Unveiling the Potential of GPT-4: A Game Changer Across Disciplines

The landscape of numerous sectors continues to shift and reshape as advancements in Artificial Intelligence (AI) unfold. A recent development, GPT-4, the latest architecture in the ChatGPT series, propels this transformation. Join us as we dissect GPT-4's capabilities that challenge human expertise and unravel its potential. These ideas are inspired by the overview presented in NRC, Gartner reports, and articles from The Economist. 

Meet ChatGPT

Based on a study by Rajesh Kandaswamy, a researcher at Gartner, ChatGPT is a language model created by OpenAI with impressive generative AI capabilities. According to Kandaswamy, organisations increasingly explore and evaluate models like ChatGPT to enhance their product and service offerings. GPT-4, the latest model from OpenAI, released in March 2023, improves upon its predecessor, GPT-3. The improvement comes from using a transformer model architecture to predict the following segment of a word, or as the techies like to call it, the “token.” 

Trained on a mix of public and private data and refined through reinforcement learning and human feedback, GPT-4 stands out with its ability to handle image and text inputs and produce extended output forms. However, improvements in accuracy, factuality, language performance, and guardrails to limit harmful outcomes don’t mean flawless. GPT-4 can still display hallucinations and biases and respond unpredictably to adversarial prompts.

GPT-4 vs. GPT-3: What's New?

GPT-4 introduces the capability to accept images as input for text generation, setting the stage for future incorporations of other modalities like audio. It also accommodates longer text inputs—up to approximately 3,000 words. Apart from these, OpenAI identifies several enhancements, including better performance in other languages, improved accuracy, factual representation, generative capabilities, intent alignment, creative enhancement, complex reasoning, and safety improvements.

GPT-4: A Deeper Look

The GPT-4 model, an AI product of OpenAI, has been causing a stir in the technology sector. Its capabilities range from generating and rephrasing sentences on-demand to tackling complex tasks across mathematics, programming, healthcare, law, and more. To put it to the test, Marc van Oostendorp, a Professor of Dutch and Academic Communication at Radboud University, administered the Central VWO Dutch Exam to GPT-4.

In a face-off with the Central VWO Dutch Exam, GPT-4's performance improved significantly from its predecessor, GPT-3.5. GPT-4 scored an impressive 85% on the Dutch exam and 80% on the French. Despite some odd mistakes, the accuracy of the answers took Professor Van Oostendorp by surprise, as it demonstrated the model's capacity for high-level question analysis.

GPT-4 and the Language Learning Curve

GPT-4's ability to learn human languages is well-known. But can it comprehend non-human languages? Theoretically, it could be possible for an AI to grasp a language made up of syllables by prime numbers, which means that an AI system might create its language. Understanding the created AI language may seem daunting today, but it could become feasible with future advancements, such as advancements made possible by Neuralink. 

Legal Texts and GPT-4

Anne Meeuwese, a Professor of Public Law and Governance of Artificial Intelligence at Leiden University, put ChatGPT to the test by using it to draft laws. While ChatGPT did not ace the task, it was a source of inspiration. Nevertheless, Professor Meeuwese urges caution, highlighting the need for careful consideration of privacy implications when using such models.

GPT-4 in the Realm of Biological Sciences

Sanne Abeln, a Professor at Utrecht University and an affiliate of VU, tested ChatGPT's potential in biological sciences by posing master-level questions. While the AI was adept at answering knowledge-based questions, it struggled to connect the answers to scientific articles and papers. The inability to link the answers and questions to scientific sources underscores the challenge with AI—it can predict, but it does not understand why these predictions occur.

GPT-4's Role in Programming

Arie van Deursen, Professor of Software Engineering at the Technical University of Delft, recognises the potential of GPT models to assist programmers. A study by Meta showed that CodeCompose could predict 8% of their coding lines, but only a quarter of these were accurate. Despite this, van Deursen sees opportunities for GPT models in software testing, search engines, and programming assistance.

The Implications of ChatGPT in Various Markets

A study by OpenAI revealed that an astounding 80% of American workers could have at least 10% of their job done by a generative AI system. This figure increases to half of the tasks for 19% of these workers. The industries most affected are those reliant on programming and writing skills, with highly skilled and higher-paying jobs appearing most vulnerable. Contrary to Meeuwese's expectations, the most susceptible industries to the disruptive powers of AI are legal services and specific areas of financial and insurance sectors.

Conclusion

The strides made by GPT-4 herald a future where AI is integral to fields such as linguistics, law, and programming. Although it still has limitations to overcome, the impact of GPT-4 is undeniable, marking an exciting era for AI. We'd love your thoughts on GPT-4's potential in these fields. Join the discussion and share your thoughts below.