Author: Zhen Kexue Source: Qin Shuo WeChat Moments (ID:qspyq2015)Recently, a chatbot model, ChatGPT, has become popular online. Its smooth responses and rich knowledge have made many netizens worry about their job security. Is ChatGPT really that powerful? This article analyzes three non-scientific aspects and five scientific aspects, pointing out the need to avoid the “non-scientific” misuse of AI and to utilize its rapid learning and imitation capabilities reasonably, as this technology has unlimited potential.Full text 5170 words, reading time approximately 14 minutesChatGPT was released on November 30 last year, and I started trying it out, writing a few small articles. At that time, it created a buzz overseas and in industry circles, while most people in China had little reaction.In spring, broker reports strongly promoted it, Microsoft invested billions in OpenAI, and Baidu followed suit, leading to an even greater hype around ChatGPT in China than abroad, exceeding expectations.However, ChatGPT does not allow user registration from Chinese IPs and phone numbers, so most people have not really used it. Some have set up relay servers with foreign accounts for friends in China. Due to these limitations, most users have only had a cursory experience and are merely spectators. Some articles even exaggerate ChatGPT’s capabilities.Therefore, I want to share my experiences from the past two months about the scientific and non-scientific uses of ChatGPT.These scientific and non-scientific aspects also reflect people’s thoughts on the scientific nature of ChatGPT in specific application scenarios.
Image: The First Pig in the Year of the Rabbit, AI Generated
First, let’s talk about the non-scientific aspects.1. Using it as a search engineWith Google and Baidu entering the fray, many people discuss the future replacement of search engines. Thus, it is natural for some to use and test ChatGPT as a search engine.This is a significant problem!If you input the question “How to make fish-flavored shredded pork?” a search engine finds existing methods on the internet and sorts the results. However, ChatGPT provides a text that looks like a “recipe.” This text does not come from a complete cooking manual; instead, its algorithm generates it word by word based on the probability of each word and context. You seem to see a complete recipe, but it has never existed, nor has anyone used it.ChatGPT merely constructs a paragraph of cooking text. To understand this, we need to return to the ChatGPT model.The ChatGPT model integrates in-context learning, RLHF, and other models.We won’t delve into the technical discussion here.One part of the model is trained using methods similar to the cloze test, which we are most familiar with in English exams.For example, a cloze test question: I like the feeling of ___.The blank can be filled with “I like the feeling of running,” “I like the feeling of cooking,” or “I like the feeling of daydreaming.”What you fill in depends on the dialogue context.If you fill in “I like the feeling of being killed,” the probability of that scenario occurring is very low.The algorithm lowers the probability of the phrase “I like” being followed by “being killed” to three words.If you fill in “I like the feeling of a table,” that is a completely incorrect answer, and the algorithm gives it a zero probability.Based on many similar questions, letting the machine answer and correcting for accuracy can train the algorithm.Thus, ChatGPT is a game of text encoding and probability combinations, not accurate factual information.The greatest challenge for search engines is finding the most relevant and accurate information. Currently, ChatGPT cannot replace search engines. In the future, it could be a great assistant to search engines, combining accurate results into more readable formats.Imagine a scenario of “search engine + ChatGPT”: based on the accurate information from the search engine, ChatGPT could produce a well-structured, beautifully written article without you having to open, copy, organize, and paste each search page.2. Using it as an expertI once asked ChatGPT: If the Chinese economy wants to return to a high growth rate in 2023, what issues need to be resolved? What suggestions do you have?
It answered: To return the Chinese economy to a high growth rate, I have the following suggestions: 1. Focus on strengthening consumer demand; 2. Enhance reform and innovation; 3. Improve the financial system; 4. Strengthen social investment; 5. Perfect the technology system.
