What is ChatGPT?

As you have probably already experienced, the use of bots to generate automatic responses is becoming more common every day.

The industry seeks to provide revolutionary technology to make those responses more “human.”

OpenAI is leading the way and they introduced a long-form question-answering AI called ChatGPT.

This technology is trained to learn what humans mean when they ask a question. ChatGPT will give you not only the answer but also human-quality answers.

How does ChatGPT works?

Like you, here at Graphic Illusion, we were assailed by this question. So, to understand how it works, we need to dig a little deeper into the concepts that shape this technology.

The way humans learn to have a conversation is through an endless exchange of sentences and ideas. We start by listening to our parents and giving them a sound that works for us as the correct response.

In the case of machines, humans provide them with a basic dictionary, or more precisely, a basic language. On top of that, humans teach them how to learn to assimilate and properly use the information they are receiving. Sort of like our parents corrected us when we used the verb or sentence construction differently.

It may sound simple but it is not.

A fundamental part of this is Language Modeling (LM), which is the use of statistical and probabilistic techniques to determine the probability that a certain sequence of words will occur in a sentence. This is what your computer or cell phone uses every time you start typing something on your screen.

LM is used in natural language processing (NLP) applications, especially in those that generate text as an output.

How Language Modeling Works?

By analyzing text data, a language model can determine the probability of the next best word. They use an algorithm that shows the path to the best context in natural language.

The model is capable of learning basic language features and can incorporate them into its database. With this new information, the language model can understand new sentences and give the most accurate answer.

There are different approaches to modeling language. It depends on the amount of data they have to analyze and the field where they are going to be used. Every day, you can see progress on this topic, and the models are improving their functionality a lot.

Language models need to be trained. Their performance is primarily determined by the amount of data they can analyze. At a very basic level, a language model requires only about 10 million or 100 million words to give a reliable answer.

In ChatGPT, according to an article published on BBC Science Focus, 300 billion words were fed into the system!

You are probably wondering how this amount of information can be used to construct adequate answers.

What is Reinforcement Learning from Human Feedback (RLHF)?

OpenAI, the creators of ChatGPT, introduced the idea of incorporating human feedback to train models to be able to interact with the environment in 2017. This method is called snapping.

It is very important to align those models with human objectives because they are becoming more and more powerful.

If you relate this to the way we learn to interact with adults growing up, you’ll see the benefit of having a “reward” from a human every time the model makes an appropriate response.

In life, in general, optimal behavior is difficult to determine. We decide if it is good or bad based on established rules. This human-provided reinforcement or reward helps Language Models to learn. In the case of ChatGPT, we are talking about a Large Language Model (LLM). That is why it was necessary to contribute 300 billion words to this LMM.

It is important to know that in traditional reinforcement learning, the reward function provided to a model is written by hand. In RLHF, the reward is self-learned and self-integrated.

This is a complex but interesting subject and if you want to learn more I recommend reading a Weights & Biases article on RLHF.

The surprising performance of ChatGPT.

The most amazing feature of this AI Chat is that it is capable of giving you very accurate and complete answers. You can even ask it for a 1000-word essay and you will get it.

However, like any AI model, it is possible to get some bugs at some points.

This LLM is so powerful that it has a sense of context and recognizes everything you have been talking about so far. So if you ask it a question related to an answer it is given you, it will know what you mean without having to give it all the explanation.

There are debates right now about how good it is to have something like this.

Because it can be used to create content with such precision, the New York City Department of Education has already banned its use for fears of student cheating.

On the other hand, the New York Times published a blog where the author recommends teaching with ChatGPT instead of prohibiting its use.

What is certain is that you can use ChatGPT to help you generate ideas and content as it analyzes large volumes of data almost instantly, saving you time and effort.

So, is this a replacement for humans?

No, ChatGPT is still far from a replacement for humans. The general idea is to treat it as a tool to be used in balance with human experience and creativity.

You can not just copy and paste the answers it gives you. It is necessary to review them and introduce the human touch. It may not always produce useful results on the first try.

You can use it to brainstorm ideas for social media, market research, audience insights, analyze big data, and other concepts you can think of to get a good search intent approach.

It is very important to get used to being very specific with the directions you try to give to ChatGPT. If you are not specific, the answer will be generic. So it is important to go in with a sense of experimentation and not rely too much on automation.