Artificial Intelligence (AI) has taken the world by storm and is developing at an exponential rate. It was only about four months ago that the San Francisco-based company Open AI made ChatGPT, the AI chatbot based on the GPT-3 family of large language models, available to the public.
Within two months ChatGPT reached 100 million users and overtook all previous app growth records. The implications of the versatile ChatGPT are far-reaching, with many people becoming concerned about their profession, while academic institutions are rethinking their assessment strategies. Meanwhile, Microsoft is embedding ChatGPT in many of its products, such as Bing, Azure, Excel and Word.
The new ChatGPT-4
On March 14, OpenAI released ChatGPT-4, a much more powerful tool, as successor to ChatGPT. Open AI describes it as “a massive leap in the field of natural language processing (NLP)” and believes that “this model, with its ability to understand and generate human-like text, has the potential to revolutionise the way we interact with machines and automate various language-based tasks.”
The “Chat” in the name emphasises that it is again a conversational computer interface, while “GPT-4” refers to the “generative pre-trained transformer version 4”. It includes the analysis of vast quantities of information from across the internet to determine how to generate human-sounding text and provide detailed responses to questions from users.
ChatGPT-4 can handle larger and much more complex interactions and queries and can solve difficult problems with greater accuracy due to its broader general knowledge and problem-solving abilities. ChatGPT-4 is also more creative, collaborative and more human, according to OpenAI.
More difficult to trick
It is known that chatbots are often easily led astray to provide answers outside of their boundaries. Since GPT-4 has been trained on numerous malicious prompts, it is much better on “factuality, steerability, and refusing to go outside of guardrails”.
A better memory
Large language models are usually trained on millions of web pages, books, and other data, but during a conversation they are limited with regard to how much they can keep “in mind”. In ChatGPT-3.5 it was 4 096 tokens or about 8 000 words or four to five pages of a book. This often led to the AI losing track of earlier answers. GPT-4 has a maximum count of 32 768 (215) tokens, which is about 64 000 words or 50 pages of text.
Until now the AI world was very much dominated by English, but GPT-4 is truly multilingual and can answer questions with high accuracy across 26 languages.
“Steerability” is an important concept in AI and refers to its capacity to change its behaviour on demand. GPT-4 integrates steerability more natively than GPT3.5 and can thus fulfil various roles, such as that of a sympathetic listener or knowledgeable expert. Users can now change the “classic personality with a fixed verbosity, tone and style” to a personality that suits their needs better.
Seeing and understanding images
One of the greatest features and most noticeable changes of ChatGPT-4 is its “multimodal technology” that enables it to handle images together with text. ChatGPT-3 and 3.5 were limited to text. The user can now submit an image alongside text and ChatGPT-4 will be able to process and discuss both. It can also explain what it sees in an image such as a cartoon.
According to OpenAI, the ability to use video as an input is also on the horizon.
CPT-4 may not be as proficient as humans in some real-world situations, but it has demonstrated human-level performance on several professional and academic benchmarks. It scored in the top 10% on a simulated Uniform Bar Exam, while its predecessor, GPT-3.5, scored in the bottom 10%. ChatGPT-4 also performed above the 88th percentile in the LSAT, SAT Math and the SAT Evidence-Based Reading and Writing exams.
Chat-GPT4 is not able to handle reasoning on current events since it has been trained on data collected before 2021. According to OpenAI it still has some “limitations such as social biases, hallucinations, and adversarial prompts”.
Unlike ChatGPT3, ChatGPT 4 is not freely accessible by the public. Access to the newest and most powerful large language model is only available to users who upgraded to and paid for ChatGPT+ (the premium version or plus option) at R368 per month. There is no way to access the latest version of ChatGPT for free, unless you are a Bing User. The “new Bing” allows users to ask real questions, get complete answers, chat and create.
Typical questions are: Create a three -course vegetarian menu for four persons; help me plan my special anniversary trip to the Greek Isles in June; what art ideas can I do with my kid using domestic items; write a rhyming poem about a dog; write me a story about a cat for children; and write code in Python to find the Fibonacci sequence.
Many future uses
Morgan Stanley is using ChatGPT-4 to systematise wealth management data, while payment company Stripe Inc is testing it to determine if it can assist in the combating of fraud, and Duolingo, the well-known language-learning app, is incorporating it to explain mistakes and to allow users to practice real-world conversations.
Be My Eyes, in co-operation with OpenAI ,is introducing a GPT-4 Virtual Assistant or Virtual Volunteer for blind people or people with poor vision to increase accessibility by providing them with powerful new resources to navigate physical environments, address daily needs, and gain greater independence.
The image-to-text generator allows users to send images to the virtual volunteer that provides instantaneous visual assistance for a wide range of tasks.
If the user, for instance, sends a picture of the interior of their refrigerator, the Virtual Volunteer will not only identify the contents but will also analyse and extrapolate which dish could be made with the available ingredients, provide some recipes and step-by-step guides to prepare them.
ChatGPT-4 can also describe the pattern on a dress, identify a plant, read a map, translate a label, and explains how to get to a particular product in the supermarket.
Google Chrome has five top ChatGPT extensions that enhance the communication experience of the user and make online interactions more efficient:
ChatGPT for Google is an open-source Chrome extension to enhance the search engine experience by displaying ChatGPT responses alongside results from popular search engines such as Google, Bing, Baidu, and DuckDuckGo. A ChatGPT account is necessary to use this extension.
The Liner AI Chrome extension powered by ChatGPT provide answers directly on the Google search results page. Just like ChatGPT it is convenient and dependable, but also offers references and suggest search terms for further exploration of a topic. It allows the user to highlight web content – including YouTube videos and images – with the Web Highlighter feature and recommends content through Content Curation based on the searches and highlights. It can also block ads on the web and YouTube for a more pleasurable browsing experience.
AIPRM for ChatGPT adds a list of 1 520 curated prompt templates for Search Engine Optimisation (SEO), Software as a Service (SaaS), and many more, to ChatGPT to make it easier to access the advanced AI capabilities of ChatGPT and get quick and efficient answers.
The ChatGPT Chrome Extension allows the user to easily access the advanced AI capabilities of ChatGPT from their web browser. Precise, to-the-point responses in real time are given to questions. It is useful for a variety of tasks such as providing assistance with language translation, personalised chatbots, or other AI-powered communication tools.
The ChatGPT for Search Engines extension allows the user to display ChatGPT responses for search results and provides prompt ideas from #SEO, #Ads, #Video, #BlogContent.
Although ChatGPT has many applications, OpenAI has not released an official desktop app yet and offers it only via a web browser. However, there are a number of non-official ChatGPT desktop apps available for the Mac, Windows and Linux operating systems.
ChatGPT-4 will probably send large waves across the world. Other tech companies like Alphabet’s Google has introduced its own AI service, called Bard, to testers. In China, Baidu will soon unveil its bot, Ernie, and will be followed by Alibaba, Meituan and several others.
ChatGPT is a powerful language model and became even more powerful with an increase from 175 billion parameters to 170 trillion parameters and the inclusion of multimodal capabilities. ChatGPT is expected to have a significant impact on the world, preparing the way for a range of exciting innovations and creations.
Professor Louis C H Fourie is an Extraordinary Professor in Information Systems of University of the Western Cape.