From the beginning of 2023, artificial intelligence became available to everyone who has a smartphone or computer. The main representative of the newest technological revolution of the XNUMXst century is ChatGPT.
In the material of the publication ForkLog it explains what ChatGPT is, how it benefits people, and how to use it.
What is ChatGPT?
ChatGPT is a large language model that was trained by OpenAI based on the GPT (Generative Pre-trained Transformer) architecture. This model is able to understand human speech, generate texts, answer questions and write articles. The platform has a huge number of parameters (more than 175 billion), thanks to which it can solve complex tasks, including text generation, language translation, question answering, content creation and much more.
A distinctive feature of ChatGPT from other dialog tools is the domain of knowledge. A chatbot is equally good at different fields such as mathematics, history, geography, art, medicine, finance and cooking. This is a universal model that you can turn to with any question.
ChatGPT is available in Russian, Ukrainian, English, Spanish, German, Japanese and other common languages.
The service was launched in November 2022, but its basic algorithm — GPT-3 — was created back in 2020. In March 2023, the developers added support for the fourth version of the model.
ChatGPT is free to register and use for free. There is also a premium version that provides priority access to new models, increased generation speed and guaranteed operation during high loads.
As of March 2023, an API was launched for ChatGPT.
How does a chatbot work?
Actually, ChatGPT is a GPT interface. If earlier it was necessary to access the API and integrate it into the application in order to use the large OpenAI language model, now anyone can use the service. Try it.
The ChatGPT page is similar to an instant messenger. In the main window, the user sees correspondence with the chatbot, and on the left - a list of dialogues. Each individual chat is saved.
OpenAI launched the service as an experiment. In the announcement, the company noted that the chatbot is based on an improved version of the GPT-3.5 language model. The engineers used reinforcement learning based on human feedback (RLHF).
To collect a dataset of dialogues, the developers engaged so-called trainers who spoke on behalf of a person and an AI assistant. They were also given access to mock sentences to help them write their answers.
During the creation of the model, the OpenAI team recorded dialogues between a person and a chatbot. The developers then randomly selected AI responses and asked people to rank them.
What is GPT?
GPT is a Large Language Model (LLM), also developed by OpenAI.
At the time of writing, the current version is GPT-4, released in March 2023. Unlike its predecessors, the developers position it as a large multimodal model. Such algorithms can analyze not only text, but also other types of content such as images, video and audio.
Originally, ChatGPT ran on an improved version of GPT-3, released in September 2020. It was trained on a dataset with a size of 570 GB, and the number of parameters was 175 billion.
GPT-4 clearly surpasses its predecessors in all indicators. However, the developers decided not to reveal the technical details of the algorithm.
All versions of the model are able to generate texts, answer questions, perform semantic searches and create extracts from texts with varying degrees of accuracy.
GPT belongs to the class of transforming neural networks.
How to use ChatGPT?
An account is required to use ChatGPT. You must have a valid email address and mobile phone number to register. You can also use Google or Microsoft accounts to sign in. After successful registration, it is possible to communicate with the chatbot.
At the same time, registration is not available for residents of Afghanistan, Belarus, Venezuela, Iran, China and Russia.
In February 2023, the service became operational in Ukraine. However, OpenAI, as before, does not provide services in the territories temporarily occupied by the Russian Federation.
To gain access to the latest GPT-4 model, a premium subscription must be purchased. Its cost is $20 per month.
Despite the claimed ability of GPT-4 to interpret images, ChatGPT does not support such a feature. The company did not say when this feature will appear in the chatbot.
What can ChatGPT do?
The possibilities of text generation in ChatGPT are practically unlimited. He can:
- compose poems and songs;
- solve mathematical and logical problems;
- write texts in different styles and genres;
- summarize information;
- choose recipes based on the list of ingredients;
- create program code;
- search for information from your own database;
- conduct a dialogue;
- adapt to the style and tone of the interlocutor.
This is by no means a complete list of ChatGPT capabilities. Some of its functions, such as the creation of simple games, were discovered by users by accident.
In the paid version of the chatbot from OpenAI based on GPT-4, there are even more usage scenarios.
Thanks to multimodality, the chatbot can process images and recognize the author, genre, style and other characteristics of pictures and art objects. It can also provide step-by-step solutions to math problems, perform calculations and analysis based on data tables and graphs, check software code, help with language learning, and much more. However, this feature is not yet available in ChatGPT on the OpenAI website.
The chatbot is able to understand humor and satire embedded in text or memes. He can also produce his own ironic content.
ChatGPT with the fourth version of the language model can process more than 25 words, unlike GPT-000, which accepts up to 3.5 words as input. Increasing the volume of text makes it possible to generate longer content and analyze massive documents.
At the request of the ChatGPT user, he can take on different "roles". For example, he can be asked to become a lawyer, financier or teacher. The image allows you to give different and more relevant answers to the same questions for the user.
Advantages of ChatGPT
The main advantage of ChatGPT is its ability to understand natural language. The user can enter a request in absolutely any wording, and the algorithm will interpret and generate a coherent and logical answer.
