Image by Editor

What is GPT?

GPT stands for Generative Pre-trained Transformer. It is a neural network machine learning model which is trained using data on the internet to generate any type of text. This sophisticated neural network is used to train large language models (LLMs) to simulate human communication. 

The model tracks words in a sequence, allowing it to learn both the context and meaning of the language. The GPT model focuses on text-only, allowing it to use artificial intelligence to analyze what the user is asking and effectively generate text. 

It has taken the artificial intelligence world by storm with its conversational abilities, contextual information, and more. The model can handle tasks such as text summarization, code generation, and provide valuable insight within seconds. 

A Little About GPT-3

GPT-3 is an autoregressive language model that trains by predicting what token is next. The model will need an initial prompt text and can continue to produce text, using that initial prompt.

Reinforcement Learning with Human Feedback (RLHF) is used to help the model achieve conversational dialogue with the user. GPT-3 is a 175-billion-parameter language model, with the following use cases:

• Text summarization
• Content creation
• Code generation
• Generate comics and poems
• App creation

If you would like to learn more about ChatGPT-3, have a read of: ChatGPT: Everything You Need to Know

What can ChatGPT-4 do?

When Microsoft Germany’s CTO Andreas Braun announced that GPT-4 was planning to drop the third week of March, there were a lot of speculations. Dr Andreas Braun said in AI in Focus: Digital Kickoff:

“We will introduce GPT-4 next week, there we will have multimodal models that will offer completely different possibilities – for example, videos”

At this time, there was not a clear announcement from OpenAI themselves, until yesterday (14th of March 2023). So what should we expect from GPT-4? GPT-4 is OpenAI’s new technology that provides the most advanced system, producing much safer and more useful responses.

Greg Brockman, President and Co-Founder of OpenAI stated in the GPT-4 Developer Livestream that OpenAI has been building GPT-4 since they opened the company, with the past 2 years focusing on perfecting the new technology. They had to rebuild the entire train stack and train the model to see what it was capable of. 

ChatGPT-4 is multimodal, meaning that it can use a variety of data types, such as image, text, speech, and numerical data along with multiple intelligence processing algorithms to produce accurate high performing outputs. It is no longer limited to being a language model.

Stating ChatGPT-4’s Role

If you haven’t had the chance to play around with ChatGPT-4 yet, there is a ‘System’ section on the left-hand side. This section is where you will tell what you want the assistant to do, with specifics and requirements. This is the instruction guide for the assistant to use to ensure it provides you with exactly what you want. 

For example:

• ‘You are ChatGPT, a large language model. Follow the users' instructions carefully’
• ‘You are an AI programming assistant. Follow the users' requirements carefully. Describe each step you took in detail. Output the code in a single code block’

Let’s start with what you can do with existing content.

Existing Content

I will first go over ChatGPT-4’s capabilities when using existing content/articles/text. 

Summarizing Context with Specifics

Greg Brockman showcases the use of ChatGPT-3.5 by copying and pasting an article and asking ChatGPT to ‘Summarize the article into a sentence where every word begins with ‘G’’. ChatGPT-3.5 naturally failed the task. However, GPT-4 successfully outputs exactly what the user asked. 

However, the assistant uttered the word ‘AI’, to which Greg Brockman responded by saying ‘AI doesn’t count! That’s cheating!’. The assistant happily responded with a word that replaced ‘AI’ and started with the letter ‘G’.

GPT-4 can specifically output what the user wants, by giving the assistant requirements to follow. 

Combining Ideas

You can also flexibly combine ideas using different articles using GPT-4. By copying and pasting your articles, you can ask the ChatGPT-4 assistant questions such as: Find one common theme between these two articles.

If the output from the ChatGPT-4 assistant wasn’t exactly what you were looking for or insightful enough, you can provide feedback, it will improve its response. 

Generating and Building with GPT-4

You can also use GPT-4 to build things! You will need to provide it with a prompt, with a little bit of detail so that the assistant can provide you with exactly what you need. For example, ‘Write me a discord bot’.

Based on the role you assigned to ChatGPT-4 in the systems section, for example, if you wanted the assistant to build you something code generated - they would be assigned as an AI programming assistant. Along with the prompt, this will successfully help the assistant output what you requested. 

You can test the code block generated by the assistant, to see if it works. If you do come across an error, you can simply send the error message to the assistant and provide you with the correct code block. You can keep on doing this and directing the assistant till your code works successfully.

Mathematical Calculations

Being able to solve complex calculations, such as taxes and high-level calculation problems can be a challenge. You can now use ChatGPT-4 to help you with these mathematical calculations. For example, if you have a tax problem you want to be calculated, you need to assign the ChatGPT-4 system as a TaxGPT, so it knows what its role is.

Providing context about your problem, the assistant will be able to perform mathematical calculations. The interesting fact is that the model is not hooked to a calculator - impressive right?

Visuals

The image feature is not currently publicly available - but it is in the works! You can input an image and ask the assistant questions about the image. At this current time, it does take a while to output, however, OpenAI is optimizing the model to make it faster.

Handwriting

You can snapshot a piece of writing that has been handwritten and ChatGPT can read the handwriting and convert it into text. Some are even joking about how it can detect doctors' handwriting - which we all have struggled with understanding in the past, and also in the present. 

Using Content After September 2021

As we already know, ChatGPT does not have any knowledge after September 2021. However, you can provide ChatGPT with the article or information as a prompt as to what you want to ask the assistant. The assistant will use it as a learning resource to provide you with an accurate output. 

Provide Feedback to OpenAI using Evals

If you wish to contribute towards ChatGPT, with feedback and opinions - you can do so by using Evals. Evals is a framework used to evaluate OpenAI models and an open-source registry of benchmarks.

Evals allows you to create and run evaluations:

• That use datasets to generate prompts,
• To measure the quality of completions provided by an OpenAI model
• To compare performance across different datasets and models.

This will help to evaluate and examine the capabilities of the model, and how OpenAI can improve it and take it to the next level. 

Wrapping it up

Since the release of ChatGPT-3.5, we’ve seen a lot of changes and advancements in ChatGPT-4. It’s exciting to see what people are planning to cook up with ChatGPT-4. If you’ve already had the chance to play around with it, let us know in the comments what you’ve learnt so far.
 
 
Nisha Arya is a Data Scientist, Freelance Technical Writer and Community Manager at KDnuggets. She is particularly interested in providing Data Science career advice or tutorials and theory based knowledge around Data Science. She also wishes to explore the different ways Artificial Intelligence is/can benefit the longevity of human life. A keen learner, seeking to broaden her tech knowledge and writing skills, whilst helping guide others.