About the beginning
ChatGPT is currently a hot topic in the tech world. Artificial Intelligence has been in the spotlight for the past decade. Here ChatGPT adds a new dimension. There are a few reasons why ChatGPT is getting noticed or talked about. A few myths have fueled this discussion such as, ChatGPT will end the days of Google, ChatGPT will cause mass unemployment across the world, etc. I am still calling these claims a lie, you will know why at the end of the tune.
Of course the main reason ChatGPT is hyped is its capabilities, it has done things no artificial intelligence has been able to do before. Currently only ChatGPT holds the record of gaining more than a million traffic in the fastest time. Even tech giants like Google, Microsoft, and Facebook could not generate so many users so quickly. However, this much traffic did not come like this. Along with the capabilities of artificial intelligence, another factor involved here is aggressive promotion. Anyway, don’t go there anymore.
Basically it is a Question Answering Search Engine based on Artificial Intelligence. It is developed by OpenAI company. ChatGPT can answer complex questions through conversation.
It is called a breakthrough technology. As soon as the tool is opened, the users are drooling. From simple to difficult, this great tool answers all types of questions in quick time.
Most of the users are now surprised with its response. Despite being a chatbot, its response is human-like which is somewhat surprising. ChatGPT suggests that the future of human-computer interaction may be seamless. Maybe computers can give people what they want.
So in today’s tune we are going to know what is ChatGPT? How does it work? With its pros and cons and some of the myths created so far.
What is ChatGPT?
ChatGPT is a Large Language Model chat bot. This is not an ordinary chat bot, but an intelligent chat-bot built with millions of data. This is version 3.5 of the Generative Pre-Trained Transformer (GPT) or Language Model. It is developed by a company called OpenAI. Its main feature is to answer user questions in the form of conversation. The response will look like someone is chatting from the other end.
It is basically a type of Large Language Model that predicts the next word according to the user’s demand from its large data or word series and tries to give the user an accurate answer.
It differs from other Large Language Models. It has an additional layer called Reinforcement Learning With Human Feedback (RLHF) which makes it more attractive to users than others. It can train itself from human feed back to give maximum correct answers. The RLHF layer was added to ChatGPT to understand what intent people are actually looking for and what answers they are looking for, and so far it’s yielding great results.
What can ChatGPT do?
Surprisingly, ChatGPT can write poems, songs, short stories and even code for you. It can also outline an article or novel. You can also create articles on specific topics with ChatGPT.
Who created ChatGPT?
ChatGPT is developed by San Francisco-based artificial intelligence company OpenAI. It is worth mentioning here that the non-profit parent company of the for-profit company OpenAI LP is OpenAI Inc. Earlier, the OpenAI company came into the limelight for its DALL·E. DALL·E is a deep learning model. The feature of DALL·E is that by giving instructions in the form of text, it can create an image according to the user’s needs.
The company’s CEO, Sam Altman, was previously president of Y Combinator. Many say ChatGPT is a Microsoft product which is completely wrong. ChatGPT was developed by OpenAI with Microsoft as an investor. Where they have an investment of 1 billion dollars. Earlier they also developed the Azure AI platform together.
Large Language Model and ChatGPT
ChatGPT is a Large Language Model (LLM) as mentioned earlier. For those who do not have an idea about this, it is better to be clear, Large Language Model (LLM) is basically an artificial intelligence model where specific tools are trained with massive data. As a result, if the user asks a question, it helps to predict which word will be in the next sentence. By now, many large companies on the Internet have their own LLM or Large Language Model. Such as Google’s PaLM, Bard, LaMDA, Huawei’s PanGu-Alpha and Meta’s OPT-IML.
The characteristic of Large Language Model is that the more words are added to it, the richer it becomes. As the amount of data in such models increases, the ability of the model to generate sentences will also increase.
Although Google has already started competing with ChatGPT, still ChatGPT is ahead of all as a Large Language Model. According to Stanford University, GPT-3 has 175 billion parameters and 570 GB of text has already been trained. Compared to its predecessor, GPT-2, its parameters were only 1.5 billion, which is about 100 times smaller.
