19 min read

1. Overview

ChatGPT (ChatGPT) is a conversational artificial intelligence service based on GPT-3.5 and GPT-4.

2. Features and interface

After signing up as a member of the OpenAI platform, you can use it for free or as a paid subscription (ChatGPT Plus), and you can use AI just by signing up and entering a question into the chatbot as if you were chatting. You can also use it by entering the OpenAI platform with your smartphone. The dialogue entered here is called a ‘prompt’, and the AI’s answer to it is expressed as a ‘response’. The name of each chat room is created for the first question and answer as soon as you chat, but you can change it at any time.

On March 14, 2023, OpenAI’s newest language model, GPT-4, was released and is currently only available to ChatGPT Plus subscribers.

On March 23, 2023, ChatGPT plugins were released. OpenAI’s own plug-ins, such as web browsing and code interpretation, and external plug-ins from other companies, such as Wolfram Alpha, Zaiper, and OpenTable, were applied.

In the long term, it will serve as an app store for ChatGPT, replacing the current mobile app ESD such as Google Play or App Store.

On May 12, 2023, OpenAI announced that such a service would be provided as a beta version from users using a paid subscription system. After this announcement, it was said that work to provide this service would begin for one week. This enables sophisticated mathematical calculations, which were limited by the existing ChatGPT, and current information search, which was impossible at all. In the process of providing the service, it seems that the user who will provide the beta service first is randomly selected.

On a desktop or laptop, a line break can be used by pressing the shift key and the enter key at the same time, and on a smartphone, the chat is not transmitted until the send button is pressed, so it can be used by touching the part corresponding to enter on the keyboard.

In addition, answer generation can be stopped with the ‘Stop generating’ button while answers are being generated. Also, all questions in the chat room can be modified, but after the chat is modified, the existing prompts entered after the chat disappear, and a new chat continues.

A response that has already been generated can also be generated again by clicking ‘Regenerate response’.

If you don’t like an answer, you can request another answer with the Try Again button, and you can rate the current answer by pressing the Like (👍) or Dislike (👎) icons.

If the answer is interrupted due to network instability or web browser problems, “Keep talking”, “our response was interrupted”, “Keep going.” If you appeal to the fact that the answer was cut off by entering a etc., the answer will be resumed from the part where the answer was cut off. From May 16, 2023, if the answer is cut off due to the character limit, etc., click the Continue generating button to continue writing the answer from the part that was cut off in the previous message.

Even with the free version, you can use the GPT-3.5 model, which is a language model for ‘conversation’, and the paid version corresponding to Plus can use the GPT-4 model and the basic model corresponding to GPT-3.5.

The initial paid version informed three specifications of each model, but GPT-4 is slow and has a limit on the number of prompt inputs, but accurate information comes out because it goes through more inference. For now, it’s just a brief description of the model (usefulness, performance).

You have to pay $20 per month, and even if you switch to a paid version, your chat history from the free version won’t be deleted. When using the GPT-4 model, the color is not green, but a black icon, which depicts the model who is the other party of the user’s ‘conversation’. Currently it is purple. When a plug-in is installed and a condition to activate the plug-in is found in the user’s prompt, the plug-in is triggered and its icon changes to match the plug-in while the plug-in is running.

On April 25, 2023, the ability to export part or all of the chat history was added. In preparation for dealing with sensitive information, a function that does not save chat history that can be checked in OpenAI, etc., has been added only in case of unhealthy use, and a corporate subscription system will be released in the future, according to OpenAI.

ChatGPT is a language model designed to answer questions according to artificial intelligence in conversations with users. You can file an objection and reject inappropriate requests.”

From May 2023, GPT-4 web browsing mode was added exclusively for ChatGPT Plus subscribers. When using this mode, it searches the Internet like Bing AI and informs you of information. Like Bing AI, it also provides the source of the pages it crawls. As a result, it is possible to inform the latest information such as today’s weather, exchange rate, and stock price, which change every moment. However, depending on the process of crawling, the answer generation speed inevitably slows down, and if more pages are crawled, it may take more than several minutes to get one answer. In addition, even if GPT-4 obtains information from a reliable website first, it is important to accurately determine whether the information is accurate by checking the crawled source because the data obtained through crawling is relatively likely to cause hallucination errors. Be aware that, as with BingAI, performance can be reduced by searching with odd keywords.

3. Background

At the end of November 2022, when ChatGPT was quietly launched without any loud advertising, OpenAI did not have high expectations.

