ChatGPT, a chat robot that knows astronomy and geography, has screenshots of conversations spread throughout the Internet. It is incomplete if you haven’t chatted with it about Qin Huang, Han Wu, and the life of the new crown virus.
Just as the Asian teams lost one after another and bid farewell to the 2022 World Cup in Qatar, ChatGPT gave an exciting answer. Perhaps, in a certain parallel world, we had the glory of being selected for the World Cup twice.
In the past few days, everyone has watched too many amazing performances of ChatGPT. Whether it is writing a thesis or writing code, it is done in one go, with ease. But its answer still has certain limitations, and there are many mistakes. If you know the correct answer, feel free to browse. If you don't know, it's better to be cautious.
The profound significance of ChatGPT is that it raises the lower limit of all articles in the world.
Most of the articles flooding various channels are far inferior to ChatGPT in terms of accuracy, knowledge and logic. However, no matter how good the article written by AI is, it is only a "short, flat and fast" data integration. Writers should not be satisfied with the "shadow writer" writing method of ChatGPT; readers should pursue more in-depth and higher-quality reports.
christmas tree story
ChatGPT is especially able to deal with some life problems, such as family disputes, emotional problems, etc. Last week, there was a screenshot of ChatGPT helping people answer "The wife is angry that the husband thinks the Christmas tree is too expensive and doesn't buy it". ChatGPT first pulls up, saying that there is something to discuss, don't be emotional, and then give a solution-or the two of them themselves Making a Christmas tree by hand is not costly; or traveling, you don’t have to pay attention to the price of the Christmas tree at home.
This routine is completely in line with psychological theory: to appease the emotions, stand on the side of the speaker, and then give some practical solutions rationally. From this perspective, ChatGPT is comparable to a psychologist who does not spend money.
"Christmas tree" was a high-frequency issue in December. Shopify executive Cynthia Savard "broke defense" against ChatGPT. ChatGPT wrote to her son: "I am writing to let you know that I am not a real person, but a character in a story your parents told you out of love."
“Your parents told you stories about me and my elves to bring joy and magic to your childhood. They wanted you to believe in the spirit and magic of the holiday.” Your love and care is real. They created great memories and traditions for you and hope your childhood was special."
The story of the two "Christmas trees" actually highlights the true ability of ChatGPT as a new generation of "chat generation pre-training converters": answering questions raised by humans as comprehensively, decently, and informatively as human beings.
Follow the human mind to answer
ChatGPT comes from the OpenAI research lab and is powered by the GPT-3.5 series of models, including model versions prior to 3.5, trained using text and code data on the Azure AI supercomputing infrastructure. It should be noted that the preview version that people are currently using is not connected to the Internet, and any responses from it will come from training the model offline. You ask it "how's the weather today", it doesn't know.
The most important change in the GPT-3.5 series of models is the adjustment based on real human feedback. This is a newly used AI training method. The marker will write the expected reply in the model, sort the marked answers according to the expected reply, and reward the model by sorting. In the process of continuously optimizing parameter iterations, the reward model is input to obtain optimized parameters.
▲ ChatGPT training steps
To use an analogy, it’s like you are taking a final exam, and you fill in the exam papers according to what you have learned throughout the semester. If you answer correctly, you will get a high score and a high ranking. When you go home, you can get the latest Super Mario Lego that your parents bought. So you made up your mind that you must study hard next year. If it’s a Chinese test, you know that the judging teacher likes Lu Xun. When writing the composition, he didn’t forget to say “In my back garden, I can see two trees outside the wall, one is a jujube tree, and the other is also a jujube tree.” classes, and the chance to unlock even more rewards.
After adding a large amount of labeled feedback that humans expect, it is equivalent to a kind of "predisposition guidance". According to the same mechanism, you can get rewards. As a result, ChatGPT has greatly surpassed its predecessor GPT3, which can only answer the "definition" of keywords dryly according to the question, while ChatGPT can already speak "following the meaning of human beings", sometimes it seems worse than a straight boyfriend intimate.
