how gpt-3 works
Written by Mark Thomas on November 3, 2021

How GPT-3 works

How GPT-3 works: Generative Pre-trained Transformer 3 is an autoregressive language model that uses Deep Learning to develop human-like texts. The GPT-3 fully loaded version has an estimated capacity of 1,5 billion machine learning parameter values.

The quality of the text generated by the model is so important that it can be difficult to evaluate whether or not it was written by a human who has both advantages and disadvantages.

For the second period on April 30, 2019, Microsoft officially renamed the Model of Power to Model2Power as its original Microsoft product.

The two scientists also showed their findings in the initial paper published on May 28, 2020.


What is GPT-3?

GPT3 is an autoregressive language prediction model employing deep learning to come up with human-readable text. This neural network technology can produce a text unable to distinguish the way people use their language.

Through this project, openAI hopes to develop a computer system that mimics the human mind with high precision.

Although not exactly the AGI, OpenAI would like this tool to serve as a good starting point in achieving this goal.

It is the third version of the prediction model in the GPT series and it is the work of the San Francisco-based OpenAI Institute.

How good is it?

GPT-3 is a scalable deep neural network for predictive programming of next words in sentences.

It is used to generate data on the corpus from over.1 billion words and can produce accurate text.

The performance is similar to the best language models for text production which is statistically much better than previous models such as Google's BERT and Stanford Natural Language Processing (NLP)'s Transformer.

OpenAI claims that GPT-3 achieved this degree in performance in the first pre-training period without having additional training evidence.

In addition, this third version of GPT may generate longer sentences or paragraphs than the previous model, Google's BERT.

How does GPT-3 work?

With 175 billion metric bits of data, the GPT3 machines learn from training data and subsequently act on problems token by token.

The processes are much harder because they are more complicated.

GPT3 working model starts with training (with training data) and learning from an example to follow (focusing on the format of the answer) and can decide the right output from a database.

To learn unsupervised rather than teaching the machine to answer using a definite input gives the machine an example and provides an example of the answer follow.

Working with GPT-3 presets

The OpenAI Playground is available for downloading and testing the GPT3 Language Models.

In this tutorial, we will only use the DaVinci model which is the most advanced at this time.

Once you start learning the Playground you can switch to the other models and experiment with those also.

There is a large variety of available pre-sets available within open AI for different use of the language model.

All the presets provided with OpenAI are easy to understand and can be described very easily, so it is a good idea right now of playing some of those presets.


Create your own GPT-3 applications

The OpenAI Playground is a tool to build your own applications running in the GPT-3 engine.

In this section, we will examine everything you can find on the platform to create your own solutions.

To follow the examples below, make sure you reset your playing area to its default settings.

To do this delete all text from the text area and click the “x” beside the name of the preset to delete it.

If you plan to create a new application delete the text to train the engine and adjust the settings on the sidebar according to your needs.

Response length

GPT-3 sometimes halts at the end of another word. When you choose to change the length of the response you need a response.

You'll have some controls if you do not set this value.

The standard reply length for response is 64.

A simple trick is to set the length to a bigger than you need and then remove the incomplete part at the end.

We will see later how to teach GP-3 to stop at the proper position in the next part of this article.

The original response to Python (a modern programming language) is given with 0 temperature and a long of 64 tokens with the token being defined either for a word or a punctuation mark.

We could also put on a second button under "Submit" for GPT.

neural network

Frequency and presence penalties

The sliders ‘Frequency Penalty’ and ’Presence Penalty’ allow you to control the level of repetition GPT-3 allows in its response.

Frequency penalty works to reduce the chance of a selected word being selected the more times that word has been used.

The penalty for absence takes into account the time taken to use the same verb but is not a case of grammatical and non-grammatical.

I am not well aware of how these two choices operate. In general, I've found that when these options are set to defaults of 0, GPT 3 is unlikely to repeat because of the randomization given by the values Temperature and/ or Top P.


The temperature setting allows for randomness of the resultant text.

A 0 value makes it deterministic because every time it outputs a new input text the same is produced. GPT-3 doesn't like strings that end inside spaces as this produces weird and sometimes unpredictable behavior.

The playground will show you a warning if you accidentally leave more than one space for an input.

If you try this several times at once the result will be different.

The temperature has also its default value of 1, so ensure the “Top P” parameter below the temperature has also some control over randomness.

Displaying Word Probabilities

This debugging tool shows what happens if the chosen token was picked. The darkest the background of that word the better that word was picked. If you type one word you'll notice a list of the words taken up in those positions in the text.

When setting this option to 'least likely' the colorization turns backward, with darker backgrounds assigned to the words regardless of the fact they are not possible.

The least likely word is colored with green for best likely words and red and blue for the least likely. This word was chosen by randomizing the Top P and Temperature setting.

Custom presets

The ELI5 software program represents the first interesting development in OpenAI Playground of all time. Use the Floppy disc icon to save the project as preset. Every preset stored in a save can be named along with a description.

If you wish to share a preset with other people you will receive a URL that you can show to friends. Note that anyone who receives this address must have access to the Playground to run the Preset. You may also copy this URL to a friend who has access to the URL.

Start text

The “Inject Start Text” option in the settings tells Playground which text automatically appends to the input before sending a requesting command to GPT3.