As the scale of the data increases, the model becomes so efficient that it can do many tasks without training. For example translation from English to French can work even if the model is not trained. But the type feature was again missing in GPT-2.
Typically LLMs work much like Autocomplete, predicting what the next word in a sentence will be. With this it can write whole one page page paragraph. But the main problem with LLMs was that it did not always understand what people really wanted.
And that’s why ChatGPT appeared! ChatGPT’s Reinforcement Learning with Human Feedback (RLHF) training makes it even more powerful. It understands what answers people want to a question.
How is ChatGPT trained?
GPT-3.5 is trained with massive amounts of code and informational data. Information is collected from all over the internet, be it Reddit discussions or other sources where a question is answered by different people. And through this ChatGPT can learn how to answer questions like a human.
As mentioned earlier, ChatGPT has an extra layer called RLHF. Through which it is also trained with human feedback. And this kind of training can give great results in predicting the next word in a sentence. AI learns what people really want to know when they ask a question. And training LLMs in this way is a breakthrough in getting the right answers.
In March 2022, a research paper entitled Training Language Models to Follow Instructions with Human Feedback was published describing this step.
According to the developers, this method will create a positive attitude of people towards their Large Language Model. In this, ChatGPT will get proper training on what words to use in writing according to people’s needs. In addition to general LLM performance, in this approach they will also value human feedback. “When people provide their own feedback on a specific question, the answers will be more helpful, truthful, and harmless,” the developers claim.
Just because a language model is big doesn’t mean it can understand user intent. For example, the Large Language Model can sometimes generate output that is not safe for users. And so the RLHF will determine which answers are not safe.
The engineers who developed ChatGPT hired a few contractors to rate its output. Their job is to rate the output of GPT-3 and the new InstructGPT. The InstructGPT model is another version of ChatGPT. The results given by contractors after rating is that they prefer InstructGPT more than GPT-3. InstructGPT is slightly better than GPT-3 in controlling provocations. Finally they said that InstructGPT is quite positive but needs further improvement.
Our results show, adding human feedback to the Large Language model, significantly improves its behavior on a variety of tasks, they say. However, more needs to be done to ensure the credibility and security of the answers to the questions.
The difference between ChatGPT and ordinary chat-bot is that it is trained in a way to understand human needs. And it will provide helpful, reliable, and safe answers. As a result of this training, ChatGPT will understand which part of the question is useful and which part is unnecessary.
Another research paper was published on how ChatGPT can understand people’s preferences
Whereas researchers have observed that the metrics on which output is rated, while yielding good results, fail to understand human intent. They said, “Many machine learning applications simply set metrics and respond accordingly. They don’t always understand human language. This can lead to problems such as YouTube’s recommended clickbait promotion most of the time.
So the solution is to have an artificial intelligence that optimizes the output answers in a way that is useful to humans.
And to do that, ChatGPT is being trained with human data sets in such a way that it can predict the best answer and give the best answer to the user.
This paper reports that various Reddit tunes and various news summaries are used in this training. We’re collecting huge datasets of these summaries and human comparisons with which to train these AIs, the researchers say.
What are the limitations of ChatGPT?
Let’s know some limitations of ChatGPT, Harmful questions cannot be answered
ChatGPT is all programmed in such a way that there are no toxic or harmful responses. So it will not answer such questions. Their initiative is to ensure that users of all ages get safe information.
The quality of the answer will depend on the direction
A notable limitation of ChatGPT is that its output depends on its input. The better the users can input, the better the output will be. So experts can use it to get better answers.
The answer will not always be correct
Its major limitation is that it is trained to look correct all the time. But it cannot be said that it will always give 100% correct answer. Many users have already noticed that ChatGPT sometimes gives wrong answers. Since it is still in experimental run, one cannot expect perfect data from it all the time.
Moderators of the Stack Overflow website have noticed that users are using ChatGPT most of the time to answer various questions. Tuned answers look true but most of them are wrong. Thousands of incorrect answers have already been removed from the platform by Stack Overflow moderators to maintain the credibility of the platform.