These answers seem well-founded and indeed align with what experts say, with some novel aspects. This is because, when asking this question, most of the texts that the model has been trained on regarding economic growth rates are spoken by professionals. In a professional context, the language produced is that of experts, not the general public’s response: “What does the growth rate matter to me? As long as I get paid, I’m fine.”Therefore, you cannot trust what ChatGPT says and cannot treat it as an expert. However, these statements can inspire you to think about whether these five suggestions are correct.
Image: Don’t Trust AI, AI Generated3. Expecting One-Time ResultsThe name ChatGPT can be misleading.Generally, when we communicate with people, we do not repeat the same question to the other party;Normal people do not provide different answers to the same question each time.Human thoughts are consistent, and what they say is consistent.A word is a promise; people should not speak carelessly.ChatGPT is different; you cannot treat it like a person. If you ask a question and feel that the answer is unsatisfactory, you can repeat the same question until you get a satisfactory answer. ChatGPT’s algorithm’s randomness means that it provides different answers each time, which can include completely opposing viewpoints.ChatGPT has a parameter called temperature. If the temperature is 0, it does not say unreliable things; for binary or unique answer questions, it does not make wild claims. If it doesn’t know, it says it doesn’t know. If the temperature is greater than zero, it provides responses that are related with a certain probability, resulting in different answers each time.I refer to the temperature parameter as the “nonsense parameter.” The higher the nonsense parameter value, the more imaginative it becomes, until the language is correct but mismatched, and the facts are completely wrong. It’s like encountering someone who occasionally speaks absurdly in real life.I have seen some friends try ChatGPT, ask open-ended questions, glance at the answers, and feel that the responses are dull, leading them to think ChatGPT is uninteresting and stop using it. This way, they cannot experience the wonders of ChatGPT.To obtain interesting results, you need to keep asking it, and you will be pleasantly surprised.For example, the same question can be set to a specific tone. If I ask ChatGPT to politely remind others to wear masks, its response is:“Please be sure to wear a mask to protect your health and the health of others.” If I ask it to remind others to wear masks in a more aggressive tone, the result is: “Not wearing a mask is like seeking death!” Completely different.This is where ChatGPT excels, creating very useful application scenarios.
Image: AI and Human Brain, AI Generated
Next, let’s discuss some scientific aspects, which are the areas where ChatGPT is suitable and can be utilized now.1. Universal OutlinesIn discussing the “non-scientific” parts, we recognized that ChatGPT imitates and produces text that resembles a tiger. When trained on logically rigorous and comprehensive scientific papers, ChatGPT gains the strongest logic and the most comprehensive thinking from a “surface” perspective, making it very suitable for breaking down a research topic into various aspects and listing research outlines or templates.For example, if I want to write a paper titled “How to Help Men Overcome Midlife Crisis,” or if a friend is experiencing midlife anxiety and you want to help guide them, ChatGPT can quickly outline the topic:
1. Definition of Midlife Crisis (1. What is midlife crisis; 2. Main manifestations of midlife crisis; 3. Causes of midlife crisis)
2. Methods to Help Men Overcome Midlife Crisis (1. Enriching oneself: boosting self-confidence, expanding interests; 2. Strengthening connections: maintaining good communication with family, friends, and colleagues; 3. Changing life: altering work status, improving lifestyle habits)
3. Psychological Care for Men Overcoming Midlife Crisis (1. Correctly understanding midlife crisis; 2. Finding ways for self-actualization; 3. Finding self-regulation methods)
4. The Impact of Social Environment on Men Overcoming Midlife Crisis (1. Family environment; 2. Work environment; 3. Social environment)
At the same time, it can provide you with the latest research conclusions and methods for overcoming midlife crisis:
The main causes of midlife crisis in men are career development and family pressure. Therefore, the best methods to help men overcome midlife crisis are to enhance their career development abilities and family skills to reduce the pressure they bear. Participating in social activities, cultivating a good mental state, regulating emotions, and finding hobbies can effectively help men overcome midlife crisis. Healthy lifestyles and good sleep habits can help men improve their psychological and physiological states, thereby alleviating midlife crisis.