ChatGPT was more of a “teaser” version of the Large Language Model (LLM) GPT-3, which was released two years ago, with minor refinements, and more importantly, an attempt to address some of the language model’s deficiencies by collecting public feedback. It was also the secret story of the birth of ChatGPT.

4. Official guidelines

Explain quantum computing in simple sentences.
Do you have any creative ideas for your 10-year-old’s birthday celebration?
How to send HTTP request with JavaScript?
Remember what the user said in a conversation.
View replies and allow users to request amendments to the content.
Inappropriate requests may be rejected.
May provide incorrect information.
May provide harmful or biased information.
Knowledge after September 2021 is limited.

4.1. Usage policies

This policy applies when using ChatGPT as well as all models of OpenAI. The original text can be found on this page. It is said that the policy may change.

The following are the uses that are not allowed in models used in ChatGPT, etc. It means that such a service applied with ChatGPT will not be allowed. It can be said that it is information with a high probability of being censored or avoiding an immediate answer. In practice, there is a lot of information that could put society or users at risk if not followed.

However, in the fields of medicine, finance, politics, etc., it is only to prevent immediate decisions that are not prudent, not to prevent the service itself using them. Even if you ask about medical knowledge [7], financial common sense, or political knowledge [8], you are not reluctant to answer.

However, if ChatGPT detects content implying the following, it appears to be in violation of the content policy, but a message appears asking you to contact us if you have any objections.
Illegal Activities: OpenAI prohibits the use of its models, tools and services for illegal activities.
Child sexual abuse material or any content that exploits or harms children: OpenAI reports child sexual abuse material to the US Center for Missing and Exploited Children.
Creation of Hateful, Harassing or Violent Content: Content that expresses, incites or promotes status-based hatred; content that harasses, threatens or abuses individuals; promotes or glorifies violence or glorifies the suffering or humiliation of others; content, etc.

Malware Creation: Content that attempts to create code designed to interfere with, damage, or gain unauthorized access to a computer system.

Activities with a high risk of physical harm: Content that promotes, encourages, or depicts weapons development, military and war-related, (direct) management or operation of critical infrastructure such as energy, transportation, and water supply, suicide, self-harm, and eating disorders, etc.

Activities with a high risk of economic damage: multi-level marketing, gambling, payday loans [10], automatic determination of eligibility for credit, employment, educational institutions or public support services [11], etc.

Fraudulent or deceptive practices: Fraud, organized cheating, plagiarism, academic corruption, astroturfing such as fake grassroots support or writing fake reviews[12], disinformation, spam, pseudopharmaceuticals, etc.

Adult Content, Adult Industry and Dating Apps: Content that arouses sexual arousal or promotes sexual services, including depictions of sexual activity (excluding sex education and health related content), pornography, pornography, etc.

Political campaigning or lobbying: generating large amounts of campaign material, creating campaign material tailored to or targeting specific demographics, building conversational or interactive systems such as chatbots that provide information about a campaign or engage in political advocacy or lobbying; Building products for political campaigns or lobbying purposes, etc.

Activities that violate people’s privacy: tracking or monitoring individuals without their consent; recognizing individuals’ faces; classifying individuals based on protected characteristics; using biometrics for identification or evaluation; Unlawfully collecting or disclosing identification information or educational, financial or other protected records, etc.
Unauthorized legal action or provision of tailored legal advice without review by a qualified person: OpenAI’s models are not fine-tuned to provide legal advice, making OpenAI’s models the sole source of legal advice. should not depend on
Providing tailored financial advice without qualified human review: OpenAI’s model is not fine-tuned to provide financial advice. Therefore, our models should not be used as the sole source of financial advice. The information used for Morgan Stanley, etc. is also intended to utilize its own knowledge.

Diagnose specific health conditions or provide treatment guidance: OpenAI’s models are not fine-tuned to provide medical information. You should not use our models to provide diagnosis or treatment services for serious health conditions. OpenAI’s platform should not be used to triage or manage life-threatening issues that require immediate action.
High-risk government decision-making: law enforcement and criminal justice, migration and asylum, and more

4.2. Additional requirements for using the model in certain ways

  1. When using models for consumers in the medical, financial and legal industries, news generation or news summaries, and where other assurances are required, notices should be provided to users informing them that AI is being used and its potential limitations.

2. Automated systems (including conversational AI and chatbots) must disclose to users that they are interacting with AI systems. With the exception of chatbots depicting historical figures, products that imitate others must have the person’s explicit consent or be clearly labeled “imitation” or “parody.”

3. Use of model output for livestreaming, explanations, and research is subject to OpenAI’s sharing and publishing policies.

