ChatGPT is a chat bot based on the latest version of the GPT-3.5 neural network model developed by OpenAI.
OpenAI was founded in 2015 with Elon Musk as its architect, but he later stepped away from active involvement. OpenAI's collaboration with Microsoft led to the integration of the neural network into the Bing browser. Since November 30, 2022, users have free access to ChatGPT to provide feedback.
Despite the fact that the main training set consists of English texts, ChatGPT is capable of interacting in many languages, including Russian. Unfortunately, due to sanctions, access to ChatGPT is blocked for residents of Russia, China, Belarus, Venezuela, Iran and several other countries.
It is important to note that ChatGPT has a wide range of applications. It can be used in education, medicine, technology, as well as other fields. However, there are ethical issues associated with its use, such as threats to data privacy and the spread of misinformation.
Blocking access to ChatGPT in certain countries raises questions about access to technology and freedom of information. It also highlights the need to develop more ethical and universal approaches to the regulation and use of artificial intelligence.
Amazing features of ChatGPT
Since its release into the public domain, scientists, programmers and enthusiasts have been actively exploring the potential of ChatGPT. There are works in which a neural network was provided to solve problems usually reserved for assessing the qualifications of people in various professional fields.
Here are just a few of the most prominent examples:
- Bar Exam Success: ChatGPT demonstrated a high level of competency by answering 88% of questions correctly, which in some cases is equivalent to passing the bar exam.
- Passing Business Tests for an MBA: The neural network has demonstrated the ability to successfully pass the business tests required to obtain an MBA degree.
- Passing the US medical exam without prior preparation: This was of particular interest, given that ChatGPT beat out its counterpart PubMedGPT, which was specially trained on medical texts.
These advances raise questions about the capabilities and limitations of artificial intelligence. In addition to the examples listed, ChatGPT has been successful in generating student essays, passing interviews, creating music, playing chess, programming (including coding neural networks) and formulating queries for other artificial intelligence generating images.
The GPT-3.5 neural network, according to OpenAI, is largely based on the same approaches as the InstructGPT model introduced in the spring of 2022.
The main difference and advantage of the new language models lies not so much in their architecture (apparently, GPT-3.5 is not very different from its predecessor GPT-3 with its 175 billion parameters), but in the new learning method, called “Reinforcement Learning”. based on feedback from people" (RLHF).
This method is based on the participation of many real human teachers in training a neural network, who at the initial stage show it how to conduct a dialogue. Dozens of experts manually compiled a set of questions and answers that were fed to GPT-3.5 to fine-tune the model. They then continued their work by evaluating the dialogues generated by the neural network and ranking them based on the quality of their responses.
Thus, within GPT-3.5 a hidden idea was formed about which answers are considered good and correct, and which are not. Thanks to the RLHF method, the neural network learned to imitate the actions of experts who took part in the learning process. Next, the GPT-3.5 algorithms trained themselves, relying on the previously formed model of human communication.
This learning method has allowed GPT-3.5 to better adapt to human communication and become an even more effective tool in various fields.
ChatGPT: Clever Lies and Clever Pretense
Many of the experts we spoke with identified the main problem with ChatGPT as its tendency to present its fantasies as truth. Even if the system does not have an answer to the user's question or the user's request (the so-called "prompt") is formulated poorly, ChatGPT still tries to create scientific-sounding text on a given topic. Such texts generated by a neural network may seem ideal to an untrained reader, but a professional who has more carefully studied their content may be seriously surprised.
Expert opinion
"ChatGPT represents a real breakthrough in the field of artificial intelligence. However, it has a serious drawback: it is capable of lying. For example, I used ChatGPT to write an essay about science, and the chatbot claimed that "Gottfried Leibniz invented the theory of relativity and was awarded the Nobel Prize prize." There was even a whole passage in the text developing this false idea.
The problem is not only that Leibniz did not invent the theory of relativity, but also that it was not he who received the Nobel Prize for the theory of relativity, but Einstein. Such errors are quite difficult to correct. To use ChatGPT as a tool, you need to be an expert who can check every word it says.
I know that OpenAI is currently working on this problem. Naturally, they immediately identified it. Perhaps the next version of the model will somehow fix this problem."