The complete aspects and viewpoints generated by ChatGPT are not something everyone can easily think of.Alternatively, if you want to know how to evaluate the investment value of a publicly traded company, ChatGPT can comprehensively list the factors to consider:
1. Financial Indicators (1. Financial Statements: Analyze the company’s balance sheet, income statement, and cash flow statement; compare the company’s revenue, profits, and financial condition with those of industry peers; 2. Financial Ratios: Study the company’s return on assets, cash flow ratios, inventory turnover, fixed asset depreciation rates, etc., to reflect the company’s financial status; 3. Financial Risks: Analyze the company’s financial risks, including short-term solvency, liquidity, and mortgages; study the company’s debt situation and compare it with the industry)
2. Market Performance (1. Stock Price: Study the historical price trends of stocks and compare them with industry peers; 2. Stock Liquidity: Study the liquidity of stocks, observing turnover rates, trading volumes, and yields, and comparing them with industry peers; 3. Dividend Policy: Analyze the company’s dividend policy, including dividend payout levels)
This is why Microsoft is optimistic about ChatGPT. Knowledge workers, even if they have learned this knowledge, have mostly forgotten it. In one minute, they cannot think of such comprehensive factors. Even if they do, they have to type it out word by word, right? ChatGPT doesn’t require typing; it automatically generates various outlines for selection and optimization, greatly enhancing efficiency.
Image: AI Assistant is Great, AI GeneratedThe training model of ChatGPT has been trained on professional paper databases, and it decomposes issues into various aspects very comprehensively and deeply. The outlines are well-structured and very beautifully presented. Those who have written papers, speeches, or research reports know the value of this.2. Text RewritingDifferent contexts require different language styles. When writing articles, if a writer wants to make the text more refined, elegant, or more accessible, ChatGPT can rewrite it to match the desired writing style.For example, the previous paragraph has a formal tone, and I want to make it more casual. ChatGPT can change it to:
Different scenes need different language styles. ChatGPT is like a magic wand that helps writers make their writing cooler, more awesome, or easier to understand!
In different contexts, different language combinations are needed. When writers are writing articles, if they want to make their writing more exciting, elegant, or easier to understand, ChatGPT is your blessing, helping you write in a style that fits your writing.
Or, if I want to change it to a more classical literary style, ChatGPT would rewrite it as:
For literary works, if one desires refinement, beauty, or simplicity, ChatGPT can assist the author in achieving the desired literary charm.
3. Detail OptimizationI am not a professional writer and struggle with writing descriptive scenes. I can ask ChatGPT to help me with details. For example, describing a forsythia flower:
This beautiful forsythia flower, with its vibrant petals, resembles a dazzling thread, radiating the enthusiasm of spring;
The brilliant stamens, like twinkling stars, convey the warmth of spring;
The fresh floral fragrance, like a gentle breeze, sweeps in the vitality of spring.
It is like a beautiful little fairy, bringing the beauty and warmth of spring to everyone.
This greatly enhances efficiency. Novelists can focus their energy on plotting while not needing to spend time polishing the text.So, can ChatGPT’s writing convey emotions? The text above about the forsythia flower carries a beautiful sentiment; I can ask ChatGPT to write it with a sad emotion:
This forsythia flower, vibrant and dripping, seems to be sorrowfully sighing about something.
Its petals, like tear-filled eyes, are moist and soft;
Its stamens, like hands that once embraced deeply, are helpless and gentle.
Its fragrance seems to whisper something quietly, yet no one can hear.
Its beauty is both a sadness and a tenderness.