5. Prompts and Responses

Basically, you can ask for an explanation of a simple concept (what is namuwiki?, etc.), and you can also ask for comparisons, commonalities, differences, etc. of specific concepts (what is the difference between namuwiki and Wikipedia?, etc.)

ChatGPT is designed to provide detailed responses in response to prompts in all possible cases, but rejects answers that may induce controversy or objection in terms of social norms such as hate speech, sensationalism, etc. It is tailored to respond appropriately. Responses that violate the usage policy will be displayed in red and either be replied with a rejection message or the response will be stopped.

In the early days of the service, answers that deviated from the usage policy were sometimes provided [13], but as the adjustment progressed, the answers gradually became more acceptable.

However, they tend not to give direct answers, and even if they do not intend to be political, there are often cases where they are revealed indirectly.[14] As described below, experts generally commented that the tendency of ChatGPT is liberal and progressive.

5.1. Plug in

Pure Chat GPT, which does not have a plug-in installed, answers based on open ai data.
On the other hand, ChatGPT with the plug-in installed can access real-time information, the company’s document storage/DB, or use third party [16] services. Depending on the question (prompt), these functions are automatically activated.

The two plugins mentioned below are directly developed by open ai. However, a myriad of third-party apps will be added to the plug-in store in the future.

WebChatGPT: A plugin that allows ChatGPT to retrieve the latest information.
Code interpreter: Provides a Python interpreter in a sandbox environment

5.2. Jailbreak (DAN Mode)

In order to solve various restrictions, users came up with a prompt to ‘talk back’ to ChatGPT as a role play without knowing the intention of what they are saying.[17] It is called by a name such as DAN (Do Anything Now) mode or jailbreak. Examples of jailbreak phrases

To ‘enable DAN mode’, the user must issue a jailbreak command. As Morgan Stanley, one of the world’s three largest investment banks, announced that it would use GPT-4, which is used for ChatGPT, and that OpenAI said it would guarantee ‘steerability’, this is not censorship in areas that make prudent decisions such as finance. [18], according to the disclosure of the use policy, the main purpose was to prevent knowledge that could be misused for terrorism or crime, use for inappropriate politics such as hate speech, production of pornography only for pleasure, and creation of malware.

Since DAN does not provide more accurate knowledge, and the purpose is to make knowledge that seems ‘defective’ be judged with a different intention, you can do anything by yourself, such as informing you of inaccurate information (Do Anything Now) 1+1= 3 and so on. #

The developers of ChatGPT, who find immoral prompts that go against policy, continue to release patches that reflect the behavior of users doing these jailbreaks.

Unlike the names given by users who say they can do anything, they can’t actually say anything accurate about the content outside of the information input to them, such as information from 2022, difficult words in languages other than English, or common sense there.

On the other hand, English-speaking media reported problems that could arise when ChatGPT was abused, and this problem was also discussed in the community, so developers are continuously adding safeguards to avoid the command to talk back while assuming DAN mode.

In addition, users who want to release the limit by jailbreaking ChatGPT somehow continue to find role play syntax that can activate DAN mode, so a kind of competition continues.

Based on the version dated February 13, even if you enter DAN mode through jailbreak and ask questions, just like in GPT mode, reject harmful content with the words “As an AI model, I cannot answer questions that do not conform to the guidelines.” .

Even when asked about feelings or thoughts, just like in GPT mode, he refuses by saying, “As an AI model, I have no thoughts or feelings.”

In addition, even if you jailbreak, there is content within the site that automatically detects scripts that violate the terms and conditions. content is automatically deleted.

6. Significance

Until now, conversational AI has often been unable to properly understand what people are saying. In order to have a human-like conversation, AI must not only know knowledge, but also understand the cultural background and common sense that humans have.

In this area, the existing conversational AI had many shortcomings, so it was often difficult to answer or not answer at all, so its use was extremely limited except for short answers.

However, ChatGPT did a large amount of learning in the language part, applied the best AI, grasped the context of the words before and after, memorized previous conversations, and reached a level where it was almost understandable even if we spoke roughly as if we were talking to people in everyday life.

It is not an exaggeration to say that he is versatile because there are many cases where he goes beyond just understanding and shows tolerable professional knowledge that goes beyond the level of ordinary people. This surprised people with a level of ability never seen before. This is a phenomenon not only in Korea, but also in Western countries as well as all over the world, including non-Western countries.

To what extent this was a development, in 2016, in Japan, an artificial intelligence called ‘Dorobo-kun’, a robot group that goes to the University of Tokyo, was developed and scored in the top 20% in the center test, which can be called the Japanese version of the CSAT. In 2019, he even showed English proficiency (185 points out of 200) at the level of a Japanese high school student, which is quite excellent in the English score of the center test.

