ChatGPT, OpenAI’s natural language generation chatbot, has been the craze for quite some time now, and now with the newest iteration of GPT models, GPT 4, launching in full swing, all eyes are on this latest model
There’s a rising curiosity among users since GPT 4 launched on 14th March 2023. Users are continually inquiring how and what the differences are between GPT 4 from the earlier models, how it is better than the earlier iterations, and what exactly it can do.
Here are the answers to all these queries and more.
What are Generative Pre-trained Transformers (GPT)?
Generative Pre-trained Transformers are a type of deep learning model that uses natural language processing and machine learning to generate human-like text. Typical uses of GPT include:
- Answering questions
- Summarizing text
- Translating to other languages
- Generating code
- Creating blog posts, stories, conversations, and different content types.
The applications of GPT models are practically endless, and you can even fine-tune them by giving specific inputs.
AI has always been aimed at generating human-like text and generating output to user queries in a way understandable to humans. This revolution with natural language became possible only after the invention of transformer models, starting with Google’s BERT in 2017. Before that, other deep learning models, like recursive neural networks (RNNs) and long short-term memory neural networks (LSTMs), were being used to generate text. The shortcoming with these models was that they failed to generate long-form content, though they were pretty good at outputting single words or short phrases.
BERT’s transformer approach was a major breakthrough in the fact that it is not a supervised learning technique. It does not require an expensive annotated dataset to train it. Google used BERT to understand and process natural language searches but could not generate text from user prompts.
Transformer architecture (Image source: GPT1 Paper
In 2018, OpenAI published a paper named (Improving Language Understanding by Generative Pre-Training) about their GPT-1 language model. This model was a concept and not released publicly.
Model performance on various tasks (Image source: GPT2 paper)
The next paper from OpenAI, Language Models are Unsupervised Multitask Learners, came out in the following year and featured their latest model GPT-2. This time, the model was available to the machine learning community and did some text generation tasks, to begin with. GPT-2 could generate a couple of sentences before breaking down.
Results on three Open-Domain QA tasks (Image Source: GPT-3 Paper
In 2020, OpenAI published another paper; Language Models are Few-Shot Learners, on their GPT-3 model. This model had 100 more parameters than GPT-2 and was trained on a vast text dataset. The model was worked on, and several iterations were published, like the GPT-3.5 series, including the conversational ChatGPT.
ChatGPT took the world by storm and became the fastest growing web application ever, with over 100 million users in just 2 months. It surprised the common users who, till now, stayed away from AI-generated text with pages of human-like content.
The awe is still not over, and OpenAI has launched its latest GPT model, GPT-4, with improvements to do way more and with more precision than GPT-3.
What’s New in GPT-4?
GPT-4 has been improved in model “alignment” – the ability to follow user intentions while also making it less offensive and generating more truthful output.
Also, GPT-4 has a maximum token limit of 32,000 (equivalent to 25,000 words), which increased from the 4000 tokens (equivalent to 3215 words) of GPT-3.5.
“We spent 6 months making GPT-4 safer and more aligned. GPT-4 is 82% less likely to respond to requests for disallowed content and 40% more likely to produce factual responses than GPT-3.5 on our internal evaluations.”, quoted by OpenAI.
OpenAI also revealed that GPT-4 outperformed the previous versions in some tests designed for humans, like Uniform Bar Examination, the SAT for university admission, and Biology Olympiad, by scoring way higher.
Image Source: OpenAI
GPT-4 has been improved to generate more factually correct answers. The number of “hallucinations,” i.e. when the learning model makes factual or reasoning errors and still presents it confidently, has reduced. GPT-4 scores 40% more on this clause than GPT-3.5.
It also improves “steerability,” which is the ability to change its behavior according to user requests. For example, you can tell it to write a particular piece in different tones, like Shakespearean, or in different dialects, like American, Australian, etc. This is a significant breakthrough because it is challenging for an AI model to teach different tonalities.
You can start by using prompts like, “You are a garrulous data expert.” GPT-4 is also better at adhering to guardrails. Meaning it is better at refusing requests when a user asks it for something illegal or unethical.
Understanding Visual Inputs
One significant development in GPT-4 is that it can understand images and can interpret them. This feature is not yet available to users; it has only been researched. Users can specify language and vision tasks by entering interspersed text and images.
Take a look at the images below, interpreting complex imagery such as charts, memes, and screenshots from academic papers.
GPT-4 Performance Tests
OpenAI took a step further and tested its product for various tests designed for humans, such as the Uniform Bar Examination and various other tests like SAT, LSAT, and Biology Olympiad.
GPT-4 was also tested on traditional benchmarks designed for various machine-learning models, and it outperformed all the existing models by a large degree.
ChatGPT is free, but GPT-4 is available to users at $20/month for Indin and USA users. GPT-4 is much higher in cost than earlier versions of GPT.
GPT-4 is already being used in different applications. Apart from Bing, which uses this model to generate more precise results, some other great apps are already using GPT-4 model to enhance their performances.
The list of apps includes:
- Be My Eyes
- Morgan Stanley Wealth management
- Khan Academy
- Government of Iceland
Though there is a great debate running regarding whether AI-generated content can be a welcome boost to creators from different walks of life or whether large language models like GPT-4 poses a threat to workers losing their job, it is no mistake that ChatGPT and then GPT-4 has taken the world by storm. Though it still needs improvements, it won’t be wrong to say that there are possibilities for making lives easier for many.
Many people have already used ChatGPT to ease their work. Be it content creation or the generation of codes and other technological uses, GPT has done it all. And when we consider all these improvements took just 5 odd years to develop and implement (from GPT-1 in 2018 to GPT-4 in 2023), it can be safe to say that it is not far from now that we will see much more improvements in GPT models, until it becomes near perfect.
While only time will tell how much GPT models can do or whether they will outperform humans, it can be easily predicted that GPT is here to stay.