Image: Sad Flower, AI GeneratedSome descriptions resemble middle school essays. While good, they may feel somewhat juvenile. ChatGPT’s strength lies in its personalized training model, which can train a person’s language style, resulting in future writing reflecting a similar language style.I previously wrote about generating artworks in the style of any artist; ChatGPT may also achieve generating writing in the style of any author.This raises intellectual property issues. Many organizations oppose AI-generated works abroad. ChatGPT is also very cautious in the countries and regions it serves because the laws differ in each place, and uncontrolled application could lead to intellectual property disputes.4. Code AssistantProgramming is also a language. Converting natural language into programming language is another area where ChatGPT can be applied. Since the audience here is primarily programmers, I won’t elaborate further.The issues here also focus on accuracy. The most time-consuming part of writing code is finding bugs. If the accuracy is insufficient, the code produced may have too many bugs, making it worse than writing it yourself. To utilize this function effectively, users need to clearly describe the purpose of the code in natural language. Only then can ChatGPT accurately “translate” natural language into code language.In the future, programmers’ work will increasingly shift towards understanding requirements, architecture, and algorithms, rather than the process of writing code.5. Inspiring CreativityWriters often struggle to produce content. When they can’t write, it can be due to either a lack of ideas or not knowing where to start. With AI, it can provide many ideas for you to choose from.
Image: The Pain of Writers, AI GeneratedOriginally, it was thought that artificial intelligence could only replace mechanical work, but in reality, AI can randomly create countless possibilities.Innovation is about cross-disciplinary combinations and wild ideas. AI can randomly generate many things, allowing you to sift through for gold, uncovering great inspiration.If it involves combinations of facts, you need to judge whether the factual basis is accurate. AI is adept at making absurd claims, so humans need to be stronger than AI to use it effectively. Just as leaders must have the ability to discern what their subordinates say to accomplish tasks effectively, AI raises the bar for humans.Some have used ChatGPT to write books or papers, while others have only engaged in casual chats with it. Using ChatGPT requires skill. Just as some make their PowerPoint presentations into blockbuster movies, while others can’t even make their PowerPoint better than a Word document. The difference lies in how people use it.Before the 1950s, literate people were considered talented; now, literacy is no longer seen as a skill. The development of AI is akin to eliminating illiteracy, making certain skills accessible to the masses and no longer exclusive to a select few.In conclusion, the “non-scientific” aspect of AI lies in avoiding its “non-scientific” misuse. Once the internet is filled with inaccurate information generated by AI, it will destroy the internet—AI will destroy itself. Improving accuracy is a challenge for ChatGPT’s competitors and a problem that ChatGPT itself must address in the future.The “scientific” aspect of AI lies in its ability to use faster electronic signals to simulate and accelerate the chemical signaling learning process of the human brain. AI is still like a three-year-old child that has just learned to speak and dares to say anything. Given time and continuous learning, this child has limitless potential.
Image: 3-Year-Old AI, AI GeneratedYou might also like these articles[Original] Comprehensive Deconstruction: The 17-Year Talent Reserve “Introduction, Cultivation, Use, and Management” in Shanghai[Original] The Experience of Developing the TOD Model in New Towns in Hong Kong and Its Implications for the Urban Sub-center of Shanghai[Original] Wang Yajie from Hualuo Think Tank: How the Yangtze River Delta Cultivates Leading Enterprises?[Original] Breaking the “Neck” Bottleneck: What Can the “Suzhou Model” of Innovation Alliances Bring?[Original] Two Years of Digitalization in Shanghai: Leading the National Industry Digitalization Services[Hot Article] The Latest Article by Hunan Provincial Party Secretary Zhang Qingwei, Deeply Grounded![Hot Article] Analysis of Investment Attraction Models and Cases in Zhejiang[Hot Article] Peking University Report: With the Decoupling of China and the US in Technology, Key Areas of Chinese Manufacturing May Encounter Bottlenecks[Hot Article] Zhejiang Regions Competing in “Sweet Potato Economy”: What Strategies Are Worth Learning?[Hot Article] Why Reform and Opening Up Was Necessary 45 Years Ago: A Shocking Memory from a Provincial-Level Cadre
Don’t forget to like and share!