ChatGPT is capable of performing a wide range of tasks such as text generation, question answering, translation, summarization, text classification and much more.
The advantages include the high accuracy of the chatbot. He processes complex requests, solves puzzles, retells historical events.
ChatGPT is scalable. It can be easily deployed in large projects. For example, a developer shared a story where a chatbot implemented a project in just 30 minutes that a specialist had worked on for three years.
In programming, ChatGPT can reduce application development time because it is already trained in the necessary skills and does not require fine-tuning.
Among other things, the ability to connect APIs to third-party products opens the experience of interaction with artificial intelligence to a larger number of users.
Disadvantages of ChatGPT
Despite a wide range of advantages, the chatbot has a number of disadvantages.
As mentioned earlier, ChatGPT understands Russian, Ukrainian, Spanish, German and other common languages. However, when processing non-English queries, the chatbot shows low speed. In addition, in Russian and Ukrainian languages, the algorithm often makes mistakes such as wrong endings or tautologies.
The accuracy of ChatGPT is not always as expected. This especially applies to historical events and statistical data. To create the language model, the developers used datasets with texts, the relevance of which dates back to September 2021. Thus, the chatbot is not unaware of some current events, such as the death of Queen Elizabeth II of Great Britain or the collapse of the FTX exchange.
Because of this, it is worth carefully checking the information provided by the chatbot. Errors can be associated not only with lack of knowledge, but also with the "hallucination effect" — mixing of different data, which leads to the generation of false statements.
ChatGPT has no limits on the number of messages per conversation. However, the basic model cannot generate texts longer than 3000 words. In GPT-4, the limit was increased to 25 words.
Also, the service often refuses to communicate on certain topics. This is due to OpenAI's internal restrictions aimed at combating the generation of offensive and harmful texts. The bot will not be able to create provocative, racist, discriminatory content, or give tips on the production of dangerous or prohibited substances.
The chatbot is often unavailable due to high load and due to the extreme popularity of the service. In just two months, it already has 100 million active users.
Part of this problem is solved by ChatGPT Plus. However, the cost of $20 does not always justify the benefits of a subscription.
What are the threats posed by the spread of ChatGPT?
Despite the enormous popularity and versatility of the service, the chatbot is often criticized.
The first concern is ChatGPT's ability to create compelling texts. Critics say it risks spreading misinformation and propaganda.
OpenAI acknowledged the dangers of the algorithm and promised to develop tools to mitigate the threat.
In addition, law enforcement agencies from around the world assume that fraudsters will also "arm" themselves with this technology.
In February 2023, OpenAI released a detector of generated texts. However, even developers have recognized its poor performance in detecting content generated by large language models.
Almost immediately after the release of ChatGPT, the academic community of many countries expressed deep concern. Schools and universities around the world have reported a dramatic increase in students using the algorithm to write essays and homework.
Academic circles believe that this reduces students' critical thinking skills. In addition, the problem of inaccuracy of the algorithm threatens the appearance of scientific works containing gross errors, fakes and unconfirmed data.
In response, some universities and schools in the US, Australia, France and other countries have banned the use of ChatGPT on the premises and devices of educational institutions.
The scalability of the algorithm also increases people's fears of losing their jobs. According to various estimates, the effect of the spread of language models will affect more than 300 million employees in developed countries.
Some companies are already actively using the tool in their work. Mass media, marketers, copywriters, analysts, programmers, lawyers, doctors are an incomplete list of professions that are affected by GPT in one way or another.
It should be noted that AI will not "take" all the work. Humanity has repeatedly passed through the stages of automation, and the labor market has been deprived of obsolete professions. At the same time, new jobs appeared.
This time everything will follow the same scenario, experts are sure. Often language models will be another useful tool. People who learn to use it correctly will only increase their productivity, which will have a positive effect on wages and employment.
Alternatives to ChatGPT
One of the most affordable alternatives to ChatGPT is Bing. The search engine from Microsoft in February 2023 received a new chat mode based on GPT-4.
To use it, you need an account and waitlist registration. However, the media reported that with the latest update, access to the search engine opens instantly.
Bing's direct competitor is Bard from Google. Its announcement also took place in February, but the public release was delayed by a month and a half.
To use Bard, you must register on the waiting list. Dialog AI is only available to US and UK residents, but VPN allows you to run the service in other countries as well.
In addition, Meta released a large LLaMA language model in February 2023. It is open source and available to AI researchers.
In March, Stanford scientists used an Alpaca chatbot algorithm similar to ChatGPT. They said that they spent less than $600 on the development of the service.
Alpaca was available online for a while, but the scientists took it offline a few days after launch for ethical reasons.
Chinese companies are also trying to keep up with Western competitors. In March 2023, the search giant Baidu introduced the Earnie Bot chatbot. The company claims that the service is capable of generating text, images, video and audio.
Earnie Bot is not yet available for public access. However, early testers praised the previous version of the chatbot.
Thus, in the last four months since the release of ChatGPT, many technology companies have entered a new phase II confrontation. We are likely to see more such tools from various developers in the near future.
To always be aware of the most important things, read us at Telegram