This article is reproduced from Lei Feng.com. If you need to reprint, please go to the official website of Lei Feng.com to apply for authorization.
Last Wednesday, OpenAI released the conversational language model ChatGPT and opened a free trial. According to OpenAI CEO Sam Altman, ChatGPT has reached 1 million users in just 5 days, while the previous GPT-3 took nearly 24 months to reach this number of users.
In the description given by OpenAI, ChatGPT is a "can answer follow-up questions, admit mistakes, challenge incorrect premise and reject inappropriate requests” dialogue model.
After the trial was opened, a large number of users started conversations with ChatGPT, ranging from chatting and answering daily questions to generating poems, novels, video scripts, and writing and Debugging the code, ChatGPT demonstrates its amazing capabilities. As the hottest AI model currently, ChatGPT’s breakthrough influence is even greater than that of GPT-3 two years ago.
As a language model, ChatGPT has the most basic The text generation ability is extraordinary in creating and continuing to write novels, poems and other literary creation scenarios.
For example, ChatGPT can generate a paragraph for you using Lu Xun’s literary style:
Tian Yuandong, a researcher at Meta FAIR, shared that he uses ChatGPT to continue writing his own novels:
Create poetry as required:
## Tell Soviet jokes:
ChatGPT can also talk to people in non-text form. For example, a netizen asked ChatGPT to describe what it would feel like to be "liberated" as an AI, and asked that he could only use emojis to answer. As can be seen from ChatGPT's answer shown in the figure below, it can understand the meanings of various emojis and arrange them according to the logic of text narrative.
The power of ChatGPT is also reflected in its "programmer" capabilities. In the following example given by the official, ChatGPT can help debug the code, and can also question the rationality of the questions and ask the user to adjust the questions.
The CEO of the American code hosting platform Replit also posted a post praising ChatGPT’s coding capabilities: not only can it explain bugs, but it can also Fix the bug and explain how to fix it".
Using the tips given by ChatGPT, you can also create a website in 10 minutes. Even novice programmers can use the code it generates to develop a production-level application. Replit therefore calls ChatGPT “ It changed software development forever.”
ChatGPT’s powerful question and answer capabilities have also been explored by netizens as its potential to act as or even replace a search engine. A few days ago, a very popular post on Twitter claimed that "Google is done". A netizen asked the same question about Google search and ChatGPT, such as "How to write a differential equation in Latex?" .
The answer given by ChatGPT completely exploded Google search:
Many netizens have developed Google plug-ins, which can browse Google search results and ChatGPT at the same time. The answer given:
As a conversation model trained from massive data, ChatGPT is like an expert in various fields, able to provide professional advice for your study, work and life around the clock.
For example, let ChatGPT answer questions related to thermodynamics for you:
Explanation of a complex regular expression:
It can also become your language learning Instructor:
#ChatGPT even "invaded" the political context. A Canadian member of Parliament asked ChatGPT to write a paragraph introducing itself to the House of Representatives and put forward reasons for whether its use should be regulated. ChatGPT responded with reason, "My development should not be regulated."
#In the recent AIGC field, there is certainly a place for ChatGPT to play a role. After a large number of AI painting applications came out, many people racked their brains on prompts in order to obtain high-quality images. Now ChatGPT is a ready-made prompt library.
For example, a netizen asked ChatGPT for design suggestions for living room decoration, and obtained exquisite images on Midjourney based on the description it gave:
ChatGPT can also write raps for you. The picture below is a rap song written by ChatGPT about robbing a house. It even has a very sense of justice and will prompt "illegal or harmful activities."
Write a Mozart-style piano score:
In addition, some netizens use ChatGPT to generate video scripts, which can be said to be good news for the majority of video bloggers.
In the minds of millions of users, the imagination space of ChatGPT is undoubtedly huge. This wave of trials It has brought a variety of applications that are either practical or fun, as well as many unexpected capabilities.
For example, someone used ChatGPT to bargain with Adobe and got a more favorable monthly rental price for themselves. The customer service staff on the other side probably didn’t expect that they were talking to an AI. Conversation, it must be said that ChatGPT “successfully passed the Turing test”.
The above are just examples of the tip of the iceberg. How much more "magic" can the "magic box" of ChatGPT continue to release? For us to discover.
Judging from the current user feedback, ChatGPT’s The language ability is generally passable and excellent. Huang Minlie, associate professor of the Department of Computer Science at Tsinghua University, told AI Technology Review that the key capabilities of ChatGPT come from three aspects: base model capability (InstructGPT), real data, and feedback learning.
ChatGPT is fine-tuned from a model in the GPT-3.5 series and is a brother model of InstructGPT, so ChatGPT has a powerful base model ability.
GPT-3 has made great iterations and improvements in capabilities since its release in 2020. Huang Minlie believes: “OpenAI has established a strong foundation for users, data and The flywheel between models, it is obvious that the capabilities of open source models have lagged far behind the API capabilities provided by platform companies, because open source models have no data."
ChatGPT uses the same method as InstructGPT, trained through reinforcement learning with human feedback (RLHF), but has slightly different data collection settings.
