Cover OpenAI released the fourth generation of its GPT language processing model on March 14 (Photo: Jakub Porzycki/NurPhoto via Getty Images)
OpenAI released the fourth generation of its GPT language processing model on March 14 (Photo: Jakub Porzycki/NurPhoto via Getty Images)
1 / 1
OpenAI’s latest version of the technology behind ChatGPT is much smarter and more collaborative than before
OpenAI has released GPT-4, the newest version of the technology forming the backbone of its viral artificial intelligence chatbot, ChatGPT.
GPT-4 is a large multimodal model, which means it can accept both image and text inputs, and produce text outputs. In its official description, the company states that while the latest technology is “less capable than humans in many real-world scenarios, [it] exhibits human-level performance on various professional and academic benchmarks”.
Sam Altman, co-founder and CEO of OpenAI, shared on Twitter that it also “hallucinates significantly less, and it is less biased”.
The company states that its team spent six months making GPT-4 “safer and more aligned” than its predecessor, GPT-3, in that it is 82 percent less likely to respond to disallowed content, such as how to create a bomb, and 40 percent more likely to provide factual responses.
What does the GPT stand for and how does it work?
GPT is short for generative pre-trained transformer, which is a programme that can write like a human. It requires a prompt—either image or text—from a user and it will provide a detailed answer in text form.
It has a wide range of applications, from language translation to generating text for chatbots such as ChatGPT. For instance, our team of editors asked ChatGPT two months ago questions about fear, physics and the World Cup, and got pretty detailed responses.
Read more: ChatGPT answers questions about physics, fear and the World Cup
What does GPT-4 differ from GPT-3 or GPT-3.5?
GPT-4 can process up to 25,000 words, which is almost eight times more than GPT-3, and is able to handle more nuanced instructions than GPT-3.5.
This means it can be used in cases such as producing long-form content, having extended conversations, and even editing and iterating with users on creative and writing tasks like composing songs.
It can also process images and respond with more accurate descriptions, classifications and analyses of them. Case in point: when OpenAI fed GPT-4 an image of baking ingredients and asked what could be made from them, it responded with a list of possible options, from pancakes to French toast and exclaimed: “These are just a few examples, but the possibilities are endless!”
GPT-4 is also more capable than GPT-3.5 in standardised tests. In the Uniform Bar Exam, for instance, GPT-4 achieved a score within the 90th percentile, while GPT-3.5 came in at the 10th percentile.
Above A comparison of various exam scores achieved by GPT-3.5 and GPT-4 (Photo: OpenAI)
A comparison of various exam scores achieved by GPT-3.5 and GPT-4 (Photo: OpenAI)
1 / 1
Where are the limitations of GPT-4?
It is not fully reliable and can “hallucinate facts and make reasoning errors”, although much less compared to earlier GPT models, as Altman mentioned.
According to OpenAI, it is still susceptible to making mistakes that are simple, such as accepting obvious false statements from a user, or difficult, such as introducing security vulnerabilities into the code it produces. Interestingly though, when one Twitter user fed it a live Ethereum contract, it could highlight some of its security flaws and areas where the contract could be exploited.
Its knowledge of events after September 2021 is limited, and it can’t update its knowledge from its experiences. Its ability to process prompts in other languages other than English is also limited.
Since launching in November 2022, ChatGPT has reportedly more than 100 million monthly active users, according to a UBS study. Citing data from analytics firm Similarweb, it was said to have an average of 13 million unique visitors daily in January—more than twice the amount in December.
Chong Seow Wei is a regional senior editor covering business, innovation, impact and people. Based in Singapore, she oversees content for Gen.T, Tatler’s platform for promising entrepreneurs and new-generation leaders, and its Power & Purpose vertical.