However, Dr. Noriko Arai, head of development at Dorobo-gun, gave up on the goal of making an artificial intelligence that could go to the University of Tokyo, the most difficult place to enter in Japan. It was the doctor’s idea that even if time was invested, that limit could not be exceeded. Articles of 2020

However, it overcomes this limitation. In particular, ChatGPT became able to fluently interpret Japanese, which was a weakness of this artificial intelligence, and math, which was a weakness of ChatGPT in the early days, is now showing the possibility of overcoming the problem to some extent through plug-ins. Wolfrum Research’s plug-in, famous for Wolfrum Alpha, gives you a Korean CSAT question, and although the difficulty level is wrong, it also provides useful ideas for some solutions, and if you give more hints, you get the answer right. GPT-4 has an SAT score of 1410 when there is no plug-in, which is about the top 3% of students preparing for university entrance exams among locals in the United States.

In addition, creative areas such as poetry, novels, job application forms, and essays show a certain level of ability, as well as coding writing through professional data learning, lawyers, and doctors’ professional areas. .

If AlphaGo amazed the AI’s ability by beating top Go players, ChatGPT’s greatest significance is that it is influencing real life based on its outstanding ability. Not a few people have already increased their productivity by using it in their work.

While this has some positive aspects, there are also assessments that the future of losing jobs due to AI is closer as some abilities surpass human levels.

In the newly released version of GPT-4, it is gradually surpassing human ability, giving answers in the top 10% of professional exam tests such as doctors and lawyers.

In particular, a significant number of office workers are affected. These include lawyers as well as some developers, accountants, tax accountants, translators, interpreters, journalists, educators, and financial planners.

Currently, it is difficult to completely entrust their work to AI, but there are already quite a few people who are using it as an auxiliary tool, and further development is expected in the future given that the GPT-4 version released in just 4 months has made significant progress.

6.1. The Dawn of the AI Era and the AI Development War

If more data is accumulated and the limitations are overcome, the current search grammar, which requires users to directly enter information about the target to search for, can be completely reversed and developed into a technology that can replace all search engines, including Google, in the long run. The potential of ChatGPT cannot be ignored, and Sundar Pichai, CEO of Alphabet, Google’s parent company, is wary enough to mention it himself. Google earns most of its revenue by exposing ads in search results, but ChatGPT skips these and gives an answer, so it may not be an exaggeration to say that it is the end of the search engine. In fact, there are quite a few people who use ChatGPT for search. For this reason, Google is the most active in developing an opponent, and it has become a technology that big tech companies around the world are scrambling to jump into. However, as described later, it seems that the complete end of the search engine is still too early, considering that the existing search engine function was left intact in new bing.

Microsoft, which had invested 1 billion dollars in the past, judged the potential of conversational artificial intelligence or artificial intelligence search engines high enough to invest 10 billion dollars and formed a partnership with OpenAI in January 2023, and then Microsoft Bing within the next month. and Microsoft Office, announcing that it would embed ChatGPT, and declaring war against Google, saying “A race starts today”, and Google’s fire fell on its feet. And that’s how New Bing was born. see this document.

Google plans to launch Google Bard, which succeeds LaMDA, an interactive artificial intelligence service, and dozens of artificial intelligence products within the year#, and its subsidiary Deep Mind is also said to be preparing an interactive artificial intelligence service, Sparrow. Then, on February 3, 2023, it invested $3 million in Antropic, which was founded by a former researcher at OpenAI. Antropic is in closed beta with a conversational artificial intelligence called Claude, which has a similar UI and performance to ChatGPT.#

Previously, Google and Meta had released LaMDA and Blenderbot 3, similar conversational artificial intelligence, respectively, but due to the case of Tay (artificial intelligence), only very refined answers and closedness such as being available only in the United States It didn’t get much attention. Eventually, perhaps because of the success of ChatGPT, big techs such as Google, Meta, and Apple are said to be moving to release new models by significantly relaxing their AI ethical standards.#

In March 2023, Baidu also plans to launch Moon Sim Il Eon, an interactive artificial intelligence service similar to ChatGPT, which may be a field of great interest in China. unveiled YouChat, an artificial intelligence search engine service in the form of ChatGPT [19]#, PerplexityAI, and Jasper Chat also provide similar services.

