GPT-4 - A categorisation of the most important innovations
OpenAI released a new version of the language model behind ChatGPT this week – GPT-4. The innovations in this model have the potential to push the current boundaries of language comprehension and elevate human-machine interaction to a new level. We immediately examined the key improvements of GPT-4 and compiled our first impressions.
At the beginning of the year, it already garnered significant attention for its impressive performance in tasks requiring natural language processing. OpenAI’s powerful text generation model is capable of generating human-like text, completing code, or creating short poems or stories, even though it is far from perfect.
According to OpenAI, the GPT-4 model is now even more creative, reliable, and capable of solving complex tasks with greater accuracy than its predecessor. In a public demo, the model demonstrated how it could generate a fully functional website layout based on a simple sketch—an example of GPT-4's ability to process not just text but also images. We take a first look at the innovations and enhancements of the language model.
#1 In addition to text, images can also be processed
What's new
GPT-4 is capable of capturing and analyzing not just text but also visual content in the form of images. The model can describe images, interpret them, and identify relationships between them. Demonstrations have shown that GPT-4 is able to provide step-by-step explanations of memes or summarize complex infographics.
Our assessment
The ability of GPT-4 to process images is currently limited to selected OpenAI partners. Nevertheless, AI systems that enable the combination of language and images have existed for some time. Especially in the field of automated information extraction from documents, expanding the model input to the visual domain is advantageous. With GPT-4, information contained in illustrations and graphs within a document can also be taken into account.
#2 Performance in complex tasks is significantly better
What's new
GPT-4 has demonstrated significant performance improvements in numerous benchmark tests compared to its predecessor. Particularly in exams typically taken by law or natural science students, the model achieves above-average scores.
Our assessment
At first glance, GPT-4's performance is undoubtedly impressive and demonstrates how well the AI can reproduce knowledge in text form. However, it is important to remember that the tests used were designed to compare human abilities in a specific field. The requirements for a specialized AI optimized for a particular domain may therefore differ from the performance expectations placed on humans.
#3 The model's response style can be adjusted depending on usage
What's new
By using a system message, both developers and later users of ChatGPT have the option to adjust the response style of the language model.
Our assessment
The ability to customize GPT-4's response behavior allows users to better tailor the technology to their specific application and, for example, improve service experiences. The advantage of this approach is that the language model can be easily optimized for downstream tasks without requiring large datasets or additional training resources.
#4 The amount of processable text has increased eightfold
What's new
The length of the context that can be used during GPT-4's text generation process increases to up to 32,000 words, or about 50 pages, in different versions. As a result, GPT-4 is now capable of generating texts with greater coherence and a stronger focus on the original topic.
Our assessment
The ability to process larger amounts of text has interesting implications for practical applications. While ChatGPT already has a very good summarization capability for short texts, it is now possible to extend this ability to complete documents.
#5 Logical errors and false statements occur less frequently
What's new
Compared to its predecessor, GPT-4's output contains fewer false statements and contradictions. Additionally, the handling of requests that violate guidelines has been improved.
Our assessment
Although GPT-4 has some improvements, the reliability of its output remains a significant issue that limits the model’s usability. The generated texts can still be influenced by biases and contain misinformation that should be carefully reviewed.
Outlook
In the world of artificial intelligence, OpenAI has undoubtedly made an exciting upgrade to its GPT-3 model with the release of GPT-4. The new features of GPT-4 are interesting as they have the potential to enhance the performance and capabilities of AI in many areas.
However, there are some limitations and concerns associated with GPT-4. Unlike previous releases, OpenAI has unfortunately been significantly less transparent regarding model details. The public has access to only very limited information about the model architecture, training approach, or data foundation. This makes it difficult, for example, to assess to what extent the model may have inherited biases that could affect its use.
GPT-4 represents an improvement over its predecessor, but not necessarily a major leap forward. Although there are some new features, they are not as revolutionary as some had hoped. It is therefore very important that we maintain realistic expectations about GPT-4’s capabilities and carefully evaluate the model’s strengths and limitations before deploying it in various applications. This ensures that the technology is used effectively within its current capabilities.