Machine learning and human brain learning are roughly the same. This is also the reason for the sudden rapid progress of artificial intelligence five or six years ago: the deep learning model convolutional neural network is actually a breakthrough in a multi-layer perception system similar to artificial neural networks.
Different neural layers receive and process different information, step by step, and get a "correct" description of something. The computer simulates this process and begins to "learn" people's "recognition" and "cognitive" abilities.
"Rewards" are similar. In most cases, all activities that humans do that are beneficial to survival will secrete some reward substances such as serotonin, oxytocin, endorphins and endogenous cannabinoids in the body to make the brain feel "happy".
And AI can get the high score judged by the system, which is a "reward" for it.
writing is a dignity
ChatGPT already has all the qualities of an excellent customer service. Compared with the foolish customer service robot that people often encounter, "I suggest you say it again", ChatGPT can definitely make all consumers happy and regain the feeling of God that they no longer have high blood pressure. Any industry that requires customer service has the potential to benefit from advances in this technology.
In addition, ChatGPT is also a good efficiency tool.
Such as search. Some people think it is enough to replace search engines like Google, maybe too absolute. First of all, ChatGPT's integration of data is based on "dialogue", and there are a lot of nonsense about succession and transformation, which will waste the searcher's time. Secondly, ChatGPT does not absolutely guarantee the authenticity, and does not provide the source of the information, and the searcher even has no way to verify it in many ways. It can be said that under the premise of fidelity, some industries that require knowledge inspection may be expected to use ChatGPT.
In other words, GPT-4 with a larger database is the big killer that can compete with the search engine giants.
Like writing. Not just ChatGPT, GPT-3 is already very good at writing papers. They will arrange the information they have collected under the very typical TOEFL essay writing ideas, which is well-organized and very standardized, but there is really no "originality".
ChatGPT is a boon when it comes to writing code. If you install a ChatGPT plug-in on the browser and ask code questions at any time, it is likely to avoid many detours on the road to Ali P8.
▲ Jonas Degrave shows how to turn ChatGPT into a Linux terminal and interact with the virtual machine directly from the web browser (manual dog head)
Like painting. ChatGPT can be linked with the recently popular Stable Diffusion. Use ChatGPT to form a paragraph of text that meets your requirements, and then input the text into Stable Diffusion. The resulting paintings are generally much better than those directly entered by yourself. It may be that AI understands AI better-ChatGPT's description is more detailed and easier to extract.
As a professional writer, the emergence of increasingly intelligent AI such as ChatGPT, I think it has played a very good role in promoting reflection.
In a sense, ChatGPT is another aspect of question-and-answer media and knowledge payment. You want to know a skill or a piece of information quickly, ChatGPT will answer you immediately without thinking. Reading and knowledge can be acquired through very convenient, minimalistic technology, not through a long, winding learning process. Unlike search engines, you don't even have to screen the information yourself to identify the source, ChatGPT will do it for you.
For creators, this is a big challenge. Fast and direct is the goal, watching a movie in five minutes and explaining a masterpiece in ten minutes. ChatGPT is a technology that meets the needs of the times, but it should not be the purpose of creation.
▲ The infinite monkey theorem states that if you put an imaginary monkey in front of a typewriter for an infinite amount of time, the monkey will eventually write Shakespeare. The monkey in the picture is writing "Macbeast" (Macbeast), a parody of Shakespeare's "Macbeth" (Macbeth)
Articles created by "shadow writers" will soon flood various media and channels, and many people will not be able to tell the difference between them and human writers. The "Space Opera House" drawn by Stable Diffusion won an award, but the judges didn't know it was drawn by AI at first.
This is a normal phenomenon, or in other words, the high-quality text of AI is sufficient for circulation and commendable. But if human creators are satisfied with this, there may never be "Macbeth" born.
ChatGPT, as well as the previously released Notion.ai, can help creators quickly generate drafts or outlines, save time and energy to focus on the details and depth of writing, and create more valuable works.
With AI, it is unreasonable for us not to create great works earlier.
#Welcome to pay attention to Aifaner's official WeChat public account: Aifaner (WeChat ID: ifanr), more exciting content will be presented to you as soon as possible.