The power of GPT-4: Understanding its capabilities

March 20, 2023

On March 14, 2023, GPT-4, the fourth version of the GPT language model, was released, fulfilling expectations regarding its potential. It has been developed by OpenAI, which previously released GPT-2 (in 2019) and GPT-3 (in 2020). GPT-4 has amazed people with its potential to produce text and graphics that appear to be created by a human, provided it is given an initial prompt. The topic is widely discussed among scientific circles, and has been covered by Nature and Scientific American in their recent articles. GPT-4 is an improved version of GPT-3.5, so what are the differences between the two models? Let us explore the key distinctions between GPT-4 and GPT-3.5.

What is GPT?

GPT (Generative Pre-trained Transformer) is a type of language processing model that uses machine learning to create a large database from text, numbers, and data. It employs a neural network to choose the right words from its database for a given sentence and can generate new texts. This AI is also able to understand human language, replicating argument and thought patterns using statistical methods.

GPT-3 vs GPT-4

GPT-4 Artificial Intelligence has a significant advantage over GPT-3, with 100 billion learning parameters compared to 175 million. Consequently, GPT-4 is more closely modelled on the neural connections of the human brain, allowing it to process and interpret more data, and more accurately replicate human language.

  1. Creativity: GPT-4 outperforms GPT-3.5 with more creative replies to prompts. ChatGPT, which uses the GPT-3.5 model, is already highly creative. But GPT-4 takes creativity even further, particularly when tackling complex, creative tasks.

  2. Multimodality: ChatGPT used to be only for text-based prompts. GPT-4, however, has a new feature: it can process both text and images. This means that it can process all kinds of images, including text documents, hand-drawn sketches, and screenshots, and interpret and understand them as if they were text prompts, showcasing a deep understanding of visual aesthetics and object relationships. However, the image feature is not currently publicly available.

  3. Versatility: GPT-4 can effortlessly adapt to different genres and styles of writing, making it a valuable asset for various: industries and applications. GPT-4 is an enhanced language model and has improved its understanding of complex and subtle prompts. OpenAI has observed it reaching the same performance levels as humans on various professional and academic tests. GPT-4 can process longer inputs of up to 25,000 words (compared to GPT-3.5’s 8,000 maximum). The increased word limit enables users to provide more detailed prompts, which in turn leads to more comprehensive outputs. GPT-4 also supports over 26 languages, including Indian languages such as Bengali, Punjabi, Marathi, and Telugu.

  4. Steerability: OpenAI asserts that GPT-4 offers a great degree of control. The AI is less likely to malfunction when used in applications that require it to play specific roles. By sending “system” messages, developers can accustom the AI’s style and tasks to their preferences, while also providing the easiest means of “jailbreaking” the model.

  5. Factuality: GPT-3.5 is prone to producing inaccurate information with confidence, which is known as “AI hallucination”. GPT-4 is 19-29% less likely to do this, according to its technical report. This is also evident from using ChatGPT, where GPT-4 results are significantly more factual.

  6. Safety: OpenAI spent six months optimising the safety of GPT-4. Compared to GPT-3.5, it is 82% less likely to generate inappropriate responses, 29% more likely to produce safe responses to sensitive requests, and 40% more likely to generate factual responses. Although GPT-4 is not perfect, it is an improvement over the flaws of GPT-3.5 which was moderated as an afterthought. OpenAI identified safety issues and tried to fix them as users interacted with it.

  7. Scope: GPT-4 has a larger context window and size than its predecessor GPT-3.5, allowing it to remember context and instructions better. This reduces the tendency for GPT-4 to go off-topic or fail to follow instructions. GPT-4 can also summarise larger texts, such as PDFs, without splitting them into smaller chunks.

How to use GPT-4?

To access GPT-4, developers must sign up for the waitlist, merge model assessments into OpenAI Evals for priority access, or subscribe to ChatGPT Plus for usage on chat.openai.com (usage limit based on actual system usage and anticipated demand). Additionally, GPT-4 can be experienced freely on Bing AI chatbot if you have access. Some of the early adopters of GPT-4 worldwide include Duolingo, Morgan Stanley, Be My Eyes, Stripe, Khan Academy etc. If someone wants to contribute to ChatGPT with feedback and opinions, you may use Evals–a framework that evaluates OpenAI models and offers an open-source registry of benchmarks.

Limitations

OpenAI’s GPT-4 language model has advanced capabilities, but it is still subject to social biases, hallucinations, and adversarial prompts. The important ethical and safety concerns are:

  • Data privacy – Protecting the personal and sensitive information of users and their interactions with the AI system.

  • Bias and Discrimination – Ensuring that the AI model does not favour or discriminate against certain groups or individuals based on their gender, race, or socioeconomic status.

  • Transparency – Clearly communicating how the AI model functions, its limitations, and the rationale behind its results or predictions.

  • Accountability – Establishing responsibility for the AI model’s actions, especially when it comes to addressing unintentional harm or negative consequences related to the model.

  • Misinformation and Deceptive Use – Preventing bad actors from utilising the AI technology for malicious purposes is essential for both users and developers.

Some strategies to help mitigate the risks associated with misinformation and deceptive AI use are (a) to develop AI transparency and ethical guidelines, (b) to ensure AI auditing and certification, (c) to implement proper monitoring and (d) to educate and create awareness among users. The company is diligently working to address these issues, though, so we can continue to take advantage of GPT-4’s capabilities.

Conclusion

It is true that advanced AI models like GPT-4 have the potential to revolutionise many fields, including science. They can help researchers generate new hypotheses, identify patterns in large datasets, and even design experiments or simulations. It is true that advanced AI models like GPT-4 have the potential to revolutionise many fields, including science. In summary, AI like GPT-4 has far-reaching implications for scientific research, ushering in a new era of accelerated discoveries and improved understanding in various disciplines.

Additional Reading

  1. Bushwick S. (2023). What the new GPT-4 AI can do. Scientific American 328(3): 16 March 2023.

  2. OpenAI (2023). GPT-4. https://openai.com/research/gpt-4

  3. Sanderson K. (2023). GPT-4 is here: what scientists think. Nature https://doi.org/10.1038/d41586-023-00816-5