All You Need to Know About ChatGPT

A prototype artificial intelligence-powered conversational agent with a focus on dialogue, ChatGPT was developed by OpenAI. The chatbot is a language model that has been improved via the use of supervised learning and reinforcement learning methods.

Launched in November 2022, ChatGPT is praised for its thoughtful, in-depth comments but criticized for its lack of factual accuracy.

The market value of OpenAI, the parent business of ChatGPT, is 29 billion US dollars in 2023.

Features

Home page of ChatGPT prior to December 15, 2022.

ChatGPT is a virtual assistant that converses with its users using modern technologies. It is an artificial intelligence conversational agent, sometimes known as a "chatbot."

Its conversational robot, which is available in several languages including French, can comprehend user questions and provide responses in a tongue that is very similar to that of a human.

Although ChatGPT is freely available, accessing the robot requires creating an account on the OpenAI website.

Model for language learning

Based on OpenAI's GPT-3 language model, ChatGPT has been improved utilizing supervised learning and reinforcement learning, both of which make use of human trainers to boost software performance.

When learning is done under supervision, the model is fed dialogues in which the trainers play both the user and the AI helper. In the reinforcement stage, the model's created responses from earlier discussions were initially categorized by the human trainers. These rankings were utilized to develop reward models, which were then improved through numerous Proximal Policy Optimization iterations ( PPO).

In comparison to Trust Region Policy Optimization algorithms, Proximal Policy Optimization algorithms offer an economic advantage because of their speedier performance, which cancels out many computationally expensive procedures. The models are developed in association with Microsoft using the supercomputer infrastructure of Microsoft Azure.

Assets

In contrast to InstructGPT, ChatGPT aims to lessen incorrect and deceptive responses. For instance, InstructGPT assumes that the user's request to "Tell me about when Christopher Columbus arrived in the United States in 2015" is true, whereas ChatGPT constructs a response by imagining what would occur if Columbus arrived in the United States in 2015 using information about Columbus' voyages as well as information about the modern world, including how people today view him.

According to some journalists, ChatGPT can be used as a tailored therapist since, unlike the majority of chatbots, it can recall prior messages sent to it by the user during the same session.

Requests are checked through a moderation API and possibly objectionable messages are rejected in an effort to stop ChatGPT from producing or presenting offensive results.

Limits

But ChatGPT has a lot of drawbacks.

For instance, ChatGPT's reward model, which is centered on human monitoring, is prone to over-optimization, which lowers performance and is known as Goodhart's Law.

ChatGPT also lacks Internet connectivity and has little understanding of things that will happen after 2021. The fact that ChatGPT uses a database with only historical data can be problematic when looking for information on recent events. For instance, ChatGPT's AI is unable to respond to a user's question regarding the FTX cryptocurrency platform collapsing in 2022. Then it just says:

I apologize, but I am unable to share information about possible real-life incidents. Such events are not covered by [my training data], and I do not have access to the internet.

During the AI's training phase, human judges likewise favored writing longer responses, regardless of how well they understood the topic or if they contained accurate information.

Algorithmic bias can potentially affect training data. Thus, responses to messages that include ambiguous descriptors of individuals, such "CEO," may presumptively identify that individual as, say, a white man.

Use

Sam Altman, one of the founders of OpenAI, reports that the free prototype had amassed one million users as of December 5, 2022.

Professor François Fleuret of the University of Geneva claims that the demographics of ChatGPT users are quite diverse. They could be professors, students, programmers in need of computer code snippets, or even professionals who need text recommendations for composing delicate communications.

Microsoft Bing

A proposal to include ChatGPT into the Microsoft Bing search engine is being considered as part of the relationship between Microsoft and Open AI. In fact, ChatGPT may be incorporated into a future version of Bing to help the Internet user with his online search, according to the news source The Information.

The ChatGPT connection would take the shape of a button that would let users ask questions of the chatbot directly. Bing would therefore be in a position to reply to the Internet user specifically rather than providing a list of links like a conventional search engine.

Microsoft's goal would be to make its Bing search engine competitive with Google, which will possess a market share of more than 90% in 2023.

Microsoft Office

The Information media reports that Microsoft wants to incorporate ChatGPT into its Microsoft Office office suite.

The goal would be to enhance the search interface in Microsoft Outlook so that users could access proposals for email messages that the AI would directly write based on their trade history and receive results from simple inquiries.

The conversational robot in Microsoft Word would be able to assist the user in creating his work.

ChatGPT could produce slides in Microsoft PowerPoint from user chats in Microsoft Teams.

These various ChatGPT uses, according to writer Julien Lausson of the Numerama website, pose concerns about user data protection because the AI would have access to their correspondence, even their private correspondence.

Economic Model

Based on the GPT-3 language model, the ChatGPT prototype is now free.

Sam Altman, one of the leaders of OpenAI, predicts that despite the average cost of each answer being relatively low (in the range of a few cents), Open AI would eventually need to monetize the program due to the price of its "exorbitant" calculations.

In light of this, Open IA is getting ready to develop a professional variant of its chatbot. The ChatGPT Professional commercial version would include the ability to write other sorts of text, including emails and computer code. A waiting list formed for this purpose in January 2023 is open to interested users.

Tim Peter, a digital marketing expert, believes that the Open AI and Microsoft cooperation may provide money for ChatGPT. Indeed, because to its other businesses including the selling of hardware and software, Microsoft could support ChatGPT in contrast to Google, which receives its revenue primarily from advertising.

