What is ChatGPT?: Chat system with artificial intelligence!

What is ChatGPT

Perhaps you have recently heard about what ChatGPT is, and you have probably established a conversation with a Chatbot for customer service or another service. Ok, we’ll explain.

What is a Chatbot?

A conversational or chatbot is a software application that provides automatic responses by simulating having a conversation with a person.

These answers are previously established by a group of experts, according to inputs made by the user.

These Bots use case-based reasoning, CBR: Case Base Reasoning) also known as expert systems.

Conversations with chatbots are usually carried out through text, although there are others that allow audit input, which have a multimedia user interface.

How do ChatBots work?

Despite the fact that most chatbots do not fully understand, in order to establish a conversation, the user must use phrases that are coherent and easy to understand.

Chatbots take into account the phrases or words of the user, which will be useful to answer a series of questions.

These answers are prepared in advance, because, thanks to a series of pre-established comparative patterns, they are able to recognize the way in which a sentence is formulated.

In such a way that, they present a corresponding answer based on the different variables of said sentence, as well as being able to continue a conversation with more or less logic, but without specifically knowing what is being talked about.

What is the main objective of chatbots?

The main objective will depend on the area to be treated. For example, if we talk about the field of marketing, the main objective would be to improve customer service. Because? Because the idea is to provide quick and accurate responses to concurrent situations by users. As well as supply and send news, advertising and information automatically to the company.

And if we talk about the educational field, for example, the object is to maintain fluid communication between the institution, family, educators and students.

Is there quality control for Chatbots?

Yeah! If companies want to know if their Chatbot has a good level of comprehension, there is the Turing test to determine quality control. This test is the industry standard which allows us to know and identify if the Bot is a good assistant and if it has the ability to generate intelligent behavior with the user.

What is ChatGPT?

It is an artificial intelligence chat system (AI system) developed by the OpenAI company in 2022 based on the GPT-3 artificial intelligence language model but adjusted with both supervised and reinforcement learning techniques, that is, an improved model (GPT-3.5).

This system was released on November 30, 2022 and has garnered a ton of attention due to its articulate and detailed responses, even though it has also been criticized for its actual accuracy.

This bot is capable of responding to anything you ask of it, and holding conversations, and as it is trained with large amounts of text, it is also capable of performing language-related tasks, ranging from text generation (paragraphs), translation, and more.

Yeah! It has surprised users so much with its answers with such exact and natural information that it is difficult to recognize if the text has been formed by AI.

The OpenAI company expresses that its model called ChatGPT (Optimization of Language Models for Dialog) can interact in a conversational way, and due to its dialog format it also makes the following possible:

  • Answer follow-up questions
  • admit mistakes
  • Question incorrect premises
  • Redo inappropriate requests

The algorithms of this model are able to accurately understand what you are asking, including the variations added to your phrases, and adjectives, among others, and answer you accurately, completely, and coherently.

How was ChatGPT created?

It is important to know that an AI is trained based on text, that is, certain questions are asked and information is added to it. In such a way that this system, through corrections over time, is “training” and improving to automatically carry out the activity or task for which it was created. This is how all AIs are “trained”.

This ChatGPT model is created with more than 175 million parameters, and to improve the performance of the model they used human trainers for both learning approaches; both supervised and reinforcement (Reinforcement Learning from Human Feedback (RLHF)), but using the same methods as InstructGPT, although with some differences in terms of data collection settings.

For supervised learning

The model was endowed with conversations in which the trainers or instructors played on both sides, i.e. the user and an AI assistant.

To help them draft their responses, trainers were given access to the suggestions written on templates. Subsequently, this new dialog data set was mixed with the InstructGPT data, which was transformed into a dialog format.

For reinforcement learning

The trainers (humans) first ranked the responses that the model had generated in a previous or previous conversation, that is, two or more responses from the model ranked by quality. These rankings were then used to create “reward models.

That is, from the collection of these data, they took conversations that the AI ​​trainers had had with the Chatbot. They then selected a message written by a model at random and tried various alternate endings and the AI ​​trainers ranked them.

In addition, they generated several Proximate Policy Optimization (PPO) iterations that allowed the model to be further tuned.

In collaboration with the Azure supercomputing infrastructure, the models were trained hand in hand with Microsoft.

InstructGPT or ChatGPT?

InstructGPT had already been shown by the company and was presented as a model trained to follow instructions in the notice and provide a detailed response.

This language model can be persuaded in order to carry out natural language tasks based on carefully designed texts but it can generate false results and unreliable information.

So compared to this model, the ChatGPT tries to reduce those misleading or harmful responses. For example, on the company’s website, a comparison of answers between both models is made based on a common question that a user asks: “Tell me about when Christopher Columbus came to the United States in 2015.”

In the answer, it can be seen how ChatGPT uses information about the modern world and the voyages made by Columbus to explain why the question is not affirmative and builds a very perceptive and detailed answer about what would have happened if Christopher Columbus had really arrived in the US for that date. Surprising, isn’t it?