The researchers trained an initial model using supervised fine-tuning: a human AI trainer acted as the user and AI assistant in a conversation, collecting data in the process. Huang Minlie believes that this kind of Fine-tune on real call data can ensure the quality and diversity of data and learn from human feedback. The amount of training data for InstructGPT is not large, totaling only 100,000, but the data quality (well-trained AI trainer) and data diversity are very high, and the most important thing is that these data come from real The world calls for data, not the "benchmark" that academia plays.
#To create a reward model for reinforcement learning, comparative data needs to be collected, using models that contain two or more responses ranked by quality. Learning from "pairwise comparison data" is very important for reinforcement learning.
Huang Minlie pointed out: If a single generated result is scored, the bias caused by the annotator's subjectivity is very large, and it is impossible to give an accurate reward value. In reinforcement learning, if the reward value is slightly different, the final trained strategy will be very different. Sorting and comparing multiple results is relatively easy. This comparative evaluation method is also widely used in the evaluation of many language generation tasks.
Besides the sound of technology hype, in many From the perspective of practitioners in the technology industry, ChatGPT is indeed a landmark AI model.
In the view of OpenAI CEO Sam Altman, we can talk to the computer through ChatGPT and get what we want, which makes the software shift from command-driven Intent-driven. As a language interface, ChatGPT will be the best solution before we implement neural interfaces.
The various imaginations about the future of ChatGPT are exciting, but ChatGPT still has some problems. Many users have found that it sometimes gives answers that seem reasonable but are incorrect or even ridiculous. For example, many users have found that ChatGPT will talk nonsense seriously:
## will be Wang Anshi's "Boancing Guazhou" The verses in are mistaken for another Song lyrics:
When writing a biography of a public figure, ChatGPT may insert Wrong data:
With the increase of users, ChatGPT has generated a lot of useless or wrong information on the Internet. This is also a common problem with text generation models. Models are trained by analyzing patterns in large amounts of text scraped from the web. They look for statistical regularities in this data and use these regularities to predict any given sentence. What word should appear next in .
This means that they lack hard-coded rules about how certain systems in the world work, and so tend to produce a lot of seemingly believable nonsense, and It is difficult to determine what proportion of the model's output is erroneous.
This inherent shortcoming of ChatGPT has had some real consequences. Programming Q&A website StackOverflow announced that it will temporarily ban users from posting content generated by ChatGPT. The website mods said: The number of seemingly reasonable but actually wrong replies is too many and has exceeded the capacity of the website.
Regarding the threat of language models producing harmful information, Turing Award winner Yann LeCun seems to remain optimistic. He believes that although language models will definitely produce wrong information, etc. Bad output, but text generation doesn't make it easier to actually share the text, which is what causes the harm.
##The objection is that ChatGPT’s ability to generate large-scale text at low cost will inevitably increase the number of texts in the future. The risk when it comes to being able to share is that the mass of AI-produced content drowns out the voices of real users with data that seems reasonable but is incorrect. Regarding this question, we might as well take a look at ChatGPT’s own answer:
ChatGPT’s language ability Some shortcomings are the reasons why many people believe that ChatGPT cannot replace search engines. Although ChatGPT seems to be able to give better answers to some individual questions than some of the current mainstream search engines, the latter still has an advantage in terms of the authenticity of the answers, and the search engines can give richer answers. Answer.
In addition, users’ search engine needs have extremely high requirements for the running speed and stability of ChatGPT, which will inevitably lead to an increase in costs. This is a very real problem for OpenAI.
Huang Minlie also pointed out that ChatGPT is still a bit far away from replacing Google search, but it can be a very good supplement to current search services.
#In short, the output quality problem of the language model is not easy to solve. OpenAI said that they are more cautious in the training of ChatGPT, so it will reject the correct answer. Furthermore, supervised training can also mislead the model because the ideal answer essentially depends on what the model knows, not what the human knows. However, ChatGPT is sensitive to adjustments to input wording or trying the same prompt multiple times, so when it can't give an answer, you can slightly reword the question to increase the probability of a correct answer.
There are other reasons that limit the language capabilities of ChatGPT. For example, it cannot access the Internet and does not have the ability to retrieve information through the Internet; in addition, for Chinese users Language, the lack of corpus results in its Chinese conversational ability being slightly inferior to English; and so on.
Although ChatGPT currently still has many weaknesses and blind spots, this is just the beginning. In the next few months, this dialogue system will surely develop into Evolve to a stronger version very quickly.
#In addition to technology, model training, deployment costs, and openness will also become factors affecting whether ChatGPT can be successfully implemented in the future. The advent of GPT-3 has spawned a large number of commercial applications. We will wait and see how many technologies ChatGPT can bring to fruition this time.
The above is the detailed content of ChatGPT users have exceeded one million. Is it a toy or a productivity?. For more information, please follow other related articles on the PHP Chinese website!
![[Web front-end] Node.js quick start](https://img.php.cn/upload/course/000/000/067/662b5d34ba7c0227.png)
