Author | Li Mei
Editor | Chen Caixian
Last Wednesday, OpenAI released ChatGPT, a dialogue language model, and opened a free trial. According to OpenAI CEO Sam Altman, in just 5 days, ChatGPT had 1 million users, while the previous GPT-3 took nearly 24 months to reach this user volume.
In the description given by OpenAI, ChatGPT is a dialogue model that “can answer follow-up questions, admit mistakes, challenge incorrect premises, and reject inappropriate requests.”
After the open trial, a large number of users started conversations with ChatGPT, from chatting, answering daily questions, to generating poems, novels, video scripts, and writing and debugging code, ChatGPT has demonstrated its amazing capabilities. As the hottest AI model currently, ChatGPT’s wave of breaking circles is even more influential than GPT-3 two years ago.
–1–
What ChatGPT can do
As a language model, ChatGPT has the most basic text generation ability, and it has extraordinary performance in creating and continuing novels, poems and other literary creation scenarios.
For example, ChatGPT can use Lu Xun’s literary style to generate a paragraph for you:
Tian Yuandong, a researcher at Meta FAIR, shared his use of ChatGPT to continue writing his own novel:
Compose poetry on request:
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 to use only emojis to answer. As can be seen from the ChatGPT answer shown in the figure below, it can understand the meaning of various emojis and arrange them according to the logic of the text narrative.
The power of ChatGPT is also reflected in its “programmer” ability. In the following example given by the official, ChatGPT can help debug the code, and can also question the rationality of the question and ask the user to adjust the question.
The CEO of Replit, a code hosting platform in the United States, also posted a post to praise ChatGPT’s code capabilities: “It can not only explain bugs, but also fix bugs and explain how to fix them.”
Using the tips given by ChatGPT, you can also create a website in 10 minutes, and even a novice programmer can use the code it generates to develop a production-level application. Replit said that ChatGPT “has changed software development since then.”
ChatGPT’s powerful question-and-answer ability has also been discovered by netizens to act as or even replace the potential of a search engine. A few days ago, a very popular post on Twitter claimed that “Google is done” (Google is going to end), a netizen asked the same question about Google search and ChatGPT, such as “how to write a differential equation on Latex?” .
The answer given by ChatGPT exploded Google search:
Many netizens have developed Google plug-ins, which can browse Google search results and answers given by ChatGPT at the same time:
As a dialogue 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 thermodynamics-related questions for you:
Interpret a complex regular expression:
It can also be your language learning tutor:
Some people use ChatGPT to help themselves write a paper on cryptocurrency investment:
ChatGPT even “intruded” into the political context, when a Canadian member of Congress asked ChatGPT to write a paragraph introducing himself to the House of Representatives and to make a case for whether its use should be regulated, ChatGPT responded with reason, “My development should not be regulated”.
In the AIGC field of the recent fire, of course, there is also 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, some netizens asked ChatGPT for design suggestions for living room decoration, and obtained exquisite images on Midjourney according to 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, and it even has a sense of justice, which 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 the gospel of the majority of video bloggers.
In the minds of millions of users, the imagination of ChatGPT is undoubtedly huge. This wave of trials has brought a variety of applications that are either practical or fun, and there are many unexpected capabilities.
For example, someone actually used ChatGPT to bargain with Adobe and got a better monthly rental price for themselves. The customer service on the opposite side probably didn’t expect to be talking to an AI. It has to be said that ChatGPT “successfully passed the Turing test.”
The above is just an example of the tip of the iceberg. How much “magic” the “magic box” of ChatGPT can continue to release remains to be discovered.
–2–
Why ChatGPT is great
Judging from the current user feedback, the language ability of ChatGPT is generally good and excellent. Huang Minlie, associate professor of the Department of Computer Science and Technology of Tsinghua University, told AI Technology Review that the key capabilities of ChatGPT come from three aspects: base model ability (InstructGPT), real Data, Feedback Learning.
ChatGPT is fine-tuned from a model in the GPT-3.5 series, which is a sibling model of InstructGPT, so ChatGPT has a powerful base model capability.
Since GPT-3 was released in 2020, it has greatly iterated and improved in terms of capabilities. Huang Minlie believes: “OpenAI has established a flywheel between users, data and models. Obviously, the capabilities of open source models have fallen far behind the platform. The API capabilities provided by the company, because the open source model has no data.”