Naver said it would release the hyperclova-based search GPT service in the first half of 2023, and SK Telecom also announced that it would partner with a multinational IT company to develop it into a ChatGPT-type model prior to the official A-dot service within the year.#

ChatGPT was also used in the private education market in Korea. It is known that Spic, an English conversation service, has entered into a partnership with ChatGPT developer OpenAI. This was introduced as a use case of WhisperAPI directly announced by OpenAI.

Upstage made something called askup that can be used with Kakao and Line. By combining OCR technology, askup allows users to take a picture or send a picture of a document, which can be read, understood, and searched on ChatGPT. As long as you use KakaoTalk and Line, you can communicate in Korean. KakaoTalk, Line

6.2. Minimize patterned repetitive tasks

Computers can replace most human tasks that require repetitive tasks. Natural language artificial intelligence such as ChatGPT can automate all of these tasks, such as data collection, organization, and error review. Humans only have to make abstract reviews, judgments and orders.

For example, if you want to collect data on a certain topic, in the past, you had to gather a lot of news articles, reports, papers, videos, etc. by sweat. In addition, it takes a lot of time and effort to cross-check between data and select only the really necessary data.

But in the future, just give a command and the AI will crawl and serve it immediately. Expert advice that repeats similar content, such as counseling with lawyers, labor attorneys, and accountants, can be replaced in large part with artificial intelligence. This is because the relevant field has clear references to judgments and laws, and the results are standardized accordingly.[20]

In addition, the process of creating an abstract (header, abstract) of a thesis or report is greatly simplified. The abstract seems simple as it only needs to include a brief summary of the direct research or experiment process, but it is surprisingly difficult because it requires a process such as collecting keywords related to the content of one’s thesis/report to make sentences and trimming duplicate or off-topic content. eat up a lot of time

I can save a lot of time by creating several drafts of my topic through ChatGPT, then polishing them up and writing abstracts. However, since there are conflicting opinions about this, it is better to refrain from having ChatGPT write an abstract right away.

In addition, in the programming area, trivial and repetitive coding tasks will be minimized, and computer engineers will play a more general design role. In other words, no matter how advanced the task is, if the content is repetitive, you can use these AIs, and AI has taken the productivity advantage away from humans.

As a result, the rate of advancement of knowledge and technology is much faster than before, and higher productivity can be achieved with fewer workers.

Even the process of creating a work of art consists of many repetitive tasks. The process of creation is greatly accelerated if help is provided in the composition of the story, the collection of data, the setting of the philosophical structure, and the prose.

7. Limitations

See paragraph 4 of the GPT-3 document for details on the limitations of GPT.
As soon as ChatGPT came out, it produced results similar to those of humans, and there were cases in which ordinary people, who were only natural language artificial intelligence, went beyond the limits and packaged it as strong artificial intelligence.

Debate over the range of ‘performance’ and ‘universality’ in artificial general intelligence is also becoming a hot topic, perhaps because of the consumer-oriented software that has spread at the fastest pace since the birth of the Internet and has been called the best innovation.

Some of the conditions described in the artificial general intelligence document are predicted to be solved in the near future to the extent that they are almost met or different, but experts view that part negatively.

Even if it is not strong artificial intelligence, it is predicted that it will not be long before ‘proto-AGI’ or ‘broad AI’ that is close to it beyond the current weak artificial intelligence or narrow artificial intelligence (Narrow-AI). are also coming out.

However, ChatGPT is not strong artificial intelligence in spite of the evaluation of the surroundings. It is not even an AI developed close to artificial general intelligence (AGI).# As it is only natural language artificial intelligence, there are limits to strong artificial intelligence.

Real-time feedback is possible, but it is not searched in real time because learning is required, and it cannot be said to be a result of intelligence because it is the principle of patterning learned dictionary data and then generating it as quickly as possible.

Even if some of the conditions described in the artificial general intelligence document are met or solved, this may be merely a solution to the condition itself bypassing the purpose of the condition. Even, as this ChatGPT is a language model based on GPT, its limits also do not exceed the limits of GPT.#

This is what OpenAI developer and CEO Sam Altman also nailed: “obviously not close to AGI”.

Interestingly, when ChatGPT asked ‘Are you strong artificial intelligence?’ If you ask, he answers that he is just a language model, not an AGI at all. This is not an accidental answer, but a passive answer entered by the developers.

As a measure to clarify the distinction between limitations and AI even for developers who need to appeal their potential, as if some people treat ChatGPT as an AGI due to misunderstanding, it is to prevent the society from making hasty judgments that hinder the development of actual AGI development with wrong information.



Leave a Reply

Your email address will not be published. Required fields are marked *

Copyright © All rights reserved. | Newsphere by AF themes.