What is GPT-4 and How Does it Differ From ChatGPT?

  ●   March 22, 2023 | Technology
Written by
March 22, 2023 | Technology

Unless you’ve been living under a rock, you’ve probably heard about the online sensation ChatGPT. 

Towards the end of last year, OpenAI, an artificial intelligence (AI) research laboratory, released ChatGPT – an artificial intelligence chatbot. From launch, ChatGPT gripped the world with its ability to tackle complex subjects and generate speeches, songs and essays off the back of information learned online. 

Well now, OpenAI has launched a new language model called GPT-4, purported to be even more creative, powerful and predictable than its predecessor, GPT-3.5 (the model behind ChatGPT). According to OpenAI, GPT-4, contains  “more advanced reasoning skills”, meaning it can handle “much more nuanced instructions”, respond to images and process up to 25,000 words (approximately eight times as much as ChatGPT). It’s also set up to be “safer and more aligned” than GPT-3.5.

Continue reading to find out exactly what you need to know about GPT-4; what’s special about this technology; what distinguishes it from GPT-3.5 and how accurate its responses are. 

How Can I Use GPT-4?

Since launch on Tuesday 14th March 2023, GPT-4 has been available to a range of companies through exclusive paid-for-access data-sharing partnerships. Household names currently adopting the technology for their products include language app Duolingo, financial firm Morgan Stanley, and payments provider Stripe.

GPT-4 is also available to subscribers of ChatGPT Plus, the OpenAI’s subscription service billed at $20 (£16.50) per month. The premium plan allows access to its chatbot during peak use periods, as well as quicker response speeds, and priority access to new features.

Can I use ChatGPT-4 for free?

At the time of writing, there is currently no way to access GPT-4 for free, unless you’re a Bing user. When GPT-4 launched on Tuesday, Microsoft concurrently revealed that its ever-popular chatbot Bing AI is being powered by GPT-4! Nevertheless, the company’s chatbot is only available in preview at the moment, meaning you have to join a waiting list to try it.

If you want to skip ahead of the queue, Microsoft has hinted that there are a few things you can do. These include making Edge your default web browser, making Bing your default search engine, and making MSN your homepage.

What’s so special about GPT-4? 

The four main improvements that distinguish GPT-4 from its predecessor and other AI technology are its abilities to:

  • Provide more creative answers
  • Deal with longer documents of up to 25,000 words
  • Deal with visual inputs
  • Make fewer mistakes

OpenAI has shared that part of what makes the new version so advanced is its performance on no less than 34 academic and professional exams. These include the Uniform Bar Exam, SAT Math, and SAT Evidence-Based Reading and Writing exams. When tested, GPT-4 performed far better than most people. In more than half of all cases, it scored higher than the 80th percentile and even reached the 90th percentile in the Uniform Bar Exam, (versus 10th previously). 

Trained on back-and-forth conversations, the software is also capable of processing follow-up questions, admitting its own failings and limitations, and rejecting inappropriate requests.

GPT-4 vs GPT-3.5

Part of what makes GPT-4 more sophisticated than its predecessor is that it was trained on more data (however, OpenAI is yet to confirm how much data was actually used). The training data used by such AI programmes is what is known as ‘parameters’ and suggests the skill of the model on a problem. We know the largest model in GPT-3.5 has 175 billion parameters (compared to competitor programmes, such as Meta’s at 65 billion parameters). Therefore, we expect the parameters of GPT-4 to be pretty substantial!

However, in certain ways GPT-4 still has the same limitations as GPT-3.5, Like GPT-3.5, the bot lacks knowledge on events “that have occurred after the vast majority of its data cuts off” (September 2021). Similarly, like the original, this version has the propensity  to “hallucinate” from time to time – a phenomenon where AI invents facts or makes reasoning errors.

Can GPT-4 recognise images?

Yes. A large part of what makes GPT-4 so advanced is its “multi-modal” ability to analyse images in addition to text. Consequently, it is capable of scanning the components of images for information and responding to these captions or answers in natural, conversational text. However, this feature is yet to be accessible to the public due to safety concerns. 

A demonstration on OpenAI’s website showed the programme responding to an image of eggs and flour with a range of meals that can be made using the ingredients. Pretty nifty!

Is GPT-4 safe?

Like the launch of GPT-3.5 and other AI systems, there have been some safety concerns about GPT-4 and its capabilities. 

AI technology is capable of generating original images, text and other content based on basic language prompts, using online information. As such, the responses produced can at times seem disruptive. 

The decision of OpenAI to hold back the programme’s image recognition feature has alerted some sceptics that there may be safeguarding concerns around the technology. The fact that this new model was developed and tested over a relatively-short six month period has cast further doubt on its security. 

For its part, OpenAI has made a point of letting users know the chatbot is safe to use, having undergone strenuous security checks. The firm consulted with over fifty experts in AI, safety and security when developing the software. Such testing purposely showed that the chatbot is “82%  less likely to respond to requests for disallowed content and 40% more likely to produce factual responses than GPT-3.5”.

Nevertheless, since AI programmes are trained using online information, there will inevitably be concerns. The National Cyber Security Centre – part of the UK’s intelligence, security and cyber agency – has warned the public that sensitive queries stored by chatbot providers could be hacked or leaked. It also warmed that such technology could be used to create more convincing phishing emails. 

Conclusion

Since GPT-4 only launched on Tuesday 14th March 2023 and is available to only a portion of the public, it is indeed too early to judge its success. What we do know is that AI is one of the most powerful content marketing trends this year and is changing the game for content marketers. As for what it means for the market industry more broadly? Time will tell. Nevertheless, the very fact that we are even talking about GPT-4 so soon after the initial launch of ChatGPT spells exciting things for the future of AI and the speed at which it’s gripping the world. 

Our Blog