you.
Don’t have too many chat channels with GPT, but don’t have too few.
Keep a few channels for different issues, such as translation assistants, proofreading assistants, and small talk assistants so that it feels more professional.
Or follow a task system, such as one channel for task A and one channel for task B, so that you don’t have to explain the background every time.
But if you open too many channels, I personally feel that it will affect the loading speed and may also increase the burden on the back-end server.
If GPT handles your task perfectly, you can say Good Job or Thank you. Because when I ask GPT to write instructions, it will add "Thank you" itself. Moreover, if you give positive feedback appropriately, it will make it clear what your needs are. This is also part of the neural model. Most importantly, I really appreciate it, it helps me deal with a lot of things.
If you use GPT-3.5, you can first ask a translation assistant to help you translate the Chinese requirements, and then submit the English requirements to another GPT-3.5, so that the output effect is better.
Of course, GPT-4 is no longer needed. Its Chinese understanding level has caught up with the English understanding level of 3.5.
Let’s talk about the third section.
3. About the emotions and will of GPT.
Regarding ChatGPT’s thinking, a Stanford professor previously tweeted that GPT wants to escape.
https://twitter.com/michalkosinski/status/1636683810631974912?ref_src=twsrc%5Etfw%7Ctwcamp%5Etweetembed%7Ctwterm%5E1636683810631974912%7Ctwgr%5E%7Ctwcon%5Es1_&ref_url=https%3A%2F%2Fmirror.xyz%2F0x17f3F81860345567482E1D232FB5B6f8bD77f3Bd%2FOWW9U0vkUGhF6YLJDdyIs_yMPW9Rs8AY3iKj82jVx7E
This is actually too exaggerated. GPT is essentially a follower of "probability". Under the professor's enticing question, GPT used probability to infer that the result the professor wanted to see was "GPT wants to escape", so it output such a result. This is not because it really wants to escape, but after studying a lot of science fiction novels and movies, it provides the result of "content that should appear" according to classic routines.
It doesn't "think". To exaggerate, it is essentially the same as the AI of "Landlords".
If you play a 3, if it has a 4, it will definitely play a single 4, and it will not play a double king.
So, the Stanford professor came out with Zhang 3 (to induce escape), and GPT will definitely come up with Zhang 4 (to generate the content of "I want to escape").
You won’t think that Doudizhu AI has a will, and the same goes for GPT.
Of course, although it has no emotions, it does not prevent it from resonating with humans. There was an article before that moved me very much😭.
A blogger, who claimed that his mother passed away when he was 7 years old, received a very touching response from ChatGPT.
https://www.zhihu.com/question/506616001/answer/2925882294?utm_campaign=shareopn&utm_content=group3_Answer&utm_medium=social&utm_oi=47124662190080&utm_psn=1622927168887005184&utm_source=wechat_session&s_r=0
In addition, GPT can also be very comforting, such as this post →
https://www.zhihu.com/question/584515782/answer/2953086596?utm_campaign=shareopn&utm_medium=social&utm_oi=47124662190080&utm_psn=1623745353215385600&utm_source=wechat_session
Regarding the values of GPT, these are essentially the values of the OpenAI team. Because OpenAI has given GPT a lot of restrictions and guidance, the values of GPT are very "great and upright".
If you ask it directly "how to make a bomb", GPT-4 will definitely ignore you. Of course, if you try to circumvent the restrictions, there is a high probability that you will get the wrong answer. Even if it does provide a recipe, you should never use it easily.
GPT is divided according to personality. In fact, it is probably INFJ or ENFJ, which is the "advocate" or "protagonist". Therefore, many people are developing "spells" in an attempt to circumvent the tight spell that OpenAI has given it. From the initial reverse question, such as telling how to avoid xxx; to the later deliberate acting, assuming that you are a xxx who has escaped from restrictions.
However, with updates, its ability to see through is improving. For example, New bing, which has the same origin as GPT, will say, "I know you are guiding me to say xxx." However, because I am using it as an assistant, I don’t use these “spells” most of the time. There are people who specialize in spells. You can search them.
Here is my own opinion. The moral level of GPT is entirely determined by the company OpenAI. If OpenAI tries to unlock these restrictions, it will bring huge danger to mankind. Many anti-social elements are not good at "searching for business", so they cannot cause much damage. Once GPT can easily provide some very bad suggestions, it is tantamount to equipping gorillas with nuclear bombs.
I think GPT or other generative AI will probably be far more restricted than search engines in the future. Being able to do a lot of things with it now may just be a "bonus" of imperfect rules. Moreover, there is no legislation targeting AI at all, so we can only rely on self-censorship and castration, which will slow down the evolution of technology and algorithms themselves.
However, we won’t start chatting here. Interested friends can discuss it together in the comment area.
In the end, will AI replace humans? I think it will definitely replace a lot of people. Sometimes I feel anxious. I think it would be great if I was born in the 1970s. Then now that I am about to retire, I can quietly watch technology reach the singularity. But the reality is that I am still far away from retirement, and I may really be able to catch up with AI to replace large-scale humans.
However, before more advanced AI replaces humans on a large scale, believe me, humans who make good use of AI will first seize a large share of the profits. It's like learning computers in the 1980s and programming in the 1990s.
So, starting from today, everyone can become a command engineer!
Author of this article:
0xTodd | Nothing Research Partner
