ChatGPT is very hot these days. If you’ve been browsing social media, you’ve probably been overwhelmed by all the news. Everyone from the bigwigs to the average Joe has shown a lot of interest in the topic, and I’ve even received 3 notices for ChatGPT-themed academic seminars. There are natural benefits to being involved in this topic, such as easier access to traffic – and to be blunt, I’m one of the beneficiaries of the ChatGPT boom. Never before have I received so much traffic for a post or answer.
But if you’ve been listening to people talk about ChatGPT, then I’m afraid you’re not the traffic getter. To put it bluntly: you are the traffic. My advice to you here is to get out of the hunt or even “stand in line” as soon as possible and use ChatGPT more efficiently to gain a competitive advantage for yourself. To achieve this goal, I think you need to do two things: first, set the right understanding; second, hands-on operation. Let’s talk about them separately.
A couple of days ago, a science fiction writer wrote a very popular article about ChatGPT, in which he made an analogy that ChatGPT is “a lossy compressed file of the full set of Internet information”. At first I thought this was a novel idea, at least the observation and comparison angle was very unique. But the further I read, the more awkward I felt.
Let’s not forget that ChatGPT is based on a large language model like GPT-3, which portrays language patterns rather than recording language content itself. If we were chasing memory for rich data, the most powerful and technologically advanced institutions in the world would be Wikipedia or large libraries and archives, and ChatGPT is amazing not because it recalls what the world has already produced, but because it creates new content based on the specific needs of users.
ChatGPT is not a search engine. It can output some code based on your specific requirements, and if you are lucky enough, this newly generated code will even work without manual tweaking. While much of the existing code is very similar and follows the same syntax rules, new code is new code, and that code has never existed on the web before. This kind of creation is something that search engines cannot achieve.
ChatGPT’s amazing advances don’t come from higher-ratio ‘lossy compression’ techniques. It looks for patterns – patterns that solve problems, patterns that describe the content, and may even contain some other patterns that we just aren’t aware of yet, or haven’t tried yet.
Is it really important to look for patterns rather than memorize content?
Absolutely. Some people like to take notes by highlighting or excerpting passages from books, but there have been many studies that prove that these practices are not really significant. Instead, it’s more important to paraphrase (elaboration) in the new context when taking notes, because it’s your own words, and it involves creativity. ChatGPT seems to cross this limit. At least in terms of programming, it can now offer new solutions directly to the user.
The new code may be less than 5% different from the code that already exists in the world, but don’t forget the old adage “there is nothing new under the sun”. Steve Jobs even put the concept of “creativity” more bluntly: “Creativity is just connecting things. When you ask creative people how they do something, they feel a little guilty because they don’t really do it, they just see something. After a while, those things seem obvious to them.”
Is it appropriate to ignore the dynamic creation capabilities of ChatGPT and just treat it as a lossy compressed file? Let’s think about it further: does the writer who made the brilliant analogy really understand the nature of large language models, does he know what has happened in the evolution of the language model, and how the model handles input and output?
If he had a clear understanding of all of these issues, he would never see the model as a giant Jpeg file that simply compresses and decompresses. Conversely, what happens if he has no idea about the technical aspects and only goes to observe a certain number of ChatGPT outputs before making a hasty judgment?
Writers can cause you to misjudge new things. For example, you might think that ChatGPT is just a memory hack because of the writer’s inappropriate analogy, and therefore think it’s not relevant to your work. You’re likely to miss out on opportunities as a result.
What we just said is just one of the directions readers may be misled. The fact that you think “ChatGPT is not relevant to me” is not necessarily not true. More passionate authors may lead you to the other extreme – thinking that ChatGPT is a sign of the advent of strong AI, or even worshiping it as a myth.
ChatGPT is not only “out of the loop”, but also out too fast. A stock speculator once said that if you see a woman in a food court recommending stocks to you, you should know that the bull market has reached its peak. One of my friends sent out a bitter laughing emoji with the message that his “chef father, who has never been interested in high technology, was forwarding him ChatGPT information”. When I saw this, I was shocked – ChatGPT has only been out for two months, is it possible that this wave has already reached the top?
Maybe, but it’s not really ChatGPT’s fault: OpenAI has not made any significant exaggerated false claims so far, but the expectations of ordinary people are really “too high”. Many people tried it and found that they could benefit from ChatGPT, so they gave it free publicity; after the publicity reached a larger group of people, the public’s perception of ChatGPT’s functions gradually went out of shape.
This is like the case of a targeted and effective medicine that has been rumored to be a miracle cure or even a “powerful pill”.
These days, many people are talking about how to use ChatGPT to make money, for example, some people use it to build websites and help people make resumes. Such applications, at least, are still considered valuable. But the amount of money they make is just small: some people directly exploit the information asymmetry to really make a lot of money. I saw a screenshot inside my circle of friends, and although I don’t fully believe the numbers in it and think it’s bragging, at least this phenomenon should be real.
Some people will go out of the reality of the current ChatGPT capabilities and make all kinds of irresponsible attempts. This frustration then spreads and brings deep disappointment to the public. For example, the following screenshot of the conversation, without source information, you take it as a joke to talk about.
For a hot technology product, these encounters may be inevitable. It’s just that we should be wary of getting too excited and following some people into a frenzy.
This is neither here nor there, so how do we get our understanding of ChatGPT right?
There is nothing mysterious about ChatGPT. You don’t have to understand the architecture of the big language model, or even how the basic units of the neural network are connected. It’s not a test. All you need to know is that it’s a successful attempt at engineering artificial intelligence.