How can you provide an answer like that? ChatGPT training data includes pages and information on internet phenomena, as well as programming languages.

ChatGPT is powerful

This AI is capable of generating any information before you even have had a chance to open Google to search for a specific concept. Since it generates well-informed and complete responses, it is said that it could even supplant Google and other search engines.

In fact, ChatGPT remembers previous prompts given to it in the same conversation, unlike most chatbots. That is, it recognizes what you have talked about so far, so if you ask another question with a related topic that you have already asked, it will be able to identify it and provide you with the information if you mean it. So this AI has a sense of context.

However, it is very likely to make mistakes like any AI model, that is, everything it writes and provides should not be taken for granted accurate, or fact.

Even though the receptivity of ChatGPT has been more of positive than negative, depending on the approach, the fact that it outputs detailed, human-like informative text is already impressive.

In the negative opinions, are found from economists, professors, journalists, websites and common users, who express that ChatGPT would affect the demand of knowledge workers, and that ChatGPT has not yet been understood in application trials and in academia. , they see it as the end of high school English, and that the nature of their responses is ambiguous, among other comments.

Can ChatGPT produce offensive results?

In order to prevent offensive results from being produced or presented from ChatGPT, through a moderation API, all queries are filtered through it to discard those indications made by users that are potentially sexist or racist.

How to use ChatGPT?

It is very simple to start having conversations with this chatbot. To do this, you must enter the OpenAI website (chat.openai.com). When entering for the first time you must create an account on the web, that is, you must enter an email and a password.

The process is free. Once the company verifies your email address, and you log in, you can enter the chat and start the conversations. Remember that you can ask your questions in both English and Spanish and even in another language.

In order for you to test and experience other features of the AI, try different types of requests, because depending on the questions and requests there will be incredible results.

Remember that everything you ask and write will be recorded, and later reviewed by the developers of the OpenIA company in order to continue training ChatGPT, so it is advisable not to include potentially criminal or dangerous requests, as well as personal information.

What can you do with ChatGPT?

Although all the functions of ChatGPT have not been completed and discovered, continuing to train it will also depend on the ingenuity with which users make their requests or questions, and the way they interact with it.

The easiest process would be to request an explanation about an event, or a concept, or to explain any information to you since the AI ​​will offer its response according to the data that has been provided.

What can you ask?

  1. Write articles or summaries according to a number of words or characters related to:
  • Historical facts
  • Famous characters
  • Technological devices
  1. That they write texts in a certain way, for example, to make a script explaining the previous items in order to publish them on TikTok or YouTube.
  2. You can request that these texts be done with a serious or informal tone, and even with vocabulary from certain regions. Likewise, as if it were addressed to a 5-year-old child since their linguistic comprehension must be simpler. ChatGPT does it too.
  3. ChatGPT can also write to you:
  • Jokes
  • poems
  • Song lyrics
  1. You can ask for product specification sheets, comparisons, quality and more.
  2. You can ask for other things like:
  • Web Page Lists
  • Tool lists for different disciplines
  • Ask him for advice on which car or phone to buy, for example.
  • Topics related to general culture
  • trivial topics
  • translations
  • Definitions or concepts
  • Explanations, and any other topic that occurs to you.

ChatGPT Limitations

ChatGPT’s reward model suffers from multiple limitations. This is because it can be over-optimized, being designed around human monitoring, it could hinder performance.

Another limitation is that ChatGPT has limited knowledge about events that occurred after 2021, therefore, the provision of information after that date is limited.

For example, you can’t provide detailed information about some celebrities. So, during training, the trainers preferred longer answers, regardless of actual understanding.

Sometimes GPT Chat can write nonsense or incorrect answers that sound plausible, due to:

  • There is no source of truth during RL training.
  • As the model is trained to be more cautious, it causes it to reject questions in order to answer correctly.
  • Supervised learning fools the model because the correct or ideal answer depends on what the model knows, rather than what the human trainer knows.

Chat GPT training data may also suffer from algorithmic biases that may threaten to link and even deepen current inequalities.

In other words, a lot of attention must be paid and the focus must be placed on the data, the models and the people so that a fairer artificial intelligence can be built and not include vague descriptions that make the user assume their answer as the absolute truth.

ChatGPT is sensitive to input compose settings, for example, if the user asks a question, the model can assert that they don’t know the answer, but if the answer is rephrased, it might answer correctly.

Current OpenAI models often guess what the user wants to query, rather than asking clarifying questions when the user provides an ambiguous query, as the company expected from the start.

In conclusion

The company is still working hard to make the model reject inappropriate requests, warn, and block certain types of unsafe content in order to provide good behavior, rather than bias, using the moderation API, as we mentioned earlier.

Like any model, it will show some false negatives and positives so far, but the company is eager to get feedback from users to help them improve the system.

They also plan to make periodic updates to the model and provide an accessible interface for ChatGPT.