You can submit another document when you use all the prefixes in the list. The prefixes will be immediately included on requests for submitting the response from the server that request.

Next, the text will be changed to a single line of text with var: foo. Press enter for the cursor at the beginning of the second line and press submit for the next variable. Tap on "Send" for Next variable.

The option “Top P”

The “Top P” argument is an alternative way of controlling the randomness and creativity of the text generated by GPT-3.

The OpenAI documentation recommends that only one of the temperatures and top P are used to make sure that one of them is set at 1. For the ELI5 bot, I've decided on using top P with 0.5 being the better response for the bot.

Obviously, this kind of application a .5 is too hot and as a result, GPT-3's answers (the generated text) become vaguer and more informal.

To see if I might improve these answers slightly I lowered it to.5

Using multiple prefixes

The ELI5 (Explain Like I'm 5) bot accepts a complex concept from the user and returns an explanation using simple words that a child can understand. GPT-3 can also be used to generate Q&A chatbots and to convert designer instructions given in English into HTML.

Here’s how we can train bots with “microphone” as their training example: We shall make sure that we use a few examples for our answer that will be used for training. We could use thing: or el5: to ensure that “thing” lines do not need to be generated.

Stop sequences

With the option “ Stop Sequences ”, you can define a series that requires GPT3 to stop when generated. Let's say we do want to use the same variable every time an engine is in motion.

Given that each line of a line is combed as var: and that a single preposition is placed in a previous input line the same type could be used to set it down as stop.

Type another var on the third line of the input text and submit again. And you get another. Add another variable to this field. Three. Line in data.


Using a short prefix for all text can be a valuable help from which GPT-3 can learn to understand what the expected response may have.

It means that you need to show a variable and allow it to generate more. The problem here is that we are not clearly saying we want a further line than this one. Maybe we just can make a prefix. But we have to clarify that despite the second line being incomplete in comparison with the first because we want something like foo on it. It really does it better.

The option “Best Of”

The ‘Best of’ option should then be used for having GPT-3 produce multiple results for a query. In fact, the play area chooses the best one to display. I know it has not actually worked for me because it seems uncertain what of the various options are the best. Also, when putting this option on a value otherwise of 1 the playground will stop generating responses because it needs to receive the complete list of responses.

Restart text

The automated ELI5 bot can now be accessed through the user interface and explain things more quickly. The option "Inject Restart Text" can be used to automatically insert text after the GPT-3 response to help us autotype the next prefix. It is easier to interact with a bot so that it tells a bot something in its response. The bot currently answers questions related to using the robot and answers them.

The OpenAI Playground

GPT III can be used to “train” the engine to produce text output. This is very simple in the OpenAI Playground. The first sentence in bold font represents what we're receiving as input. The second part uses the identical text: prefix, which appears in bold. This second appearance of the prefix is one of the final parts in the input. Hit the 'Submit button in the bottom part on a page twice to generate text. Whenever a person presses submit again the engine runs to generate another bit of text.

Celebrating its strengths: Potential applications of GPT-3

GPT 3 can give humans reading capabilities. Often they become more automated if the software is needed for their business systems. Companies may use it to expand customer service that won't send a customer to a computer. Currently, the company operates a private beta program with invited users. Organizations with access to the program can develop some of them including Text-based customer support applications that can be developed by the program. For example, the client service app is developed.

Language Translators

GPT-3 is programmed to Translate between Spanish and French using contexts. It can rephrase whole English paragraphs for simple content. Technologies are very useful in some branches such as medical industries and legal firms where words are not readily understood in ordinary people. Revtheo is an example. This tool helps people to find what is meant based on the way that word is used. More importantly, technology can be used to rephrase whole English phrases into simple text.

Mathematical Decoders

GPT-3 can understand mathematics but it can't generate accurate mathematical answers. The software shows how the answer was received correctly when adding two numbers. The results may not have helped. It seems that the system is configured to handle English inputs in structured data formats such as eXtreme Markup Language (XML) or JavaScript Object Notation (JUNN). It can be used to learn how to use structured data to learn more advanced.

What we will build

The OpenAI Playground is a web-based interface that allows you to play and prototype solutions based on GPT-3. We don't intend on building one single arbitrary project. Instead, we'll implement several prototypes in response to various problems. As part of this course, we will also take care of the transfer of work you do on your Playground to a standalone Python application.


In order to use it only a GPT3 license is required. If you want to write Python applications for yourself, you can also install Python 3.6 or higher. It is completely optional you may skip a part about Python if you aren't interested. The same can be done by requesting a beta license directly via OpenAI to follow some examples.

Crazy hype

GPT3 altered the entire Artificial Intelligence landscape in both the industry and the education arena. In this manner, many people associated humans with certain features, some developed products for them. Some companies built on top of this system. It made headlines everywhere and a group of scientists began to build similar algorithms. We can find some examples of fake news generated by GPT 2 and GPT 3. There are dozens of businesses that have built AI copywriting assistants that are now widely used by the copywriting industry.

Potential dangers

GPT3 is an excellent Artificial Intelligence but is capable, like any other powerful technology being malicious; fake news is a good example.

Article written by Mark Thomas