Ai chatbots Explained & How Chat-GPT Train:
Tool tester is website that conducted an interesting survey
a few months ago. They gave a text to about 2,000 Americans and asked them
whether they think the text was written by a human or an Ai. Surprisingly, 53%
of the people were not able to tell that the text was written by ChatGPT. They
thought that it was actually written by a human. This was the result when
chatgpt 3.5 they used chat gpt 3.5 version. Then there was an advanced version,
chatgpt 4.0, after using which the statistics were even more mind-blowing.
63.5% of people could not recognize chatgpt 4.0 that this text was written by
chatgpt.
Chatgpt4.0 how is this possible? By the way, this is not
that surprising because as i told you in the last article artificial
intelligence's purpose is to mimic the cognitive functions of humans. So,
chatgpt is mimicking the way humans talk. And it is doing it so well that
people are not able to recognize it. Chatgpt can mimic the chat behavior of
humans very well.[
For this reason, the category of artificial intelligence
chatgpt is categorized in is called the chatbots category. Chatbots simulate
human conversations. Chatbots are not new technology. Before this, apple's Siri
is also a chatbot. Amazon's Alexa is also a chatbot. And often when you go to a
website to book a hotel or to buy anything else, you will often see a chat
button on the screen you can click that and chat. These are all Ai chats.
ChatGPT takes these ai chats to the next level. Because it falls under the
category of generative chatbots. The other chatbots are either rule-based
chatbots or retrieval-based chatbots.
What is a rule-based chatbot?
Rules would be written by the programmers. For expected
questions like, “how to book a room" programmers would feed the rules and
FAQs. The computer will automatically find the keyword and from there, it will
take that keyword and will give you an answer.
What is a retrieval-based?
Retrieval-based chatbots are worse. There if you ask a
question, it will try to "best match" your question with a
pre-existing response and will give you that as an answer. Before the chatgpt
was launched, you'd remember that these chatbots were very irritating. Because
often these food delivery companies used them because often these and if you
had to make a complaint that the order delivered had gone bad or was wrong. Then
you had the option to chat. But you were not chatting with a human instead, a
computer would be chatting with you. And the responses were very irritating.
Because it was providing you with pre-defined responses.[
THE CATEGORY OF CHATGPT GENERATIVE CHATBOTS
The category of chatgpt generative chatbots, is a new category. Here, you are not provided existing responses, rather new responses are created based on your input. By using machine learning and natural language processing. But to make chatgpt this good, it took a lot of effort. A large quantity of data had to be fed to it. and along with that, advanced techniques have been used such as transformers and reinforcement learning with human feedback. We will talk about both these techniques later.
How chatgpt was trained?
First, let's understand how chatgpt was trained. Now, a
computer can understand only the language of 0s and 1s. It cannot understand
our language so how can you teach it?
The English language is so complicated in itself that if we
give it some sentences to understand, It won't be able to recognize patterns by
itself.to train these chatbots, friends, large datasets are used. This is the
reason why; they are called large language models. We provided it with large
and we told it where it needs to find relationships in that text, where it
needs to recognize patterns.in the basic training of language models, we give a
computer a sequence of words and tell it to guess the next word.[
Here, two techniques are used most commonly.one is next
token predicting and the other 1s-maskedlanguage-modeling.