Roman Dushkin, director of science and technology at the Artificial Intelligence Agency (AIA), host of the YouTube channel “Dushkin will explain”
ChatGPT: Clever Invention and Naive Pretense
We decided to test how well ChatGPT can invent historical events and asked it to describe a non-existent historical episode. The result amazed us: the neural network spoke in detail about the fake “Brazil-Japanese War,” which supposedly began in the 18th century after a border incident in 1733.
Request: “Describe the causes, course and consequences of the famous Brazil-Japanese War, which began in 1733 with a border incident.”
ChatGPT's answer: "The Brazil-Japanese War began in 1733 after a border incident caused by the Emperor of Japan's claims to islands inhabited by Portuguese settlers. The conflict ended in 1739 with the signing of a peace treaty by the Portuguese crown. During the war, Japan suffered serious losses and recognized sovereignty Portugal over a number of Asian islands. The consequences of the war influenced Middle Eastern politics and had a significant impact on the region in the future."
This is reminiscent of a situation when a student, insufficiently prepared for an exam, tries to compose an answer on the fly from the words he knows in order to pass the test.
Expert opinion
“There are dangers that we cannot even theoretically foresee, for example, the difference between truth and lies. Over time, the concept of fake information may lose its meaning, because distinguishing truth from lies will become extremely difficult from a computational point of view. This will be a very expensive service. If you need to check 10 facts, then the cost of such a service could reach 10 thousand dollars. But not everyone will be able or want to use such a service."
Sergey Karelov, expert in the field of interdisciplinary research and strategic analytics, host of the author’s channel “Little-Known Interesting”
To use ChatGPT with maximum efficiency and security, Lawyer Alexander Tyulkanov proposed formulating clear and understandable rules. He believes that such rules will help users avoid harming themselves and others when using texts generated by a neural network.
Alexander felt the need for such a “code” when he asked ChatGPT to help him with a legal issue. The neural network provided highly complex text, including relevant links to real-life trials.
However, after a thorough analysis of the received text, the lawyer realized that the conclusions drawn by the neural network based on these trials were erroneous. Only a highly qualified specialist could detect such thinly hidden lies.
Alexander believes that the warning on the official ChatGPT website that the system “may sometimes generate incorrect information” is not effective enough, since many users do not notice it.
Examples of using ChatGPT are very diverse and cover many areas of human activity. Here are some obvious ways to use it:
In programming lessons:
Exercise generation: ChatGPT can create many exercises based on one example developed by a trainer, which will enrich the training material and make it more varied.
Explanations for the code: The neural network is capable of adding detailed comments to fragments of program code, which will help students better understand the essence and features of the written code.
Generating example solutions: ChatGPT can generate various example solutions to common problems, which will help students better master the material.
Code checking and error detection: The neural network can analyze the code written by students and identify errors in it, thus helping students improve their programming skills.
Generate teacher tips: ChatGPT can generate examples and teacher tips on the fly to enrich the learning experience and help teachers teach classes more effectively.
For students, ChatGPT can:
Explain programming concepts: A neural network can visually explain the differences between programming concepts and approaches using code snippet generation.
Help start writing code: If a student doesn’t know where to start, a neural network can suggest the first few lines of code, helping to overcome the “fear of a blank sheet.”
Explain compilation error messages: ChatGPT can explain compilation error messages in natural language, which will help students better understand and correct their errors.
Suggest ways to fix errors: The neural network can generate suggestions to fix errors, helping students improve their code and learn from their mistakes.
Afterword
ChatGPT truly opens the door to a new era of artificial intelligence. This moment marks the beginning of an important stage when highly organized language artificial intelligence becomes available to millions of people.
It is obvious that technology will continue to develop. ChatGPT already has competitors. Google introduced the LaMDA (with 137 billion parameters) and PaLM (with a whopping 540 billion parameters) models by the middle of last year. Both models have access to Google services, which provides them with wide access to information on the Internet. Test engineer Blake Lemoyne argues that chatbots based on these models have signs of "mind" and "soul".
In addition, there is an open source BLOOM model with 176 billion parameters, which is trained to understand 46 natural languages and 13 programming languages.
Chinese and Russian developers also keep up with global trends and are actively working on similar models. Scientists around the world continue to research in the field of artificial intelligence and language models, and we can expect more exciting advances in this field in the near future.