OpenAI's latest release, GPT-4, now accessible through ChatGPT and Bing

March 15, 2023 by
OpenAI's latest release, GPT-4, now accessible through ChatGPT and Bing
DxTalks, Ibrahim Kazeem

OpenAI's latest release, GPT-4, now accessible through ChatGPT and Bing

OpenAI's latest AI model, GPT-4, has been recently launched and is found to perform at human-like levels on several professional and academic benchmarks. 

GPT-4 is a robust multimodal model that enables users to input both text and image and receive text output. 

This article examines GPT-4's capabilities, restrictions, and potential risks involved in its use.

 By the end of this article, readers will have a better understanding of GPT-4's potential impact, its strengths and weaknesses, and what it can and cannot accomplish.


GPT-4 exhibits superior capabilities compared to its predecessor, GPT-3.5, in terms of handling nuanced instructions, creativity, and reliability. 

OpenAI conducted tests on several benchmarks, including simulated human-designed exams, and determined that GPT-4 surpasses other large language models. 

Moreover, GPT-4 performs well in languages other than English, including low-resource languages such as Welsh, Latvian, and Swahili.

-Visual Queries

GPT-4 has the ability to process text and images as inputs, allowing it to generate text outputs based on both types of input.

Although the model’s visual input capability is currently in the research preview stage, it has demonstrated comparable performance to text-only inputs.

-Steering ability

OpenAI has been actively developing ways to enable developers and API users to steer and customize the behavior of GPT-4. 

This is done through the "system" message, which allows developers to prescribe the style and task of their AI. API users can also customize their users' experience within certain limits, providing significant personalization options. 

The goal is to improve the usability and applicability of GPT-4 in various fields and industries.


Despite its impressive capabilities, GPT-4 has some noteworthy limitations. 

It is not able to accurately process data from after September 2021, which can lead to incorrect reasoning and false statements being accepted as true. 

Additionally, GPT-4 can be overconfident in its predictions, and its accuracy can be reduced after post-training fine-tuning. It is also unable to solve challenging problems like humans, such as introducing security issues in its code. 

Consequently, caution should be taken when using language model outputs, particularly in high-stakes contexts.

-Risks and Mitigations

Despite the impressive capabilities of GPT-4, it can also bring about new risks, such as the generation of dangerous advice, flawed code, or incorrect data. 

To address these risks, OpenAI has sought the help of over 50 experts to conduct adversarial testing on the model, as well as collecting new data in order to refine GPT-4's ability to reject dangerous requests. 

As a result, OpenAI has made numerous improvements to GPT-4, making it 82% less likely to produce improper content than its predecessor, GPT-3.5. 

The model is also better programmed to observe use rules when it comes to sensitive topics such as medical advice and self-harm. 

Even though OpenAI has made the model more resistant to bad behaviour, the possibility of creating content that goes against usage rules still exists. 

OpenAI has voiced its concerns that GPT-4 may be a double-edged sword, and are thus collaborating with other researchers to explore the potential impacts it may have on society.

-Training Procedure

The GPT-4 base model was trained to predict the next word in a document using publicly available data and data licensed by OpenAI. 

To make it more aligned with the user's intent, the model was further fine-tuned using reinforcement learning with human feedback (RLHF) within guardrails.

-Predictability Scale 

OpenAI has placed a major emphasis on constructing a deep learning stack that performs reliably at differing sizes. 

Through the utilization of various optimization measures and infrastructure, OpenAI has been able to accurately forecast GPT-4's ultimate training loss.


Microsoft has confirmed that the new Bing search experience is now running on GPT-4, OpenAI's powerful natural language processing (NLP) model. To access the API, users must sign up for the waitlist in order to see if they can get access. 

Currently, the usage limit is capped, but OpenAI may adjust it depending on demand and system performance. 

Those who want more GPT-4 usage can sign up for a ChatGPT Plus subscription, which gives them access to additional features. OpenAI is also exploring the possibility of adding another subscription tier in order to allow more GPT-4 usage.

Rounding Up

The unveiling of GPT-4 marks a major step forward in OpenAI's mission to expand the capabilities of deep learning. 

Despite some imperfections, this new language model has been able to demonstrate near-human level performance on a range of academic and professional tasks. 

While this provides immense benefits, it is important to be mindful of the potential risks associated with using such models in critical situations. 

OpenAI is actively pursuing solutions to reduce these risks, as well as developing a deep learning stack which is more reliable and predictable.

About DxTalks

DxTalks is an events, media, and consultancy firm specialized in the Digital Transformation, Technology, Blockchain and Web3 field focused on the MENA region.