OpenAI’s language model, ChatGPT, has taken the world by storm and has been used in many creative ways. From writing essays to coding websites, the program has changed everyone’s perspective on what the internet can do. Recently becoming even more powerful, GPT-4 can be considered the new processor for ChatGPT, with many new features and capabilities.
With the introduction of the updated version of GPT-4, the program now has the additional processing power to complete more intricate and specific tasks with higher accuracy. However, the base of the language model remains quite similar. As a result, simple prompts given will still yield identical results compared to the previous versions.
The most significant addition is the ability of GPT-4 to process and complete tasks related to images that users can attach along with the text. This opens up a realm of possibilities. To illustrate, when given a prompt, GPT-4 was able to create a website based on a drawing; simultaneously demonstrating the new ability to write code in all major programming languages. Additionally, the image-processing tools are sufficiently prompt that an image of the ingredients inside a fridge was enough for GPT-4 to come up with meal ideas using only those ingredients. Finally, a glimpse of the updated language model’s reasoning capabilities has come to light and revealed that it is now possible for the program to interpret jokes which are given in a picture format, also to explain why they are funny. All of these examples foreshadow only a fraction of all the possible implementations of the new and improved software in situations where it has not been used before.
The upgrade has also brought along with it a higher word capacity for GPT-4’s answers, increasing the limit from around 4000 to 25000 words. Users can now receive responses on much more detailed prompts, as the new model operates using a much wider base of information.
However, issues from the previous versions still persist, with “hallucination” being the most significant one. In essence, this refers to the tendency of GPT-4 to “confidently state things that are entirely made up as if they are facts” as described by OpenAI’s CEO. This is present now more than before, due to the more specific replies from the language model.
Nevertheless, GPT-4 has performed exceptionally well in many tests originally made for humans, surpassing the scores of previous models outstandingly. An example of this is the Bar exam (used in law schools) in which GPT-4 placed in the 90th percentile of test takers. In direct comparison, the previous model placed in the 10th percentile. This again highlights the technical improvements in the new program and proves that this AI can pass examinations at the “human” level.
The new and strengthened model has invited a wave of criticism and concern. The CEO of OpenAI, Sam Altman, has warned the public of the potential dangers that could ensue with the introduction of the new software. According to him, the company needs to “be careful” and that users should be “happy that we are a little bit scared of this”. The main worry he expressed was that there exists a possibility of widespread disinformation, but also that the model could be used for cyber attacks due to its enhanced ability to write code. Although he went on to assure the public that GPT-4 is fully in human control, the main problem was whose hands it ended up in and the purposes it was used for.
Similarly, Tesla’s CEO Elon Musk has also expressed skepticism regarding the safety associated with the new capabilities of the language model, warning that the misuse of artificial intelligence can potentially be more dangerous than nuclear weapons. He further called for AI safety regulation, saying back in December 2022 that he has been requesting it for “over a decade”. Additionally, he recently complained about the situation on Twitter, asking “what will be left for us humans to do?”
On the other hand, some safety issues from GPT-3.5 are no longer a threat in the new model. The ethics component of the software has been revamped, with GPT-4 sometimes refusing to answer offensive questions and prompts. Furthermore, OpenAI has improved the software in the sense that many harms GPT-3.5 could cause are no longer a danger with GPT-4’s defenses. All in all, the future of GPT-4 and all subsequent artificial intelligence programs is filled with both instability and potential, seeming limitless.
