It has been almost two years since OpenAI released ChatGPT 3.5 in November 2022. According to the latest Gartner Hype Cycle from last month [1], Generative AI is over the peak of inflated expectations. I have to agree, as it turns out that ChatGPT and other LLMs have just become an easier form of googling, but with additional risks, disappointments and frustrations.
First of all, the common misconception that I often see from people with a less technical background is that they believe LLMs are telling the truth. No, it is trained with a large amount of data and tries to connect your prompt with the statistically most likely outcome. To give you an example, if you type in “Second World”. The most likely output you will get is “War”. This was also an issue back in 2022, as students started to copy what ChatGPT was outputting for their assignments without checking if the sources were actually real and if the output actually made sense.
Now, 2 years later, there have not been really major improvements. If you give ChatGPT a list of tasks to do, it tends to forget what the first task was. If you ask ChatGPT to give you sources for the text that it outputs, it will now give you actual sources that exist, but if you check the content of these pages you see that it does not match at all. But to be not overly negative, I used ChatGPT for coding in the past, and it sometimes does give decent feedback, but it probably would have been quicker for me if I just googled it myself. Something that LLMs also have helped me with is rewriting my own text in a more clear manner, as well as summarizing pdfs with ChatPDF. In case you did not know about ChatPDF, you just upload your pdf and start asking questions [2].
What I am actually impressed about is the generation of images and videos. I have used Bing Image Creator in the past for creating images about my mood or visualizing my thoughts and I have to say that it does a pretty accurate job. While for video generation OpenAI’s “Sora” has not been released yet, I was amazed about one example [3] that you can see below.
References:
[1] Gartner.: Gartner 2024 Hype Cycle for Emerging Technologies Highlights Developer Productivity, Total Experience, AI and Security. https://www.gartner.com/en/newsroom/press-releases/2024-08-21-gartner-2024-hype-cycle-for-emerging-technologies-highlights-developer-productivity-total-experience-ai-and-security (2024)
[3] OpenAI.: Train Window. https://vimeo.com/913132696/e459e90255 (2024)
Featured image created with Bing Image Creator
The blog gives a good overview of the current state of Generative AI. The author mentioned well that AI’s not always tell the truth. I also experienced this myself. It is always important to double check the information that is provided by AI. I agree that there are limitations, but I think that AI’s can be really valuable, more than is being mentioned in the blog, they will get better and eventually will almost always produce truthful answers. For example integrating AI’s with more reliable databases or real-time data verification tools. Overall, there is a balanced view on where AI’s can still grow and where are the strengths.
Thank you for the insightful blog post. I agree with many of your points regarding ChatGPT, and I appreciate how you shared your experiences using it in the past and present. Looking ahead, incorporating ChatGPT into routine tasks like rewording text might be more practical than for in-depth research. I also fully agree with your praise of OpenAI’s Sora, and I enjoyed the inclusion of the short clip. Keep up the great work!