OpenAI itself has stated some limitations. OpenAI says,
ChatGPT can sometimes give wrong answers. And solving this problem is quite challenging because during training it cannot be 100% sure, the sources are true.
During training, so much care is taken that the questions that can be answered correctly are sometimes left out.
Sometimes this training can even confuse the model, because the correct answer depends on what the model has, not what the human wants.
Is ChatGPT free?
Since it is not yet complete and research is being done on it, it is now free for anyone to use. It is now available to users for free. It is being improved based on user feedback. The official announcement said that although we try to reject malicious requests, sometimes it may give wrong or malicious replies.
They also said that although we will block certain harmful content using the Moderation API, there will still be some negative or positive content. We are interested in receiving user feedback so that we can further improve our ongoing work. At the moment, ChatGPT creators are hosting a contest for up to $500 ChatGPT credits to entice users to rate the response.
Language Models Will Models Replace Search Engines?
Google has already released their own AI chat-bot. Google’s chat-bot is called LaMDA. Google engineers claim that LaMDA will be able to carry out human-like conversations.
Currently, where companies like OpenAI, Google, and Microsoft are creating their own chat bots, there is a question on everyone’s mind and that is, will these chat bots take the place of traditional search engines?
Many people on social media are already claiming that ChatGPT is the next Google.
The answer to the question whether chat-bots will take over the search engines will be most disturbing to those who are engaged in the profession of search engines like SEO, ranking etc.
Technology researchers, however, are not dismissing such claims as baseless. Since technology is improving day by day, chat-bots can also replace search engines, but in that case, more things need to be added here, hybrid search engines have to be created.
Some Myths About ChatGPT
Ever since people came to know about ChatGPT social media has been heating up about it. There is nothing surprising about it in the viral or trending era. So now we will talk about some of the myths that are being spread around the internet.
Google’s days are over here
A tune is constantly being shared on various social media, including Twitter, “The days of Google may be over,” but are the days of Google really over? I don’t think so. Because you must first understand what ChatGPT is and what services Google is providing. If I compare it with Google’s search engine, the current version of ChatGPT lags behind. We are getting results from thousands of websites by searching something on Google, we can create our own topics from there and we can verify the truth about something from multiple sources. Does this matter get you to ChatGPT? In google search we are getting images, videos, news, articles about a particular topic which is so far impossible in ChatGPT. Moreover, Google will definitely not sit on this result of ChatGPT.
A few days ago I saw in a tune what is the national bird of Bangladesh ChatGPT says Haas. Maybe it has been corrected now but if you googled this same question you would get hundreds of answers from which to verify its authenticity. We all know what is the national bird of Bangladesh, I gave an example.
ChatGPT will replace Quora
If you search on Google about ChatGPT, a question would come up before, ChatGPT will replace Quora? My opinion on this question is that ChatGPT may not replace Quara even if it upgrades itself further. ChatGPT basically collects data from Q&A websites like Quora, analyzes it and provides results. As unlike any question on Quora that gets human answers, ChatGPT gets machine answers. Not going to say that ChatGPT can give wrong information, but to say that a few people’s answer will definitely be ahead of ChatGPT’s answer alone.
ChatGPT will put millions of people out of work
Admittedly, artificial intelligence will take over many human jobs in the coming days, and we are already seeing it. But my question here is how accurate it is for ChatGPT. So far we know that with ChatGPT we can code, write music, write poetry. If we talk about coding. Naturally, those who do freelancing know how much revision to give after designing a website or developing a program. The client repeatedly sends corrections, fix this fix that. Now how much customization can one ask for by coding with ChatGPT?
It is known that you can translate through ChatGPT. Google Translator has been working since 2006, but still can you translate from English to Bengali accurately? That is possible to do by humans?
last word
ChatGPT is currently being talked about a lot on the internet, although it is free now, it looks like it will have to be paid for in the future. More than 1 million users registered on it just five days after its opening.
