Just months after its launch, ChatGPT has taken the AI world by storm with its cutting-edge language capabilities powered by GPT-3.5 from OpenAI, which was integrated into several popular platforms like Bing, Skype, and Edge. However, the company has now revealed that it has upgraded to the latest and most powerful language model yet – GPT-4.
With its enhanced ability to process and understand language, GPT-4 promises to take the already impressive capabilities of ChatGPT to the next level. Users can expect even more accurate and natural language responses, making interactions with ChatGPT an even more seamless experience. This upgrade reaffirms the commitment of OpenAI to push the boundaries of what is possible with AI language models.
Microsoft has been quietly using GPT-4 for the last five weeks in its Bing Preview program, a fact that has only recently come to light. Microsoft has been a long-time partner of OpenAI and its use of GPT-4 in Bing reaffirms its commitment to leveraging the latest AI technology to enhance the user experience. While it’s unclear when the customized version of GPT-4 will be rolled out to the public, it’s clear that Microsoft is investing heavily in AI language models to deliver more powerful and accurate search results to its users.
To start with, GPT-4 is expected to be a “multimodal” language model, which means that it will be able to process not only text but also other types of input such as images. This could potentially allow users to attach images to their queries and receive more comprehensive responses. Although GPT-4 will not be able to generate images like other AI models such as Stable Diffusion or MidJourney, it will still be able to provide explanations and information in text format.
For instance, GPT-4 might be able to explain any situation using both textual and visual cues. However, the output generated by GPT-4 will be limited to text only, without any accompanying images.
OpenAI: ChatGPT is getting much smarter
In addition to its multimodal capabilities, GPT-4 is expected to be even smarter than its predecessor, GPT-3. According to reports, the OpenAI team has already tested the model using practice exam books from 2022 and 2023. It’s worth noting that these exams and their answers were not included in the model’s training data, as the model’s knowledge cutoff is September 2021.
Despite this limitation, the results of the tests reportedly showed that GPT-4 performed exceptionally well, providing accurate answers and demonstrating an impressive level of understanding of the questions.
GPT-4’s enhanced capabilities have also been demonstrated through its performance on standardized exams. In contrast to its predecessor, GPT-3.5, which scored in the bottom 10% on the bar exam (a legal exam that lawyers must pass), GPT-4 scored in the top 10%, indicating a significant improvement in its ability to process and analyze legal information.
Furthermore, GPT-4’s performance on other standardized tests was equally impressive such as 88th percentile on the LSAT exam (a standardized test required for admission to law school), while GPT-3.5 only scored in the 40th percentile. In SAT Math, GPT-4 scored in the 89th percentile, compared to GPT-3.5’s 70th percentile.
One of the most significant improvements in GPT-4 is its “steerability” feature, which allows users to exert greater control over the model’s responses. Unlike previous versions, which required users to prepend certain rules in order to direct the AI’s behavior, GPT-4 is designed to be more flexible and adaptable, allowing users to guide the model’s output more directly.
Previously, it was possible to trick the AI into revealing its underlying rules, as Microsoft did when it instructed “Sydney” to act as Bing without revealing its Sydney code name.
This increased steerability is expected to make GPT-4 more useful in a variety of applications, from customer service chatbots to language translation tools. Importantly, this feature also reduces the risk of “cheating” or “hacking” the model, as users can no longer exploit specific rules or patterns in the AI’s responses.
In previous versions of AI language models, developers had to resort to appending certain rules to direct the model’s behavior. However, this approach came with certain limitations and risks, including the possibility of “jailbreaking” the model.
Fortunately, with the development of GPT-4’s steerability feature, companies can now exert greater control over the model’s style and task with a system message. This improved method of controlling the AI’s behavior is expected to make it easier for companies to deploy natural language processing tools in a variety of contexts, from chatbots to customer service interactions.
Previously, it was possible to trick the AI into revealing its underlying rules, as Microsoft did when it instructed “Sydney” to act as Bing without revealing its Sydney code name.
By enabling developers to control the AI’s output more precisely, this new method is also expected to reduce the risk of hacking or cheating the model. Microsoft and OpenAI have been at the forefront of this development, working together to refine the technology and ensure that it meets the highest standards of performance and security.
While GPT-4 represents a major leap forward in AI language models, it’s important to acknowledge that the model still has some limitations, particularly when it comes to factual accuracy. Like its predecessor, GPT-4 has the potential to generate false information, known as “hallucinations”.
However, GPT-4 is significantly better at avoiding these mistakes than previous versions, such as GPT-3.5, which scored 40% lower on internal testing. This improvement is a testament to the incredible pace of progress in natural language processing, as GPT-3 was only released in mid-2020, and GPT-3.5 arrived in early 2022.
Source: Bing Blogs, OpenAI