The excitement and thrill comes from the fact that a new AI product is actually available and can help you. 2022 has seen a lot of things happen in AI that could go down in the technology chronicles. But the real excitement is not about breakthroughs in AI theory or even research paradigms, but about the success of ‘engineering AI’. What is “engineering”? The 1970s personal computers were engineered, the 1990s Internet was engineered, and what we saw in 2022 was DALLE2, Midjourney, Stable Diffusion, and ChatGPT.
Yann Lecun said that ChatGPT is not a new technology and OpenAI is not as good as Google, and there is nothing wrong with that from a purely academic point of view. But history will remember that it was OpenAI, not Google, that made ChatGPT first and thus led the AI breakthrough.
We often have a special admiration for the inventors of some hot things, thinking that they are on the right path every step of the way and that their strategic planning is clear and exceptional. But if you look behind the scenes, you’ll see that many successes are not the result of sophisticated design and planning. In the case of ChatGPT, OpenAI’s initial idea was to test it internally before releasing the chatbot. But OpenAI encountered tremendous difficulty in training the model, and even thought about giving up or at least making significant adjustments to the goal. Why?
Because the ChatGPT internal testing process was uneventful, and OpenAI used a technique called RLHF (Reinforcement Learning from Human Feedback), which relies on feedback from real people to train and calibrate the model.
However, the internal testers that OpenAI found sat in front of ChatGPT and didn’t know what to say. As a result of the awkward conversation, the project was not progressing well, and OpenAI simply decided not to wait and opened up the beta testing in a half-baked form. That’s right, every ChatGPT user in the “Research Preview” is actually a free beta tester for OpenAI. In the past two months or so, “beta testers” around the world have been contributing to the improvement of the model night and day. The size of the beta testers grew rapidly until OpenAI had to charge to keep its servers from being overwhelmed.
The release of ChatGPT was actually a risky decision. The selection of seed users would generally avoid being so hasty. For one thing, it’s difficult to recover a model from a word-of-mouth flop in the initial group of enthusiastic users; for another, user input and feedback data is likely to contaminate the model. This was the case when a large factory’s dialogue robot learned a “dirty mouth” within a few days.
But apparently, the decision turned out to be a safe one – ChatGPT not only didn’t flop in word-of-mouth, but also caused other major companies to issue red alerts and even saw significant stock price fluctuations.
You see, ChatGPT is just a big language model that relies on the help of many, many users to make rapid progress. We don’t need to tout it too much, nor do we need to follow others in identifying it as a gimmick. Once we have set our perceptions right, we can start the next step: getting up and running.
My second piece of advice is to not follow the crowd. Trust your own eyes, trust your own hands, try ChatGPT, and then summarize and iterate more based on your practice results to really use it to your advantage.
I am currently using it for two main purposes, one is writing and the other is programming.
Let’s start with writing. After writing for a long time, you will easily encounter the creator’s bottleneck like Writer’s Block, and I try to use ChatGPT to break through it.
For example, let it help you think of some examples to support your assertions: while writing, you suddenly remember that the example you gave is not quite sufficient, is there a better example? Previously, this meant going to a search engine, typing in keywords, and sifting through a sea of results. But now you can tell ChatGPT to “add an example to prove the above assertion”.
I think the examples here can’t be used directly, but they are still inspiring to the author. Face recognition, for example, has become a “daily use but not a knowledge” thing. I use it almost every day to verify payments, but it’s really not the first thing I think of when writing as an example of AI engineering. If you don’t think it’s a good enough example, let it be followed by others.
Sometimes, I even found that ChatGPT has the ability of “generalization”. For example, if you mention a certain phenomenon, you can ask it to find a similar situation. For example, in the example below, I analyzed the ChatGPT example, but wondered what other historical events rhymed similarly. Just ask ChatGPT. I think the answer uses the Internet bubble to describe the “high expectations of the public”. But you need to be careful not to use the results of ChatGPT without screening them. For example, if you look at the answer below, do you see any problems?
The problem is still very obvious. Who launched DALL-E, Facebook or OpenAI? I call this incorrect answer to ChatGPT “serious nonsense”. You’ll find that it’s not really sure of the answer, but it’s answered with a tone of “categorical”. If you don’t choose the right answer, you will be embarrassed when you publish your article or book in the future. So you should be careful when choosing ChatGPT answers.Another application is programming – an example of step-by-step requests and continuous improvement. You can refer to my “How to write a crawler with ChatGPT” video, which is not repeated here.
I also recently found something more interesting that I want to share with you. In the above example of writing a crawler, I was still required to explicitly state “how” ChatGPT needed to be done, i.e., describe the steps of the process, but you can actually communicate with it in another way – by directly telling it “what” to do. Show it an example, and then ask it to write the program according to the example. This is actually asking ChatGPT to “actively learn”.
For example, in the example below, I asked ChatGPT to replace a specific piece of text with a specific piece of content that appears in the text. Interestingly, instead of telling it exactly what to do, I give it a relatively vague example. What appears in the example … is not really text, but represents an arbitrariness.
After trying it, I found that the code ran smoothly and the text replacement worked. I was very excited, not just because ChatGPT helped me solve the specific problem at hand. It’s because with this discovery, I’ll be more comfortable to let ChatGPT help me with some mechanical and repetitive text processing tasks in the future, because I can just give examples.
There are many different voices in the ChatGPT discussion, so don’t get caught up in it and be misled. You may miss out on the competitive advantage if you think ChatGPT is uninspiring and useless, or you may go crazy and be deeply disappointed if you think ChatGPT can’t do anything. Get it right, think of it as an interesting big language model, a successful start to AI engineering, and focus on trying to use it to improve your own workflows to reap the real benefits. Way to go and happy using ChatGPT, or Google’s Bard.
Published by YooCare Editor on February 17, 2023 1:21 am, last updated on April 10, 2024 7:51 am
