Since ChatGPT’s first release in late 2022, it has undergone many improvements and has become increasingly stable with every new update. However, as most new technologies do, ChatGPT eventually plateaued in popularity and showed a small dip in total website traffic. Things stayed this way until OpenAI made a major update to their AI model by unveiling their newest and 10 times more powerful version of ChatGPT earlier this year. Able to consider 1.5 trillion more parameters than its predecessor, as well as single handedly passing the SATs and placing itself in the top 10% of students in the Simulated Bar Exam, GPT-4 shocked users worldwide with its incredible capabilities. OpenAI has only allowed access to GPT-4 for people who want to pay $20/month to put this model to the test. Fortunately for you, I’ve purchased a month’s worth of access to this powerhouse in order to report my findings. After a few days of testing out this AI model, I have compiled its positive and negative qualities and examined the possible effects it could have on the world.
Let’s start with the good: ChatGPT has a new ability to browse the whole web instead of being limited to information from late 2021. GPT-4 uses Microsoft’s Bing search engine to browse the web and answer present-day inquiries. It first puts your question into Bing and reads what the search engine initially has to say. If it doesn’t find anything there, it opens an article that it thinks is relevant to your question, reads the article thoroughly, summarizes it and finally it spits out a professional response. It may sound like a long process, but it takes place within two minutes, depending on your inquiry.
Another exclusive feature to GPT-4 is the ability to listen and respond to human speech. Using GPT-4’s multimodal system, ChatGPT can now listen to your inquiry and respond through your choice of 5 scarily-human-like voices. This mode of ChatGPT is limited to data from no later than January 2022. This feature can be very helpful to people who might not be able to type out their full question; instead, they can simply talk to ChatGPT and have it speak back in a voice they can understand.
The final shocking feature of GPT-4 is how it can “see” an image without actually seeing it. ChatGPT combines different analyses between shapes, lines, edges, light, and colors within your image in order to summarize what it thinks is going on. To put this mode to the test, I gave ChatGPT a photo of a show at the Hollywood Bowl:
After about 15 seconds of thinking, GPT-4 described exactly what it thought was happening, including aspects of the lighting, audience, and architecture of the venue. It even knew that the show was taking place at the Hollywood Bowl! Here’s ChatGPT’s response:
Finally, let’s take a look at the performance of GPT-4 compared to GPT-3.5. As I previously mentioned, GPT-4 is 10 times more powerful than GPT-3.5 and is able to consider over one and a half trillion more parameters, making it the most powerful AI model in the world. GPT-3.5 had about 175 billion parameters, while GPT-4 totally blows that out of the water with 1.76 trillion parameters. Additionally, GPT-4 performs at least 10% better, and up to 80% better in standard exams ranging from AP tests to the SAT. Here’s a graph from OpenAI that shows GPT-3.5’s performance in blue and GPT-4 in green.
I think it’s very interesting to see that GPT-4 didn’t make any significant improvements on exams in humanities-related disciplines. This might be the case due to the fact that fields like English and history call for interpretation, creativity, and originality.
GPT-4 is a total game changer for the AI industry. From the crazy features to mind-blowing benchmarks, GPT-4 is both a blessing and a curse. I think OpenAI has made a good choice by limiting access to this model to paying users. However, it shouldn’t be only for the rich. I believe OpenAI can make their future, more powerful models restricted through invite-only access or individual authentication to make sure you’re going to use the model for good.