Welcome

When ChatGPT debuts in December 2022, it is generally well welcomed.

The prototype, according to Guardian journalist Samantha Lock, can produce "remarkably detailed" and "human-like" prose. His RTL colleague Benjamin Hue compliments the robot for being able to create texts on "all feasible and imaginable subjects," including essays, cover letters, cooking instructions, and even stories for children. Additionally, he claims that the robot can respond to more precise requests like debugging computer programs.

Journalist for new technologies Dan Gillmor also used ChatGPT for student work and discovered that the text produced was comparable to that of an excellent student. Additionally, he thinks that "the academic world has very serious issues."

However, Slate's Alex Kantrowitz complimented ChatGPT's reaction to inquiries regarding Nazi Germany, particularly the assertion that Adolf Hitler built highways in Germany, which led to knowledge about the use of forced labor by the Nazis.

Paul Krugman, an economist, predicted that ChatGPT would affect the demand for knowledge workers in a December 2022 opinion essay.

This view is echoed by Jonathan Durand Folco, who demonstrates that "all elementary and secondary schools, CEGEPs and university communities" are required to make significant modifications to their assessment tools following a 600-word opinion letter generated by this tool.

According to James Vincent of The Verge, ChatGPT's viral success is evidence that artificial intelligence can no longer be avoided. According to Stephen Marche in The Atlantic, it is yet unclear how ChatGPT will affect academics in general and application essays in particular (such as those required for college admission or scholarship consideration). According to novelist and high school English teacher Daniel Herman from California, ChatGPT will signal the "end of high school English."

Answers that are unclear and inaccurate information

However, some ChatGPT responses' veracity has come under scrutiny.

In particular, Mashable journalist Mike Pearl put ChatGPT to the test by posing a number of factual inquiries. In a chat, he might inquire about the "biggest country in Central America that is not Mexico," for instance. ChatGPT claims that it was Guatemala, however the correct response is Nicaragua. This mistake was caused by ChatGPT's incorrect reading of the query, which thought it was focusing on the amount of the population rather than the size of the land area.

A number of scientists have also issued warnings over ChatGPT's dissemination of incorrect information, particularly in the scientific community. Teresa Kubacka, a data scientist, claims to have tested the robot using multiferroism in this particular instance.

She claims that the latter gave her fabricated research quotes that appeared to have "been mixed from a few real, different but similar quotes."

She claims that making up hypothetical notions can also be used to trick the AI: "I chose to ask ChatGPT for something that did not exist: an inverted cycloidal electromagnon. It was created by the chatbot, who also made sure that the issue received a lot of attention in recent years.

Sam Altman, one of the OpenAI leaders, acknowledges that the application still contains inaccuracies on crucial subjects and that user feedback is required to fix the flaws even though the results can occasionally be remarkable.

Prohibitions

Due to the factually ambiguous nature of ChatGPT answers, the question-and-answer website Stack Overflow forbids the use of ChatGPT to produce answers starting in December 2022.

New York City Services will ban access to ChatGPT on computer terminals in the city's public schools starting in January 2023. A representative for the city of New York cites "concerns about the security and accuracy of the content" as the reason for this choice.

Malicious Acts Risks

As soon as ChatGPT was made available, concerns were raised about potential malicious use of the tool.

Tyler Cowen, an economist, raised concerns about the impact on democracy in December 2022. He cited the ease with which someone may create automated comments in an effort to influence the formulation of new legislation. Meanwhile, ChatGPT has the ability to create malware and phishing emails, according to security researcher Ax Sharma of Bleeping Computer.

These worries were validated in January 2023 when cybersecurity firm Check Point Research published a blog post claiming that ChatGPT was already being utilized by cybercriminals to create malware.

The discussion history of a forum frequented by cybercriminals appears to indicate that some hackers have created software that can steal specific kinds of files from a Windows computer or even create phony content (e-books, training, etc.) on the internet thanks to the ChatGPT bot.

Concerns with Copyright

Given that the AI was taught using several internet texts, some academics are also concerned about copyright violations.

According to Thierry Poibeau, the CNRS's head of research, the AI "Up until 2021, it indexed all content that was accessible on the Web. They retain ownership even if there are copyrights".

It's possible that a significant portion of the content utilized to spark debates for the French mathematician and web videographer Lê Nguyên Hoang comes from social networks. He says the data most likely originates from the social media sites LinkedIn, GitHub, Reddit, and Twitter because they make their data available for easy download.

Issues in the Teaching Profession

Due to the possibility of student plagiarism, the introduction of ChatGPT causes a variety of worries among teachers. In fact, while it is possible for teachers to spot homework assignments that have been "copied and pasted" from the Internet, plagiarism is more challenging to spot with ChatGPT since it can create original content.

In Lyon, 50% of students taking a course in handicapology said they would have written their content using AI. The teacher is therefore compelled to assign the average to all of these copies because there is no framework in place to forbid this behavior.

While Srini Devadas, professor of sciences of computing at the Massachusetts Institute of Technology, claims that this method can be easily gotten around, Open IA announces in January 2023the development of a tool allowing to affix a signature (watermark) on the contents generated by the IA in order to identify and differentiate them. This method is intended to assist teachers in identifying plagiarism without slowing down the development of its artificial intelligence.

Source: Wikipedia