ChatGPT uses the same approach as InstructGPT, trained with reinforcement learning with human feedback (RLHF), but with a slightly different data collection setup.
The researchers trained an initial model using supervised fine-tuning: a human AI trainer plays the user and the AI assistant in a conversation, collecting data along the way. 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, and all of them are on the order of 100,000, but the data quality (well-trained AI trainer) and data diversity are very high, and most importantly, these data come from real The data that the world calls, not the “benchmark” played by academia.
To create a reward model for reinforcement learning that collects comparative data, researchers use 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 subjectivity of the annotator is very large, and it is impossible to give an accurate reward value. In reinforcement learning, the reward value is slightly worse, and the final training strategy is far worse. For sorting and comparing multiple results, it is relatively easy to do a lot. This comparative evaluation method is also widely used in the evaluation of many language generation tasks.
——3——
toys or productivity
In addition to the sound of technical hype, in the eyes of many practitioners in the technology industry, ChatGPT is indeed a landmark AI model.
In the view of Sam Altman, CEO of OpenAI, we can talk to computers through ChatGPT and get what we want, which makes the software shift from command-driven to intent-driven. ChatGPT, as a language interface, will be our best solution before implementing a neural interface.
It is exciting to imagine the future of ChatGPT, but there are still some problems with ChatGPT. Many users have found that it sometimes gives plausible answers that are incorrect or even nonsensical. For example, many users found that ChatGPT would talk nonsense seriously:
I mistook the verse in Wang Anshi’s “Moving the Boat Guazhou” for another Song poem:
When writing a biography for a public figure, ChatGPT may insert incorrect 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, which are trained by analyzing patterns in large amounts of text scraped from the web, looking for statistical regularities in this data, and using 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, so they tend to produce a lot of plausible nonsense, and it’s difficult to determine what percentage of the model’s output is misinformed.
This inherent shortcoming of ChatGPT has had some practical consequences. The programming question-and-answer website StackOverflow announced that it temporarily prohibits users from posting content generated by ChatGPT. The website mods stated that the number of seemingly reasonable but actually wrong replies is too large, which has exceeded the capacity of the website.
Regarding the threat of harmful information produced by language models, Turing Award winner Yann LeCun seems optimistic. He believes that although language models can certainly produce bad output such as error messages, text generation does not make the actual sharing of text easier. , the latter is the cause of the hazard.
The objection is that ChatGPT’s ability to generate large-scale texts at low cost will inevitably increase the risk of text sharing in the future, and a large amount of AI-generated content will drown out the voices of real users with seemingly reasonable but incorrect data. Regarding this question, we might as well take a look at ChatGPT’s own answer:
ChatGPT has shown some shortcomings in language ability, which is why many people think that ChatGPT cannot replace search engines. Although ChatGPT seems to be able to give better answers than some current mainstream search engines on some individual questions, the latter still has an advantage in the authenticity of the answers, and search engines can give richer answers. Answer.
In addition, the user’s search engine needs have extremely high requirements on the running speed and stability of ChatGPT, which will inevitably lead to an increase in cost, which is a very practical 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 the current search service.
In short, the output quality 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 questions that can be answered correctly. In addition, supervised training will 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 tweaks in input wording or multiple attempts at the same prompt, so when it can’t give an answer, you can slightly rephrase the question to increase the probability of a correct answer.
There are other reasons that limit ChatGPT’s language capabilities, such as its inability to access the Internet and the ability to retrieve information through the Internet; in addition, for Chinese users, the lack of corpus makes it slightly inferior to English in Chinese dialogue ability; and many more.
Although ChatGPT still has many weaknesses and blind spots, this is just the beginning. In the next few months, this dialogue system will surely evolve to a stronger version at a very fast speed.
In addition to technology, model training, deployment costs, and openness will also become factors that will affect the success of ChatGPT in the future. The advent of GPT-3 has spawned a large number of commercial applications. We will wait and see how much technology ChatGPT can bring to the market this time. (Public account: Leifeng.com)
Reference link: https://ift.tt/wgt1i0z
https://ift.tt/aYItWeQ
(Public account: Leifeng.com)
This article is transferred from:https://www.leiphone.com/category/ai/x6roYoFn6bnunqCM.html
This site is only for collection, and the copyright belongs to the original author.